From 054bc9f5e26028cd84aa816c5c06ed131959fdde Mon Sep 17 00:00:00 2001 From: raphael couturier Date: Sat, 13 Dec 2014 20:51:44 +0100 Subject: [PATCH 01/16] end of english correction --- Heter_paper.tex | 352 ++++++++++++++++++++++++++++-------------------- 1 file changed, 204 insertions(+), 148 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index 50fc783..9d743db 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -514,7 +514,7 @@ function has the following form: where $N$ is the number of nodes and $F$ is the number of available frequencies for each node. Then, the optimal set of scaling factors that satisfies (\ref{eq:max}) can be selected. The objective function can work with any energy model or any power values for each node -(static and dynamic powers). However, the most energy reduction gain can be achieved when +(static and dynamic powers). However, the most important energy reduction gain can be achieved when the energy curve has a convex form as shown in~\cite{Zhuo_Energy.efficient.Dynamic.Task.Scheduling,Rauber_Analytical.Modeling.for.Energy,Hao_Learning.based.DVFS}. \section{The scaling factors selection algorithm for heterogeneous platforms } @@ -555,34 +555,41 @@ and the computation scaling factor $Scp_i$ as follows: \label{eq:Fint} F_{i} = \frac{Fmax_i}{Scp_i},~{i=1,2,\cdots,N} \end{equation} -If the computed initial frequency for a node is not available in the gears of that node, the computed -initial frequency is replaced by the nearest available frequency. In figure (\ref{fig:st_freq}), -the nodes are sorted by their computing powers in ascending order and the frequencies of the faster -nodes are scaled down according to the computed initial frequency scaling factors. The resulting new -frequencies are colored in blue in figure (\ref{fig:st_freq}). This set of frequencies can be considered -as a higher bound for the search space of the optimal vector of frequencies because selecting frequency -scaling factors higher than the higher bound will not improve the performance of the application and -it will increase its overall energy consumption. Therefore the algorithm that selects the frequency -scaling factors starts the search method from these initial frequencies and takes a downward search direction -toward lower frequencies. The algorithm iterates on all left frequencies, from the higher bound until all -nodes reach their minimum frequencies, to compute their overall energy consumption and performance, and select -the optimal frequency scaling factors vector. At each iteration the algorithm determines the slowest node -according to the equation (\ref{eq:perf}) and keeps its frequency unchanged, while it lowers the frequency of -all other nodes by one gear. -The new overall energy consumption and execution time are computed according to the new scaling factors. -The optimal set of frequency scaling factors is the set that gives the highest distance according to the objective -function (\ref{eq:max}). - -The plots~(\ref{fig:r1} and \ref{fig:r2}) illustrate the normalized performance and consumed energy for an -application running on a homogeneous platform and a heterogeneous platform respectively while increasing the -scaling factors. It can be noticed that in a homogeneous platform the search for the optimal scaling factor -should be started from the maximum frequency because the performance and the consumed energy is decreased since -the beginning of the plot. On the other hand, in the heterogeneous platform the performance is maintained at -the beginning of the plot even if the frequencies of the faster nodes are decreased until the scaled down nodes -have computing powers lower than the slowest node. In other words, until they reach the higher bound. It can -also be noticed that the higher the difference between the faster nodes and the slower nodes is, the bigger -the maximum distance between the energy curve and the performance curve is while varying the scaling factors -which results in bigger energy savings. +If the computed initial frequency for a node is not available in the gears of +that node, it is replaced by the nearest available frequency. In figure +(\ref{fig:st_freq}), the nodes are sorted by their computing power in ascending +order and the frequencies of the faster nodes are scaled down according to the +computed initial frequency scaling factors. The resulting new frequencies are +colored in blue in figure (\ref{fig:st_freq}). This set of frequencies can be +considered as a higher bound for the search space of the optimal vector of +frequencies because selecting frequency scaling factors higher than the higher +bound will not improve the performance of the application and it will increase +its overall energy consumption. Therefore the algorithm that selects the +frequency scaling factors starts the search method from these initial +frequencies and takes a downward search direction toward lower frequencies. The +algorithm iterates on all left frequencies, from the higher bound until all +nodes reach their minimum frequencies, to compute their overall energy +consumption and performance, and select the optimal frequency scaling factors +vector. At each iteration the algorithm determines the slowest node according to +the equation (\ref{eq:perf}) and keeps its frequency unchanged, while it lowers +the frequency of all other nodes by one gear. The new overall energy +consumption and execution time are computed according to the new scaling +factors. The optimal set of frequency scaling factors is the set that gives the +highest distance according to the objective function (\ref{eq:max}). + +Figures~\ref{fig:r1} and \ref{fig:r2} illustrate the normalized performance and +consumed energy for an application running on a homogeneous platform and a +heterogeneous platform respectively while increasing the scaling factors. It can +be noticed that in a homogeneous platform the search for the optimal scaling +factor should start from the maximum frequency because the performance and the +consumed energy decrease from the beginning of the plot. On the other hand, +in the heterogeneous platform the performance is maintained at the beginning of +the plot even if the frequencies of the faster nodes decrease until the +computing power of scaled down nodes are lower than the slowest node. In other +words, until they reach the higher bound. It can also be noticed that the higher +the difference between the faster nodes and the slower nodes is, the bigger the +maximum distance between the energy curve and the performance curve is while + the scaling factors are varying which results in bigger energy savings. \begin{figure}[t] \centering \includegraphics[scale=0.5]{fig/start_freq} @@ -660,18 +667,21 @@ which results in bigger energy savings. \subsection{The evaluation of the proposed algorithm} \label{sec.verif.algo} -The precision of the proposed algorithm mainly depends on the execution time prediction model defined in -(\ref{eq:perf}) and the energy model computed by (\ref{eq:energy}). -The energy model is also significantly dependent on the execution time model because the static energy is -linearly related to the execution time and the dynamic energy is related to the computation time. So, all of -the works presented in this paper is based on the execution time model. To verify this model, the predicted -execution time was compared to the real execution time over SimGrid/SMPI simulator, v3.10~\cite{casanova+giersch+legrand+al.2014.versatile}, -for all the NAS parallel benchmarks NPB v3.3 -\cite{NAS.Parallel.Benchmarks}, running class B on 8 or 9 nodes. The comparison showed that the proposed execution time model is very precise, -the maximum normalized difference between the predicted execution time and the real execution time is equal -to 0.03 for all the NAS benchmarks. - -Since the proposed algorithm is not an exact method and does not test all the possible solutions (vectors of scaling factors) +The precision of the proposed algorithm mainly depends on the execution time +prediction model defined in (\ref{eq:perf}) and the energy model computed by +(\ref{eq:energy}). The energy model is also significantly dependent on the +execution time model because the static energy is linearly related to the +execution time and the dynamic energy is related to the computation time. So, +all the works presented in this paper are based on the execution time model. To +verify this model, the predicted execution time was compared to the real +execution time over SimGrid/SMPI simulator, +v3.10~\cite{casanova+giersch+legrand+al.2014.versatile}, for all the NAS +parallel benchmarks NPB v3.3 \cite{NAS.Parallel.Benchmarks}, running class B on +8 or 9 nodes. The comparison showed that the proposed execution time model is +very precise, the maximum normalized difference between the predicted execution +time and the real execution time is equal to 0.03 for all the NAS benchmarks. + +Since the proposed algorithm is not an exact method it does not test all the possible solutions (vectors of scaling factors) in the search space. To prove its efficiency, it was compared on small instances to a brute force search algorithm that tests all the possible solutions. The brute force algorithm was applied to different NAS benchmarks classes with different number of nodes. The solutions returned by the brute force algorithm and the proposed algorithm were identical @@ -684,22 +694,27 @@ vector of frequency scaling factors that gives the results of the next sections. \section{Experimental results} \label{sec.expe} -To evaluate the efficiency and the overall energy consumption reduction of algorithm~ \ref{HSA}, -it was applied to the NAS parallel benchmarks NPB v3.3. The experiments were executed -on the simulator SimGrid/SMPI which offers easy tools to create a heterogeneous platform and run -message passing applications over it. The heterogeneous platform that was used in the experiments, -had one core per node because just one process was executed per node. -The heterogeneous platform was composed of four types of nodes. Each type of nodes had different -characteristics such as the maximum CPU frequency, the number of -available frequencies and the computational power, see Table \ref{table:platform}. The characteristics -of these different types of nodes are inspired from the specifications of real Intel processors. -The heterogeneous platform had up to 144 nodes and had nodes from the four types in equal proportions, -for example if a benchmark was executed on 8 nodes, 2 nodes from each type were used. Since the constructors -of CPUs do not specify the dynamic and the static power of their CPUs, for each type of node they were -chosen proportionally to its computing power (FLOPS). In the initial heterogeneous platform, while computing -with highest frequency, each node consumed power proportional to its computing power which 80\% of it was -dynamic power and the rest was 20\% for the static power, the same assumption was made in \cite{Our_first_paper,Rauber_Analytical.Modeling.for.Energy}. -Finally, These nodes were connected via an ethernet network with 1 Gbit/s bandwidth. +To evaluate the efficiency and the overall energy consumption reduction of +algorithm~\ref{HSA}, it was applied to the NAS parallel benchmarks NPB v3.3. The +experiments were executed on the simulator SimGrid/SMPI which offers easy tools +to create a heterogeneous platform and run message passing applications over it. +The heterogeneous platform that was used in the experiments, had one core per +node because just one process was executed per node. The heterogeneous platform +was composed of four types of nodes. Each type of nodes had different +characteristics such as the maximum CPU frequency, the number of available +frequencies and the computational power, see Table \ref{table:platform}. The +characteristics of these different types of nodes are inspired from the +specifications of real Intel processors. The heterogeneous platform had up to +144 nodes and had nodes from the four types in equal proportions, for example if +a benchmark was executed on 8 nodes, 2 nodes from each type were used. Since the +constructors of CPUs do not specify the dynamic and the static power of their +CPUs, for each type of node they were chosen proportionally to its computing +power (FLOPS). In the initial heterogeneous platform, while computing with +highest frequency, each node consumed an amount of power proportional to its +computing power (which corresponds to 80\% of its dynamic power and the +remaining 20\% to the static power), the same assumption was made in +\cite{Our_first_paper,Rauber_Analytical.Modeling.for.Energy}. Finally, These +nodes were connected via an ethernet network with 1 Gbit/s bandwidth. \begin{table}[htb] @@ -736,12 +751,14 @@ Finally, These nodes were connected via an ethernet network with 1 Gbit/s bandwi \label{sec.res} -The proposed algorithm was applied to the seven parallel NAS benchmarks (EP, CG, MG, FT, BT, LU and SP) -and the benchmarks were executed with the three classes: A,B and C. However, due to the lack of space in -this paper, only the results of the biggest class, C, are presented while being run on different number -of nodes, ranging from 4 to 128 or 144 nodes depending on the benchmark being executed. Indeed, the -benchmarks CG, MG, LU, EP and FT should be executed on $1, 2, 4, 8, 16, 32, 64, 128$ nodes. -The other benchmarks such as BT and SP should be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. +The proposed algorithm was applied to the seven parallel NAS benchmarks (EP, CG, +MG, FT, BT, LU and SP) and the benchmarks were executed with the three classes: +A, B and C. However, due to the lack of space in this paper, only the results of +the biggest class, C, are presented while being run on different number of +nodes, ranging from 4 to 128 or 144 nodes depending on the benchmark being +executed. Indeed, the benchmarks CG, MG, LU, EP and FT had to be executed on $1, +2, 4, 8, 16, 32, 64, 128$ nodes. The other benchmarks such as BT and SP had to +be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. @@ -907,32 +924,38 @@ The other benchmarks such as BT and SP should be executed on $1, 4, 9, 16, 36, 6 \end{tabular} \label{table:res_128n} \end{table} -The overall energy consumption was computed for each instance according to the energy -consumption model (\ref{eq:energy}), with and without applying the algorithm. The -execution time was also measured for all these experiments. Then, the energy saving -and performance degradation percentages were computed for each instance. -The results are presented in Tables (\ref{table:res_4n}, \ref{table:res_8n}, \ref{table:res_16n}, -\ref{table:res_32n}, \ref{table:res_64n} and \ref{table:res_128n}). All these results are the -average values from many experiments for energy savings and performance degradation. -The tables show the experimental results for running the NAS parallel benchmarks on different -number of nodes. The experiments show that the algorithm reduce significantly the energy -consumption (up to 35\%) and tries to limit the performance degradation. They also show that -the energy saving percentage is decreased when the number of the computing nodes is increased. -This reduction is due to the increase of the communication times compared to the execution times -when the benchmarks are run over a high number of nodes. Indeed, the benchmarks with the same class, C, -are executed on different number of nodes, so the computation required for each iteration is divided -by the number of computing nodes. On the other hand, more communications are required when increasing -the number of nodes so the static energy is increased linearly according to the communication time and -the dynamic power is less relevant in the overall energy consumption. Therefore, reducing the frequency -with algorithm~(\ref{HSA}) have less effect in reducing the overall energy savings. It can also be -noticed that for the benchmarks EP and SP that contain little or no communications, the energy savings -are not significantly affected with the high number of nodes. No experiments were conducted using bigger -classes such as D, because they require a lot of memory(more than 64GB) when being executed by the simulator -on one machine. The maximum distance between the normalized energy curve and the normalized performance -for each instance is also shown in the result tables. It is decreased in the same way as the energy -saving percentage. The tables also show that the performance degradation percentage is not significantly -increased when the number of computing nodes is increased because the computation times are small when -compared to the communication times. +The overall energy consumption was computed for each instance according to the +energy consumption model (\ref{eq:energy}), with and without applying the +algorithm. The execution time was also measured for all these experiments. Then, +the energy saving and performance degradation percentages were computed for each +instance. The results are presented in Tables (\ref{table:res_4n}, +\ref{table:res_8n}, \ref{table:res_16n}, \ref{table:res_32n}, +\ref{table:res_64n} and \ref{table:res_128n}). All these results are the average +values from many experiments for energy savings and performance degradation. +The tables show the experimental results for running the NAS parallel benchmarks +on different number of nodes. The experiments show that the algorithm +significantly reduces the energy consumption (up to 35\%) and tries to limit the +performance degradation. They also show that the energy saving percentage +decreases when the number of computing nodes increases. This reduction is due +to the increase of the communication times compared to the execution times when +the benchmarks are run over a high number of nodes. Indeed, the benchmarks with +the same class, C, are executed on different numbers of nodes, so the +computation required for each iteration is divided by the number of computing +nodes. On the other hand, more communications are required when increasing the +number of nodes so the static energy increases linearly according to the +communication time and the dynamic power is less relevant in the overall energy +consumption. Therefore, reducing the frequency with algorithm~(\ref{HSA}) is +less effective in reducing the overall energy savings. It can also be noticed +that for the benchmarks EP and SP that contain little or no communications, the +energy savings are not significantly affected by the high number of nodes. No +experiments were conducted using bigger classes than D, because they require a +lot of memory (more than 64GB) when being executed by the simulator on one +machine. The maximum distance between the normalized energy curve and the +normalized performance for each instance is also shown in the result tables. It +decrease in the same way as the energy saving percentage. The tables also show +that the performance degradation percentage is not significantly increased when +the number of computing nodes is increased because the computation times are +small when compared to the communication times. @@ -944,65 +967,81 @@ compared to the communication times. \subfloat[Performance degradation ]{% \includegraphics[width=.33\textwidth]{fig/per_deg}\label{fig:per_deg}} \label{fig:avg} - \caption{The energy and performance for all NAS benchmarks running with difference number of nodes} + \caption{The energy and performance for all NAS benchmarks running with a different number of nodes} \end{figure} -Plots (\ref{fig:energy} and \ref{fig:per_deg}) present the energy saving and performance degradation -respectively for all the benchmarks according to the number of used nodes. As shown in the first plot, -the energy saving percentages of the benchmarks MG, LU, BT and FT are decreased linearly when the -number of nodes is increased. While for the EP and SP benchmarks, the energy saving percentage is not -affected by the increase of the number of computing nodes, because in these benchmarks there are little or -no communications. Finally, the energy saving of the GC benchmark is significantly decreased when the number -of nodes is increased because this benchmark has more communications than the others. The second plot -shows that the performance degradation percentages of most of the benchmarks are decreased when they -run on a big number of nodes because they spend more time communicating than computing, thus, scaling -down the frequencies of some nodes have less effect on the performance. +Figures \ref{fig:energy} and \ref{fig:per_deg} present the energy saving and +performance degradation respectively for all the benchmarks according to the +number of used nodes. As shown in the first plot, the energy saving percentages +of the benchmarks MG, LU, BT and FT decrease linearly when the number of nodes +increase. While for the EP and SP benchmarks, the energy saving percentage is +not affected by the increase of the number of computing nodes, because in these +benchmarks there are little or no communications. Finally, the energy saving of +the GC benchmark significantly decrease when the number of nodes increase +because this benchmark has more communications than the others. The second plot +shows that the performance degradation percentages of most of the benchmarks +decrease when they run on a big number of nodes because they spend more time +communicating than computing, thus, scaling down the frequencies of some nodes +has less effect on the performance. \subsection{The results for different power consumption scenarios} \label{sec.compare} -The results of the previous section were obtained while using processors that consume during computation -an overall power which is 80\% composed of dynamic power and 20\% of static power. In this section, -these ratios are changed and two new power scenarios are considered in order to evaluate how the proposed -algorithm adapts itself according to the static and dynamic power values. The two new power scenarios -are the following: +The results of the previous section were obtained while using processors that +consume during computation an overall power which is 80\% composed of dynamic +power and of 20\% of static power. In this section, these ratios are changed and +two new power scenarios are considered in order to evaluate how the proposed +algorithm adapts itself according to the static and dynamic power values. The +two new power scenarios are the following: \begin{itemize} -\item 70\% dynamic power and 30\% static power -\item 90\% dynamic power and 10\% static power +\item 70\% of dynamic power and 30\% of static power +\item 90\% of dynamic power and 10\% of static power \end{itemize} -The NAS parallel benchmarks were executed again over processors that follow the new power scenarios. -The class C of each benchmark was run over 8 or 9 nodes and the results are presented in Tables -\ref{table:res_s1} and \ref{table:res_s2}. These tables show that the energy saving percentage of the 70\%-30\% -scenario is less for all benchmarks compared to the energy saving of the 90\%-10\% scenario. Indeed, in the latter -more dynamic power is consumed when nodes are running on their maximum frequencies, thus, scaling down the frequency -of the nodes results in higher energy savings than in the 70\%-30\% scenario. On the other hand, the performance -degradation percentage is less in the 70\%-30\% scenario compared to the 90\%-10\% scenario. This is due to the -higher static power percentage in the first scenario which makes it more relevant in the overall consumed energy. -Indeed, the static energy is related to the execution time and if the performance is degraded the total consumed -static energy is directly increased. Therefore, the proposed algorithm do not scales down much the frequencies of the -nodes in order to limit the increase of the execution time and thus limiting the effect of the consumed static energy. - -The two new power scenarios are compared to the old one in figure (\ref{fig:sen_comp}). It shows the average of -the performance degradation, the energy saving and the distances for all NAS benchmarks of class C running on 8 or 9 nodes. -The comparison shows that the energy saving ratio is proportional to the dynamic power ratio: it is increased -when applying the 90\%-10\% scenario because at maximum frequency the dynamic energy is the most relevant -in the overall consumed energy and can be reduced by lowering the frequency of some processors. On the other hand, -the energy saving is decreased when the 70\%-30\% scenario is used because the dynamic energy is less relevant in -the overall consumed energy and lowering the frequency do not returns big energy savings. -Moreover, the average of the performance degradation is decreased when using a higher ratio for static power -(e.g. 70\%-30\% scenario and 80\%-20\% scenario). Since the proposed algorithm optimizes the energy consumption -when using a higher ratio for dynamic power the algorithm selects bigger frequency scaling factors that result in -more energy saving but less performance, for example see the figure (\ref{fig:scales_comp}). The opposite happens -when using a higher ratio for static power, the algorithm proportionally selects smaller scaling values which -results in less energy saving but less performance degradation. +The NAS parallel benchmarks were executed again over processors that follow the +new power scenarios. The class C of each benchmark was run over 8 or 9 nodes +and the results are presented in Tables \ref{table:res_s1} and +\ref{table:res_s2}. These tables show that the energy saving percentage of the +70\%-30\% scenario is smaller for all benchmarks compared to the energy saving +of the 90\%-10\% scenario. Indeed, in the latter more dynamic power is consumed +when nodes are running on their maximum frequencies, thus, scaling down the +frequency of the nodes results in higher energy savings than in the 70\%-30\% +scenario. On the other hand, the performance degradation percentage is smaller +in the 70\%-30\% scenario compared to the 90\%-10\% scenario. This is due to the +higher static power percentage in the first scenario which makes it more +relevant in the overall consumed energy. Indeed, the static energy is related +to the execution time and if the performance is degraded the amount of consumed +static energy directly increas. Therefore, the proposed algorithm does not +really significantly scale down much the frequencies of the nodes in order to +limit the increase of the execution time and thus limiting the effect of the +consumed static energy. + +Both new power scenarios are compared to the old one in figure +(\ref{fig:sen_comp}). It shows the average of the performance degradation, the +energy saving and the distances for all NAS benchmarks of class C running on 8 +or 9 nodes. The comparison shows that the energy saving ratio is proportional +to the dynamic power ratio: it is increased when applying the 90\%-10\% scenario +because at maximum frequency the dynamic energy is the most relevant in the +overall consumed energy and can be reduced by lowering the frequency of some +processors. On the other hand, the energy saving decreases when the 70\%-30\% +scenario is used because the dynamic energy is less relevant in the overall +consumed energy and lowering the frequency does not return big energy savings. +Moreover, the average of the performance degradation is decreased when using a +higher ratio for static power (e.g. 70\%-30\% scenario and 80\%-20\% +scenario). Since the proposed algorithm optimizes the energy consumption when +using a higher ratio for dynamic power the algorithm selects bigger frequency +scaling factors that result in more energy saving but less performance, for +example see Figure (\ref{fig:scales_comp}). The opposite happens when using a +higher ratio for static power, the algorithm proportionally selects smaller +scaling values which result in less energy saving but also less performance +degradation. \begin{table}[htb] - \caption{The results of 70\%-30\% powers scenario} + \caption{The results of the 70\%-30\% power scenario} % title of Table \centering \begin{tabular}{|*{6}{l|}} @@ -1031,7 +1070,7 @@ results in less energy saving but less performance degradation. \begin{table}[htb] - \caption{The results of 90\%-10\% powers scenario} + \caption{The results of the 90\%-10\% power scenario} % title of Table \centering \begin{tabular}{|*{6}{l|}} @@ -1060,7 +1099,7 @@ results in less energy saving but less performance degradation. \begin{figure} \centering - \subfloat[Comparison of the results on 8 nodes]{% + \subfloat[Comparison between the results on 8 nodes]{% \includegraphics[width=.33\textwidth]{fig/sen_comp}\label{fig:sen_comp}}% \subfloat[Comparison the selected frequency scaling factors of MG benchmark class C running on 8 nodes]{% @@ -1080,12 +1119,14 @@ They developed a green governor that regularly applies an online frequency selec To fairly compare both algorithms, the same energy and execution time models, equations (\ref{eq:energy}) and (\ref{eq:fnew}), were used for both algorithms to predict the energy consumption and the execution times. Also Spiliopoulos et al. algorithm was adapted to start the search from the initial frequencies computed using the equation (\ref{eq:Fint}). The resulting algorithm is an exhaustive search algorithm that minimizes the EDP and has the initial frequencies values as an upper bound. -Both algorithms were applied to the parallel NAS benchmarks to compare their efficiency. Table \ref{table:compare_EDP} presents the results of comparing the execution times and the energy consumptions for both versions of the NAS benchmarks while running the class C of each benchmark over 8 or 9 heterogeneous nodes. The results show that our algorithm gives better energy savings than Spiliopoulos et al. algorithm, +Both algorithms were applied to the parallel NAS benchmarks to compare their efficiency. Table \ref{table:compare_EDP} presents the results of comparing the execution times and the energy consumptions for both versions of the NAS benchmarks while running the class C of each benchmark over 8 or 9 heterogeneous nodes. The results show that our algorithm provides better energy savings than Spiliopoulos et al. algorithm, on average it results in 29.76\% energy saving while their algorithm returns just 25.75\%. The average of performance degradation percentage is approximately the same for both algorithms, about 4\%. -For all benchmarks, our algorithm outperforms -Spiliopoulos et al. algorithm in term of energy and performance tradeoff, see figure (\ref{fig:compare_EDP}), because it maximizes the distance between the energy saving and the performance degradation values while giving the same weight for both metrics. +For all benchmarks, our algorithm outperforms Spiliopoulos et al. algorithm in +terms of energy and performance tradeoff, see figure (\ref{fig:compare_EDP}), +because it maximizes the distance between the energy saving and the performance +degradation values while giving the same weight for both metrics. @@ -1123,16 +1164,31 @@ Spiliopoulos et al. algorithm in term of energy and performance tradeoff, see fi \section{Conclusion} \label{sec.concl} -In this paper, a new online frequency selecting algorithm has been presented. It selects the best possible vector of frequency scaling factors that gives the maximum distance (optimal tradeoff) between the predicted energy and -the predicted performance curves for a heterogeneous platform. This algorithm uses a new energy model for measuring -and predicting the energy of distributed iterative applications running over heterogeneous -platform. To evaluate the proposed method, it was applied on the NAS parallel benchmarks and executed over a heterogeneous platform simulated by Simgrid. The results of the experiments showed that the algorithm reduces up to 35\% the energy consumption of a message passing iterative method while limiting the degradation of the performance. The algorithm also selects different scaling factors according to the percentage of the computing and communication times, and according to the values of the static and dynamic powers of the CPUs. Finally, the algorithm was compared to Spiliopoulos et al. algorithm and the results showed that it - outperforms their algorithm in term of energy-time tradeoff. - -In the near future, this method will be applied to real heterogeneous platforms to evaluate its performance in a real study case. It would also be interesting to evaluate its scalability over large scale heterogeneous platform and measure the energy consumption reduction it can produce. Afterward, we would like to develop a similar method that is adapted to asynchronous iterative applications -where each task does not wait for others tasks to finish their works. The development of such method might require a new -energy model because the number of iterations is not -known in advance and depends on the global convergence of the iterative system. +In this paper, a new online frequency selecting algorithm has been presented. It +selects the best possible vector of frequency scaling factors that gives the +maximum distance (optimal tradeoff) between the predicted energy and the +predicted performance curves for a heterogeneous platform. This algorithm uses a +new energy model for measuring and predicting the energy of distributed +iterative applications running over heterogeneous platforms. To evaluate the +proposed method, it was applied on the NAS parallel benchmarks and executed over +a heterogeneous platform simulated by Simgrid. The results of the experiments +showed that the algorithm reduces up to 35\% the energy consumption of a message +passing iterative method while limiting the degradation of the performance. The +algorithm also selects different scaling factors according to the percentage of +the computing and communication times, and according to the values of the static +and dynamic powers of the CPUs. Finally, the algorithm was compared to +Spiliopoulos et al. algorithm and the results showed that it outperforms their +algorithm in terms of energy-time tradeoff. + +In the near future, this method will be applied to real heterogeneous platforms +to evaluate its performance in a real study case. It would also be interesting +to evaluate its scalability over large scale heterogeneous platforms and measure +the energy consumption reduction it can produce. Afterward, we would like to +develop a similar method that is adapted to asynchronous iterative applications +where each task does not wait for other tasks to finish their works. The +development of such a method might require a new energy model because the number +of iterations is not known in advance and depends on the global convergence of +the iterative system. \section*{Acknowledgment} -- 2.39.5 From 8c643c738580db77edb260d118e0a2dc4af78ab2 Mon Sep 17 00:00:00 2001 From: raphael couturier Date: Mon, 15 Dec 2014 10:37:00 +0100 Subject: [PATCH 02/16] small modif --- Heter_paper.tex | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index 9d743db..ca66e17 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -63,8 +63,7 @@ Arnaud Giersch } \IEEEauthorblockA{% - FEMTO-ST Institute\\ - University of Franche-Comté\\ + FEMTO-ST Institute, University of Franche-Comte\\ IUT de Belfort-Montbéliard, 19 avenue du Maréchal Juin, BP 527, 90016 Belfort cedex, France\\ % Telephone: \mbox{+33 3 84 58 77 86}, % Raphaël @@ -83,8 +82,7 @@ platforms many techniques have been used. Dynamic voltage and frequency scaling consumption. However, lowering the frequency of a CPU might increase the execution time of an application running on that processor. Therefore, the frequency that gives the best tradeoff between the energy consumption and the -performance of an application must be selected. - +performance of an application must be selected.\\ In this paper, a new online frequencies selecting algorithm for heterogeneous platforms is presented. It selects the frequency which tries to give the best tradeoff between energy saving and performance degradation, for each node -- 2.39.5 From 0a891245b5a70f9d2c9408a0b9eaeb03a5beede2 Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 10:42:32 +0100 Subject: [PATCH 03/16] Foating tables and figure should be placed at the top. --- Heter_paper.tex | 32 ++++++++++++++++---------------- 1 file changed, 16 insertions(+), 16 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index ca66e17..b8c513f 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -238,7 +238,7 @@ nodes to finish their computations (see Figure~(\ref{fig:heter})). Therefore, the overall execution time of the program is the execution time of the slowest task which has the highest computation time and no slack time. - \begin{figure}[t] + \begin{figure}[!t] \centering \includegraphics[scale=0.6]{fig/commtasks} \caption{Parallel tasks on a heterogeneous platform} @@ -484,7 +484,7 @@ normalized execution time is inverted which gives the normalized performance equ \end{multline} -\begin{figure} +\begin{figure}[!t] \centering \subfloat[Homogeneous platform]{% \includegraphics[width=.33\textwidth]{fig/homo}\label{fig:r1}}% @@ -588,7 +588,7 @@ words, until they reach the higher bound. It can also be noticed that the higher the difference between the faster nodes and the slower nodes is, the bigger the maximum distance between the energy curve and the performance curve is while the scaling factors are varying which results in bigger energy savings. -\begin{figure}[t] +\begin{figure}[!t] \centering \includegraphics[scale=0.5]{fig/start_freq} \caption{Selecting the initial frequencies} @@ -715,7 +715,7 @@ remaining 20\% to the static power), the same assumption was made in nodes were connected via an ethernet network with 1 Gbit/s bandwidth. -\begin{table}[htb] +\begin{table}[!t] \caption{Heterogeneous nodes characteristics} % title of Table \centering @@ -760,7 +760,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. -\begin{table}[htb] +\begin{table}[!t] \caption{Running NAS benchmarks on 4 nodes } % title of Table \centering @@ -787,7 +787,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \label{table:res_4n} \end{table} -\begin{table}[htb] +\begin{table}[!t] \caption{Running NAS benchmarks on 8 and 9 nodes } % title of Table \centering @@ -814,7 +814,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \label{table:res_8n} \end{table} -\begin{table}[htb] +\begin{table}[!t] \caption{Running NAS benchmarks on 16 nodes } % title of Table \centering @@ -841,7 +841,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \label{table:res_16n} \end{table} -\begin{table}[htb] +\begin{table}[!t] \caption{Running NAS benchmarks on 32 and 36 nodes } % title of Table \centering @@ -868,7 +868,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \label{table:res_32n} \end{table} -\begin{table}[htb] +\begin{table}[!t] \caption{Running NAS benchmarks on 64 nodes } % title of Table \centering @@ -896,7 +896,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \end{table} -\begin{table}[htb] +\begin{table}[!t] \caption{Running NAS benchmarks on 128 and 144 nodes } % title of Table \centering @@ -957,7 +957,7 @@ small when compared to the communication times. -\begin{figure} +\begin{figure}[!t] \centering \subfloat[Energy saving]{% \includegraphics[width=.33\textwidth]{fig/energy}\label{fig:energy}}% @@ -1038,7 +1038,7 @@ scaling values which result in less energy saving but also less performance degradation. - \begin{table}[htb] + \begin{table}[!t] \caption{The results of the 70\%-30\% power scenario} % title of Table \centering @@ -1067,7 +1067,7 @@ degradation. -\begin{table}[htb] +\begin{table}[!t] \caption{The results of the 90\%-10\% power scenario} % title of Table \centering @@ -1095,7 +1095,7 @@ degradation. \end{table} -\begin{figure} +\begin{figure}[!t] \centering \subfloat[Comparison between the results on 8 nodes]{% \includegraphics[width=.33\textwidth]{fig/sen_comp}\label{fig:sen_comp}}% @@ -1129,7 +1129,7 @@ degradation values while giving the same weight for both metrics. -\begin{table}[h] +\begin{table}[!t] \caption{Comparing the proposed algorithm} \centering \begin{tabular}{|l|l|l|l|l|l|l|l|} @@ -1152,7 +1152,7 @@ degradation values while giving the same weight for both metrics. -\begin{figure}[t] +\begin{figure}[!t] \centering \includegraphics[scale=0.5]{fig/compare_EDP.pdf} \caption{Tradeoff comparison for NAS benchmarks class C} -- 2.39.5 From 6c8d2570f50b55f5c02486e7faad65ce42024d17 Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 11:16:34 +0100 Subject: [PATCH 04/16] Put several tables in a single float, and go below 10 pages. --- Heter_paper.tex | 26 +++++++++++++++----------- 1 file changed, 15 insertions(+), 11 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index b8c513f..4897f65 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -785,9 +785,10 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \hline \end{tabular} \label{table:res_4n} -\end{table} +% \end{table} -\begin{table}[!t] +\medskip +% \begin{table}[!t] \caption{Running NAS benchmarks on 8 and 9 nodes } % title of Table \centering @@ -812,9 +813,10 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \hline \end{tabular} \label{table:res_8n} -\end{table} +% \end{table} -\begin{table}[!t] +\medskip +% \begin{table}[!t] \caption{Running NAS benchmarks on 16 nodes } % title of Table \centering @@ -839,9 +841,10 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \hline \end{tabular} \label{table:res_16n} -\end{table} +% \end{table} -\begin{table}[!t] +\medskip +% \begin{table}[!t] \caption{Running NAS benchmarks on 32 and 36 nodes } % title of Table \centering @@ -866,9 +869,10 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \hline \end{tabular} \label{table:res_32n} -\end{table} +% \end{table} -\begin{table}[!t] +\medskip +% \begin{table}[!t] \caption{Running NAS benchmarks on 64 nodes } % title of Table \centering @@ -893,10 +897,10 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \hline \end{tabular} \label{table:res_64n} -\end{table} +% \end{table} - -\begin{table}[!t] +\medskip +% \begin{table}[!t] \caption{Running NAS benchmarks on 128 and 144 nodes } % title of Table \centering -- 2.39.5 From e3218738c399bcbe6245c791778bc81810bed71b Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 11:26:20 +0100 Subject: [PATCH 05/16] Spell check. --- Heter_paper.tex | 33 +++++++++++++++++---------------- 1 file changed, 17 insertions(+), 16 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index 4897f65..4c6df55 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -81,15 +81,15 @@ platforms many techniques have been used. Dynamic voltage and frequency scaling (DVFS) is one of them. It reduces the frequency of a CPU to lower its energy consumption. However, lowering the frequency of a CPU might increase the execution time of an application running on that processor. Therefore, the -frequency that gives the best tradeoff between the energy consumption and the +frequency that gives the best trade-off between the energy consumption and the performance of an application must be selected.\\ In this paper, a new online frequencies selecting algorithm for heterogeneous platforms is presented. It selects the frequency which tries to give the best -tradeoff between energy saving and performance degradation, for each node +trade-off between energy saving and performance degradation, for each node computing the message passing iterative application. The algorithm has a small overhead and works without training or profiling. It uses a new energy model for message passing iterative applications running on a heterogeneous platform. The -proposed algorithm is evaluated on the Simgrid simulator while running the NAS +proposed algorithm is evaluated on the SimGrid simulator while running the NAS parallel benchmarks. The experiments show that it reduces the energy consumption by up to 35\% while limiting the performance degradation as much as possible. Finally, the algorithm is compared to an existing method, the @@ -125,7 +125,7 @@ processor by lowering its frequency the number of FLOPS executed by the processor which might increase the execution time of the application running over that processor. Therefore, researchers use different optimization strategies to select the frequency that gives the best -tradeoff between the energy reduction and performance degradation ratio. In +trade-off between the energy reduction and performance degradation ratio. In \cite{Our_first_paper}, a frequency selecting algorithm was proposed to reduce the energy consumption of message passing iterative applications running over homogeneous platforms. The results of the experiments show significant energy @@ -193,7 +193,7 @@ efficiency than other clusters only composed of CPUs. The work presented in this paper concerns the second type of platform, with heterogeneous CPUs. Many methods were conceived to reduce the energy consumption of this type of platform. Naveen et al.~\cite{Naveen_Power.Efficient.Resource.Scaling} -developed a method that minimizes the value of $energy*delay^2$ (the delay is the sum of slack times that happen during synchronous communications) by dynamically assigning new frequencies to the CPUs of the heterogeneous cluster. Lizhe et al.~\cite{Lizhe_Energy.aware.parallel.task.scheduling} proposed +developed a method that minimizes the value of $energy\cdot delay^2$ (the delay is the sum of slack times that happen during synchronous communications) by dynamically assigning new frequencies to the CPUs of the heterogeneous cluster. Lizhe et al.~\cite{Lizhe_Energy.aware.parallel.task.scheduling} proposed an algorithm that divides the executed tasks into two types: the critical and non critical tasks. The algorithm scales down the frequency of non critical tasks proportionally to their slack and communication times while limiting the performance degradation percentage to less than 10\%. In~\cite{Joshi_Blackbox.prediction.of.impact.of.DVFS}, they developed a heterogeneous cluster composed of two types @@ -712,7 +712,7 @@ highest frequency, each node consumed an amount of power proportional to its computing power (which corresponds to 80\% of its dynamic power and the remaining 20\% to the static power), the same assumption was made in \cite{Our_first_paper,Rauber_Analytical.Modeling.for.Energy}. Finally, These -nodes were connected via an ethernet network with 1 Gbit/s bandwidth. +nodes were connected via an Ethernet network with 1 Gbit/s bandwidth. \begin{table}[!t] @@ -1016,7 +1016,7 @@ in the 70\%-30\% scenario compared to the 90\%-10\% scenario. This is due to the higher static power percentage in the first scenario which makes it more relevant in the overall consumed energy. Indeed, the static energy is related to the execution time and if the performance is degraded the amount of consumed -static energy directly increas. Therefore, the proposed algorithm does not +static energy directly increases. Therefore, the proposed algorithm does not really significantly scale down much the frequencies of the nodes in order to limit the increase of the execution time and thus limiting the effect of the consumed static energy. @@ -1117,16 +1117,16 @@ degradation. \label{sec.compare_EDP} In this section, the scaling factors selection algorithm, called MaxDist, is compared to Spiliopoulos et al. algorithm \cite{Spiliopoulos_Green.governors.Adaptive.DVFS}, called EDP. -They developed a green governor that regularly applies an online frequency selecting algorithm to reduce the energy consumed by a multicore architecture without degrading much its performance. The algorithm selects the frequencies that minimize the energy and delay products, $EDP=Enegry*Delay$ using the predicted overall energy consumption and execution time delay for each frequency. +They developed a green governor that regularly applies an online frequency selecting algorithm to reduce the energy consumed by a multicore architecture without degrading much its performance. The algorithm selects the frequencies that minimize the energy and delay products, $EDP=Energy\cdot Delay$ using the predicted overall energy consumption and execution time delay for each frequency. To fairly compare both algorithms, the same energy and execution time models, equations (\ref{eq:energy}) and (\ref{eq:fnew}), were used for both algorithms to predict the energy consumption and the execution times. Also Spiliopoulos et al. algorithm was adapted to start the search from the initial frequencies computed using the equation (\ref{eq:Fint}). The resulting algorithm is an exhaustive search algorithm that minimizes the EDP and has the initial frequencies values as an upper bound. -Both algorithms were applied to the parallel NAS benchmarks to compare their efficiency. Table \ref{table:compare_EDP} presents the results of comparing the execution times and the energy consumptions for both versions of the NAS benchmarks while running the class C of each benchmark over 8 or 9 heterogeneous nodes. The results show that our algorithm provides better energy savings than Spiliopoulos et al. algorithm, +Both algorithms were applied to the parallel NAS benchmarks to compare their efficiency. Table \ref{table:compare_EDP} presents the results of comparing the execution times and the energy consumption for both versions of the NAS benchmarks while running the class C of each benchmark over 8 or 9 heterogeneous nodes. The results show that our algorithm provides better energy savings than Spiliopoulos et al. algorithm, on average it results in 29.76\% energy saving while their algorithm returns just 25.75\%. The average of performance degradation percentage is approximately the same for both algorithms, about 4\%. For all benchmarks, our algorithm outperforms Spiliopoulos et al. algorithm in -terms of energy and performance tradeoff, see figure (\ref{fig:compare_EDP}), +terms of energy and performance trade-off, see figure (\ref{fig:compare_EDP}), because it maximizes the distance between the energy saving and the performance degradation values while giving the same weight for both metrics. @@ -1159,7 +1159,7 @@ degradation values while giving the same weight for both metrics. \begin{figure}[!t] \centering \includegraphics[scale=0.5]{fig/compare_EDP.pdf} - \caption{Tradeoff comparison for NAS benchmarks class C} + \caption{Trade-off comparison for NAS benchmarks class C} \label{fig:compare_EDP} \end{figure} @@ -1168,19 +1168,19 @@ degradation values while giving the same weight for both metrics. \label{sec.concl} In this paper, a new online frequency selecting algorithm has been presented. It selects the best possible vector of frequency scaling factors that gives the -maximum distance (optimal tradeoff) between the predicted energy and the +maximum distance (optimal trade-off) between the predicted energy and the predicted performance curves for a heterogeneous platform. This algorithm uses a new energy model for measuring and predicting the energy of distributed iterative applications running over heterogeneous platforms. To evaluate the proposed method, it was applied on the NAS parallel benchmarks and executed over -a heterogeneous platform simulated by Simgrid. The results of the experiments +a heterogeneous platform simulated by SimGrid. The results of the experiments showed that the algorithm reduces up to 35\% the energy consumption of a message passing iterative method while limiting the degradation of the performance. The algorithm also selects different scaling factors according to the percentage of the computing and communication times, and according to the values of the static and dynamic powers of the CPUs. Finally, the algorithm was compared to Spiliopoulos et al. algorithm and the results showed that it outperforms their -algorithm in terms of energy-time tradeoff. +algorithm in terms of energy-time trade-off. In the near future, this method will be applied to real heterogeneous platforms to evaluate its performance in a real study case. It would also be interesting @@ -1218,6 +1218,7 @@ Babylon (Iraq) for supporting his work. %%% End: % LocalWords: Fanfakh Charr FIXME Tianhe DVFS HPC NAS NPB SMPI Rauber's Rauber -% LocalWords: CMOS EPSA Franche Comté Tflop Rünger IUT Maréchal Juin cedex +% LocalWords: CMOS EPSA Franche Comté Tflop Rünger IUT Maréchal Juin cedex GPU % LocalWords: de badri muslim MPI TcpOld TcmOld dNew dOld cp Sopt Tcp Tcm Ps -% LocalWords: Scp Fmax Fdiff SimGrid GFlops Xeon EP BT +% LocalWords: Scp Fmax Fdiff SimGrid GFlops Xeon EP BT GPUs CPUs AMD +% LocalWords: Spiliopoulos scalability -- 2.39.5 From 7cbdba88cb895d373a2bb5a27084fc687d7f7476 Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 11:33:23 +0100 Subject: [PATCH 06/16] Fix figure again. See commit 44e73bc744e4e46f7377334b777c1bf8988d6ba4 for details. --- fig/three_scenarios.pdf | Bin 13526 -> 14890 bytes 1 file changed, 0 insertions(+), 0 deletions(-) diff --git a/fig/three_scenarios.pdf b/fig/three_scenarios.pdf index 4e2a400873dd5b3d4306ac1e0be1d80faad69869..9f36c43425dd73f646c6b9f488188088f6d86246 100644 GIT binary patch literal 14890 zcmb_@1yJ0}(r*YFBuH>w++i1Y5AN>1xbET*2o@kX1Phwr5+t|>f_p+B5Zv8e0&mI5 zx#!$-zxv*LRd=g)|1&+^Jw4si)Bmmd4Xvt_G%JXW3zfEaH}*X$HvkB5Hn&3+5@J^b zJ6XHg0C->&b#`f62RASTz%K1z<_4AmTR2;SMMO}cZV<4UBdS-%u>OMctP~bLE(%D? z07pF`86Y~ZB~V@6eu4Ld@@Vp$hMYkqwC*)q5)S$G0cUe{Gx-LaVbR4q`8R%b);$X` z?Nz-{&xuGlAMAkR`|l}pD(r^$=h^!ukuw9qTa-3mOz!9I@5%xeA}CRcYxbF6-d#J= z=gUa$NCZq@<)ApB_+X+WxjVh?w>+J7bIw}Q6LlLi$*YdM>#m5VYTq%Ya>r69+^mNy zUo3$au)e#v{&?>lJ$)#=>Ho7J7eSl?cO5l@RHD6Y_fFTZu8To>`Cu6uV(h#2d3hcZ z+MN09{jLYn>~bWwf1_c)c=+PgQ3XTlk-zKVnL0*?!5)5?!dNixi0in{fy|6E4gzvA zv_W(qyz#>Q#o;6BTy^T}F5u<&DXJTV94$yW?4YaZu?^cd_sipG}- zTiJJeO8>?<%E`-cpo7_(0J#^L!{eqPJaol`Oj{=%; zE`yR1;Z|lB_~m&K4>;l<;OX;jyTxCO9_-e<4%o?(TH;2=d}fzxX+`YVg7L z0NVTy=1yfC#1*7DR8cVr44Wc?cZZex{xwnjDH5zPPm^cT5U|2k+f&1GIA*cF$x|V0 zb>SxVZ^|VJ62f&9`IDFu)$=P2d2@DV1M+>o+xLEmmXZv|NfMDe-RPFa3sVoxYupGsdLZ zj35g}s+WQLil(P$!2MB~j}v1bX@qAY1!s&w^E1K}Q0~)JInJ2RH`WR{<5I)VhBFf> z@;PcnuV{0=t&+R=G)ck%cEIVlw9JzcuNv!0*%==OQL%t(pN@_^9^-y2Zo+V483>6e zR(F9<9NzprDFe1|{C(RKrs!ccD`B&)H2vH{z8Osud9hqojAuEsG;qzhIox(qQ_AWZWvJHnL7FGSfYB%Tr?Jo$*D@1%0IViaLKP&3wt6V_3A9EUcri`%!!! z6RsUoTu=J-Xxpt&#BKgb(OkOHM??$|AKWF~59eU%G}7iT_zvme&(aNwt9h!>rqy09 z1`S$v$m>i=rSc=$^A90)vwG%0gKKt$KLhp7_SeU+6m^-@WNBns@+Y!BJE+b=X83t! z2VadZ?Xb5`T70Gz*361Rq9(XSunJz57CIQ}>d4`HHhHyMEbb^jci3sV)``$zmZ+vZm2qdH zrH?}|F>pfIK(d{F{MMFrcOOoYa_9<}JgIET!{F3-lTMNcULYtaou9s3Vs7Xrs2F_O zFI5=;{g5wy*lMiiTPTzo}Q1VbijWHDQqQ`$To zi9Ag?B+WA8DKHh|^$ed;243EoVf=^d$qp)PWsRp)@qjAD6;Y&WaaR(V*VREatddpc_g^b2@Ljx`>%527TTgyB#*w%=)!eW-CZSFy--9+>+(o{maIHKDLX@a+T zMxH(la?e5|6;1@G9vtS{(EH%Al}|3(ag({7Gv#Ndao7&^!vafn68H%JYy9OI%lb27l?tu{=lR-QJ>KTUSQkj`$) zGnMc1_I--Q=OILhVqy=Hh;Ff9m{n~}K1%d@)bfM`Ths)(!;}Z0%oZ*{sG=+TJY!iK zD0gmL2)`8xh*@z+IC%PgQZ6JE+mS}vi#*9T2Z@OV>1wiM`vZ4pQ)P`f+W3_4>O>JX zelYGj|J^NG%;gjh*vayDHGs7qD#%0KczA-ixqzs@XlCVBZBNXsZ?bK0Wq~G>E zOz}tcgK2{q2>4|o6u_=vYY7Dypg!2Ij%oyeX>fl~`8UxA?Y|4>ACi)0Ze|Y7)_#w___yJ| zR{DE-bzam5UFv_S{0pLI3Kt_AhkOQ|>rwP#%Xa`mCz^(3%kJ}0~X-FS1Js<=8_3pqyb8s1+a?Z`jSxg{R> z1q6N#RR7Uoe(1PV3i=xt7}xI)LzT8~b#>U%SyNbJ*@iVQo$r&qzi8CvUBh9TR9Zdk zz<b>{IG}Fb;R{Gm6ezkZOy?DGdEjjCk-&f){0fa*}?MHCYJ(3Eg-fo zZq5(@|3i@3HJ!DbY#(0M0H9wR@Yk;X#R|&`TL{!m(#8w|0I@5Y{cdss*>!9!-E5!+ z0B&v`02db*fPGXle|$l}2c`!_ zzXu`Jy!sI8OrFJ}Tcd0>=>2*O^Y!Laj<;~yFL(Eq)vKVG!| zo`I;pQxF6@Gr0MHeRrfKj!X_J3ro1N0=iQG$suXtbA=qaCceP4%S z7{(2MM{@Yo^YW?}39hr9CR#3?k0-){S82|R!#0CUDU(hEZNs`=qgQUO<-6^yPr=q{ zV#C>x+mAE?4eNp}_K2lDDDAC%0q4%J!On*>wx)~y>QRxKjE5g64VE+t?K}RioVv+# zG0KGUugr&2e0(4v9(&3VRfs3X{@b+}w3)e=UIdiy3Dbx(`<}F(*uM-2j8hkBvN({< znD3-oJ&wY_L=Qc#kEPT+|Nf?*5Kzaur$v=la4T%_tlD?^6zqODRZGpm7q?k~Lf+Z+ zDXTd1c$5Rh_9ewK9Wvdkd;Xo*XanLQni?AMo4fUK)mxhlL%$&RQpU>oWc=Lu9YrqdDvM1A_GxX` z<>@BsE3NKmV5iPAbYiueY@jek?pTgm4Z&X6?{{8`j_gkcU zEw_EhcFDl&iPC{Vd4<<#L2v7Y&?7BbwFm+?$Ct#I^Fpaf5`ZA5`4FUqGwM z?owF>$)0YEfAoEX1W0a~j@RL%H<6)9eV zX4Q>#<&~^6qem1aP)}&p^CwQprh3VmvOz(Eh^%}d_p+GCC4c3pi38$vf}5R}3qjf% zJLP1Pt2bjo(r7vOIl3{vc5oY-W={b8%O!|GnsPNWgr6hfC9CSrjLb9V*0x73K16TL zxvS%qBkC=La9dX%u(>(?F+~|d_Zg+ zz}aYiC2c<9e}lDKgIzwj|N6;X<)^v*@+GGd8^F8S!Ldw*ZIHB`U=#7oT=b4_(*O>9 zgpR5QhjB}Ebn_s&H9Ql6#t~C_^^1yBKQn8gVejcIWbJ}otoPtUon?3Ajl-ai=|l$0n((=PnN7Ce^u&y45B2^tiUY-TAUf* z40lFz%dYKoA39Fh>y$KlCpA~S$l~y%C!`|0`a$@fiRBH^-q-!Yn`tA_N9lwfH8t5A z1PFoqYoI2_jXC?U^|?JGJREM_IF;#xyccPz)K0ra{TG3tQazUy_#FZSzns>>9rUhe zvw<7LVt1D53D~CqPs;W$qlasKyXN?fDo~NG{5;qWB2ug4% z!6TIpE|eqfEnh@gv~h)-zQH$I4BT^DFa|%E)sg^1iH;? zp`|jofuemI=nCAwtNL+w>+~L>zSaq{GG!jQnByOPy}wYfuAofxp6h#p*JlGNzjR*Z z6rgeqcK^ygZWY5La#GOXM`kQ)9Qj~gW9RPynaD^3cL<7c=t{JvNUka3e+eG>?z|p_U?jNN!6>c&(>i*I9Z9eVjOO+m*H86!49Z8`P zc-p2$#9ZMw=ze^rJh_YkD(ECbPN^r zA%_hr>iRlo;sqWdPM{!p+Uk!m>mC_9D#2HiQpjQSVXv$k{Pxf7VkQHqAS;K8Z_H6| zJJbV~FmNKM?(3}X`*Td`a#y}gFA<5ws6aCBJVd|q00o8k2Qd_%x{`e-sf&qq2eB z$IYGx=a@9%$bplhglQwUhm)aKK;9)NF4BdtU6@G#73?;WwFZ?WC zPPd56m6(EnjC`HphT)Kch&(y-J%A$i7`jFmKeb8met^73sDy$n3>lqy?<|DEp# zHP+Q98dzqdIj-R~y{4iN24sx)hF%2k0k=tyYv3>>**e|hSS?ZW+`_lQVl00y=8OL zL+%BS!lj&lmnVhu!uDbe1YB|vr zZ6?e7aHZ34;>Pl7D2hTn=DB(s9PX%<%2MW!$ZA*cMclw!>2*d#FStN=*u~im-wGCq z;2JR%zMd;gsoBP?RK8kyG$$m|WS;wEb}v_tXz;_yRlqdwGfb{zqfLX&XKe4pe#WLm zW{P4OZQtr`&W$YQ*C{|#%u_5=fa6ze7VgsiFK_%^&XDuYkQ=TscKud1oa4p^H$vDm z$7v?&TPN>-e6t^#ekDXN5Di~(-s*EXkyvqetL42)y?cAO9Y%A!)P@;xNEy=-g~85m zLKUA6#^zN!7*AF}H2x5{}ru4?HC5$t!k|H%2Hem5iNj@m_qhedE)qBwH^YrNPr zInynJKIQh^HFV~14Q_TpXLUEYZX_0-Li7p*J+_FdGdC+0DE41S{%nvn+v%`>Bmgf= zcy@XbM=gACfn|-U{4`Bq4S-5?$OTj`cxwx$8Zwi9XJ@hjse@~l9c)Yf%&G_to@p!( zgc5h^qS#}codE%o!PS9zVSAaAarxun7?}=zV{*XlF3N92Ko*kljxp;gYg65sXi}RINwX%Mqg}_S^2qC^Zwm5HdR) zygsvbU3`~uC-c7{31OPj!VHvG@C)tL@^@d|<6VNq<<2OLe?DD|ryv8>Vkf}(zH${ymc zW9gGC`o69y*>DR`uJ%1%f5bS@OG&|VQ#l*vNLR+QM1KA>8K@kV!x=&r6ZAo#1OY?e zf^bWhj8ecTpH@}RXj;Et>`i(#o?d0X0dG0AJAXC*#hPCn_4ryI(~U^<5latMPd?#F zf~DsI9lb}vfU0|uC$>`qzS7!SYb)PaGa;(sCimL`_!q77-JLHlB*d}K&X(sdlSt)V zf$ut*)D2g4a0!w^P|0YzH%HDbJPWdPhn|LZm(lMliuPx14%DMKro*!)x{nMjH6pLS zV!q!qL49@8{1MQYW?M>9BwsI7H*lb#kg<<~ zdx2aZ6`D;H(8&wKDtECFwZB2La%2ukFS>*_givGY+2eFFzo!uN&!l*A9MrZPY#0q| zG$&yY>`j&1oO#88Y+EO`Z9+*4e5$U0Ajb7l;VGN3i?v3!h6c_D0s7DivIC8P&Nxgg zjJ2CY0`74Ns7caAeJY0)GZ)SnthDxMubOpKLe;lt>Q@h5n7uzl4p=#w;TBDz-vBNQ z5P8`hAl03hw{axa>18{~;q3FbZg}2{WPIn+GZ379BjZks>^8C5A0c`ueCuR(*gPV z3fgDX`=$dtRHmC=(EIR`mp<93QjX;n_MNj+87PTQ z=DzV;>)+~s`@`2bt@VP$cYyUe|Nab|sm45d@WQgl!ruUtF4eoO-zJfjFc-ONlc*}v zv#2#BKq#0C-i=Bzi^HqetC}dWAo&J`D2Db2@_`}VKCe2*O0;&+8T)szhs;2lWKEU2 zHe6;{-`s4fu((X$D(C4a7n->d=&Z!iTDp=9Plhlp>JZm|clCVXvGIj>f#nE1X|<^^ z$(1TRkuG6*^LCfw&Sm$9&g3^6+)UzXaq8(bk4j_%)9CSd0y6Bc8VRTD`cUE?gBIc8r^G;NDpNES^B5E4xeU^9u}aFGrc z4@J~U-+upEEx$I;fQ6Njp?)H+%c_pDN$lP8>p)S}qpj}w`LqwM<2hLKA6nmGF?4G; zo#Z19zOK-Don73O1CUn5%g_Z*LV%YXPL+obX zt(VQZ_>Zbqb!3$_Ux%fQXP4L8Lyup)GAW$&^P;RpYUjLa+%NlP2{A6cs>`!S!Wz=M z8t;oX!Ad6Ka1BNq%+$@7)xzS=T{}9K5%P}^9epNv54h(WCG@&QyMxb~MGW2WE(q=| z^}ZFl$H)?r?~u;w^Ws!#yD++=&Jz0E-t~-?i>7@Mluu7~ivpEm=SJZQi&21U`vLbea&F*Pqb~A{n zrxIFun)Xc*P^0;9H(B%K$sV8WXJ%!cj|>fYiDO1cBEBfa^zv+!qoljPeNCwR9QQn>Ob>lFPzm%(7M@110V^{QEg?8 z)Jed}@%fZB?`@e`bw2HcDX>13#iypM$4;ET@KK6svuX2~fi7Wv4qQcwz$;s{i*0oDG;}nH6S(O|`@vptfNgoodBo;~LLQ^Xp){T$Cqbuv@LSf zCikt&6^ejY0~<28x?_7ke;!yHXO$*0Yel6fe`>sC)Gkz$F+Y0*X^U9}mGTrr;RVkwa`_K%}mc6)Ielasr z@vueHAEZBHQdd5%ya=1!`lJ!FqAIWvhe$d#=AuMtB z)cQ0kYrM*=RS}J?K)mOzp1DPF0lHm6CrkQPe^)wZ70Y?eNfGie{|6cvNg zB$Dp; zaKLZvMyE2Pjep66*Ax0!>|oeDf*Dtqb``P1%j}sNxQd2HRV`WYWPR9e`;}qxu^>$c z%bN3a5PfPE-WS80lA83#fDC(kCkK|OA%(LcyfHjp*C#Pt@XAV^dT((s`qfj4`svn+ z68e(xX^Rsr_*38Le0u+_zcNZcMgQdo!Z7SQrYiMfF(w4@d!*-_z|bPa4=Ur&T5&3M zqbpHqtK_5-yGk8x>kLV9EgWoZQ5Cye(ba^t5B2eB%hE9jyt8;;b>(%*8--iSY?Eu-BEsC$)D> zt#q~cam%RO;jXQg@q- zI0>{zhu^};Hb#$9IueWDB&1YZt%it)lgp&D@(au5e`3{-jw(ogrz-n3YwSI?82S$m zjJGR$ne+0|K*Pfvk)2o7ubYdu8nWN7!Z|bGq6Ai$FU_fNJo)+4k3ZdePs zPh94kj7a9WJmj`o zH|0>|(%NIHRa3+olkCMLmkB5JBvSJ=7v#B7&Zjruu5!AhM24*k0-BgagAYRx$M3^H|50>hgDPP-E@r_7@fl^G~xE zYc&yq-APZHK0s{tYU$POJLL?0AjT^;-Jcw4(}2h`y!@Y4Uij^PUziIzSWrH(afU`uQ^X)`d1& zb}b)|*ln26TgXC#Dr|TGSg}h5Q8hmC+uu&*mG!4cf?BW6Ptj?pc=HP)2+h}O;M(8F zYNhz6zSZe*1veHKH=bsu{HQiRcD=OXiu-avrI5^&%FH3iBx$ckuaDcd%-Z1_j2xUd z7zTZ@)>bPby;6PES6=p0Ps*(BZS_!;elp!)5~z`I1~_RxJ1@2iM#{wKN8zZVGqho}_{ZpRAI!M&OG^9L(5 z_+dw)1J!VeqYPS(R%YS9-|1NHPoTGKier#WqH2!$<@pX%uI$=kFT#|sSqI#^OdSl z3VAQ*DO1oE`Dj8NS3{@9nqU^;s@=egw*fAre3{^xY{HT0Ma6Eb5o_O#!9vt?zm2N2 zLK(E(4sN}I$JvQ3`WlU=A7)=R78y8X>TMV|)s$N)(ozj}gjm>5t2LpDn2BJ{rDQZgT?0y>8XcXVy$c5$&`qL`n@y;N8m1&C9@*~dX15x z_x4vvt5TFx3E9|9-Kxm{IYuyQ)Z+ z8UUUlf+O90jkPF;+E$M9tJuPm^qjX2j*dA(b~;G!e{k7$o_Z-B<3zd+?_};HX6D+- zsNk*Gg$<*svP9~|%fiY3^m_g?bq4I>dlf-8^SZf&W8rE-@SuT-HyM3aP3Sy@I=Ht- zp`56?6#j0)sdc{f2Sw2ckAVA#s)shvs8k8l3LKUpu+Ezp+dSM^AiFmWUGQtA`)C`T zb{|B~!i>(F?iul-rd!oy=lN^;H%ev~&V5slWs&b!S+p}&Mx!m6M$v?oGPJZB+M zM}kj>jiy*rSENwYG+!||k5IW{QW8CdNn;QHkW@{!f63gNJ1>;-+!sq9!GXZ^y5A{x zFlcV-Y?7|v=Ep%r!m_|nY;vB%@UXl*gsXBy4=1I?-`mP)SeXGI1@5era4L z8&z)qxkAx?bKVS#FsH)F@oO@H%v?V93f#Ud{;JcXPAj1hK^IPP-RrMk`;h~DYVy}M zRVe&6zgA(8#286HPCWedKblzv%; z0XycI*K-H_BV;OnjXa^$jc(^~OYIgFKTS(%sSLVuyeDpyETZ&T*zMFDD&dzR*rT$; z3SB;~xqVa5D`YLH?mF4K7a~?uFQ3H^g=gdUqwKXA(s&3U{Rrjiv!D`@@+HTXBU_8D zrAj95JR~Quy%_SxR*6VGCdW?W0_o7mQISWFF5+QR;KoLbMhrOw4eAR-Oi-ysUQ%P< zOHc(Wlj9y0<`GSP38WKfqFb)t(hTf{QZ*`ckuAr`ipd^Xa`oi`qnM!Cq!X&BWh z50O{vUT)-&Vo@|L`y0UxC-LBmN`^g&?c8b1Pe|uy^B((Vy;V<$0cDEf{RQ6hye;0ZVrjps* zA^hSjJ;jR_)E`I0jl=1GXm1iZ&J(P3)70|Byi`)+%N>Vb02VVmV<7TlMFPQ|eY3Qm z2-X{K^6aKvB0d0?`F20?^Mrj){#>NCD$AiW45j}eB865Qqs&2vr5hJ=Mrh<4hge%DfG zQZkK`kjJZM=$C1^$H1>SRDqI-VJjxGD91dFish=MaS8c_X?BPSqa4NRm83ywZwK-d z^LM^feN4rSZ@>O-e;sgz*LJeh|1FwCBW;1p{f?!LPI&M>B)c1~$t>j5sOP8jM`qgq zoas`bA{yVcU~~{^X2NKHM8<$|sT}Iy8h-VswQC4~6n9gw0BdT*amymR%M zP6H?NRx`&69be+)7g8#+^-B$JvY{^_J&1waACXqd49$`U@tE?g4sq?(0i8I3HSyF^ z%Hh)tv3=+LPY0C;j1&98(n+&5Xdra8N_j*$M$Is>$8cL>09k^pWS^}k;__y2t0yHY zRU2qQQ=1%3uU-2HKaLH(!);ED$zL$Mp`1S6zsrB}csIG==%;S$sc+dG&dF47tRX}nAs#CnCn`d#c)-?qVQtnpvrIS5Q&zK>katb<_#L?QDZ!3Q-O~fj zC7%iDiSXBqT1(Lb*lTRG)!7}GSl7aO;+?^{qR{2dw8{Jl_f{^N{{uTepL*nI`Z zo)FS(V!R^D| zzP$x*DAe7r^?mEil-T%e{YpOzm&dztUFiP0#PGRIi~0DwcM19XwO}6Sw9h8J>8-Vw zms8^!r92|B$?eyu+J3dH+}>L^&UpnoyWVGP_WB)*U-Aps{i~o&Bjw`*6xT<&!>+9} zC_Da-EmI?2A&P^$kZZi80!eVwar2}CmHN}t>f~x&v|7vPWQR-SJ__jVxgtLE0Z`ke z*eX*fso{KgQgm+@YMZO7fy-m*>tVx3DnB?fG&H1b_vT%>bV_D48j6T*uRDsLGhpH` zZxCs=Bio&;vk*?1AJlJN8AFp?jdyms+AS=&81d9{DtDW5s8mt1t;n!x_O$BF3Tw(2 zYCu0-!jq#w{bFqPGw`vWSlFODpG??A$zdw0rvhUSToC87btg8(2HJJZWz3DaNMpbC zG2T%;HHyrbh#Mh`EqR$Iux;M9qsY*E#>eD>GP0GI-4o(>V3tTrJ_DAMVLuvY&_91> z48hCqX2Fxt1D~eFPj67uri)x$Una|t@OaO`Uj<@JG`?Pxy(;*eyT->%TR%HjqdfZ! z$102H1;p2OMkhHLe_r?cXG*6w8k-k2t1#Eev31n?@|Si(o3tY6+taB~Vftm!=A(C! z4M-H9M5&*zvhWYys-yvKe6R4Y@W2trAGbv&Dkcu>d}?G3ie4b>4x4O8o>E_1WAq(z zm6}t;=Mmp0mGj@-kfZlJW@Zp01p3P->E}++Zef&tdE-#0tD!#b0DJ52rT_V8^S(lX z?&n4l`kk1_fnajVQpDD?0gf#>F>{XaPk77cGXZOC*<>9ibDL@EjL#Hi^H^L#SP>his#HL-K-=Ysf*^vm=me&FF7VCi+vlx|E?FTx$hVV3=UXJfh9xf**KBjNV?Ap z6Tk>J35j~Sm|56^-2mobYg;E_`h(_ndVsB^ zFue}95>UxS5^Q5D>jMF6_$X^y_}E+UThfb)pbCka!LY|*RxdLbR;yn(R-gbFUQd`x zQe9D+>Vd&az{}Ca5eC8qcsV*aK?S^o>3@j`z}gRFc6z`s7B_ogdR-+|fSa?kgT1XA zfP;;n4a5rKWC2(~VE95$XNWxj#KuDpQwg!O5>SWXkN;2z>j~4_xVgCqu(Nx5da`+P zvN=Pn*+KmL{Omvub`B0!7zZoV+sVz$i`5BA4V&+mA-1~z5_5vG{j!P8!r77CR+k;b z24w$pqCXLyFzhGbVPZ3LXLmPYdUtnQO94)9AP)y8zd5V91rG-+2n4obg~2U>teij& zE-*Krl^Gu=_n*_b{LRdNn-4Ysn~)Rc2Utr02C%ktw*W(g>19+E0WvnuP`5u2rT~66 zAU8ex@8$outb)bnfA#__pugQ91mh7I~6@Z3z)ACwnJnPbYB~7a0EY7Zmr8 z`F;ug<`H6lh|u4B^WXNCnTwLr|L!kG$A=C5z$YN(Y~k(*L$1n6{r_y2|8|5eE&gS& zi#x>OS8iBZu!9}I54J&JK?cDrwzLqidH}q`cDC7H9%hH_NO}<=*r1!On*&(nSLTU3 zxc$*^uzkoR0S7ZDYhijXR!gvznY)7Fv#;mj=)Zshb&=` z5_m{45C@1G1mfiQm0k2eAU%u?mWBT-$G>hTctOBcs6YUS8};El`r`xO;pXP#23P@p z(?EPYup1E%AAr+u8i zd;jnGJ?A`|$z+n5Og@wB!h}XaRGbCK%8pLcmEYCe^}Q<#9Yg`3ur;ti=i_6OF}5*< znNfh@93?h!b1Rs!1Dm)N6lN@HY-np_EFggH2y-xoTBEz9|JIa_TI9fNy;a|Nt6!8$ zJ%|q=q~LARqNroXI9E8IDHH=}WjAabC1Lj`_OM~CaldB4=cTn)x)Y57$!d|m_q2Ra z)8yarSN$F=QbNI%-}bP4cPDp0y(2P)(<~*Sp!aANzAHKA9GtzJDv;d_AfS<1S??n- zJ~4704%G+th~3Y5c>OdJ^A^1C3u#?_L}kC+WStR!koONDu5AoZ$(@|kzQ9NI8_6En zx7`aSFiLm3zZ+Ea=a|nV`xPQlv@q#UIj}ExN#{JZiWDi={4?sxT|9=a`P}21#BVw5 zhc9{x3*dUyU#3uuo}M84XayH(0A0!YTiG`^ z{vys!&pj1ocpZ@3Uee0M4@|KQ&o>@e7MY0o5RBhI`7z6TU%jpQmE)foS@e@_Evl>| zyzmOJr60Cx=IW}m7&?>WH8RvT);N)SVgLh2_6FvM0NDrY2nB|;;{hYNRT!BEQatFJ zH~kmE9Kvi-aX)4kI9Pfg&*m~mJuDWV-*0_-L(eVao$wZM*-(@3r8!XC@~0^w&=(XP zBi$$bxtEF6U{K=YO3#sDg~dX<^$J`47AAC$c%5%a{hf+oC)#^4RzL7^tWm$%muq?w z=lZ7E68nfMmLtDuMMF1!ACWK1TMB{Fzl1qAai^@}v3@T8&`q%P(0~n3S_!X`vS1wZ zWt_3IiL}W0TLna5nK)~yh8q1|BbyWv#ADI_Hq8XOW*n^OtL9_}uc?nM@f2@l6j+53HQsH;{maeq0A zNIsg|a9V$h%aU;HE@uL!Vy9oHeL-nFJ0yTWNPZc z4aM-3q>WK*AU2V>>_JX;<*!d1C?}mP9m-{H`ew$IYp1fzm~>gKq)uNq=zE7I56J}T zS8p(ZEQjj(9>VJ1CU6S7OE5V(eapc9P51J5iYXd~HU6|%Uu2F|8ACsPEJMXDRG!I6 z%)Tj|7em#a@Kxd#SYtWqFO%<{Oh)HozJ<0T6=nt(IUJQyF39Mzm1k=Q9;m(ZEb=jvUCQ%Hcb9hW40)ls3=ITd(LSqtEA;e|4O zsFiOC3g6iGS8|)GFI&ImpXnhx@_u!uNq6J^O6?-E=?~*6N7^W+^t_TmQhz|uEd-fL zSkkHF(^9Xnr@t%ofr->aoUVnAgV_j)dU(n2RZf1RYW1OEPU~mwst3!>mvv^yY?n^R z=y5OJUBWQL17i1rBtCspSmDz>?z1XUxv9df5Fs;Z_=seKKPG8yNHPmOzj=AWa+YXP zLNTZWkRJw_RqM-r)uK5Rz~5VF$2im5eJ-uyK%#~TCPI|N{-&>Crm)0leOaZjvfpA% z7yq3hZxD;Xr+!ou@Gf4ZPHNthd5RNeIn#!9au%Ptrc^(3vdxgSCS#@^jZway$?RfN z`@L+#T+1PE<3jiSD=h!2KG}!@=XVA}8eT%LgEj-QN)<<=)lTXePV+F|QM#`Df7&(Y*{8dv`u^vu+uN76xnr(f|t?#=+t z{VXo<)f<8GrrZ?W*v9DZJ3E|uIt2H zhozv-m*UBLyWCnQ%W?Cl*XE^qe{&IqVq*FUlZrG|1QQp94yEv=oNa~(T1y}f2i9Tw zW+cYKB7&ZH2)%frIGuy=(H%bJ7AG3RX5PYOc=K)YAIbpsBdU$la4c-hh^SkPSEngL z;{Aj;-VgP@MQboySLMFSL&eqPn4&p`%)?Tdwgza+Ec&Km$$z zg0w_|yEbl&g!53@*}&Gohj5YP>?S+Vo{$K=-=Xjs7Yxj)8=98=TX{#vEK4-F@-(i% zfMBFkk|IZ9?e=w}&%-VrQPKFv^M^7+gx~aT#QgHR^<+l;er=&?$2fi2vBfPmH`Izo zOdfB8+&MkuaMH7N1M$ubYeLg$+C(MB)bY+H^O)*iW<~(I)bXAXE8gm%@ncnwWh>U8 z{`Oy@YgZDL# z{V12f`AV1_q_1!!S<$5!=}4ct2m( zgLOt!s$(EKaW%xXQpM4Ogc@b?u%DNXRizKf%!8yaZy65B5IWms10D!b^KZ0V4Yb?3 zdzJqkOAmOoRy@=V{^G9cJnTr^*-{*LiW#Oct6^*cq!FfQ|8Y&7+I zZPuNR&|*U;ZOltYgHo^U+2&b2)KfVNNbTV~G-4%K&Y}jv0x@i}zDRB@osk)lT0+}C zmO>Sml%wZ(tzDti+B!+Y)VZPYzI7Vd(j;dgTpQb$oXb%dFY2%k+uNaVo^ZB#=$K@l zpE#PvY6wQ*aQcyqxWhXda(0H@9loql-o%>d-pi&8-RuHpHOYm$7Ekp!{p>Yw22w}% zyUq*`KazDgEMlnJ6kt-*qi5dp=j!u+arSCm9-bb(9K}x8Oq(zv{n%&_s2Lc+@X z&bs5|wrh#{yXpqD$b|i@<8(ek2iW=3(eU|u-$#qOlA%M;C$_Vidx>GR z*NA%v8I9VoRaAKlGYpEVhFTr_bP=tCqG+7+PrV-mt-0$K4R-Tc90yr`HeEJBIY}0p zwhTdD7tvZsq~vcyhOH6Z=cVd8hkMSwE1MS2XsOi64ia6slc)D`>{)3?pVhvvNaOV> z7-Kyv4cK|a@n)Wd`Y`8(GCBJbGD^Tc--kYC7#36boZT8gZ9VpR8x4d>5jx)+M0Lkq z3+({Mu(jBA>8$mbcQV6LmLNYbss&C7Chxq#2y=oN$)z7ve3the zaH#mr<1Z|oB#&?Zhd|^Dn$u536HAs^)kLlp2s^IhpM~IG`4M(HpVVb`g0K!hIb%61 zi#*5L)D2qIo(KUdEsRgk9@=OxOsXa(e`xj%i#u7fHdxg8Q%?wJNDfoC*XbUCJ*UDz zDknp5mPPjHiEl=?!GF!FSX?}#Vy813nb-g&uF{r!TP}U9c!!_kU?Eq#YHW@k=9fNS zxitMdOwWtl*Mt=aI9E_C2ueyLB;!LILko%dppsNPwbj6SS=qF}-}58jJ8L7yS2D@L z7selWT%Fz31G}-vHkAMH?r6|X zg^=)C+zHs3K1kx?JavLiCyEAu}Y!#K}EPY{==)t zcPbM+?^STy?!$e?(x({X&~OD@h^aygH4CBaE~(w-2cQT9S^T0GxDmRQ(#3|^hQ}-K zI7KhF%MYurA8ahv`_PGfz29~nGLA(y?NQUB>&UE=m4})yLE?2~cc<+e{U#FhiDtv@ zTY0xWj^E#^w7q_-NE#7+D%aNJX!A36Hqr2k<9-?6)K&nRLJLx3OoP?c#Xex|U|NOov5do_| z&<5-U84c@B{sXl{fLw4kIDjXe8jl{_)pTI3GEw?H&qNwhh3um64 zC?ok9)0jS&e(NC67qjc2l;C-3!1uW-#Gaw2o_cPnt|K&~7^FhFr{hCXj;!qJYRmTW z4{LxQfm=l3slHx?a}HpqMIk!s#?9XLPU>X=SF!^14FfAWeb{Fy;lP)%1~609N?v7o zW|QDfgBd($6CWda*foprLXZBSQAV|c%b{<>>VWsHY=@n9zf$20nK$9bL(K`iksE z+sF1qHc>{Hi5BjG{QbnU$sZng#@TP#DQga-+$kasrVb}@{d;#QncYQw-&~~Atx&l{ zx*KFVk_Gn05e$`#><@n4_&w}Sc~skcw5cp=o|#2|_psvr?9;IPWK z8}>Ftc4kJ7V^dMf_a;b#l2biTpu;o#mmy)1emHS-dV*d&*r7q3$!wt}U(a2eWD=xl zK}&=z1}(zxt=X38^ib{7@^2v{o4H z2ddJd?#1_S?h>Jkl&F<~^QkMuDuV(QoOU{ebH)RN6en{6g)4o`8>4l(b%Mrqv?D|k zqkmYlYvC38=$S97^)6kfkV>q0RCgzROv_>WIIaIj#TG5by5yVpv454No~~-1BFth4 zC1S&LzJuJ#k6%tg7bf~az6x6V$0u=!sgn3%k~mv@>!Vkgsa5j^@{}Yr^O8zX+-qj^ zIwisS;j>rl5y74H1H+@e_mKBUg(wkMUPpGP-&a{D#UEcLG$&XR6JN(KSbQ)5FW@X? zk3G8&76hE^1}-nB47W_`<17!iwWO>XRm-Tx@N|tuJHq}=0JIVzeHb zT*R(b$;ik=|LXOD`Pjs!`H9#@n@F zhII?;c6Ig4^mz$YL_^kozE0nS5~15ugdH&lwd0sin~B0-9-kKwUDkTTOnCg%Ed< zPqcoMe$$wi2H|D4P(>0-khwy_s~Xh9;RU*4G<6^o(1k;zVZx=WM7I)R`<8{3L2K8- z+@jL+Vcz+u{y91IW9qn4R@>FVQt8qU-CKMe#_PBZyBB*JRcPpk%V#EOnfrnz<}Pk4q@eC*up8CVr0?{>;h;!>0&M z8#DUO2geLU200Sdz}FFxM%CKix9d8+z1E(9n`nEw_Ow)XIS~rc)EfHZ(v8;fsCN?@ zcv_zoeXxq3l})Fi$t(IpiNEzwq@7ht#6{lCOcqcA_O~2lFdwLkbr&z!$w(UXbcQ^3 z)qSRw8{(N^5B%N4vF1pHI=_TZw;}ary#T|G z^JDzBHt8zRW=F%gh(FJhg&i2i0)q2p&~xuwlep`sy*) zu{dfXxBlT~b>!{aZBkoaBkPv{LWx!pJAiehm&B0$+)1m*A?z1JuYLOnw~-Jk2fOw( zvpL#CC2?-dUCDPEFNcwN4QZ1W!I80-IDp?9VFlt1? zc29XgCrz2eN4yhQ6g_$}|J< z20_Cq*`J4QL8GRv84Q+#-PD50ebm+|_bF9FgDFR*0~skUjAvZ>+aTi(8SQwrB=HFa zspnFrlcNQCGD!udHC$K>r5r6kOvbf_vT@avUL-3)wHXl3mrE1tg?!b7brr%1UcQ3h zORpikF>V0c3dj`M>==))P%@Z|x0&N*>boVXI{s#wFh}=uC@*WYuLOYDX29z0YKz@8 zToagFkQb)rQtZ^yI=QqY@p)43)HUPt?&MwGi9tgB#YtiZlg>(9d*m99l^+KlQ|OM4uRdlD7?7=>pta!z1rc7% zdk4Gr8_N41!m-JCUTjB_|4{-HkCbLID;X1~lIxV{~BjK9%0JoRqV zDjBrgVM0kO_4Byk=c;BJUwQGPi|P30jmG!R=FV!r@p`LrK7zddO5+VGl>{DR+9De$>hBlWGv<*Zj=^9JVCS9+yCk zXUOk-)4jMm>bmhfkcW?aeVezBJk2P)r6rt{WD`5ol{V+cQq(3|-?g;?HJW$JgKk71 z5N#znAxTY*A$mP6?M7oKvJtIn#|(K>cHvyFEJ`ZXm3lu67p}Mt^5uDL&YAYKnn#Y5 zc#Da0%!X%VsFs)$|TC^J?>YJnC)@DwH1(;&wmf zQIJq~pg0yHjB|h|V-^gmb2@8@{Z+?vzjJA>6;NZLMAI-UuuS-&#WQ}if5g$)$3W6w!eHP)AE+k z5!vl&E=_%Zd(+BKD$rqB+&5aWp8V4~L8PRDKT{J=Zfzp&j97vR#iV37fP1X9!+h>! zR4PNfD!CKKvC>@3l!RJPCCclnQW|20lKN&Q^=FNFq*Q zX>sSgQM`t7J-gk&ZSHMjIy{&fFE%fcNGhIzSQDbJ8z!e}{xrr^VPTTxMCUS2`fy#5 zX*NznF}v&r&Y&&Hv2>`C(+O&^tMOa)#I)2@=`7b`{y|pqU|ZVfB>6F4jH>(^-c{#; zYKWb3kyi(oahvHL-wW)Xu4?Nw*w#jkUs(x!16%6hD<+q1F@y6G3;G?5WmtPS!fOfZ zIC0!lgiJ+1eln$@$;7Gro1y^fU~JBHJMdQOS(>C4xA>+NxB5u*x2@CDEzmpoe^S~L zP&ycJDBtpswIh;%c8mt%XCOIu=-R8GQv323q=qP4ZGH51NAIK zgMbc)m}wIP5niL96DoY8NWXf!sGJ)l#0bAgU6!vzh zW1#_fySnA9KpDBI?S5xSIS={Io$ANJXIXL0n#v7V!;9GsU$m?;KCbIF*A$wl=4t2z za&kP>SdTXH1u6}THp?4Zmub42N@OZ68-0);AN^g?=q_rfUULmXf390H@#gEG&Yd`5 zrHYo5=9eVmsu4ADgq%z}>Z~P3`mORW>4*>s)PQzGe8KFlq34XvSsfrFanT53BiibI zLxUPxCwzy&}1CABuY&3dBkaUfCoC{mt9U2#gdhgwc47& zr&BJ_)fqN1?+5J9M?3YPwhC&IyngzV>p?D3FVoB2CiNH2S#vG>(@g<_j|=TxTBm1q ze0-Ig)gpm`B3TKoTOv-Ce;#S|g45}dBu*!VlQfeGsn8=-MZVIO-Nv3EqR@fWkaR<{ z``_gosER`R?ZvLeknEWyxLC7zJ4v?<(ib(6U`*9Kwx=Ov`*B2G-@kKckkvBzD_W=4 z=-;@>gyCY?5AJ6iBB$k&Jj(I6Yhc~}uTOU;zqP`ezJgY+ zFL;mY8Gj~WTCtGbF2r;guG9PspPTMzm#Wh}(j`JuGbPF3F{S5a64WV?9UUEb$?Y_B zL`&|N7KjxyU%lOW;wEd<#Rz`nR4%);Vy2&kjn=m0;xyvnh(iZ-8@Tm3vaE<{8UOm*;$2+YrcwGYRzC?G zoA)GRtmg=^z=OeWQS5U#MubByR%}IP-yj94io@FwhvBfKGl2o`h7SLP%-8b%rvBR! znR@Q8Euvi7uDDO6-%Zj)hUN=(+q|Mu>lW>PWL+ZA=ma!&q@}&kQbmR2M-KFAzLF+|xGgP%xpo-lwm|KIWQSVee!y%M*XTCmP-q(1@0!j2J&a{q_9T z&9-U7=56PTV2x*0?_C!f?kug14(ZP(wktOIt`FjRvplv!+3qzgFYUE$zG`Nq4%!Pe zsr|WF9~HRXP#)zwIghfj1r18z3!bg*o9c02L`+tc6Bl2)>`ACihX+-E>&*K((tLh1 zr!JU7`Ec0wik#Xr|Lq9(EhvWS91_06Np+BheBG}=1i=LhTuoQ>xTbX!dCD#$HB3|H znCTI<3X(JVGjyQ($q?}N9GawfBAIT-=UdF#g_cpQaUHaCa|CxKc70LY_yCeCJcQsa zS~@sfqsykpS>-MY@`zSf+t2l7gG@ zk~m4(GQH0+d2=sfhEYgcTQl%84a?-g{D&3Z58OCL;zK_z{$wt5JA4MSCPlCd6*yluoIzwtQ^NBxU-1jdR`P-BS2CTya8*En8Twmo;!#UVLVH z!oU0R#`F&J<-~KU{3G^5aBPO;ZEPReCL+yT*IWIN)tz_`Ik^p$OHElN4c5A3NR;9O z^6ntEt?j$ec)d~RnmL#FA$r^JR)w9Z+JvcbkE|W7bsj{Nr}U)glt&})MFhq(OCzkF zj}4v7NlYpqhJP{4uCce0mk>OLOzRnu-CwO@nb}!u zk$NF00be^-B}xR&^2INbp7xt?HVA8_g zjx>=_M;TR@>p{q_SZVB9q{|lTTtloI-j+xI_cqcL9@pTp>#>xaM7jVh-EwZW%I}Qo zuZSyV?abg`;M|>3hPAfPrcO+3Sp)?BOihJdjOG+nH2b~LGnbaty;~+J-hI=g*D{%^ z25^B@?PBQdq?cFk*PNny-+gdgJe=uj%|akG$%P6j5S@6jkSYE``GJDWps26xU{YwU z#GKqNqapP)mZs?|Ni{hZ;NxN^gt|3DEt2j=He2 zmdCDmq}t!H55}`@b@%G&JRXmzOjksLVwaJs7*X6F_A`{qD zfNLn&L3HBW`HFmd@RIa|^zw&?L)_>F(;AfoPGok=4!~uCZ}#Z!%ByHoVsl{O&#fEg z$DJ0}rnq%e>@?<+MOJCI3tVp!u@i*V0tH3QAtjo4Rq-sc6i-&AZ>!Z6QMPXRQ2(F& zzp%?oMpQSV;%&M3G$X5(90r$UExWN;pj*4+bl(%cd%D`6Y><5RzB8D$=s>p3F9MB5{g*jzA-voZXH0#_HJ=z!^q-GOAs+cG zyl^hYqy)$Le*N|Fx#!(6vu_jfAy2)t0-?!O)L7lCUnccgC?ENcWh5^2ON=LG1z@j! z{<^<1%@0YZF->`-#4{3%w~f_POr&QkIaIz^yDu-4_*@e>AKVfAP}cfEFxR`s2UEU^ z@ekCTMlPBr>aiR5r69Y*MvrO2spI=SB;Cdl;+>cZ$ zp&H;nUL}Jy%l(ZT(XbmBi6j#pRxH(yiJN?b#4O@E)QZF`u`9!j$fQrcBGoQ+j(8^Y zOfBc_oGor8als&5m6)-E-m{q;+0z`P^Z+kf!6cuH9o8QN7F5daCOs0pCJpHw+T{Sj zws)26%={Kid0lD*f}`!$bZ5=>{>;M-4GzV6mL6BBdN z0S)Ei;sf&{48_a=ldoq;!RzBWCqg`$>~@|)wJ|Bh9}IM>yLf{ewtVO2R?UnYtUi6R z5Cha({w!~Q@8^o+^4vcr^p~Dx+7cMT$NrX-JR@Kewa0|+J7-4#Z64r+bRq>io}ZF2+CTnfMF9d6FG5v!~5)>z^T%2`{Wh zHAxDqbOx<*V~%Dh=?%;wwOB->O{nF(^wCnpQ0S$3K`h}b$MnYqpFLAMVM88g?!8iLY4g2?xm)t+4Yj&J6 zpC!{w8z7utODfs)9Y50L-T!2Nvl6E$IZ#T?JHpM{e~1+MDYQg(PZ9!Qir`?dck<{A zDFdvbofo;aba2!6XCm{laz8uP5Z=}re`#K(_$&WD@h5&@%%el_7wZupCvlref{Z0O z^C6Q^0j4=(Ym&H1hMH&ZlyQE?K0Ir8EZOJ?SoA@XMc95cZ~Mr_tsF}hTtG^WynJq<9zqo>(lHrJ17{g0&86WW*e;w;t);3z**vlzTnrUAej@2}lgl<13Q< zYU^J1SX=xvxf;t=4n~Zs_uKvAxP-0w=1hGvCaUtRtAsqZFW6m>x-*#T6vWZI;SnmT z+c`p@rKjx{bG>&4k?RU=q3gq-JDz%>?fI-a${}h?a6pDXOwgNsiRe!IV!O}ukldm5 z7{{w}dRO%Faz#FJF)GS~cm;#L$$vq}b=c`~HvL0Fcy{FiSz|Q_1*VPFgIhf5OeB6ZDN{w#%XbI*riINA!7R%OvqP7;#lj92OBzj{adcgg%7WZjxS;n|f#ln_<%7+PfHWd$qVctoR~o=b32$VJcAitm4jLvvk415Jxns@2F4Chn7OTu zvay4?35&3;l@a<=p0lYV1v{L8iwg_ex@fTgxj+;wU;vl`2;h490C4KCNy4C3=7vHx zrdGxj0CYAXM?+&97zGzM5S{Hw5^!az;)iky&a-12jC0}^g5+_Bu z-M&>mF0|1^Q~ccKW?szxOWBsa--E#d z;v4z~KvC_f-jgGl;Ge<2j2ZRCw;hA!XUN8{6A|DgX->ii9{{|YFIr+8CBXH&6N zwK0FX7eMg@+yB5Z5b!Vehyy8}ob{jWz<+Cp`v)E)e~GXuIM^C08^g5N;P({RRE%9< zI{zLX|JLg-_bNb5joB2S4o^|X{xkralCh($lY^nLBL(nDScy&6*vK6E6wOZx!0>pFH4`QBDPLXzW6s4 zz<;})@s|z}D0~C5HHFXG6h6tnCXLP}Z)a>HWcakBY5i?1HjR&(6gEy)R)6~dz{i$> z+L+QC+pwrAGyI23|B>hmpM;PT%*@t-Ph3n~901^egM)+P@1@wk(ZX<=3jhEC1mFh4 zZ-cr-#Dk?>smSgv0hfc|Zs}kp7bg z1aW}iKL2ky?*ASS2xNx`=YPt(8Dg`Y+aPcT+CcQb~&icQJZ7LLQe@Q%(VX=7qb0sQMy5WFHvp+&(b z#tw#vfgnQM!s4RroT8!-b`F3r2M7!hVh3?^f#J~jzeSz^FKTNjVrFb;>F8ul!3Gv$ zXBXmTXBQR|2Z=yHV0LbHAP0xIm Date: Mon, 15 Dec 2014 11:45:15 +0100 Subject: [PATCH 07/16] Fix underfull hbox. --- Heter_paper.tex | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index 4c6df55..5d28377 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -283,7 +283,7 @@ vector of scaling factors can be predicted using (\ref{eq:perf}). \textit T_\textit{new} = \max_{i=1,2,\dots,N} ({TcpOld_{i}} \cdot S_{i}) + MinTcm \end{equation} -Where:\\ +Where: \begin{equation} \label{eq:perf2} MinTcm = \min_{i=1,2,\dots,N} (Tcm_i) -- 2.39.5 From 706e9b5431b0c6d9cb8b184ac36e3956bfb3abcc Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 11:46:17 +0100 Subject: [PATCH 08/16] Symbol for watt is W, not w. --- Heter_paper.tex | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index 5d28377..1d92130 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -725,16 +725,16 @@ nodes were connected via an Ethernet network with 1 Gbit/s bandwidth. type &GFLOPS & Freq. & Freq. & Freq. & power & power \\ & & GHz & GHz &GHz & & \\ \hline - 1 &40 & 2.5 & 1.2 & 0.1 & 20~w &4~w \\ + 1 &40 & 2.5 & 1.2 & 0.1 & 20~W &4~W \\ \hline - 2 &50 & 2.66 & 1.6 & 0.133 & 25~w &5~w \\ + 2 &50 & 2.66 & 1.6 & 0.133 & 25~W &5~W \\ \hline - 3 &60 & 2.9 & 1.2 & 0.1 & 30~w &6~w \\ + 3 &60 & 2.9 & 1.2 & 0.1 & 30~W &6~W \\ \hline - 4 &70 & 3.4 & 1.6 & 0.133 & 35~w &7~w \\ + 4 &70 & 3.4 & 1.6 & 0.133 & 35~W &7~W \\ \hline \end{tabular} -- 2.39.5 From 6fbe9073516c2e43184d29b41cfcef71c8478a36 Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 11:57:48 +0100 Subject: [PATCH 09/16] Improve alignment in tables. --- Heter_paper.tex | 37 ++++++++++++++++++------------------- 1 file changed, 18 insertions(+), 19 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index 1d92130..dc87dc8 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -8,7 +8,6 @@ \usepackage{algorithm} \usepackage{subfig} \usepackage{amsmath} -\usepackage{multirow} \usepackage{url} \DeclareUrlCommand\email{\urlstyle{same}} @@ -764,7 +763,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \caption{Running NAS benchmarks on 4 nodes } % title of Table \centering - \begin{tabular}{|*{7}{l|}} + \begin{tabular}{|*{7}{r|}} \hline Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ @@ -792,7 +791,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \caption{Running NAS benchmarks on 8 and 9 nodes } % title of Table \centering - \begin{tabular}{|*{7}{l|}} + \begin{tabular}{|*{7}{r|}} \hline Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ @@ -820,7 +819,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \caption{Running NAS benchmarks on 16 nodes } % title of Table \centering - \begin{tabular}{|*{7}{l|}} + \begin{tabular}{|*{7}{r|}} \hline Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ @@ -848,7 +847,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \caption{Running NAS benchmarks on 32 and 36 nodes } % title of Table \centering - \begin{tabular}{|*{7}{l|}} + \begin{tabular}{|*{7}{r|}} \hline Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ @@ -876,7 +875,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \caption{Running NAS benchmarks on 64 nodes } % title of Table \centering - \begin{tabular}{|*{7}{l|}} + \begin{tabular}{|*{7}{r|}} \hline Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ @@ -904,7 +903,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \caption{Running NAS benchmarks on 128 and 144 nodes } % title of Table \centering - \begin{tabular}{|*{7}{l|}} + \begin{tabular}{|*{7}{r|}} \hline Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ @@ -1046,7 +1045,7 @@ degradation. \caption{The results of the 70\%-30\% power scenario} % title of Table \centering - \begin{tabular}{|*{6}{l|}} + \begin{tabular}{|*{6}{r|}} \hline Program & Energy & Energy & Performance & Distance \\ name & consumption/J & saving\% & degradation\% & \\ @@ -1075,7 +1074,7 @@ degradation. \caption{The results of the 90\%-10\% power scenario} % title of Table \centering - \begin{tabular}{|*{6}{l|}} + \begin{tabular}{|*{6}{r|}} \hline Program & Energy & Energy & Performance & Distance \\ name & consumption/J & saving\% & degradation\% & \\ @@ -1136,17 +1135,17 @@ degradation values while giving the same weight for both metrics. \begin{table}[!t] \caption{Comparing the proposed algorithm} \centering -\begin{tabular}{|l|l|l|l|l|l|l|l|} +\begin{tabular}{|*{7}{r|}} \hline -\multicolumn{2}{|l|}{\multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}Program \\ name\end{tabular}}} & \multicolumn{2}{l|}{Energy saving \%} & \multicolumn{2}{l|}{Perf. degradation \%} & \multicolumn{2}{l|}{Distance} \\ \cline{3-8} -\multicolumn{2}{|l|}{} & EDP & MaxDist & EDP & MaxDist & EDP & MaxDist \\ \hline -\multicolumn{2}{|l|}{CG} & 27.58 & 31.25 & 5.82 & 7.12 & 21.76 & 24.13 \\ \hline -\multicolumn{2}{|l|}{MG} & 29.49 & 33.78 & 3.74 & 6.41 & 25.75 & 27.37 \\ \hline -\multicolumn{2}{|l|}{LU} & 19.55 & 28.33 & 0.0 & 0.01 & 19.55 & 28.22 \\ \hline -\multicolumn{2}{|l|}{EP} & 28.40 & 27.04 & 4.29 & 0.49 & 24.11 & 26.55 \\ \hline -\multicolumn{2}{|l|}{BT} & 27.68 & 32.32 & 6.45 & 7.87 & 21.23 & 24.43 \\ \hline -\multicolumn{2}{|l|}{SP} & 20.52 & 24.73 & 5.21 & 2.78 & 15.31 & 21.95 \\ \hline -\multicolumn{2}{|l|}{FT} & 27.03 & 31.02 & 2.75 & 2.54 & 24.28 & 28.48 \\ \hline +Program & \multicolumn{2}{c|}{Energy saving \%} & \multicolumn{2}{c|}{Perf. degradation \%} & \multicolumn{2}{c|}{Distance} \\ \cline{2-7} +name & EDP & MaxDist & EDP & MaxDist & EDP & MaxDist \\ \hline +CG & 27.58 & 31.25 & 5.82 & 7.12 & 21.76 & 24.13 \\ \hline +MG & 29.49 & 33.78 & 3.74 & 6.41 & 25.75 & 27.37 \\ \hline +LU & 19.55 & 28.33 & 0.0 & 0.01 & 19.55 & 28.22 \\ \hline +EP & 28.40 & 27.04 & 4.29 & 0.49 & 24.11 & 26.55 \\ \hline +BT & 27.68 & 32.32 & 6.45 & 7.87 & 21.23 & 24.43 \\ \hline +SP & 20.52 & 24.73 & 5.21 & 2.78 & 15.31 & 21.95 \\ \hline +FT & 27.03 & 31.02 & 2.75 & 2.54 & 24.28 & 28.48 \\ \hline \end{tabular} \label{table:compare_EDP} -- 2.39.5 From 8635fccbf588d4a9dd25d07624be34eae4ab66ad Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 13:34:38 +0100 Subject: [PATCH 10/16] Misc corrections. * remove double spaces * use \dots{} instead of ... * tidy \ref to figures, tables, etc. * fix overfull hboxes in tables * reformat some parts of source code (use git diff --word-diff to see changed words) --- Heter_paper.tex | 281 +++++++++++++++++++++++++++--------------------- 1 file changed, 158 insertions(+), 123 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index dc87dc8..014ad60 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -118,7 +118,7 @@ This heterogeneous platform executes more than 5 GFLOPS per watt while consuming Besides platform improvements, there are many software and hardware techniques to lower the energy consumption of these platforms, such as scheduling, DVFS, -... DVFS is a widely used process to reduce the energy consumption of a +\dots{} DVFS is a widely used process to reduce the energy consumption of a processor by lowering its frequency \cite{Rizvandi_Some.Observations.on.Optimal.Frequency}. However, it also reduces the number of FLOPS executed by the processor which might increase the execution @@ -151,24 +151,24 @@ Finally, in Section~\ref{sec.concl} the paper ends with a summary and some futur \section{Related works} \label{sec.relwork} DVFS is a technique used in modern processors to scale down both the voltage and -the frequency of the CPU while computing, in order to reduce the energy -consumption of the processor. DVFS is also allowed in GPUs to achieve the -same goal. Reducing the frequency of a processor lowers its number of FLOPS and -might degrade the performance of the application running on that processor, -especially if it is compute bound. Therefore selecting the appropriate frequency -for a processor to satisfy some objectives while taking into account all the -constraints, is not a trivial operation. Many researchers used different -strategies to tackle this problem. Some of them developed online methods that -compute the new frequency while executing the application, such as -~\cite{Hao_Learning.based.DVFS,Spiliopoulos_Green.governors.Adaptive.DVFS}. Others -used offline methods that might need to run the application and profile it -before selecting the new frequency, such as -~\cite{Rountree_Bounding.energy.consumption.in.MPI,Cochran_Pack_and_Cap_Adaptive_DVFS}. The -methods could be heuristics, exact or brute force methods that satisfy varied -objectives such as energy reduction or performance. They also could be adapted -to the execution's environment and the type of the application such as -sequential, parallel or distributed architecture, homogeneous or heterogeneous -platform, synchronous or asynchronous application, ... +the frequency of the CPU while computing, in order to reduce the energy +consumption of the processor. DVFS is also allowed in GPUs to achieve the same +goal. Reducing the frequency of a processor lowers its number of FLOPS and might +degrade the performance of the application running on that processor, especially +if it is compute bound. Therefore selecting the appropriate frequency for a +processor to satisfy some objectives while taking into account all the +constraints, is not a trivial operation. Many researchers used different +strategies to tackle this problem. Some of them developed online methods that +compute the new frequency while executing the application, such +as~\cite{Hao_Learning.based.DVFS,Spiliopoulos_Green.governors.Adaptive.DVFS}. +Others used offline methods that might need to run the application and profile +it before selecting the new frequency, such +as~\cite{Rountree_Bounding.energy.consumption.in.MPI,Cochran_Pack_and_Cap_Adaptive_DVFS}. +The methods could be heuristics, exact or brute force methods that satisfy +varied objectives such as energy reduction or performance. They also could be +adapted to the execution's environment and the type of the application such as +sequential, parallel or distributed architecture, homogeneous or heterogeneous +platform, synchronous or asynchronous application, \dots{} In this paper, we are interested in reducing energy for message passing iterative synchronous applications running over heterogeneous platforms. Some works have already been done for such platforms and they can be classified into two types of heterogeneous platforms: @@ -179,30 +179,44 @@ Some works have already been done for such platforms and they can be classified \end{itemize} -For the first type of platform, the computing intensive parallel tasks are executed on the GPUs and the rest are executed -on the CPUs. Luley et al. -~\cite{Luley_Energy.efficiency.evaluation.and.benchmarking}, proposed a heterogeneous -cluster composed of Intel Xeon CPUs and NVIDIA GPUs. Their main goal was to maximize the -energy efficiency of the platform during computation by maximizing the number of FLOPS per watt generated. -In~\cite{KaiMa_Holistic.Approach.to.Energy.Efficiency.in.GPU-CPU}, Kai Ma et al. developed a scheduling -algorithm that distributes workloads proportional to the computing power of the nodes which could be a GPU or a CPU. All the tasks must be completed at the same time. -In~\cite{Rong_Effects.of.DVFS.on.K20.GPU}, Rong et al. showed that -a heterogeneous (GPUs and CPUs) cluster that enables DVFS gave better energy and performance -efficiency than other clusters only composed of CPUs. +For the first type of platform, the computing intensive parallel tasks are +executed on the GPUs and the rest are executed on the CPUs. Luley et +al.~\cite{Luley_Energy.efficiency.evaluation.and.benchmarking}, proposed a +heterogeneous cluster composed of Intel Xeon CPUs and NVIDIA GPUs. Their main +goal was to maximize the energy efficiency of the platform during computation by +maximizing the number of FLOPS per watt generated. +In~\cite{KaiMa_Holistic.Approach.to.Energy.Efficiency.in.GPU-CPU}, Kai Ma et +al. developed a scheduling algorithm that distributes workloads proportional to +the computing power of the nodes which could be a GPU or a CPU. All the tasks +must be completed at the same time. In~\cite{Rong_Effects.of.DVFS.on.K20.GPU}, +Rong et al. showed that a heterogeneous (GPUs and CPUs) cluster that enables +DVFS gave better energy and performance efficiency than other clusters only +composed of CPUs. -The work presented in this paper concerns the second type of platform, with heterogeneous CPUs. -Many methods were conceived to reduce the energy consumption of this type of platform. Naveen et al.~\cite{Naveen_Power.Efficient.Resource.Scaling} -developed a method that minimizes the value of $energy\cdot delay^2$ (the delay is the sum of slack times that happen during synchronous communications) by dynamically assigning new frequencies to the CPUs of the heterogeneous cluster. Lizhe et al.~\cite{Lizhe_Energy.aware.parallel.task.scheduling} proposed -an algorithm that divides the executed tasks into two types: the critical and -non critical tasks. The algorithm scales down the frequency of non critical tasks proportionally to their slack and communication times while limiting the performance degradation percentage to less than 10\%. In~\cite{Joshi_Blackbox.prediction.of.impact.of.DVFS}, they developed - a heterogeneous cluster composed of two types -of Intel and AMD processors. They use a gradient method to predict the impact of DVFS operations on performance. -In~\cite{Shelepov_Scheduling.on.Heterogeneous.Multicore} and \cite{Li_Minimizing.Energy.Consumption.for.Frame.Based.Tasks}, - the best frequencies for a specified heterogeneous cluster are selected offline using some -heuristic. Chen et al.~\cite{Chen_DVFS.under.quality.of.service.requirements} used a greedy dynamic programming approach to -minimize the power consumption of heterogeneous servers while respecting given time constraints. This approach -had considerable overhead. -In contrast to the above described papers, this paper presents the following contributions : +The work presented in this paper concerns the second type of platform, with +heterogeneous CPUs. Many methods were conceived to reduce the energy +consumption of this type of platform. Naveen et +al.~\cite{Naveen_Power.Efficient.Resource.Scaling} developed a method that +minimizes the value of $energy\cdot delay^2$ (the delay is the sum of slack +times that happen during synchronous communications) by dynamically assigning +new frequencies to the CPUs of the heterogeneous cluster. Lizhe et +al.~\cite{Lizhe_Energy.aware.parallel.task.scheduling} proposed an algorithm +that divides the executed tasks into two types: the critical and non critical +tasks. The algorithm scales down the frequency of non critical tasks +proportionally to their slack and communication times while limiting the +performance degradation percentage to less than +10\%. In~\cite{Joshi_Blackbox.prediction.of.impact.of.DVFS}, they developed a +heterogeneous cluster composed of two types of Intel and AMD processors. They +use a gradient method to predict the impact of DVFS operations on performance. +In~\cite{Shelepov_Scheduling.on.Heterogeneous.Multicore} and +\cite{Li_Minimizing.Energy.Consumption.for.Frame.Based.Tasks}, the best +frequencies for a specified heterogeneous cluster are selected offline using +some heuristic. Chen et +al.~\cite{Chen_DVFS.under.quality.of.service.requirements} used a greedy dynamic +programming approach to minimize the power consumption of heterogeneous servers +while respecting given time constraints. This approach had considerable +overhead. In contrast to the above described papers, this paper presents the +following contributions : \begin{enumerate} \item two new energy and performance models for message passing iterative synchronous applications running over a heterogeneous platform. Both models take into account communication and slack times. The models can predict the required energy and the execution time of the application. @@ -228,14 +242,14 @@ network. Therefore, each node has different characteristics such as computing power (FLOPS), energy consumption, CPU's frequency range, \dots{} but they all have the same network bandwidth and latency. -The overall execution time of a distributed iterative synchronous application -over a heterogeneous platform consists of the sum of the computation time and -the communication time for every iteration on a node. However, due to the -heterogeneous computation power of the computing nodes, slack times might occur -when fast nodes have to wait, during synchronous communications, for the slower -nodes to finish their computations (see Figure~(\ref{fig:heter})). -Therefore, the overall execution time of the program is the execution time of the slowest -task which has the highest computation time and no slack time. +The overall execution time of a distributed iterative synchronous application +over a heterogeneous platform consists of the sum of the computation time and +the communication time for every iteration on a node. However, due to the +heterogeneous computation power of the computing nodes, slack times might occur +when fast nodes have to wait, during synchronous communications, for the slower +nodes to finish their computations (see Figure~\ref{fig:heter}). Therefore, the +overall execution time of the program is the execution time of the slowest task +which has the highest computation time and no slack time. \begin{figure}[!t] \centering @@ -387,7 +401,7 @@ $Tcp_{i}$ might be different and different frequency scaling factors might be computed in order to decrease the overall energy consumption of the application and reduce slack times. The communication time of a processor $i$ is noted as $Tcm_{i}$ and could contain slack times when communicating with slower -nodes, see figure(\ref{fig:heter}). Therefore, all nodes do not have equal +nodes, see Figure~\ref{fig:heter}. Therefore, all nodes do not have equal communication times. While the dynamic energy is computed according to the frequency scaling factor and the dynamic power of each node as in (\ref{eq:Edyn}), the static energy is computed as the sum of the execution time @@ -495,12 +509,12 @@ normalized execution time is inverted which gives the normalized performance equ \caption{The energy and performance relation} \end{figure} -Then, the objective function can be modeled in order to find the maximum distance -between the energy curve (\ref{eq:enorm}) and the performance -curve (\ref{eq:pnorm_inv}) over all available sets of scaling factors. This -represents the minimum energy consumption with minimum execution time (maximum -performance) at the same time, see figure~(\ref{fig:r1}) or figure~(\ref{fig:r2}). Then the objective -function has the following form: +Then, the objective function can be modeled in order to find the maximum +distance between the energy curve (\ref{eq:enorm}) and the performance curve +(\ref{eq:pnorm_inv}) over all available sets of scaling factors. This +represents the minimum energy consumption with minimum execution time (maximum +performance) at the same time, see Figure~\ref{fig:r1} or +Figure~\ref{fig:r2}. Then the objective function has the following form: \begin{equation} \label{eq:max} Max Dist = @@ -518,29 +532,33 @@ the energy curve has a convex form as shown in~\cite{Zhuo_Energy.efficient.Dynam \label{sec.optim} \subsection{The algorithm details} -In this section, algorithm \ref{HSA} is presented. It selects the frequency scaling factors -vector that gives the best trade-off between minimizing the energy consumption and maximizing -the performance of a message passing synchronous iterative application executed on a heterogeneous -platform. It works online during the execution time of the iterative message passing program. -It uses information gathered during the first iteration such as the computation time and the -communication time in one iteration for each node. The algorithm is executed after the first -iteration and returns a vector of optimal frequency scaling factors that satisfies the objective -function (\ref{eq:max}). The program applies DVFS operations to change the frequencies of the CPUs -according to the computed scaling factors. This algorithm is called just once during the execution -of the program. Algorithm~(\ref{dvfs}) shows where and when the proposed scaling algorithm is called -in the iterative MPI program. - -The nodes in a heterogeneous platform have different computing powers, thus while executing message -passing iterative synchronous applications, fast nodes have to wait for the slower ones to finish their -computations before being able to synchronously communicate with them as in figure (\ref{fig:heter}). -These periods are called idle or slack times. -The algorithm takes into account this problem and tries to reduce these slack times when selecting the -frequency scaling factors vector. At first, it selects initial frequency scaling factors that increase -the execution times of fast nodes and minimize the differences between the computation times of -fast and slow nodes. The value of the initial frequency scaling factor for each node is inversely -proportional to its computation time that was gathered from the first iteration. These initial frequency -scaling factors are computed as a ratio between the computation time of the slowest node and the -computation time of the node $i$ as follows: +In this section, Algorithm~\ref{HSA} is presented. It selects the frequency +scaling factors vector that gives the best trade-off between minimizing the +energy consumption and maximizing the performance of a message passing +synchronous iterative application executed on a heterogeneous platform. It works +online during the execution time of the iterative message passing program. It +uses information gathered during the first iteration such as the computation +time and the communication time in one iteration for each node. The algorithm is +executed after the first iteration and returns a vector of optimal frequency +scaling factors that satisfies the objective function (\ref{eq:max}). The +program applies DVFS operations to change the frequencies of the CPUs according +to the computed scaling factors. This algorithm is called just once during the +execution of the program. Algorithm~\ref{dvfs} shows where and when the proposed +scaling algorithm is called in the iterative MPI program. + +The nodes in a heterogeneous platform have different computing powers, thus +while executing message passing iterative synchronous applications, fast nodes +have to wait for the slower ones to finish their computations before being able +to synchronously communicate with them as in Figure~\ref{fig:heter}. These +periods are called idle or slack times. The algorithm takes into account this +problem and tries to reduce these slack times when selecting the frequency +scaling factors vector. At first, it selects initial frequency scaling factors +that increase the execution times of fast nodes and minimize the differences +between the computation times of fast and slow nodes. The value of the initial +frequency scaling factor for each node is inversely proportional to its +computation time that was gathered from the first iteration. These initial +frequency scaling factors are computed as a ratio between the computation time +of the slowest node and the computation time of the node $i$ as follows: \begin{equation} \label{eq:Scp} Scp_{i} = \frac{\max_{i=1,2,\dots,N}(Tcp_i)}{Tcp_i} @@ -552,25 +570,25 @@ and the computation scaling factor $Scp_i$ as follows: \label{eq:Fint} F_{i} = \frac{Fmax_i}{Scp_i},~{i=1,2,\cdots,N} \end{equation} -If the computed initial frequency for a node is not available in the gears of -that node, it is replaced by the nearest available frequency. In figure -(\ref{fig:st_freq}), the nodes are sorted by their computing power in ascending -order and the frequencies of the faster nodes are scaled down according to the -computed initial frequency scaling factors. The resulting new frequencies are -colored in blue in figure (\ref{fig:st_freq}). This set of frequencies can be -considered as a higher bound for the search space of the optimal vector of -frequencies because selecting frequency scaling factors higher than the higher -bound will not improve the performance of the application and it will increase -its overall energy consumption. Therefore the algorithm that selects the -frequency scaling factors starts the search method from these initial -frequencies and takes a downward search direction toward lower frequencies. The -algorithm iterates on all left frequencies, from the higher bound until all -nodes reach their minimum frequencies, to compute their overall energy -consumption and performance, and select the optimal frequency scaling factors -vector. At each iteration the algorithm determines the slowest node according to -the equation (\ref{eq:perf}) and keeps its frequency unchanged, while it lowers -the frequency of all other nodes by one gear. The new overall energy -consumption and execution time are computed according to the new scaling +If the computed initial frequency for a node is not available in the gears of +that node, it is replaced by the nearest available frequency. In +Figure~\ref{fig:st_freq}, the nodes are sorted by their computing power in +ascending order and the frequencies of the faster nodes are scaled down +according to the computed initial frequency scaling factors. The resulting new +frequencies are colored in blue in Figure~\ref{fig:st_freq}. This set of +frequencies can be considered as a higher bound for the search space of the +optimal vector of frequencies because selecting frequency scaling factors higher +than the higher bound will not improve the performance of the application and it +will increase its overall energy consumption. Therefore the algorithm that +selects the frequency scaling factors starts the search method from these +initial frequencies and takes a downward search direction toward lower +frequencies. The algorithm iterates on all left frequencies, from the higher +bound until all nodes reach their minimum frequencies, to compute their overall +energy consumption and performance, and select the optimal frequency scaling +factors vector. At each iteration the algorithm determines the slowest node +according to the equation (\ref{eq:perf}) and keeps its frequency unchanged, +while it lowers the frequency of all other nodes by one gear. The new overall +energy consumption and execution time are computed according to the new scaling factors. The optimal set of frequency scaling factors is the set that gives the highest distance according to the objective function (\ref{eq:max}). @@ -651,7 +669,7 @@ maximum distance between the energy curve and the performance curve is while \If {$(k=1)$} \State Gather all times of computation and\newline\hspace*{3em}% communication from each node. - \State Call algorithm \ref{HSA}. + \State Call Algorithm \ref{HSA}. \State Compute the new frequencies from the\newline\hspace*{3em}% returned optimal scaling factors. \State Set the new frequencies to nodes. @@ -678,28 +696,33 @@ parallel benchmarks NPB v3.3 \cite{NAS.Parallel.Benchmarks}, running class B on very precise, the maximum normalized difference between the predicted execution time and the real execution time is equal to 0.03 for all the NAS benchmarks. -Since the proposed algorithm is not an exact method it does not test all the possible solutions (vectors of scaling factors) -in the search space. To prove its efficiency, it was compared on small instances to a brute force search algorithm -that tests all the possible solutions. The brute force algorithm was applied to different NAS benchmarks classes with -different number of nodes. The solutions returned by the brute force algorithm and the proposed algorithm were identical -and the proposed algorithm was on average 10 times faster than the brute force algorithm. It has a small execution time: -for a heterogeneous cluster composed of four different types of nodes having the characteristics presented in -table~\ref{table:platform}, it takes on average \np[ms]{0.04} for 4 nodes and \np[ms]{0.15} on average for 144 nodes -to compute the best scaling factors vector. The algorithm complexity is $O(F\cdot (N \cdot4) )$, where $F$ is the number -of iterations and $N$ is the number of computing nodes. The algorithm needs from 12 to 20 iterations to select the best -vector of frequency scaling factors that gives the results of the next sections. +Since the proposed algorithm is not an exact method it does not test all the +possible solutions (vectors of scaling factors) in the search space. To prove +its efficiency, it was compared on small instances to a brute force search +algorithm that tests all the possible solutions. The brute force algorithm was +applied to different NAS benchmarks classes with different number of nodes. The +solutions returned by the brute force algorithm and the proposed algorithm were +identical and the proposed algorithm was on average 10 times faster than the +brute force algorithm. It has a small execution time: for a heterogeneous +cluster composed of four different types of nodes having the characteristics +presented in Table~\ref{table:platform}, it takes on average \np[ms]{0.04} for 4 +nodes and \np[ms]{0.15} on average for 144 nodes to compute the best scaling +factors vector. The algorithm complexity is $O(F\cdot (N \cdot4) )$, where $F$ +is the number of iterations and $N$ is the number of computing nodes. The +algorithm needs from 12 to 20 iterations to select the best vector of frequency +scaling factors that gives the results of the next sections. \section{Experimental results} \label{sec.expe} To evaluate the efficiency and the overall energy consumption reduction of -algorithm~\ref{HSA}, it was applied to the NAS parallel benchmarks NPB v3.3. The +Algorithm~\ref{HSA}, it was applied to the NAS parallel benchmarks NPB v3.3. The experiments were executed on the simulator SimGrid/SMPI which offers easy tools to create a heterogeneous platform and run message passing applications over it. The heterogeneous platform that was used in the experiments, had one core per node because just one process was executed per node. The heterogeneous platform was composed of four types of nodes. Each type of nodes had different characteristics such as the maximum CPU frequency, the number of available -frequencies and the computational power, see Table \ref{table:platform}. The +frequencies and the computational power, see Table~\ref{table:platform}. The characteristics of these different types of nodes are inspired from the specifications of real Intel processors. The heterogeneous platform had up to 144 nodes and had nodes from the four types in equal proportions, for example if @@ -765,6 +788,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \centering \begin{tabular}{|*{7}{r|}} \hline + \hspace{-2.2084pt}% Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ \hline @@ -793,6 +817,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \centering \begin{tabular}{|*{7}{r|}} \hline + \hspace{-2.2084pt}% Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ \hline @@ -821,6 +846,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \centering \begin{tabular}{|*{7}{r|}} \hline + \hspace{-2.2084pt}% Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ \hline @@ -849,6 +875,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \centering \begin{tabular}{|*{7}{r|}} \hline + \hspace{-2.2084pt}% Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ \hline @@ -877,6 +904,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \centering \begin{tabular}{|*{7}{r|}} \hline + \hspace{-2.2084pt}% Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ \hline @@ -905,6 +933,7 @@ be executed on $1, 4, 9, 16, 36, 64, 144$ nodes. \centering \begin{tabular}{|*{7}{r|}} \hline + \hspace{-2.2084pt}% Program & Execution & Energy & Energy & Performance & Distance \\ name & time/s & consumption/J & saving\% & degradation\% & \\ \hline @@ -929,9 +958,9 @@ The overall energy consumption was computed for each instance according to the energy consumption model (\ref{eq:energy}), with and without applying the algorithm. The execution time was also measured for all these experiments. Then, the energy saving and performance degradation percentages were computed for each -instance. The results are presented in Tables (\ref{table:res_4n}, +instance. The results are presented in Tables~\ref{table:res_4n}, \ref{table:res_8n}, \ref{table:res_16n}, \ref{table:res_32n}, -\ref{table:res_64n} and \ref{table:res_128n}). All these results are the average +\ref{table:res_64n} and \ref{table:res_128n}. All these results are the average values from many experiments for energy savings and performance degradation. The tables show the experimental results for running the NAS parallel benchmarks on different number of nodes. The experiments show that the algorithm @@ -945,7 +974,7 @@ computation required for each iteration is divided by the number of computing nodes. On the other hand, more communications are required when increasing the number of nodes so the static energy increases linearly according to the communication time and the dynamic power is less relevant in the overall energy -consumption. Therefore, reducing the frequency with algorithm~(\ref{HSA}) is +consumption. Therefore, reducing the frequency with Algorithm~\ref{HSA} is less effective in reducing the overall energy savings. It can also be noticed that for the benchmarks EP and SP that contain little or no communications, the energy savings are not significantly affected by the high number of nodes. No @@ -971,7 +1000,7 @@ small when compared to the communication times. \caption{The energy and performance for all NAS benchmarks running with a different number of nodes} \end{figure} -Figures \ref{fig:energy} and \ref{fig:per_deg} present the energy saving and +Figures~\ref{fig:energy} and \ref{fig:per_deg} present the energy saving and performance degradation respectively for all the benchmarks according to the number of used nodes. As shown in the first plot, the energy saving percentages of the benchmarks MG, LU, BT and FT decrease linearly when the number of nodes @@ -1004,7 +1033,7 @@ two new power scenarios are the following: The NAS parallel benchmarks were executed again over processors that follow the new power scenarios. The class C of each benchmark was run over 8 or 9 nodes -and the results are presented in Tables \ref{table:res_s1} and +and the results are presented in Tables~\ref{table:res_s1} and \ref{table:res_s2}. These tables show that the energy saving percentage of the 70\%-30\% scenario is smaller for all benchmarks compared to the energy saving of the 90\%-10\% scenario. Indeed, in the latter more dynamic power is consumed @@ -1020,8 +1049,8 @@ really significantly scale down much the frequencies of the nodes in order to limit the increase of the execution time and thus limiting the effect of the consumed static energy. -Both new power scenarios are compared to the old one in figure -(\ref{fig:sen_comp}). It shows the average of the performance degradation, the +Both new power scenarios are compared to the old one in +Figure~\ref{fig:sen_comp}. It shows the average of the performance degradation, the energy saving and the distances for all NAS benchmarks of class C running on 8 or 9 nodes. The comparison shows that the energy saving ratio is proportional to the dynamic power ratio: it is increased when applying the 90\%-10\% scenario @@ -1035,7 +1064,7 @@ higher ratio for static power (e.g. 70\%-30\% scenario and 80\%-20\% scenario). Since the proposed algorithm optimizes the energy consumption when using a higher ratio for dynamic power the algorithm selects bigger frequency scaling factors that result in more energy saving but less performance, for -example see Figure (\ref{fig:scales_comp}). The opposite happens when using a +example see Figure~\ref{fig:scales_comp}. The opposite happens when using a higher ratio for static power, the algorithm proportionally selects smaller scaling values which result in less energy saving but also less performance degradation. @@ -1120,12 +1149,18 @@ They developed a green governor that regularly applies an online frequency selec To fairly compare both algorithms, the same energy and execution time models, equations (\ref{eq:energy}) and (\ref{eq:fnew}), were used for both algorithms to predict the energy consumption and the execution times. Also Spiliopoulos et al. algorithm was adapted to start the search from the initial frequencies computed using the equation (\ref{eq:Fint}). The resulting algorithm is an exhaustive search algorithm that minimizes the EDP and has the initial frequencies values as an upper bound. -Both algorithms were applied to the parallel NAS benchmarks to compare their efficiency. Table \ref{table:compare_EDP} presents the results of comparing the execution times and the energy consumption for both versions of the NAS benchmarks while running the class C of each benchmark over 8 or 9 heterogeneous nodes. The results show that our algorithm provides better energy savings than Spiliopoulos et al. algorithm, -on average it results in 29.76\% energy saving while their algorithm returns just 25.75\%. The average of performance degradation percentage is approximately the same for both algorithms, about 4\%. +Both algorithms were applied to the parallel NAS benchmarks to compare their +efficiency. Table~\ref{table:compare_EDP} presents the results of comparing the +execution times and the energy consumption for both versions of the NAS +benchmarks while running the class C of each benchmark over 8 or 9 heterogeneous +nodes. The results show that our algorithm provides better energy savings than +Spiliopoulos et al. algorithm, on average it results in 29.76\% energy saving +while their algorithm returns just 25.75\%. The average of performance +degradation percentage is approximately the same for both algorithms, about 4\%. For all benchmarks, our algorithm outperforms Spiliopoulos et al. algorithm in -terms of energy and performance trade-off, see figure (\ref{fig:compare_EDP}), +terms of energy and performance trade-off, see Figure~\ref{fig:compare_EDP}, because it maximizes the distance between the energy saving and the performance degradation values while giving the same weight for both metrics. -- 2.39.5 From 3d2e5df308624df6df07bd686fe741155c3a952b Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 15:00:20 +0100 Subject: [PATCH 11/16] Use macros for variable names and other corrections in equations. --- Heter_paper.tex | 184 +++++++++++++++++++++++++----------------------- 1 file changed, 97 insertions(+), 87 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index 014ad60..0491692 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -23,33 +23,46 @@ \newcommand{\JC}[2][inline]{% \todo[color=red!10,#1]{\sffamily\textbf{JC:} #2}\xspace} -\newcommand{\Xsub}[2]{\ensuremath{#1_\textit{#2}}} +\newcommand{\Xsub}[2]{{\ensuremath{#1_\mathit{#2}}}} -\newcommand{\Dist}{\textit{Dist}} +\newcommand{\CL}{\Xsub{C}{L}} +\newcommand{\Dist}{\mathit{Dist}} +\newcommand{\EdNew}{\Xsub{E}{dNew}} \newcommand{\Eind}{\Xsub{E}{ind}} \newcommand{\Enorm}{\Xsub{E}{Norm}} \newcommand{\Eoriginal}{\Xsub{E}{Original}} \newcommand{\Ereduced}{\Xsub{E}{Reduced}} +\newcommand{\Es}{\Xsub{E}{S}} \newcommand{\Fdiff}{\Xsub{F}{diff}} \newcommand{\Fmax}{\Xsub{F}{max}} \newcommand{\Fnew}{\Xsub{F}{new}} \newcommand{\Ileak}{\Xsub{I}{leak}} \newcommand{\Kdesign}{\Xsub{K}{design}} -\newcommand{\MaxDist}{\textit{Max Dist}} +\newcommand{\MaxDist}{\mathit{Max}\Dist} +\newcommand{\MinTcm}{\mathit{Min}\Tcm} \newcommand{\Ntrans}{\Xsub{N}{trans}} -\newcommand{\Pdyn}{\Xsub{P}{dyn}} -\newcommand{\PnormInv}{\Xsub{P}{NormInv}} +\newcommand{\PdNew}{\Xsub{P}{dNew}} +\newcommand{\PdOld}{\Xsub{P}{dOld}} +%\newcommand{\Pdyn}{\Xsub{P}{dyn}} +\newcommand{\Pd}{\Xsub{P}{d}} +%\newcommand{\PnormInv}{\Xsub{P}{NormInv}} \newcommand{\Pnorm}{\Xsub{P}{Norm}} -\newcommand{\Tnorm}{\Xsub{T}{Norm}} -\newcommand{\Pstates}{\Xsub{P}{states}} -\newcommand{\Pstatic}{\Xsub{P}{static}} +%\newcommand{\Pstates}{\Xsub{P}{states}} +%\newcommand{\Pstatic}{\Xsub{P}{static}} +\newcommand{\Ps}{\Xsub{P}{s}} +\newcommand{\Scp}{\Xsub{S}{cp}} \newcommand{\Sopt}{\Xsub{S}{opt}} -\newcommand{\Tcomp}{\Xsub{T}{comp}} -\newcommand{\TmaxCommOld}{\Xsub{T}{Max Comm Old}} -\newcommand{\TmaxCompOld}{\Xsub{T}{Max Comp Old}} -\newcommand{\Tmax}{\Xsub{T}{max}} +\newcommand{\Tcm}{\Xsub{T}{cm}} +%\newcommand{\Tcomp}{\Xsub{T}{comp}} +\newcommand{\TcpOld}{\Xsub{T}{cpOld}} +\newcommand{\Tcp}{\Xsub{T}{cp}} +%\newcommand{\TmaxCommOld}{\Xsub{T}{Max Comm Old}} +%\newcommand{\TmaxCompOld}{\Xsub{T}{Max Comp Old}} +%\newcommand{\Tmax}{\Xsub{T}{max}} \newcommand{\Tnew}{\Xsub{T}{New}} +%\newcommand{\Tnorm}{\Xsub{T}{Norm}} \newcommand{\Told}{\Xsub{T}{Old}} + \begin{document} \title{Energy Consumption Reduction for Message Passing Iterative Applications in Heterogeneous Architecture Using DVFS} @@ -268,7 +281,7 @@ factor S which is the ratio between the maximum and the new frequency of a CPU as in (\ref{eq:s}). \begin{equation} \label{eq:s} - S = \frac{F_\textit{max}}{F_\textit{new}} + S = \frac{\Fmax}{\Fnew} \end{equation} The execution time of a compute bound sequential program is linearly proportional to the frequency scaling factor $S$. On the other hand, message passing @@ -293,16 +306,15 @@ operation. Then the execution time for one iteration of the application with any vector of scaling factors can be predicted using (\ref{eq:perf}). \begin{equation} \label{eq:perf} - \textit T_\textit{new} = - \max_{i=1,2,\dots,N} ({TcpOld_{i}} \cdot S_{i}) + MinTcm + \Tnew = \max_{i=1,2,\dots,N} ({\TcpOld_{i}} \cdot S_{i}) + \MinTcm \end{equation} Where: \begin{equation} \label{eq:perf2} - MinTcm = \min_{i=1,2,\dots,N} (Tcm_i) + \MinTcm = \min_{i=1,2,\dots,N} (\Tcm_i) \end{equation} -where $TcpOld_i$ is the computation time of processor $i$ during the first -iteration and $MinTcm$ is the communication time of the slowest processor from +where $\TcpOld_i$ is the computation time of processor $i$ during the first +iteration and $\MinTcm$ is the communication time of the slowest processor from the first iteration. The model computes the maximum computation time with scaling factor from each node added to the communication time of the slowest node. It means only the communication time without any slack time is taken into @@ -323,28 +335,28 @@ Rauber_Analytical.Modeling.for.Energy,Zhuo_Energy.efficient.Dynamic.Task.Schedul Rizvandi_Some.Observations.on.Optimal.Frequency} divide the power consumed by a processor into two power metrics: the static and the dynamic power. While the first one is consumed as long as the computing unit is turned on, the latter is only consumed during -computation times. The dynamic power $Pd$ is related to the switching -activity $\alpha$, load capacitance $C_L$, the supply voltage $V$ and +computation times. The dynamic power $\Pd$ is related to the switching +activity $\alpha$, load capacitance $\CL$, the supply voltage $V$ and operational frequency $F$, as shown in (\ref{eq:pd}). \begin{equation} \label{eq:pd} - Pd = \alpha \cdot C_L \cdot V^2 \cdot F + \Pd = \alpha \cdot \CL \cdot V^2 \cdot F \end{equation} -The static power $Ps$ captures the leakage power as follows: +The static power $\Ps$ captures the leakage power as follows: \begin{equation} \label{eq:ps} - Ps = V \cdot N_{trans} \cdot K_{design} \cdot I_{leak} + \Ps = V \cdot \Ntrans \cdot \Kdesign \cdot \Ileak \end{equation} -where V is the supply voltage, $N_{trans}$ is the number of transistors, -$K_{design}$ is a design dependent parameter and $I_{leak}$ is a +where V is the supply voltage, $\Ntrans$ is the number of transistors, +$\Kdesign$ is a design dependent parameter and $\Ileak$ is a technology dependent parameter. The energy consumed by an individual processor to execute a given program can be computed as: \begin{equation} \label{eq:eind} - E_\textit{ind} = Pd \cdot Tcp + Ps \cdot T + \Eind = \Pd \cdot \Tcp + \Ps \cdot T \end{equation} -where $T$ is the execution time of the program, $Tcp$ is the computation -time and $Tcp \le T$. $Tcp$ may be equal to $T$ if there is no +where $T$ is the execution time of the program, $\Tcp$ is the computation +time and $\Tcp \le T$. $\Tcp$ may be equal to $T$ if there is no communication and no slack time. The main objective of DVFS operation is to reduce the overall energy consumption~\cite{Le_DVFS.Laws.of.Diminishing.Returns}. @@ -355,19 +367,19 @@ process of the frequency can be expressed by the scaling factor $S$ which is the ratio between the maximum and the new frequency as in (\ref{eq:s}). The CPU governors are power schemes supplied by the operating system's kernel to lower a core's frequency. The new frequency -$F_{new}$ from (\ref{eq:s}) can be calculated as follows: +$\Fnew$ from (\ref{eq:s}) can be calculated as follows: \begin{equation} \label{eq:fnew} - F_\textit{new} = S^{-1} \cdot F_\textit{max} + \Fnew = S^{-1} \cdot \Fmax \end{equation} -Replacing $F_{new}$ in (\ref{eq:pd}) as in (\ref{eq:fnew}) gives the following +Replacing $\Fnew$ in (\ref{eq:pd}) as in (\ref{eq:fnew}) gives the following equation for dynamic power consumption: \begin{multline} \label{eq:pdnew} - {P}_\textit{dNew} = \alpha \cdot C_L \cdot V^2 \cdot F_{new} = \alpha \cdot C_L \cdot \beta^2 \cdot F_{new}^3 \\ - {} = \alpha \cdot C_L \cdot V^2 \cdot F_{max} \cdot S^{-3} = P_{dOld} \cdot S^{-3} + \PdNew = \alpha \cdot \CL \cdot V^2 \cdot \Fnew = \alpha \cdot \CL \cdot \beta^2 \cdot \Fnew^3 \\ + {} = \alpha \cdot \CL \cdot V^2 \cdot \Fmax \cdot S^{-3} = \PdOld \cdot S^{-3} \end{multline} -where $ {P}_\textit{dNew}$ and $P_{dOld}$ are the dynamic power consumed with the +where $\PdNew$ and $\PdOld$ are the dynamic power consumed with the new frequency and the maximum frequency respectively. According to (\ref{eq:pdnew}) the dynamic power is reduced by a factor of $S^{-3}$ when @@ -377,7 +389,7 @@ The new dynamic energy is the dynamic power multiplied by the new time of compu and is given by the following equation: \begin{equation} \label{eq:Edyn} - E_\textit{dNew} = P_{dOld} \cdot S^{-3} \cdot (Tcp \cdot S)= S^{-2}\cdot P_{dOld} \cdot Tcp + \EdNew = \PdOld \cdot S^{-3} \cdot (\Tcp \cdot S)= S^{-2}\cdot \PdOld \cdot \Tcp \end{equation} The static power is related to the power leakage of the CPU and is consumed during computation and even when idle. As in~\cite{Rauber_Analytical.Modeling.for.Energy,Zhuo_Energy.efficient.Dynamic.Task.Scheduling}, @@ -390,17 +402,17 @@ to the frequency scaling factor, while this scaling factor does not affect the c The static energy of a processor after scaling its frequency is computed as follows: \begin{equation} \label{eq:Estatic} - E_\textit{s} = Ps \cdot (Tcp \cdot S + Tcm) + \Es = \Ps \cdot (\Tcp \cdot S + \Tcm) \end{equation} In the considered heterogeneous platform, each processor $i$ might have -different dynamic and static powers, noted as $Pd_{i}$ and $Ps_{i}$ +different dynamic and static powers, noted as $\Pd_{i}$ and $\Ps_{i}$ respectively. Therefore, even if the distributed message passing iterative application is load balanced, the computation time of each CPU $i$ noted -$Tcp_{i}$ might be different and different frequency scaling factors might be +$\Tcp_{i}$ might be different and different frequency scaling factors might be computed in order to decrease the overall energy consumption of the application and reduce slack times. The communication time of a processor $i$ is noted as -$Tcm_{i}$ and could contain slack times when communicating with slower +$\Tcm_{i}$ and could contain slack times when communicating with slower nodes, see Figure~\ref{fig:heter}. Therefore, all nodes do not have equal communication times. While the dynamic energy is computed according to the frequency scaling factor and the dynamic power of each node as in @@ -411,9 +423,9 @@ heterogeneous platform during one iteration is the summation of all dynamic and static energies for each processor. It is computed as follows: \begin{multline} \label{eq:energy} - E = \sum_{i=1}^{N} {(S_i^{-2} \cdot Pd_{i} \cdot Tcp_i)} + {} \\ - \sum_{i=1}^{N} (Ps_{i} \cdot (\max_{i=1,2,\dots,N} (Tcp_i \cdot S_{i}) + - {MinTcm))} + E = \sum_{i=1}^{N} {(S_i^{-2} \cdot \Pd_{i} \cdot \Tcp_i)} + {} \\ + \sum_{i=1}^{N} (\Ps_{i} \cdot (\max_{i=1,2,\dots,N} (\Tcp_i \cdot S_{i}) + + {\MinTcm))} \end{multline} Reducing the frequencies of the processors according to the vector of @@ -460,9 +472,9 @@ scaling down the frequencies of some processors) and the initial one (with maximum frequency for all nodes) as follows: \begin{multline} \label{eq:pnorm} - P_\textit{Norm} = \frac{T_\textit{New}}{T_\textit{Old}}\\ - {} = \frac{ \max_{i=1,2,\dots,N} (Tcp_{i} \cdot S_{i}) +MinTcm} - {\max_{i=1,2,\dots,N}{(Tcp_i+Tcm_i)}} + \Pnorm = \frac{\Tnew}{\Told}\\ + {} = \frac{ \max_{i=1,2,\dots,N} (\Tcp_{i} \cdot S_{i}) +\MinTcm} + {\max_{i=1,2,\dots,N}{(\Tcp_i+\Tcm_i)}} \end{multline} @@ -470,13 +482,13 @@ In the same way, the energy is normalized by computing the ratio between the con while scaling down the frequency and the consumed energy with maximum frequency for all nodes: \begin{multline} \label{eq:enorm} - E_\textit{Norm} = \frac{E_\textit{Reduced}}{E_\textit{Original}} \\ - {} = \frac{ \sum_{i=1}^{N}{(S_i^{-2} \cdot Pd_i \cdot Tcp_i)} + - \sum_{i=1}^{N} {(Ps_i \cdot T_{New})}}{\sum_{i=1}^{N}{( Pd_i \cdot Tcp_i)} + - \sum_{i=1}^{N} {(Ps_i \cdot T_{Old})}} + \Enorm = \frac{\Ereduced}{\Eoriginal} \\ + {} = \frac{ \sum_{i=1}^{N}{(S_i^{-2} \cdot \Pd_i \cdot \Tcp_i)} + + \sum_{i=1}^{N} {(\Ps_i \cdot \Tnew)}}{\sum_{i=1}^{N}{( \Pd_i \cdot \Tcp_i)} + + \sum_{i=1}^{N} {(\Ps_i \cdot \Told)}} \end{multline} -Where $E_\textit{Reduced}$ and $E_\textit{Original}$ are computed using (\ref{eq:energy}) and - $T_{New}$ and $T_{Old}$ are computed as in (\ref{eq:pnorm}). +Where $\Ereduced$ and $\Eoriginal$ are computed using (\ref{eq:energy}) and + $\Tnew$ and $\Told$ are computed as in (\ref{eq:pnorm}). While the main goal is to optimize the energy and execution time at the same time, the normalized @@ -491,9 +503,9 @@ execution time following the same direction. Therefore, the equation of the normalized execution time is inverted which gives the normalized performance equation, as follows: \begin{multline} \label{eq:pnorm_inv} - P_\textit{Norm} = \frac{T_\textit{Old}}{T_\textit{New}}\\ - = \frac{\max_{i=1,2,\dots,N}{(Tcp_i+Tcm_i)}} - { \max_{i=1,2,\dots,N} (Tcp_{i} \cdot S_{i}) + MinTcm} + \Pnorm = \frac{\Told}{\Tnew}\\ + = \frac{\max_{i=1,2,\dots,N}{(\Tcp_i+\Tcm_i)}} + { \max_{i=1,2,\dots,N} (\Tcp_{i} \cdot S_{i}) + \MinTcm} \end{multline} @@ -517,10 +529,10 @@ performance) at the same time, see Figure~\ref{fig:r1} or Figure~\ref{fig:r2}. Then the objective function has the following form: \begin{equation} \label{eq:max} - Max Dist = - \max_{i=1,\dots F, j=1,\dots,N} - (\overbrace{P_\textit{Norm}(S_{ij})}^{\text{Maximize}} - - \overbrace{E_\textit{Norm}(S_{ij})}^{\text{Minimize}} ) + \MaxDist = + \mathop{\max_{i=1,\dots F}}_{j=1,\dots,N} + (\overbrace{\Pnorm(S_{ij})}^{\text{Maximize}} - + \overbrace{\Enorm(S_{ij})}^{\text{Minimize}} ) \end{equation} where $N$ is the number of nodes and $F$ is the number of available frequencies for each node. Then, the optimal set of scaling factors that satisfies (\ref{eq:max}) can be selected. @@ -561,14 +573,14 @@ frequency scaling factors are computed as a ratio between the computation time of the slowest node and the computation time of the node $i$ as follows: \begin{equation} \label{eq:Scp} - Scp_{i} = \frac{\max_{i=1,2,\dots,N}(Tcp_i)}{Tcp_i} + \Scp_{i} = \frac{\max_{i=1,2,\dots,N}(\Tcp_i)}{\Tcp_i} \end{equation} Using the initial frequency scaling factors computed in (\ref{eq:Scp}), the algorithm computes the initial frequencies for all nodes as a ratio between the maximum frequency of node $i$ -and the computation scaling factor $Scp_i$ as follows: +and the computation scaling factor $\Scp_i$ as follows: \begin{equation} \label{eq:Fint} - F_{i} = \frac{Fmax_i}{Scp_i},~{i=1,2,\cdots,N} + F_{i} = \frac{\Fmax_i}{\Scp_i},~{i=1,2,\dots,N} \end{equation} If the computed initial frequency for a node is not available in the gears of that node, it is replaced by the nearest available frequency. In @@ -620,41 +632,40 @@ maximum distance between the energy curve and the performance curve is while % \footnotesize \Require ~ \begin{description} - \item[$Tcp_i$] array of all computation times for all nodes during one iteration and with highest frequency. - \item[$Tcm_i$] array of all communication times for all nodes during one iteration and with highest frequency. - \item[$Fmax_i$] array of the maximum frequencies for all nodes. - \item[$Pd_i$] array of the dynamic powers for all nodes. - \item[$Ps_i$] array of the static powers for all nodes. - \item[$Fdiff_i$] array of the difference between two successive frequencies for all nodes. + \item[$\Tcp_i$] array of all computation times for all nodes during one iteration and with highest frequency. + \item[$\Tcm_i$] array of all communication times for all nodes during one iteration and with highest frequency. + \item[$\Fmax_i$] array of the maximum frequencies for all nodes. + \item[$\Pd_i$] array of the dynamic powers for all nodes. + \item[$\Ps_i$] array of the static powers for all nodes. + \item[$\Fdiff_i$] array of the difference between two successive frequencies for all nodes. \end{description} - \Ensure $Sopt_1,Sopt_2 \dots, Sopt_N$ is a vector of optimal scaling factors + \Ensure $\Sopt_1,\Sopt_2 \dots, \Sopt_N$ is a vector of optimal scaling factors - \State $ Scp_i \gets \frac{\max_{i=1,2,\dots,N}(Tcp_i)}{Tcp_i} $ - \State $F_{i} \gets \frac{Fmax_i}{Scp_i},~{i=1,2,\cdots,N}$ + \State $\Scp_i \gets \frac{\max_{i=1,2,\dots,N}(\Tcp_i)}{\Tcp_i} $ + \State $F_{i} \gets \frac{\Fmax_i}{\Scp_i},~{i=1,2,\cdots,N}$ \State Round the computed initial frequencies $F_i$ to the closest one available in each node. \If{(not the first frequency)} - \State $F_i \gets F_i+Fdiff_i,~i=1,\dots,N.$ + \State $F_i \gets F_i+\Fdiff_i,~i=1,\dots,N.$ \EndIf - \State $T_\textit{Old} \gets max_{~i=1,\dots,N } (Tcp_i+Tcm_i)$ - \State $E_\textit{Original} \gets \sum_{i=1}^{N}{( Pd_i \cdot Tcp_i)} +\sum_{i=1}^{N} {(Ps_i \cdot T_{Old})}$ - \State $Sopt_{i} \gets 1,~i=1,\dots,N. $ - \State $Dist \gets 0 $ + \State $\Told \gets max_{~i=1,\dots,N } (\Tcp_i+\Tcm_i)$ + \State $\Eoriginal \gets \sum_{i=1}^{N}{( \Pd_i \cdot \Tcp_i)} +\sum_{i=1}^{N} {(\Ps_i \cdot \Told)}$ + \State $\Sopt_{i} \gets 1,~i=1,\dots,N. $ + \State $\Dist \gets 0 $ \While {(all nodes not reach their minimum frequency)} \If{(not the last freq. \textbf{and} not the slowest node)} - \State $F_i \gets F_i - Fdiff_i,~i=1,\dots,N.$ - \State $S_i \gets \frac{Fmax_i}{F_i},~i=1,\dots,N.$ + \State $F_i \gets F_i - \Fdiff_i,~i=1,\dots,N.$ + \State $S_i \gets \frac{\Fmax_i}{F_i},~i=1,\dots,N.$ \EndIf - \State $T_{New} \gets max_\textit{~i=1,\dots,N} (Tcp_{i} \cdot S_{i}) + MinTcm $ - \State $E_\textit{Reduced} \gets \sum_{i=1}^{N}{(S_i^{-2} \cdot Pd_i \cdot Tcp_i)} + $ \hspace*{43 mm} - $\sum_{i=1}^{N} {(Ps_i \cdot T_{New})} $ - \State $ P_\textit{Norm} \gets \frac{T_\textit{Old}}{T_\textit{New}}$ - \State $E_\textit{Norm}\gets \frac{E_\textit{Reduced}}{E_\textit{Original}}$ + \State $\Tnew \gets max_\textit{~i=1,\dots,N} (\Tcp_{i} \cdot S_{i}) + \MinTcm $ + \State $\Ereduced \gets \sum_{i=1}^{N}{(S_i^{-2} \cdot \Pd_i \cdot \Tcp_i)} + \sum_{i=1}^{N} {(\Ps_i \cdot \rlap{\Tnew)}} $ + \State $\Pnorm \gets \frac{\Told}{\Tnew}$ + \State $\Enorm\gets \frac{\Ereduced}{\Eoriginal}$ \If{$(\Pnorm - \Enorm > \Dist)$} - \State $Sopt_{i} \gets S_{i},~i=1,\dots,N. $ + \State $\Sopt_{i} \gets S_{i},~i=1,\dots,N. $ \State $\Dist \gets \Pnorm - \Enorm$ \EndIf \EndWhile - \State Return $Sopt_1,Sopt_2,\dots,Sopt_N$ + \State Return $\Sopt_1,\Sopt_2,\dots,\Sopt_N$ \end{algorithmic} \caption{frequency scaling factors selection algorithm} \label{HSA} @@ -1143,7 +1154,7 @@ degradation. \subsection{The comparison of the proposed scaling algorithm } \label{sec.compare_EDP} -In this section, the scaling factors selection algorithm, called MaxDist, +In this section, the scaling factors selection algorithm, called $\MaxDist$, is compared to Spiliopoulos et al. algorithm \cite{Spiliopoulos_Green.governors.Adaptive.DVFS}, called EDP. They developed a green governor that regularly applies an online frequency selecting algorithm to reduce the energy consumed by a multicore architecture without degrading much its performance. The algorithm selects the frequencies that minimize the energy and delay products, $EDP=Energy\cdot Delay$ using the predicted overall energy consumption and execution time delay for each frequency. To fairly compare both algorithms, the same energy and execution time models, equations (\ref{eq:energy}) and (\ref{eq:fnew}), were used for both algorithms to predict the energy consumption and the execution times. Also Spiliopoulos et al. algorithm was adapted to start the search from the @@ -1253,6 +1264,5 @@ Babylon (Iraq) for supporting his work. % LocalWords: Fanfakh Charr FIXME Tianhe DVFS HPC NAS NPB SMPI Rauber's Rauber % LocalWords: CMOS EPSA Franche Comté Tflop Rünger IUT Maréchal Juin cedex GPU -% LocalWords: de badri muslim MPI TcpOld TcmOld dNew dOld cp Sopt Tcp Tcm Ps -% LocalWords: Scp Fmax Fdiff SimGrid GFlops Xeon EP BT GPUs CPUs AMD +% LocalWords: de badri muslim MPI SimGrid GFlops Xeon EP BT GPUs CPUs AMD % LocalWords: Spiliopoulos scalability -- 2.39.5 From adcfb91e645a60b4373b27438bcc01e994c40fff Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 15:20:32 +0100 Subject: [PATCH 12/16] Rewrite formulae to be shorter. --- Heter_paper.tex | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index 0491692..9ac2e0c 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -648,7 +648,8 @@ maximum distance between the energy curve and the performance curve is while \State $F_i \gets F_i+\Fdiff_i,~i=1,\dots,N.$ \EndIf \State $\Told \gets max_{~i=1,\dots,N } (\Tcp_i+\Tcm_i)$ - \State $\Eoriginal \gets \sum_{i=1}^{N}{( \Pd_i \cdot \Tcp_i)} +\sum_{i=1}^{N} {(\Ps_i \cdot \Told)}$ + % \State $\Eoriginal \gets \sum_{i=1}^{N}{( \Pd_i \cdot \Tcp_i)} +\sum_{i=1}^{N} {(\Ps_i \cdot \Told)}$ + \State $\Eoriginal \gets \sum_{i=1}^{N}{( \Pd_i \cdot \Tcp_i + \Ps_i \cdot \Told)}$ \State $\Sopt_{i} \gets 1,~i=1,\dots,N. $ \State $\Dist \gets 0 $ \While {(all nodes not reach their minimum frequency)} @@ -657,7 +658,8 @@ maximum distance between the energy curve and the performance curve is while \State $S_i \gets \frac{\Fmax_i}{F_i},~i=1,\dots,N.$ \EndIf \State $\Tnew \gets max_\textit{~i=1,\dots,N} (\Tcp_{i} \cdot S_{i}) + \MinTcm $ - \State $\Ereduced \gets \sum_{i=1}^{N}{(S_i^{-2} \cdot \Pd_i \cdot \Tcp_i)} + \sum_{i=1}^{N} {(\Ps_i \cdot \rlap{\Tnew)}} $ +% \State $\Ereduced \gets \sum_{i=1}^{N}{(S_i^{-2} \cdot \Pd_i \cdot \Tcp_i)} + \sum_{i=1}^{N} {(\Ps_i \cdot \rlap{\Tnew)}} $ + \State $\Ereduced \gets \sum_{i=1}^{N}{(S_i^{-2} \cdot \Pd_i \cdot \Tcp_i + \Ps_i \cdot \rlap{\Tnew)}} $ \State $\Pnorm \gets \frac{\Told}{\Tnew}$ \State $\Enorm\gets \frac{\Ereduced}{\Eoriginal}$ \If{$(\Pnorm - \Enorm > \Dist)$} -- 2.39.5 From ee26b603924e0c91a3e8bfda8acb626fe1502394 Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 15:31:12 +0100 Subject: [PATCH 13/16] Use \mathit between $'s --- Heter_paper.tex | 36 +++++++++++++++++++++++------------- 1 file changed, 23 insertions(+), 13 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index 9ac2e0c..c48c927 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -210,14 +210,14 @@ The work presented in this paper concerns the second type of platform, with heterogeneous CPUs. Many methods were conceived to reduce the energy consumption of this type of platform. Naveen et al.~\cite{Naveen_Power.Efficient.Resource.Scaling} developed a method that -minimizes the value of $energy\cdot delay^2$ (the delay is the sum of slack -times that happen during synchronous communications) by dynamically assigning -new frequencies to the CPUs of the heterogeneous cluster. Lizhe et -al.~\cite{Lizhe_Energy.aware.parallel.task.scheduling} proposed an algorithm -that divides the executed tasks into two types: the critical and non critical -tasks. The algorithm scales down the frequency of non critical tasks -proportionally to their slack and communication times while limiting the -performance degradation percentage to less than +minimizes the value of $\mathit{energy}\times \mathit{delay}^2$ (the delay is +the sum of slack times that happen during synchronous communications) by +dynamically assigning new frequencies to the CPUs of the heterogeneous +cluster. Lizhe et al.~\cite{Lizhe_Energy.aware.parallel.task.scheduling} +proposed an algorithm that divides the executed tasks into two types: the +critical and non critical tasks. The algorithm scales down the frequency of non +critical tasks proportionally to their slack and communication times while +limiting the performance degradation percentage to less than 10\%. In~\cite{Joshi_Blackbox.prediction.of.impact.of.DVFS}, they developed a heterogeneous cluster composed of two types of Intel and AMD processors. They use a gradient method to predict the impact of DVFS operations on performance. @@ -1156,11 +1156,21 @@ degradation. \subsection{The comparison of the proposed scaling algorithm } \label{sec.compare_EDP} -In this section, the scaling factors selection algorithm, called $\MaxDist$, -is compared to Spiliopoulos et al. algorithm \cite{Spiliopoulos_Green.governors.Adaptive.DVFS}, called EDP. -They developed a green governor that regularly applies an online frequency selecting algorithm to reduce the energy consumed by a multicore architecture without degrading much its performance. The algorithm selects the frequencies that minimize the energy and delay products, $EDP=Energy\cdot Delay$ using the predicted overall energy consumption and execution time delay for each frequency. -To fairly compare both algorithms, the same energy and execution time models, equations (\ref{eq:energy}) and (\ref{eq:fnew}), were used for both algorithms to predict the energy consumption and the execution times. Also Spiliopoulos et al. algorithm was adapted to start the search from the -initial frequencies computed using the equation (\ref{eq:Fint}). The resulting algorithm is an exhaustive search algorithm that minimizes the EDP and has the initial frequencies values as an upper bound. +In this section, the scaling factors selection algorithm, called MaxDist, is +compared to Spiliopoulos et al. algorithm +\cite{Spiliopoulos_Green.governors.Adaptive.DVFS}, called EDP. They developed a +green governor that regularly applies an online frequency selecting algorithm to +reduce the energy consumed by a multicore architecture without degrading much +its performance. The algorithm selects the frequencies that minimize the energy +and delay products, $\mathit{EDP}=\mathit{energy}\times \mathit{delay}$ using +the predicted overall energy consumption and execution time delay for each +frequency. To fairly compare both algorithms, the same energy and execution +time models, equations (\ref{eq:energy}) and (\ref{eq:fnew}), were used for both +algorithms to predict the energy consumption and the execution times. Also +Spiliopoulos et al. algorithm was adapted to start the search from the initial +frequencies computed using the equation (\ref{eq:Fint}). The resulting algorithm +is an exhaustive search algorithm that minimizes the EDP and has the initial +frequencies values as an upper bound. Both algorithms were applied to the parallel NAS benchmarks to compare their efficiency. Table~\ref{table:compare_EDP} presents the results of comparing the -- 2.39.5 From ce971d72d98fa648a7ed9ac0c615adea4de78b84 Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 15:31:47 +0100 Subject: [PATCH 14/16] Use abbreviations for months in bib file. --- my_reference.bib | 34 +++++++++++++++++----------------- 1 file changed, 17 insertions(+), 17 deletions(-) diff --git a/my_reference.bib b/my_reference.bib index 5ceb5d6..a21de7b 100644 --- a/my_reference.bib +++ b/my_reference.bib @@ -88,7 +88,7 @@ author={Rountree, B. and Lowenthal, D.K. and Funk, S. and Freeh, Vincent W. and booktitle={Supercomputing, 2007. SC '07. Proceedings of the 2007 ACM/IEEE Conference on}, title={Bounding energy consumption in large-scale {MPI} programs}, year={2007}, -month={November}, +month=nov, pages={1-9}, keywords={Clustering algorithms;Delay effects;Dynamic voltage scaling;Energy consumption;Frequency;Government;Laboratories;Large-scale systems;Linear programming;Processor scheduling}, doi={10.1145/1362622.1362688} @@ -110,7 +110,7 @@ author={Kessler, C.W. and Melot, N. and Eitschberger, P. and Keller, J.}, booktitle={Power and Timing Modeling, Optimization and Simulation (PATMOS), 2013 23rd International Workshop on}, title={Crown scheduling: Energy-efficient resource allocation, mapping and discrete frequency scaling for collections of malleable streaming tasks}, year={2013}, -month={Sept}, +month=sep, pages={215-222}, keywords={cores;microprocessor chips;optimisation;power consumption;resource allocation;scaling circuits;scheduling;ILP;crown scheduling;data flows;discrete voltage-frequency scaling;dynamic discrete frequency scaling;dynamic rescaling;energy-efficient resource allocation;energy-optimal code;integer linear programming;malleable streaming tasks;many-core processor;mapping;optimization;pipelined task graph;power consumption;processor cores;streaming task collections;Dynamic scheduling;Optimization;Processor scheduling;Radio spectrum management;Resource management;Schedules}, doi={10.1109/PATMOS.2013.6662176} @@ -121,7 +121,7 @@ author={Kimura, H. and Sato, M. and Hotta, Y. and Boku, T. and Takahashi, D.}, booktitle={IEEE Cluster Computing, 2006}, title={Emprical study on Reducing Energy of Parallel Programs using Slack Reclamation by {DVFS} in a Power-scalable High Performance Cluster}, year={2006}, -month={Sept}, +month=sep, pages={1-10}, keywords={directed graphs;parallel programming;power aware computing;AMD Turion;PC clusters;PowerWatch;Transmeta Crusoe;control library;directed acyclic task graph;dynamic voltage scaling;energy consumption;energy reduction;frequency scaling;high performance computing;microprocessors;parallel programs;power consumption;power monitoring tools;slack reclamation;Clustering algorithms;Concurrent computing;Dynamic voltage scaling;Energy consumption;Energy efficiency;Frequency synchronization;Gears;Libraries;Microprocessors;Monitoring}, doi={10.1109/CLUSTR.2006.311839}, @@ -163,7 +163,7 @@ author={Huang, S. and Feng, W.}, booktitle={Cluster Computing and the Grid, 2009. CCGRID '09. 9th IEEE/ACM International Symposium on}, title={Energy-Efficient Cluster Computing via Accurate Workload Characterization}, year={2009}, -month={May}, +month=may, pages={68-75}, keywords={parallel processing;power aware computing;workstation clusters;cluster computer;eco-friendly daemon;energy consumption reduction;energy-efficient cluster computing;power consumption reduction;processor stall cycles;workload characterization;Application software;Clustering algorithms;Energy consumption;Energy efficiency;Frequency;Grid computing;Hardware;High performance computing;Runtime;Voltage}, doi={10.1109/CCGRID.2009.88} @@ -222,7 +222,7 @@ author={Dong Li and De Supinski, B.R. and Schulz, M. and Cameron, K. and Nikolop booktitle={Parallel Distributed Processing (IPDPS), 2010 IEEE International Symposium on}, title={Hybrid MPI/OpenMP power-aware computing}, year={2010}, -month={April}, +month=apr, pages={1-12}, keywords={message passing;parallel algorithms;power aware computing;HPC environment;dynamic concurrency throttling;dynamic voltage-and-frequency scaling;high performance computing;hybrid MPI-OpenMP computing;hybrid programming models;large-scale distributed systems;message passing interface;parallel programs;power-aware computing;power-aware performance prediction model;Concurrent computing;Discrete cosine transforms;Dynamic programming;Dynamic voltage scaling;Frequency;Heuristic algorithms;Large-scale systems;Multicore processing;Power system modeling;Predictive models;MPI;OpenMP;performance modeling;power-aware high -performance computing}, doi={10.1109/IPDPS.2010.5470463}, @@ -389,7 +389,7 @@ author={Yongpan Liu and Huazhong Yang and Dick, R.P. and Hui Wang and Li Shang}, booktitle={Quality Electronic Design, 2007. ISQED '07. 8th International Symposium on}, title={Thermal vs Energy Optimization for {DVFS}-Enabled Processors in Embedded Systems}, year={2007}, -month={March}, +month=mar, pages={204-209}, keywords={circuit optimisation;embedded systems;integrated circuit design;low-power electronics;microprocessor chips;nonlinear programming;thermal management (packaging);DVFS-enabled processors;application peak temperature;cooling costs;dynamic voltage voltage;embedded systems;energy consumption;frequency scaling;nonlinear programming;power optimization;run-time thermal emergencies;system thermal profile;thermal optimization;thermal-constrained energy optimization;Cooling;Cost function;Design optimization;Dynamic voltage scaling;Embedded system;Energy consumption;Frequency;Power system planning;Runtime;Temperature}, doi={10.1109/ISQED.2007.158} @@ -400,7 +400,7 @@ author={Fen Xie and Martonosi, M. and Malik, S.}, booktitle={Low Power Electronics and Design, 2005. ISLPED '05. Proceedings of the 2005 International Symposium on}, title={Bounds on power savings using runtime dynamic voltage scaling: an exact algorithm and a linear-time heuristic approximation}, year={2005}, -month={Aug}, +month=aug, pages={287-292}, keywords={approximation theory;energy conservation;low-power electronics;power consumption;power supply circuits;DVFS policy;discrete voltage/frequency voltage level;dynamic voltage scaling;dynamic voltage/frequency scaling;energy reduction technique;exponential algorithm;linear-time heuristic approximation;power reduction technique;switching cost;Approximation algorithms;Costs;Dynamic voltage scaling;Energy consumption;Frequency;Linear approximation;Power system modeling;Runtime;Semiconductor device modeling;Upper bound}, doi={10.1109/LPE.2005.195529} @@ -411,7 +411,7 @@ author={Lizhe Wang and von Laszewski, G. and Dayal, J. and Fugang Wang}, booktitle={Cluster, Cloud and Grid Computing (CCGrid), 2010 10th IEEE/ACM International Conference on}, title={Towards Energy Aware Scheduling for Precedence Constrained Parallel Tasks in a Cluster with {DVFS}}, year={2010}, -month={May}, +month=may, pages={368-377}, keywords={environmental factors;parallel processing;power aware computing;scheduling;workstation clusters;dynamic voltage frequency scaling technique;energy aware scheduling heuristics;green service level agreement;high end computing;precedence constrained parallel tasks;Computational modeling;Concurrent computing;Costs;Dynamic voltage scaling;Energy consumption;Frequency;Grid computing;High performance computing;Power engineering computing;Processor scheduling;Cluster Computing;Green Computing;Task Scheduling}, doi={10.1109/CCGRID.2010.19} @@ -450,7 +450,7 @@ author={Kappiah, N. and Freeh, Vincent W. and Lowenthal, D.K.}, booktitle={Supercomputing, 2005. Proceedings of the ACM/IEEE SC 2005 Conference}, title={Just In Time Dynamic Voltage Scaling: Exploiting Inter-Node Slack to Save Energy in {MPI} Programs}, year={2005}, -month={Nov}, +month=nov, pages={33-33}, keywords={Computer science;Dynamic voltage scaling;Energy consumption;Energy efficiency;Frequency;Gears;Jitter;Microprocessors;Performance loss;Permission}, doi={10.1109/SC.2005.39} @@ -505,11 +505,11 @@ author = {Wei Liu and Wei Du and Jing Chen and Wei Wang and GuoSun Zeng} @inproceedings{Le_DVFS.Laws.of.Diminishing.Returns, author = {Le Sueur, Etienne and Gernot Heiser}, - month = {October}, + month = oct, year = {2010}, title = {Dynamic Voltage and Frequency Scaling: The Laws of Diminishing Returns}, booktitle = {Proceedings of the 2010 Workshop on Power Aware Computing and Systems (HotPower'10)}, - address = {Vancouver, Canada} + NOaddress = {Vancouver, Canada} } @inproceedings{Cochran_Pack_and_Cap_Adaptive_DVFS, @@ -533,7 +533,7 @@ author={Dhiman, G. and Rosing, T.S.}, booktitle={Low Power Electronics and Design (ISLPED), 2007 ACM/IEEE International Symposium on}, title={Dynamic voltage frequency scaling for multi-tasking systems using online learning}, year={2007}, -month={Aug}, +month=aug, pages={207-212}, keywords={Linux;computer aided instruction;multiprogramming;power aware computing;program compilers;system monitoring;Intel PXA27x;Linux 2.6.9;dynamic voltage frequency scaling;multitasking systems;online learning;processors runtime statistics;Batteries;Computer applications;Delay;Dynamic voltage scaling;Embedded system;Energy consumption;Frequency estimation;Linux;Power engineering computing;Statistics;dynamic voltage frequency scaling;online learning}, doi={10.1145/1283780.1283825} @@ -573,7 +573,7 @@ author={Dhiman, G. and Rosing, T.S.}, journal={Computer-Aided Design of Integrated Circuits and Systems, IEEE Transactions on}, title={System-Level Power Management Using Online Learning}, year={2009}, -month={May}, +month=may, volume={28}, number={5}, pages={676-689}, @@ -611,7 +611,7 @@ ISSN={0278-0070} Parallel and Distributed Processing with Applications}, year = {2014}, - month = {Aug}, + month = aug, pages = {225--230}, publisher = {IEEE Computer Society}, address = {Milan, Italy} @@ -672,7 +672,7 @@ author={Kai Ma and Xue Li and Wei Chen and Chi Zhang and Xiaorui Wang}, booktitle={Parallel Processing (ICPP), 2012 41st International Conference on}, title={GreenGPU: A Holistic Approach to Energy Efficiency in GPU-CPU Heterogeneous Architectures}, year={2012}, -month={Sept}, +month=sep, pages={48-57}, keywords={energy consumption;graphics processing units;parallel architectures;AMD Phenom II CPU;CUDA framework;GPU-CPU heterogeneous architectures;GreenGPU;Nvidia GeForce GPU;energy consumption;energy efficiency;high performance computing;holistic approach;Algorithm design and analysis;Computer architecture;Frequency conversion;Graphics processing unit;Green products;Heuristic algorithms;Time frequency analysis;GPU;dynamic frequency scaling;energy efficiency;workload division}, doi={10.1109/ICPP.2012.31}, @@ -686,7 +686,7 @@ author={Rong Ge and Vogt, R. and Majumder, J. and Alam, A. and Burtscher, M. and booktitle={Parallel Processing (ICPP), 2013 42nd International Conference on}, title={Effects of Dynamic Voltage and Frequency Scaling on a K20 GPU}, year={2013}, -month={Oct}, +month=oct, pages={826-833}, keywords={energy conservation;graphics processing units;parallel processing;power aware computing;power consumption;DVFS schedulers;GPU computing;K20 GPU;Nvidia K20c Kepler GPU;application performance;compute-bound high-performance workloads;dual Intel Sandy Bridge CPU;dynamic voltage and frequency scaling;energy efficiency;high-throughput workloads;power consumption;power-aware heterogeneous system;Benchmark testing;Computer architecture;Energy consumption;Graphics processing units;Market research;Measurement;Power demand;DVFS in GPU Computing;Dynamic Voltage and Frequency Scaling;Energy-Efficient Computing}, doi={10.1109/ICPP.2013.98}, @@ -746,7 +746,7 @@ author={Spiliopoulos, V. and Kaxiras, S. and Keramidas, G.}, booktitle={International Green Computing Conference and Workshops (IGCC)}, title={Green governors: A framework for Continuously Adaptive DVFS}, year={2011}, -month={July}, +month=jul, pages={1-8}, doi={10.1109/IGCC.2011.6008552} } -- 2.39.5 From 2457bc79b2e872cd82b05ecf6b8a290c614cf43f Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 15:39:23 +0100 Subject: [PATCH 15/16] Move around floats to keep under 10 pages. --- Heter_paper.tex | 58 ++++++++++++++++++++----------------------------- 1 file changed, 24 insertions(+), 34 deletions(-) diff --git a/Heter_paper.tex b/Heter_paper.tex index c48c927..db8a304 100644 --- a/Heter_paper.tex +++ b/Heter_paper.tex @@ -1139,6 +1139,24 @@ degradation. \label{table:res_s2} \end{table} +\begin{table}[!t] + \caption{Comparing the proposed algorithm} + \centering +\begin{tabular}{|*{7}{r|}} +\hline +Program & \multicolumn{2}{c|}{Energy saving \%} & \multicolumn{2}{c|}{Perf. degradation \%} & \multicolumn{2}{c|}{Distance} \\ \cline{2-7} +name & EDP & MaxDist & EDP & MaxDist & EDP & MaxDist \\ \hline +CG & 27.58 & 31.25 & 5.82 & 7.12 & 21.76 & 24.13 \\ \hline +MG & 29.49 & 33.78 & 3.74 & 6.41 & 25.75 & 27.37 \\ \hline +LU & 19.55 & 28.33 & 0.0 & 0.01 & 19.55 & 28.22 \\ \hline +EP & 28.40 & 27.04 & 4.29 & 0.49 & 24.11 & 26.55 \\ \hline +BT & 27.68 & 32.32 & 6.45 & 7.87 & 21.23 & 24.43 \\ \hline +SP & 20.52 & 24.73 & 5.21 & 2.78 & 15.31 & 21.95 \\ \hline +FT & 27.03 & 31.02 & 2.75 & 2.54 & 24.28 & 28.48 \\ \hline + +\end{tabular} +\label{table:compare_EDP} +\end{table} \begin{figure}[!t] \centering @@ -1151,7 +1169,12 @@ degradation. \caption{The comparison of the three power scenarios} \end{figure} - +\begin{figure}[!t] + \centering + \includegraphics[scale=0.5]{fig/compare_EDP.pdf} + \caption{Trade-off comparison for NAS benchmarks class C} + \label{fig:compare_EDP} +\end{figure} \subsection{The comparison of the proposed scaling algorithm } @@ -1188,39 +1211,6 @@ because it maximizes the distance between the energy saving and the performance degradation values while giving the same weight for both metrics. - - -\begin{table}[!t] - \caption{Comparing the proposed algorithm} - \centering -\begin{tabular}{|*{7}{r|}} -\hline -Program & \multicolumn{2}{c|}{Energy saving \%} & \multicolumn{2}{c|}{Perf. degradation \%} & \multicolumn{2}{c|}{Distance} \\ \cline{2-7} -name & EDP & MaxDist & EDP & MaxDist & EDP & MaxDist \\ \hline -CG & 27.58 & 31.25 & 5.82 & 7.12 & 21.76 & 24.13 \\ \hline -MG & 29.49 & 33.78 & 3.74 & 6.41 & 25.75 & 27.37 \\ \hline -LU & 19.55 & 28.33 & 0.0 & 0.01 & 19.55 & 28.22 \\ \hline -EP & 28.40 & 27.04 & 4.29 & 0.49 & 24.11 & 26.55 \\ \hline -BT & 27.68 & 32.32 & 6.45 & 7.87 & 21.23 & 24.43 \\ \hline -SP & 20.52 & 24.73 & 5.21 & 2.78 & 15.31 & 21.95 \\ \hline -FT & 27.03 & 31.02 & 2.75 & 2.54 & 24.28 & 28.48 \\ \hline - -\end{tabular} -\label{table:compare_EDP} -\end{table} - - - - - -\begin{figure}[!t] - \centering - \includegraphics[scale=0.5]{fig/compare_EDP.pdf} - \caption{Trade-off comparison for NAS benchmarks class C} - \label{fig:compare_EDP} -\end{figure} - - \section{Conclusion} \label{sec.concl} In this paper, a new online frequency selecting algorithm has been presented. It -- 2.39.5 From 4eb9ad4870b37f5f73dbc34ee177a966250404c2 Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Mon, 15 Dec 2014 16:15:37 +0100 Subject: [PATCH 16/16] Fix compare_EDP.pdf too. --- fig/compare_EDP.pdf | Bin 17668 -> 18711 bytes 1 file changed, 0 insertions(+), 0 deletions(-) diff --git a/fig/compare_EDP.pdf b/fig/compare_EDP.pdf index 1aaef32784aa33acf2d183e082a98e9db5f49192..62b710d6a43fe6018694ad1717b63046b5ef93a3 100644 GIT binary patch literal 18711 zcmb`v1yEeu(gunXGz51B3(hdOy9aj-!6C@N;BG;JOK^902=2jxdvFM@fnb3Uctg%P z_ns^7|KF?nchxX^_v&8OU-zb_S2wMSq%;eNl@pbA@Hi$9l?%WI0Gqr;6%+(2nA=-A zTLHMC6g8ljcIitlvd1*xQDJALOsSTbwWY9s&=Z&aSo0QENyaF}|g7dOME!={i~RZfaYqH2($zR!WMpAlUu$+eAFbHuK`zWWR~StD~P922mt|1&a^T z`R|3s-m*v~V8=VgISI09WbL5Seec|p8b5z?loXz^8B$|kKE#*Bu~Hm)BIKTW#4SzE zwi?6__Dq}gB({r+o%Ia2vW=wFIJAaz;Shv{4Ru4ivTuivnv7CO?YZ7dYUK%P zUflBw(Pr_f{(OaMW&I^PgBBhRf*Q3~Cl{3GKqOj_tO?=S$;kL2{Y~g(Cvk|8Tc5I1 z3XnykiW%aas2}ods0qF#udC*M`Fe*bt>LlZMoJzo{zF%gMo^(<(44SdvX|OOt17=s zeL2LYU|B$%;lviJl$zF<+)$vkz&GI)wm`bqj8#KVVTghfM6i*>E2vWA$|OTRd=^PM zYkR-u7Nc_WoZsBu>`!BX-aQ%ElUY4|^RjWE{z>Ber|oHYIG6)fj4jQ9isok4#^PXi zfIb_vnS+-dz|O{Jhze8#J3}oIzy{@lnA*_R zp(YLhYM8q_L*=C0on_R2i-Oqy5RnE$#X*0_Lv0!;0eyz{=mdE31!|~J(x3BtV)?^2 zKzoBu5b%43P5|I5YcnT+KI+r_)ldxq&>mc#y8N3c=pUkgN=g_z8{2{{|LWq;*(;hm z8=FB}03hCf8=L1}JNV}^P=Vr#KqWB5&e#?esH~z09TuqZ3g8TJfjX!^C;oe-e|j}F z9@MA4)c(1w-%J0il^6B*$Urp^8!G5&$N$dwm(HMj`-eaLyF2{LU$XOngoXd%7!Biu z?B9u^OM5ASc)CMi@vV0(W^Ci=iD%|vIy zKAQP!pA7jUNh+K9S4(=a=bY!k6HV@eBhCE5Qt(77N18a$KcOk(&&pEpoQNAFAsoA$ zCGYhN$E;OwTEI$G1U#;M2v;W+BabqX*Cny4ylvNdF6`2&QP0S(2&emXibz^%c7ICnx8StrH z(%i`uV(s7zh5-1U^aIoYYuZ~sy^sK)Ki*ZpjpvCK6{;qPle2`CF$4etDj5HH%gzqe zwl;INa?%Izu(Jbrcz6N492@{{5a{GRa~MEm^={pRB0{`&{@>G_H0FFHH4 zpQn2c4i2cDsk@jsKN+zG6o^j~`WxQgeIEcl2*yy;qw1JnxSy10lABDl`jNFC5U3Pnl!X5|A&7I7W z6YB_}zIo4;1LJCCYWukJo;QsJ<%FV8MNi+Cg<{-X>5{~H)c=?{R{a5?{sP!kT{<)%m;PanQ{yixrV>|Qz4gKQKSAjk^faM9R z+)w!9G<C>wKN`!iGV+UDtYfGy?AD@PLYJ&=*0_C8-*V>WSivqn)(xRb57tX!hzm74Vq>t=UL~?JL$r!ZWQm`YD7p zQFdp6altG#y3C8%2;$fZaIol{X3EO=DYwqP*uH@Z&r&95as_4%F4AeWrx|r5R=7-Gm{=6X7$ypX zb`PaH^MURjhUD_?&a=%#ArkSg(K`eda9?rW4H+`Qm(TAskNZVoB?QH$={LI0e5UvL zO#4oO&qo+sU0|gBUd5OMV0j5^V!jQ_C9X%;!0~fFn0sK`Fm$l4@qrwy@S0RGMD7tRUBCl#1YP3Ur^y?*lL7?=eoEqu;{GL_N9^Ev*MA899KR z!1+I$D~H~3GfXVAx^l})yHRsK;MuKuODE+$mM2Sle09zU?vyfGn?85Cg!Lo~o?6r| zb>trJXLP5_d$mt;IdAk8^&X*J6c(7{w4c1+*d;168Sw+@4nrxFPZ$=fUo-cj8ST;o z#o=h`wKo~K7vC9+Byo*uqjK>0`B_=_);^Wh4Pk2-c^}t46@l4w?iSB>@;;FTlLvAp zY$icmhd+r7MSfE)&O#o5KO{?=5J7BhO`nhb*)K9rR3Ueqkq+!zhNeVP8I&zQ@xdu& zyyv!J)Fob!NzHq%wdd!_(Om8(N`Bl^rB$X{30v=0+;NGY=W9B6or42BfgF37POrm% zQ2V4v!~`OB4JukuIC#)P*dfnYWMqb9O7MO$?Hl&dZVL;>Gh41>ubg2K*MkI3<-{yldqnB=*P{2( zcg-K3;`e<#WwWWDpTnE{IEVJ^8%x+}%`^I&6Hen=eM%f~@rMA&4%r9D2eqt%Ptvud zi3Hmbjas04r6^6(J~Bl+m~WM}A0swLP}@}kbknE9LPPrDs6Q<=bP|Ii+xnMkyd>6T z?c6ZVNan1|E9Zpm#Dt0G4q;|tamQqGwE`r5QSh`l>d&EZQ|9|F&4>I+AEA+}_Kex^bbcf@Rd*`1qN+ z0DB#E#+&DRdN#c+HwN)M+rCbfwFAzq7zORU~({l5Is9 zYv>MuUm&h-ti`=((sYh-j&UCx2-Vh+%6Bp2WVNSJK_i!rvSWK|dw;dMTjQnd_2tWI zsO4a!3$W@r9hb|;V7Vmy;|n;XPED;5g?idtpTMGXpbNDFLNmPg+gg=Ax0qooM9$tm zMZ2kwVOl9MREx&o=Lz;pPM&##v+`acbe=Cm-SmyQkJ6W=du`|MU)!+!aBt!fY=`@5 z&6SB-4yt%eSTrG%Q^2z1Tf)T$2^GaW*3Rd)bV7J6|_Q3zente*%1F`vCFQyG3 z!!}cZFI%H>>0y`KT-H{Hw}n5l)ypKmH+qSfQMZEZpz3{@sE2O)k`mRzkNzo>srxcC zy!>lo`3yA-NAq$r>mx%V{MNCKnrW?YJBwF3_Ms^DO&@QUtaAcFA;3W12r7_5nf}sY zdgYZOnH`5|xTFBR0C)K{CigtI(+7-q2096aFP6hnt&zp40g^m0^DR4P=;igC4kSIB zs7WD*E09`@^eF_ck{9E6Ha=V+{-w?+2ARInR{Oy8aP9zQm(0a!`}eOAZxLGiQ{E8f z<|`E|+-W0tYj4e+Y7Gvotl@JK{H^ z=;nCkiDeSGE{`?w^9U+iD{~*6?q`Wiu|!){T#ZqGXyF>b3fR5q3k#_cMZ;~vj$-Ux=f?!9V?5L(6=tdF5LuEBGbUD6xSkhsM7hR= z$(mU5ZI4W%CypSzYt-;E`A-KVb*ms7EnH0EkW{>6x7qr#qr6Xa1aSP9ZW=D+;3!K)G0 z^l8I=dk{GIMyQNNU5wwt!z%M78Hg|+-9jEcge5waD2Pa7cnFjt#W(s3lp>*?x@TI0 zx|m}0=?&Li$gL^*d=BY#0^w+WDJe;ii`7fIfeqe@%}`uvb$F3=}~6C4d@$SAG{nxsC<+iTr84cUQc2isU=SJIQaN z)nC~8YK(V2Kagh!9-m_bw~HP>zY16?UwD3n?Df;Dq~XnG0oU9^uD?i;)=*Oq9qVO? zcxR4Z4iFc0)Rp%X`a%~6M>Zo{p?TdOk@{5v8yGf@t`L7}WB6YM`{cUV5crG3i>Y{b zP`zkqnG3Rigi)AYqF3xByu6kCEd1-gI%xrq6K( z_nIAHA445%*EvTzAp3LXR+6&fglmj(gssqP${diOt2k2*4q+yi@(^9CGJ=xBW#i)~ zFS@F%je#2a!3Ee=o7L*Lpz4krV%U#$4_ujRcD#pt^3tg#;|C?Er%2624bQU(IhB*0 zY+LL%!?tAS;^1X;`d%hi&Gwyh3?)4^mM$WtIgq z&)#<1&TizvnXJ4M(Ejs_G~y!ED2HO6ZSL6XkOw685S$$YwJmR>=&I{_YE z*kakDc?4T4F8L6xKvgE`tzJSU-nYA(<{y|g>^Mg(OgJMfiuvAHAW9_mn0_&NgMk#9 z+q1!Q%wfP5NiwT?j0calQ;)p4i95OpYWHg>T(qyIP>+HB2UJcX(C2Q+Tb?SsQxskE z){?N%)%jzT$$kFyiPueD%gkY)nv;IBm!6kTGXYF0TptW!aa2)G%Sn$4TZ1rRoeYpP zcoKn%&_KBs*E}1bI@$Av9#?#dQp^i$G3a{I(;NW_3(+(`=tBl1!UyCd?$sewP?Ptr z2dl7Fo&F^{gY0MN7Ey6F7SE)T02$x{s#v9TaDL74NKDz{L2{ug-i_R?Y6wi2G*S32 z{auNkdBaExtdulQ2hJs@Jq+ab&4wEb=7eumMtU4pUreR*oS=YAp>0k z12$8E2c}!=!-cr)z;yOt{h~U76W;O}>rY7eZ)l{aOxe>FD~9;^s35B=Vy9GYi9u){ z&T=J(bt6+;Iy(?Tps9@Z!Vc?Jv44xfcEfUgfxO!11EQl#1ri1ZrHr_Pcy?Jv?Pr8S zAX-bTftB{s4rUb7+?)6HUP_XkeA7#JA~kpWLyL4?MbcGAmQ7%Tnq9CbY4n<vvXn8kM$FMN+p7IvuaiJZ*mK@g9g{5Z9Bl@`nJN<31A=8{9 zMvIr@kQGiY$9ZPtjgl5n&?{!zq<39!8x+?zY|^kj0xFzOy$}2_v%>F>n8xg?P5R!S zSZ1FRP-)|gEH5lPCp}Iu_=PMY`|*qV`h79UIzglM-bzDzvzN@s*ZUCma+6!a#*C#- zG|VJ?dnQQy_*O7grkbPeeO^U=te8Z$=iA%n>G@?_qF}~`QHIR8v-MYm-;&PoQIh4Z zritG>%6?(qw88GA+Uye*!<21SUt&HbcyPO;a9<~UcQS?9%@*fSn%B*eAYGfEWohfl!GB$O8w@3JGqVbV-B zu+|97l@Mz!7|KrV-=ns?w9dnR3^xb6RlS1>spAVsYesq`<#E5bt5-rq!v>7fGB|1lj^JT9y%Il< zoF*z{Y|K*S&?<@@{%BrYVDSd0fc%vjl9HlarmEiXhnKq8yc`mQCi2rXF9|i{Zm1b0 z#Hq3~dfDDZdB?aJEfgJoy}WbRG%z5&AzSNbcC2Kfk0ahSAKK}cTj;XDP6!{_iS)pz zx8&EhH1d1D?D>l+m%6&7;EVO=iF%@x%WqAezGev_I8NjqK(Nd0HB5^n-Z`O)r5Fp& z`X!f_Y_&#y!{@UOX-IM1wEd3Jbfq^zOV~vO88BoMCYu|fl_MiasP4TBT`iwlv9o~N zHRi}o{~~r_%~5{JR&s_n*BlBb|MrIc;XSMx|8r}Hohu*zm&E@sw3G~-~ojx{BuBY*KnF?E1d*lxe zZdco_yBay!m@G-J(9yMmwmj_G3BN@qZxKbu5S5?1^NP^?KwIG*G~8qxJg-3J@E(Qd zbaKIXtwr@rFiv#KTID=+7`x4_mSsT%pbaKWrCphS6}QJKb^RJ{K%?`;_EC>Sgi>C! z&NZhqmBXv1RL?NHQC{6u%C{NBd8!_y!@Rol`iJoeb8_TaszGq!-d@@oo!FXEN@jx_ z*yX`WCU%dSak4Bv{ znSl{{H;*ALU#!il>Fbv7S%oBo6Gx0fy!~JL$^8?*3--RrBC2f7N z-qCag*_0fo(hcEE4c%DwM_^P-e8|jnMJaihUCbf=AckR9VTpart`7{3B_1E$CGv^l zFry370Eonb29F>f#YPM0anJG?#aQ>NlZT>gwe#YvJT8=bl?-L8{|19~+^D%skUg+z z{4_H=ltTBy6b#!mxAw8abW6F2mZFEkgP8i!&iY+;OAZ@f9ry5*>=q^7*I#5c>WUfP z@f*o;w5m;O{V*b2Dl%6`Pc5=7TG+{KECd$0j*YKz%Nf-tEx%kaTOy^g7g};$`V7A| z^}u(_pxhfye#a+W6Fr-t^qte#EkS)7y~2pEx;dJ5_}%`HFL}Lp*!loR=nvC-9>G3! z%V@v52^X6mV}jP3jzon++r!J+9`}@f@9B!&+oSvL)|?jA>{E;Pd{dJy zg07&sCHV5^kNSJWEV1?>fm@h~vKL32gcTL&C=yFopRg~#FRNdi=-y6X88&TuAzC9x zn)3NzU>N$MWsS6L*bG_^1ZRu2PZw&LfItS@O>w%%j@OGeLdW&VtPCDS=-+ct-xi}b zoYHVU&`$(w+JkzjtEsgQu^^lIiie__t(ZYtliXH9dR<8~FXapZ=s1AD1TIhBe5F}F zmq>NGqcEjGF^dAkJjXm&wfV`hUIsmWhKOf%=B`>uTk#ejx<3!IZp()=$yK=b;`FP( zMH(z!iFw#;5qkV#8iS%7!DPGCu>BtX`(fPO_o$yyrkSUi>j^{*L-@kZ>d)>9yne+~8lz?%$0U8!g>f zIS8~AfK{wmk)D**-UkE2q@%)!HqkPgby3E=Q6_Xf2F#%~TlH8vrs~ZTS%gZz`F21ZOwWLNfrPH(|29g_e@&;9r zZue$xfBrUZLM;RB)d+Ef0>ijey*zGob%w6{m|Lsj1w=UwWIJE+7(98lQve(X&y z_2q8<^GB18nU)8FSqjaZE!#60OwSFPx=15$jKjB6Cb?HdJt?nw#%zs?P}M07?!$`0 zO!Vc*KF?rGhEXn+s?P-eWCxnXGo|u!%+QAvl}3tH$u}vxi47V|8S92px=ts4Q4OT~ z=+;828gVl9L(XLp#~`6?=mgd*vr0`rzI!I}D@N~l)!Oy$S;$$@l6yu~s#g0O=|`X5 zid}5CP^{b5FzeS zHf*8X7I{YFsKt`}I!2=8x}XEy#(vHTzVF*xZtIQh^?ms<_hR})-X2>j?xs6=UgYaI z8S=iD7LO5LJQ$e6BLeNOO)se7HEEeB4A-0-21B|hGbZrXc?Y`QIEgi}H@vQD!Z1r# zdhttYQ9V4o$c%#%!07B-S-~@BZCQX^tN_~O`GWZ+Y|H{h)HIz#L(#yp}=Xli392&x{oi9b+Uc6r%->bb_wFauDvlDVR9X_bv zMBo1^LM68>VdhAe_X&tE+VNmY+F^0o^J+_~nvxWbv01(A{q7ZmIX09b(qMIP_La0f z|L35MWlK=ZY%#&?M9O?y*;}!y2ct-jtNP??=O)@4!RZrvqZTLQowRvW_;MDT3noi1 zy?k`>l}}Wxd%HgZx5@m zGejs4HTe_v>UO)d&8ZrWQ0Q`9rDq00D4seb92?GkA#dyzYPrNjgOBlNhI-vsPSY(z zx4a%_biV6O@bz4=9;_1i`V>xfYyynew5z(#D)RBRH=yVDc>*V^y6aLuA^=n<=Jg zT&gf35lRHrUwL~ctyAZ_N&SNs=G?tO2Q!y1(`JZKt(5v&)uR3QZ8vNC6|y6bJZZkaSi@NiAj|d( ziP-3P{w>m$1~V^0FjRZx$LT;DmdH;4Lwk2e<$cy1V4^!Cy*;-V7Oyj>m+^J!I_7ud zEZa>-o)|=Oqs?L$1vYw7Y#gw93x-k=`q5e5UTx>bp;?kf=-Sw-b@wM^up zg>k|}gb$z|S3J|kYJn4AtQoINRLIC@xcDVVI@Bkg1wW4C0sY-NLG8z6IGxn6ntTuZ zoKD4V%L<+lKBhDe>rcrYC#OBe!qV`U^m;$p5*lOD9vlk7lZ4TkMq6%l9Of==8dDxv zSgY$=J}nx6I!d^Bi!JoC4aoVA!Cr4o6VCn0IbN-Jbab>_l`|evkXQ8YM4TEP4waBu zeP&Ea&7Dh)sQVuNP9uo2dVIarLER?6e62f=VX+E5vs{qLFKzfu!va|4yZhabg>7T| zml-^IW(32zz7Y`@-BYZ(588^u>1w3% zE42|TkM*+KmON{AB?~35zhX^Yo`FDxthZ3E-v?%9H=6)VZx-{L{ zC}!?5b;gJkR89#Onbqz4$nIh@pXnJIw6DEUkX4Z9F1(+lFnM@v z*J%*-sGAkhRv#EV=&GEiYtY@1>o#Tn8jGq`gY@xwu0>rRDBcv3b8rz zx>jr!!3}Rke=`fYNg@{|57K;PhQEaTjsH8gytdiKJD^OT%@m=v;DWw zJKMj8-k%=-5qjtQe~G;F@%%4=_umJx{uX%u+ab;WDe%t6&i#+LI~Okx*Wcsr{5BdS z3#m8bvf{u=80NZbz(8GxSF!mmVQa26nWi7y{WxAxOmFv zH6C4%`t9fGP3!vBn!uN5)bAiuf_bXSZrLMI1rZUwb_+igXySS*(Z!qn?p4Bt7>>96 z&$z`um{5WE)OS3Nnqay@RieEik5li_yI_Ab?iHKF?RzdOpzQyI3Bynq-8T|S3yM92 z+xS?TIvZEf(G|Kp@az0s7OBY~Q2$n8Th^rMo|TxAYhU8n*6igq5RZV_e}S?PMT zmQGGCS1td*huRVuvN(Uo>z9JZ)SFikh=xRluty~tW+3w{f;sk{=X;o~j)^IF8`t0I z>q{ldJ_duGPv~t#cCNE^!qIpWA03ae$qzALz1yQm%=+TzyP zMB`DT(hU!^6qZD|lyE1&zb8F1dIpqCn!Ye49dSU5&^ra3m|9D23u>W|U zGxV<5!O6AmA(PC+_}gi$?76QiBOLCJj}A#uFJC=u5Gzkm5KCh{dxrdJ)`Swr&q_|z z)IaQ^BNu%<4>V&<(|jC6h4+$r#D+QqmYuze8=@&VD#k5vCytgE366mXhn-8jaP&n# zvj{YTA#0#P}WU z%S+5zEFs}>f^pyHii!Sk^QufpC#kbHp^~{6FqH9Qap2l@W0;^U<`Rx6#*m+qO zugB;dM|f5V;w9@$@;WsoH_t0R?0;{3F;W4hTuE{a*#GJ&VHr#3n;(^F5Dz>q>#F8a#_VR*k;G)M5YWBZ~b805vPUD;bm ze0ewD+=Ee@cGbUB=Z(-MR+{la-Nr#NJuoMyn9HgTU5`CEKPk|d+lG+l15HU!BV}4> zB8(MoNt`TA5^cfmA{N0GQVW9?uBW(}J&mB+hx(paldLh#M~p8nJ%MbLXz10HG%0!O z_N~IrCZfI7c|GYN%L^fkh_nk+O`;#1bc3G>f(c$Q7#o{*${qSnPh^O|DWa3HTwV0$ z0b*ZgQQ5V1H_8w+cTQToC((*+Q)t4`M=XJNyYwP!tV!`8;JX+7mgi5gPlB2maDnhs z)Nh|NxOBe8&0z}3G?#<<7!Ny zA~@Sr;^AJxT4xV@Xh3`|E!di7@`0IIg6q;pS0`HD8idBAg7-@ge_V`WlETcGwk-aa z4Zc>c8ec*%msNf{H=4~9xd$iOUY^lNa=in2tNMB4y!+SvmCJ37@Jixpi&YnwsN->+ z4;a#FjkZ6Jn-f+)`iU@sXck*^*ty;4@m=aNPhPK&+x%vZNnIWs zptJw*ErUE_OiFF$T=yf8n@4Cf>RXKctLQl{Krbc(b;9EwYK#7<#EAL8gU-2%^om8y7DY5)*t7ZG%Y0I z8uyYqqHL14ETouZuBZ4#P9{ClH*99qyk`!xaI=_T{H-CBN$v&;;bRx#cx?P2OZ2Qz z8r|MS6t&me@OH~8BHgPcXeBe)ev*1KD5&y6> zZcxrs2&y2U|0(85qCqUe$ra zJ&KK=ZUH-A>W78`npL84Yq=X3hf6sFp?I428uE4AOpaIUy(T<5889B$3pN>B@u#>O zL}QNr#dVKglo{zEv zFY6o9J9IC025#?-mWnu#qZ$x9BJCYfjx@XYjKIDN7vtU4Lg44K9Wxi>_X1x9>fG3b z&bSTDqPwDHjJkgC&;8=JzeVAG++NLAd39jE2~L`8dB{rc1=7CGGeVzDnZkDMV7u$3 z24;9}t`cmYE7#n9?;p{)|6V)t!GUP-5%VaEM?|?ldvJHJHT#u_jNhYfweHPjZE>l( z^46!B<{j%%!5(4N8-z@EIKSI-wfdu1EJk^q4&YhuL*x5f`+A~PlFnp)N1*UI-@UGA zYS6J5Bc4)vLcX;V-0|Qh4~cC0f`>d3MDWr-P$~U=cCl6_`Qvv&VTzv3DU4m!ST4WA zyUU~O>0;r2D+Dy^GK0rjp_t>_J95vP*wtg=UMqS*$xOmlPRZ0zhuq>fqL{g{0ZL?c zMf&^p{7NF?OdoufS$4@)2)^;pU&#`5PV$^1abA0%Q=YAsSa<;c9r8sx3lG z9w%}qw3P}Y-AB(fH6p`PzqXp>>J0Q&7W1kK8hPirxBO76t@Fp+E1PwuINS=bLlLB3fl49<#Gb_qOjOyxu(!F^U=Cp)TMVo4LG?9Eq`Wbd3S8+G-CxeWET(pEhvjCi_Eoc7(u&c#Gc}{<8ul^ zp*Itux}Nh3i!!0Z!A;ZYYzh-;GCSRuLj_DnEI%q72TzCZLZZ@{4Krc4lB@5(UN4_Z z%yuQZ+#YFlIq{g*v1z|%Irn^35+mQP*4c4@b}2F?5`v>DzCl+mRl?m}rL(aT zx!ojz#|lQ&2bV!h8Sco62Rqf=!M2$Wjy0{dByY&ViCw_JDi_{s%k`?M{9lcMt2HfE zRZ)e0wkz0nUEk-h!EDJaEENYSE2MIb`wVwIoy(&kL5~^vnoL>|*7DyDZOQiow&QrO z3xPw|LoxYkW-@f|rcJ1|tjT9p&`>J^zqR_4Q)X@A2-Bzu0nBr6r|JrMZvQrm1rQ=kDuvO`^g_P%=1-6Y`cVByL+eq@;FM zQv+Kd_!janFQl}5AMg3R_Tb;{#HO!f(aEQ1nyyH47`pX&8WbSX$y zKAhqH{CUkAEiAC*`^%EMSnd~-hcVtSh~_qlj2nBUTckbnfW?)N=hRM2+4sE7ALgLTi1xj)bfc!~Cs_u0*9Nb$LISOt;l5-QgycI@G9H{l^axDkIYJa+04W zX2(+{%Va*B7d+3&5H_T4)x-KGP!Ig^kePMt2jr#R2dYgGlNXC@o?qNE37UVU_}bMw z!%yBd=+~|82J;~C)Q9`PTAl|=^p5{i%S?H@2>6Xj5W->viB8ikl4OBlhf zDZ96%4sPsTA`>08d885aPUS=lA}rGJvFrsusMG`xeu(C|$oI|=8OZ*Mr%2#2iq?6( z!AIWp$@G(mTVQd)`1Mj|j}dkdoribips{aOmt>JW(z;^gcnQ<<$vxgt$B&Sz5ff4K z)dtn#Z!er9QyoE^9M!!w$nD&=G(L*10p7=s7;V}s_Q}Zd-aonn;ICU+_~M<1e!;@* z%EnOe!bqaTW1*0>U}jM_g80SH#H%_%>V!KQwE@Sm1>B!&2D{B&joi#6`sPT`1bxD3QWCn!$lc)B8uC3yN!0 zHE}r!NA;ccYiveJS!VGmlTqGs5xc2a4me(bILuv1wyRikJ=X?sol!29aP;H_l$V6n z@dp+nYoxg4j)ZLEu18UZ?2R1Tj%y_HjIl|qV@8$Q+1J$#l={SVo*#)8CX831s+}4e zYs8wv5|-SM^*xt49Jtembl|lvg;XXjBh7PNaxDZ47+iUsqW_5T>aew+y9;pu3ANV6w?|vdTbrUeM{lf=*{Q>Jie6CEx6mS{DUwIbXpL{T z!I@4k83&~JN?a3}i8nzbHJtWEO*0almAQ%ynMJBEP8oWor1W0qq#@6#urlLDDzE}s z)c0Gt7%QuqYY5^JCuhF$p((=c;e~~UHVmJ^9;5*1j*Nu0*WrQ{8a((c+aW%NDnx(CJcyW|yw}2U^ zv$Piq5Kb(St086V4{gvf($UfNDZ5LaBPh3a616+VN8NK#D6CBWaGsbg&!rIUJ_>xJ zOMURQnrutfde>6k$!_GWRY9oy$CEF*F6$98-1Z{ZmuFRU9X-SYw)M$c_UmR$_F4N& zt`#3kYZ_EmJ#U_?W#3N4y_8rx?DBh+UL7Y8(nWY;L|ASK8LrZd*qxI3QC*69h|(%M z&}%8yDb|_8#)!SxwPyJcST)h^WlDXIi;!cLY7m#NQOdQnqTdnmDAHwCqT4zm>&?1T zG~+P1joGZws*VLBil;_w=VLVfrA>`7ooCP+qYU$FqT>TkUh-t&CC0nPM-NBX#$h+g zB;jb{9@uK;_jfSR6UtIP1P4PR^D2r%V#&Fc{`zm)<04;u=Y9Mobw2xC#wA@uayPPP zVVq6bkR!em23)QTTh0CxEVkbTyXp;m{a?ZYf6G1nossuf^6@`&H+2*h{>obead7@K zZ%t6d-ND$@#@rcTVs2?|FGTmPt&0v|Z6-vg&85hu=pbQkWi9IoF<19g(lGV3G37I( z6A?xg6fuToew(wn8#}OA{4S^91b}Ac2~kO?DM(X2F}U-)+d0@lbD#n4cDD9T{O&?@ zzeV_=_fKRX9pE>MvyBj)j-m>{84R|yv33Tqv+}WmSU?=i01F5-htLfSu>pWsx#^%? zLd-1q)u8F$e{=|a5~8zmc6Q(g0^Qu)Slu{S!4OLzh>wpC$i@z2XJ>(OusC_xI~%*R z*gH`}$NN1IYn^|H**mfRo)fDn*bZo|0|c?M0sk84ub>JvXBY4^u(1i)#aW2X#l_l; z-`JFm-5k1!EGDMBe9$aJa|;$W=*7Zi&c4a1a1MW)=D&;wZGaBQ0o@1a z9X~X`-OR<*93n&~qoM$iu>w0e|B)ID;A3Uuq67Z9{Qs+2L3Q)rcLA!f7^zcDbU>3{Aq4ZP$h$)Gd45j zw|J^I07bU(KX(`iMI@cDAhglh+S%4z__yVW**gDmV{83nN&L3P_Lf3)?kr~J7RD~N z&U8>s{pL3NLmRf%!hhWSmIMB?um7T}-z2DIL)8ZShu;uJ1^gdf|92hz(^*2Z|Apw> zo?PL7)B$)ZMe%2L^j}s@|KE(70h-wSWFt@y^JFFTlKf9524V+sfw)1Ozpaaojg1b< z2DQTfk>el5A>1M67N~3h5FhH3Kl26(wRxBwP_KWQKkJG8XG(*>~qlLq2torg-_-u8qG7Rzc8cL`MoQY`rf1~C46d0#d*#WQv?{K-}#qj6TJ^M&)p)*;24e2i1dRlO& z&8@vLdiC%-)Y_x_7Q*K2lQ>B%86NYIaxt#;HcQwfZbmsiL_bB8!bT*dSrJ3;mJtj? z>Dm1jw9!wo6xGSBto?Z+Q~z8bq{X`yJ3Ka^`Mp9`0!!}-6U0K$^4{L_yPFLfI1nV* zQP;J`LQ%L@_zzZ_7v@u?uMV-z2Z7eoN@}I17alN$v-Yk!r=5`Ud)fTgdPEh49X!cD zcvnOttn{jo=SARGmYz+|;;y8=fDb1(b0R~N6Ev)370b|BBj?;KZn#rl^Z>E?a;GFS zv;>r!zdo43>7C(Lq{ip+r@vQDgVsf1CRL4x$$1h#Z$+oqglmYJWga2oLLe(fgc-h6 ztfMQ%_n`XX52_aB$vLF!e$@%%`~m`>?7?<_x4>Z@%t#Lwjx@^&U5v?ttfuF+!$bns zltedpn;;CbrCQcI^VkQ=BvqC!f!s#s89+)6$W`C#24234I8d@sFJE9E!i`|vwI{NM zmS+`dSjwaQ6Vxn#)0Y1(&E*~=@IstXZ* zo(1KJ9%2mgI?9R|8eE^H4fO^lgbi0?-h2Z+>1YCu`O;YyvQoc3bCWOKQkQ?{(E`I& zt^(iUb5ChO`Yg@Ccasmy;qx-s%U@-Es?1utuBhDYicE% z-V^rNTI$+Uh8bc94`1AYlj=GAL9ro+Yt{d!?ZNlN7Z&~f#^76oA3C~(pM9co7@2-u zZk2p*Eiy!&$CWRaQBRtp*B{JKnTGpdT#_;Si|YUfb6&EGkz=@2`rDX>tdR=~6P?VY zPqIjbtOl#8U(6|=ff+sA=;#E$NyYcHJeHIV>SK>!3gd=-=@|ZH`XSLs>VlL3_7l0$aJZG!$3^TjRk!yP&{l?cR<#3Nm<<|q zi>Q!-sM4t5*k%K=a+oRZPCBR*0u+$a)6-wMR`=ayPz93)uOWn!>xK`63E zYbdCVmEVDA8{x3AEvir+^V{FYe1lNZqhrzzwOh<*Ax@t+(E^=%VN)F*ywLlSzyug4-Q4T@8t~Wz_fP z1$FVyDlU)J9Zxy-Zi(~v;qC&yM$50U3pV-6^+O1?cPeakXa_$Np&0PO;0GhOZN5{U zNiqi<4dkN&8~nd(c_*If)~ftbc! z)hGl(-+-Gu7+ZrTHB!Ex4RKpk`oE+1*dIbS{piggon>U3!#cVi_2dc-5wp~%vtOX<|yeDa@$ZIRFH&$>|+UkP3#5+k-MK-QXOgkzU{4{`sa{hu_Usz$r+C zzsa@;E?gUs>c&-oknrs!Cm}E4;sZHr*miL%v2Iyy*w04G@=fd$D;pR$x$=4;1YOk) zD8R^4f-JWa9cBv2T*38hnEnd+T+k}sgXlrCh2MIytd?Q%j!iecOD?PF|PguEd}p8%>o=-a4>@8!9k`IbFgth}vLG`Qtt zdG_{Xew=4O3yTm9!!5owW26>W*T6_^Z#S0__1oNFJH^YX8{nWrxVKJcr*7gbsEjOZQ2k`=&(PSx zeQj0s(cq+;SYSF!Wxpm2@7gvZa%G^Krc82|1;XKrS%bP5Dq4?56jG2XWckkZRk;2b zw2hxeqwHt#uS%j3l439Yds{U)kPE1&y&9@|LiaNohs_4Qcn?#b{nTTKDjLO-+v|Dg z=D~?Q4K0<~4xBm>Y_aGRe9LRR^g~!sWE)iFDkxRHRdf_+r=q4~%2vSTRh%r8RpnJA zNn##~He2M;$m`7Q+{6#C_}JH}Gj|*a_7>+Qp%%^Xe_A{cUGh{73i{}Xd{n1m(6We{ z#HHuQRd(Etglo3(btRm5ELR*MK5gzM7}HY7(e!a7g>({ZC~CB{B_OhX72g$KU3lS8>0+DtV>bu_}Cnuk+r8#hA zS+3Dmle~w& zTASf~9VrI|!!l=Mi82Rnr^1;S-%dYSIeY4(bw{b*Tn-e|O}DLxH6DL))*1g&U(ojH zYcPUC_B|mD?R%Du;JE{TFL9IEJk{1bcjZ1VP0nSxu4ToQ$t9lzJ;iC2RZtq7;W4F% zn;Mn6U^Th~7wNIQYb$0$YrF-0MChq({12TD`j^*@HSA^~C`hXf*(+Gz%dj;w93P>rzB=R4MQa$1nYef57OYK~rhTCPr;b%CX$s3bBQMYqX z;;Br3FB`UEj%M?+K9_%zB)P1iR{CyC(=r`Zo5UHppobpOL**fDQX7iFK`AC&5zaNB z{j0*kjwJdcoQk#!M-sAX6Mo??~u|<+~}>>LXmgAQc7t`shqyC{z<89XCrD32p2# z1x1o6LI}a40%EdH+A1swttE!R!*Dc&PF9zsgY%sp_EkJt86U5M>X6E4VxlEWlegmPnpI2mD}dfn()7b`SJb|i z53M3y#)<)Tq6YY=tWtMrftBKSK=~L5R2>@YBo_&3GV(6|ocOcK_Lws$I6nGJa9oqh zK)lbv%9EfL;oNCCiQWAxGE{f*F3S%`UzNYBzt@X7n@5F)(n&refm3?!VfZ-*_r5GJ zJ1p#7nxQhv8lJH%RJ93p>B+}XKof;|E~>A9b6*Z91wGeeOH_oIYK7~g{`~-XDF(K) za96Za3AWLkv_0S}w20h(PsT zMM5z#JU9)W^t9$fxw>rS4}o#LE8H+0TJ&E=uHY>8h|knEXFJ!F0RYSo{3%k3n`My6 zw75G-ftxr5Op+=VJb5N%(j@WL!?F^iofP6cN%;sFX>pu+b#c@(H6QFuf`L5=`4ZB= zQqofq<={YJUXCDrSiF0r)8-19+1~x@^V09%*-x}u<28;nwd6@z#ztd?!{aM_dp}TM zBoE^_;+c*+RgASiqc&bYBnz7AC`eRKCuEy40r%IfLwk8VVK=jcrnf(+uIBei7k$Zy zkpOkO6~BB`Q|PB>YghV}k27Vk6k0ntzcOdJnymM1GJxay z`iQm(g&rX*+T;;45pyRS{>%0Xr`5JKWl%LTA!X6LRa#wSJ#tsxtNE`GXE#c=Q(EN~$ja>%af4SraTEk)JXTm39Qr$YY^ASJ8`feZ9(xo4@ zrGad6mjYw9qsi;05lVa}3?5k|&`U%iRd#HoRoNj|u?A!eZz@kF9oG&!YJ!X?@gf_T z4+z1`61J~CGFK1lq}Mh%AW;Qw895gf+gp4Q-M0U*Wu#$SA}6a#>Fy?q2%x$+9?rW+ zET(ZF<$8@sdM62D8HZk0KtL_KVeb>)I_7=)fbBrxu`RDW#{J$87&6oIG7C8s=*zw- zEFyIju%I0)_L)W-pdx3RL#V?{fky>24_YkFb`Bv&ubGVILFjFN{GHq3jBEns4#%bV z+QOYVhi7R$@8}Yaw(yoyGzigi>bJS)bPBCGJN;u0ky7g7hVyW8|?EKtX3eIxfb zRNHp!`&MU{@mB8Nduc=-@ZZaCZmcu$e3|L*VV@_a(eH_*NPmyq2Mxy7OGHuNBA1|? z+6H7=Y$h}oU#C=g5QINjs+SoaADOsdJu(j`?yEDkBxrXNn#fg`GD3L=^Nc%~t+VA2xWO)gs7(yP3Lug25Q5=1A2+Qz$qTR)qulkeZOMhi{B%GnIuqi&%VT(ES-c zEAW1rd+?XmJ>iDOm)(lS6g}rkdQ7_b?iZc9v@|vWb({ z<4!U134l=;S0HAmdztT*>PWsW87803sC#HvEyyv^x%~c)Mv0tYy!Ab-Q82m zCcAN*bH_>yGdMNGoH&sf$>>oc(dy%CLToT)C#NJ9g=6SzYZ*xleCWA`g5M!yRI)&{9ZVedyu3jri@b?B6qR zs?@OFQPpS=Hsh`hMMGIy1tE$_z=;$Xe|W)kbU*5 zyWwy^r+MU)I;ib?AwTdSsP%#zkJ1s?LUe;4U225aZHL<{>))h~q|bvCQC-GR+K(-1 z$kM%8ZU`?zHD>-ro z6qExKv=RMWAq=90WsH;>s?tnqK(aX@smGZjNuh`-IdsQQ;V6@wl&w9yFBm5tXS1{4 zqLJpuy*KLp)+Wm@NwW*tu@aJ4Fj5sW;7t=hCMXV$`w-l3+I4G(RPT?3uwUP&rR|bs z``-27{SXw}i|i-7zxOC|d>Sv|E9LKB+=#}%>Xb0T>A1t4na%ALJd_ClwfUqic1dFt z3Vy$PFY~^)AHkBv-QBAqsHAq57 z`Lxw}ka9<>+YZ9?4PslAv&M3EGVLNZ$i#y#rF)b_4`*vhb!KD?hZtv?&ug@m^6x@RkS*=>_-`@58-3LKZZZ^w@kijo&n zt{h`cYa}gBIm&vvRUe*M7b->QYhng59P9B7O@~sTgMYe@zSBosh#oh9HpKXfBtFqM zF9aRAr7l`9wSgiTfDu@FVogo*UMS{!r*iUn4J4R<5`*gx5`WsSFdWmWt=t!Y8w2_Z zj{e_APwU8}OuWQ$DVfzI$d5pH&dRFfTZ(IX9;()$QRgV`>z-BDa`mPCNip6@RcPHccJwv% zrn4+c7nq_rqSncm%`BM!p^Grt%K3$hJ${wJ(tcGZqLp?MAoDI$FI&|4_~NIDa)k0^ zJS&LB0wrNzg`0%jTm{81AJj?__mK?{tOm6j0rwWL5DN1J@Oxd& z)0|NzvHGjLt1X;1jD!hxpH0DezW7PFkZ0x#vQxqJxpCYmw)5a0eHW(gI;l%8mcceRVut79-La&Dj8tIJ){0OvH&+rC2=Ag zV_}4mr$8>?FOfwh2h8`R;xfd*XvSV7owGZ4vU{|iOd%aQm( zVYKgMpX*k;U+I~~fdc=2XMxe?FgKZWonRTCFvH&4W_xEVY;3z~K4dYd7;-i@vp;<5 zj-6Q-8?S@L)Gy8A!AEb(jUbIrXsW3wd(4|H`PYuDYsX1nq)i4RTVk+p1$8+uOTE`e z%6`1VrYRB4;~5%~-q;#)V_QIaT-*frY>M_Q6^A^4=cuzU<&!~k*GDtG4Xrps4y*Un z9!LrYmTqFoKjYMM(?89djcQxu?%HWJs6@waR2fLhDhR$i{q)`4USDP}^=4d#!c{S)P(TqIRRNY7!a_7}w!{hB0$ntqp8MzQJD`OsJdQ}enOo0c z#3Qkc;_d-)NBy zc!hb?Hj{P*sX&&jYiyu%o6B*r@q}+R(XN2t9X0%6_K3cC;{WPYc;{ z^;WH%{d?})O@i3S9s#G=-LoOQ^))w`O;d&sL0I1#6HaS!w{cP&GA zN$r+{8$QdOoGca{6flXE?45hjbiYGWP!dK#rTGYqua8divCj`n;zghxX}VRfy1jqY zlzG8SR$Eluv?@u_nZe9ZdX&%pM96t$yWUZ~c;`ElN`BIsm9-MQ78Wc9eQ0;z<(bUR zcmSPN3otP;bKc1zY|;!wQ4xyfx;CpCAlY6rdjU9_+J4ktL)`u-XQXQan566FaO5&sHrU$sz2 zJg1Vy?Ro*m5O`#O)w-hxu`G#NMqZ(qJZM})F>%ZRaYNS^a^mf_t91Ocas6XuS{z#e zd%|~b%BGj1*a11)*6D>Dg|@gDl=8tGa3jlFz7;k6 zzb4g4Q}1FiNSuw0NJ^XTYYz!Jv+rdpRnrC9@3k7&pwrPq9+6F za0mdGf}1RbfE1(cCwZB7haK$6HA%9vlENGWN;^a!RDy|>r-fn3ElseP zBa1@s{am2Kyf*{}K%z(dppz7h?$blMn7#z5L)GtAW0#5E=IHk7QPG`cV2_aT*o6P! zocQW_(Ln|{Pr}BJ{(ZNNT_%=jL?a#&k9{Q@KZYHl0L?@Wjh3%((x(Tl!svitqJZS+ zi?B1HymOCTCdj(83sR44H+e=0V4(_mvPRgMRE~=L!O0wtW@v}6*;n7c;dTXbrGq}4 zY%3MB{Ou+Z~uzeo+W7i z+(BQUBdN-55nR(^1LLp@-#E!juM<5l067Qcq7<&LrQUj3k>6c+mkdhwrwPyWAzz4{YF@wF%)0qho6c1EVW*Z7B3d}g8hBAA9@s-eRj*LP6 z7X&jh;yV=*e#Z&TyDyD)urxeG+}>v^IU@xLm-?zKzoJBdMHm~RbtqL&fHo{m*y1++ zL!rYBj=;HhW6`QlBKMd2&ypfY1p6=#VA$4vYwym=gP%21IAAtxa&YCpBZGyo%#Ji1BSQQf(Ic$()%*ACYSk`n((P;jK?)(9pNE;yS% zlFtjEo`Oa%ojZ+PfJIgN;dL;Xp0ElyMY+n$czSo++uDX(uC_@Ao8Yt-Xe|))yU0?= z&;d4g$~UC3*aZ(g6&~G7JQTz^S9Aq##@ZT*zh|?p1k2p$coKC^ubkt|f1jO+=Vp44!)hQV8%r9$mJcqeZ{uVpAec8O`Ryq+ai&*H6HREDeI8? zu5=3qLOtl;x$@nu7P+;^|G1%Hz3IlifpGff9E%fjPK?L-VnJKlS`)>*mU*@w5_i#R-gcEcA%^`qe3zZ-OqVB_cNB5l zaJTNLW_NqUD5eh=leN3$YHT7YMU#EjqRVwso@N|-l_OLCBkkcgA^G(CfraXFwTYe z{lQq(KcRl3WrW4cbD&m9tgR`3In~buB*Ofc2>F!Y1>F^#b+L2V`3&*eq+e zK8O>T0cd#xo;s8|;<}Z0tS~^dPFcJ58-- z+H#p=gC*rEPMwtH?N~@;ZZ)Sl_~DhsuBov|!J-y%@i^3F_K~`3)af15D!&qgFGWzp z^c?p#P2`50D{)gR;c1S6JAEaBXmwY8E22}A3+v*2cDWgcW2ie#(?K4MA}R2qzK771 z8|Z^GNuLp|MF%HDffn6npRl{(#74QfBOIW(Bd@qH@=~AcuzP2UIsak~_yJil9NiV1y(d6x>_{SLb*vM}c$N{D-F6nHymrOgeD9rT^dZEckFZ5-(ZZLN#|Z_8Iq9f_G= z6R+BWg0^m2^dNR-VtOVZ6EO(L`pT(+OibDg5>EP7=7s_`rdGzpKmdb)qoJ{l6EO#n z1;FrT@$H755d;D-2JO?*nNsMFlh_*--#yuEAf4ZJ8Y>pZ_A)t_X<{p47R zPY_q-P&KILW|&j_8FwrkDydgJ27`k-#hT&A^$tLy*L}2fxY*~zngJMjzD4g7`fNz9;cwHU>V@l4zcV{~}4Eu>m7+Y<*=Jlk#; z&gRdpXZy>g&eINk%J9-VF9?C3S#tly$2k5o3jcN|1|??$r#IhHc5pWSLn^56X#6&m z|F>6}J2*NCndv*cIg7OZKhuni00uR4BPTOQEhbhLVm3xbVm3B5Vs<7bVpb6Fk9$^D zR^qqw?b(~`kNejv3k&O?Pr%##o6f)FjIZUqrI}vK(*B39{U;KL-vUJuz@TiaVq^Zc zP>cA@;r`-*uO|M!$_oT~Q~EDn`0phg96*l0uHfLXRL0v)TxTPLBZ0)_M~?Z7W6s8~bpzIlpAi*#nHXoVvQ>JSTO#wjOd7?qb-Ky4^-})zhLPn;R{VFF zDI)6(54J%L&$Rq=m)cz?wn7)z_>7)V{ZlqnpZxtTtJ{pHYBFZW)2EZRb#I?_Hc}o> z;;1?^q)Bo7^eBqefU80?apVg$rkSSehZ1cc6Ana@?I0;@Fhu~E^v-!m{r70Bmv0}WTNWNu7P?~NH zdc*hg&oyHnEd!Vlh#btpl;3CMVo0h)ajL}^jf%Y>Gi-y5_~Eui^M}`LzwZ?(P;-V( z3_lP=JN0JG~4 z>v_|X+?&p0usg!wN9Lf~neYkFON7houl8xq)*g(;BVYn*1!-Ki>P%Txfb!rGq99x^ z2wEnGZ0}=C62o?%oKKL0+B3Y9wE!QOAWSn9K?}_}Nz!f(i`h2Kap5=CzT!taP=)Y1 zKT)oWF^LZ*$s`{(;oj+TdwFA#r={9%Qj|P!!`AlF8dHnSdZntpz+GF(ushA8aaco5 zTZMYTJw1_r5{t3f|2MZx1*cen^2CO9W;U`gdkG^cHO;;JfP@_{AaKt)@JkOEwH;Y> zkBAZLiKy&*_I|!zY7_R}Zy6~+S#uD_ILF2vrxsnZw0eY=yWSC5?!s`GO9VF#Rply* zUVv3=3iw|pOE);8M%npL+u1p+5t+ky!@H-kw$oSG&wYiGA^hzAPGZ)}0Zh%di;-XD zQ45nQe~fZkvTh@~FMrKi-?ydLg1~9cSsco%g*$M`C^4>7W++z0L8-^-buo^-tI2bC zIKmQvy_umGT{A7+q(V$)(bMInt5gtDNw96Qg^f=UBzLrwvsa6*14?=3a^={xyR{@8 zHB|8K&brtJQisz|QAuVg1-OyNFQ36+Pk(!T-UYvmz2Dm$TCvG8yt{`gcj4U+fz4Y3 ziqlv8-s25NzuU~|5vH)0np^C=ku7ms9;*_D1G8ZSpx;?;mmkv2b^TDuTH|AlAVxt! z`cg70iQi$XAlw2LyVDTF#2(_JB}+PajGa&ni^EaNX@E^k=J<_Ft(T13sJ51Yn}Iqo z3i~t!dmSrr$3=+en51L_r(^>nJTaSr=2zNk03!p8s&h{vyMzXuK3K^iRF2}80d6uf zC?X?5Da@3dnyN3BI|7K4Tf0xON9#WZQRO>14)bHk;vb%?J~TiOi`H^nFwC?FkmCBG z%hV6=$FF!PPF=QtyZ>NvU-ZybBd#hv)hI#5n3Nc|^6bLMw3{7WXDUtHd}sZ0dZ5L@ zN1xu*y)S$&_lHZDl~q7<>D73M;T9@hbx+Q8UsF{tO=v*Oh5#=<(T5=D_oW=j(;C}2 zf!^KvGfG`m2gHjlX|ZLrgEH+32S+PC#(e zQ4jG~3IW7&q)G5HNd7$jNNLh)Iue%gpnG<{0Xc`b1w4xO*s}qJF=(MEcUHA0HobW$ z_^!Y*U~5>6o+Mbv_QQsx08U3(FfYqWHRs6axBIbn_uPze8eu*6kU(*`@IwH-!)?6J z4upc4!Jb)D_AZui?A65w^JVvTGS>0iAhv;(D*qLerqmIYKu&wpj&Ti`@KjsB{#z?8 zlye?$p`X(T_+L-N1Pc`dCBH!agcW4Ah7qQ})RGN{9UaUA zq!Mo#hGP#hU(+DwRg2IF8LFD2PoU$ZQB4m(Y0#twVs}3r?;3pIYebSA6N8qBBqI}s ze69+-mq8)$nFcRZ)|BY(_f{l4CA(4;v>lyL3Cw~>^ov~3aJ}T-+FfU`A4(^)pybMT zcqiM+F;M?G*0XyMA4>Crt~kNghFwDGT`p8wOk$uHqgW8%q8%z)bK`_2l{$R+F?yp_ ze>9yO4a`Tr*$QG)=m=tmAqsXwoQdK+rRW{r7^!VcWQF^-C5~gmB`M9IBtLm2Q8Y$S zp+~%#qMh9#3%QJA;Z`A2p%jJldxF97*@@m6DjLCZsVvn>*TI00TP#eLmXLLP-r-Mv z%V{2#LP{jIO8Y?Mn=_S9`ULwK0D_!pyg>=_`DwJBU8}y6wrQ zXvym{F(q(hYlo)Tw z%21K<{@8LZnAAiXWHPy8SxB~%fB#*BF&mW4^eyJ&m#c`ZDTpnm?FJP0FmjQ;n6(M* z-0A(6a0Dnba;)7j4o4d3`n2)1aU@Ufo*(2#F6_a{+})32w_Jc9Qm1Ji{d>TsGYStL z=vyl&Wu)(785I^)o3%UTT$H%D^W}{$u27@8Z5KnMY+H}Tz6_^GGYr1BE`0{Z6|Ksf zmx=O3Nq4`z9#ub1ya>KPs^mdmCAj7HePI!;ja^a zP9qOMfwzN8HPgwAZ_7TPGhl5t!S@@SZeO5caL5pLeW3mg*l!WFBDDmT6iB=X>ht^m>{T^6~-R@w+ z^z>flSw=g?I4%)$U0ETtTGG(kcf3Bv=9$u^s9kgkT%A(*4P8rUspSW>Q$3do9v>a1va?V7FsRFnpZg8+9c< zL4WcXY?U|^cOD9?w)a9MF$?KCP%&duwIcOSFq&wsL+5aO57Q5WirOt*gM z|4>yU9I}rk&!IeDx8PlTb)tBnx2El$FgCVC9n4YDRr=1fvN8Oo|0wP$Lfg1;zS&aC zQp^&qUd&Q%hFAA-cqzQT$xHSu14>Jo$Ayb#WZi^W^~wmaGo*2rz>Y*Oi0N0LAMKek z16va1V8vqe#8fsXo(K2PyPQ|GB^)p+u^ zq{b(i4tav7x!7arW%|6su!{!F1e}ZBWh`Xhv+VPwiqwXN)C$c_L$<4Ni3f7!FOW`^ zd(Fu|TrEM~jkxPVhEP*vG+0m=A`hgV?=#hul?#fS^A37xny-6H=Sz!eYDk%F5n_wf zwT@vkEB%A&B+J5-GfOzt3VxelV1BdHke|7-7&AAEED`Hh_Y|LE+7Gsh5BdQ{oy=uq zF6=^x30uaHquSDS2%`C+pXK(k-w;hC`%~>M7YtgooIAH?d#1-DQS$nP!(oPV%cZlpNjnCENUb5@_{OPlU(`^7Kn(`*+H@MhZa>;H&G#Mcu zeFda7yIHR!<~J2`8k*!Ww6mdpa357Sg&evr@2+`BP09_|z+1=zO#Up?Js+@2tTUyq zXs81ca2G&=-~{{(Z3hZ3W-(=;DHIJdWQi-J^j(P$iKy~VSE%2)XKDQ9lAMrb?~-(3 zZDCIVS3+HA;M*uJfX2^dg0CjjNlI@_#F0|=EIS{obBv%6Ok~@VQ@xA$;SD}uI>4t8 zdm85@`Y#Y6)dNzd;L3nB9|H5b-~?v96gvI)PGK|A&Mv`4yO{c+Bu}ieyD=WfQ_nzP zL>P9$DQF+x&udhzcI}NchY@>6^>f=+ZnTs`E3XNqJ3r_s^jAfse@$)rgy7jZ+-)*7 zZ-hnFqjOtjI=E%;`L!qH_H_L;B{P#;B4p(hMlb2=?)iy7fdnUZGWKvoLlG~J43c37 zFO4xsK_C^uDC9%ELR1?xZ*9ES)AQrJXfz`3>evu?HFq)#4pi;Q{Vq|PgHMLdty*t? zJLAsRM2@h_=_bzMKC_LrL0Zw8k{F-QhLFa7&OoeYUV58p+x=PmdHG#t_4qt01-&3R z!w7%zL8@_T^MV?Gj76QKu&U^4RH<6Jqy?rU_av9v@*@bka#~}lk||ySe`~>zgw;BM zwY#ssL#mHzTW^JKJwxEzDvz^+pya+c;Fx$b8>NHsaaDg}Mv_{(I3%w>WVg|lk&r-i zpG!h0KM7@b;=x%E&$QKS;%|fJX(#m7wffSyK|rrQnQL3!IDz|ZQPZ-3Wqaph$w5;B zhsrOvJ2E~}xzpdWR78}hv_g~w!u)E=;hwr;m?ani+9l71%*B?wI(CWbbtj3nP8Map zp}jxPkYZVXj+1LRi=#_nu~x=3;k@D!$J9+~4Y0#1n7&3s=Ttx3$w-zTsXwm&_%i`2 zV*J8@r_tPbsbFP^e{fJ0*qk2?R;D(_ua)kuV$xx7p!_Lk$Z7Q4>51*=t;)dxE}Moe zE|v9>Wl6MCGoH&TE&~4L{rBGoERh8yzV4R|!G)^?_1|5Fe;hk6bqSJ-y6z200XVhT zlQU_jID|YK8swmZa$DEOeiPd2&aFAVE36c{>)@8tJ=3#c3g4^ITl&AWEH7IRJvxJ6un*|WgT=`a~Ui& zvSv5;M5tn74O$MKPihcaB~wC41R0wSjU`eK4Yg=wh7Qsmvf=>{9*$8^-A_YY&fFMV z{QBcHU_`ga#tFgo>)eIH$QK<#jV!x99c!((TXKyJ_+cm%|{vU zR%pAW5gO2N2nE^7?IxZMYHywpRckML9ixP9d;7p{Dg~v99R6VRk`KS`?Nwo+wD8ez ztPVB_z;%pP=PoiVJz5sW(Y;Uo)cxyAe=GPfu0*okF}9dwEl?@#o0h5}8jo^~`{B(& zOBPqF&F~~;h%5I2{48VHs)B00aog$)<67IjnYN?7Ms?xw4N)QA?N;|_u>bbUT|)y< zZXf^Iq0HCE6(T zhDJa%A&q`TK3`W%oFS`I`0>M3r%9EglJQ&IVk%%FxQre>!)wJ1c1=+g9czO=QhxX> z>f4QQ#Zs^zQ(O>rKK~${v=Bj2Z58-hKjBc9WXg7nhB_2wvXTZn%xZ*BY|yND^pgfV z7RJU8x(par=}%9m{_wse@Vj}CIs;u$8*DJOh9V#JOp9ab5Ke=gL@*0u($*4!%s))| z(UtMnUukQohYp!h!+(=!d2+xU6F8kbHP&Y-RsGEhznjarmxz!NGDm$_Q-yh~*jHo6 z`%r%&jrozkv9x_MNt`e>Gc85han_I}qo}<-beS;P{;OGY)L{yaqka1A@s?3AIvbu` zuPJh$;m3+^-Mj1UO&xhf>FJX83Lkm!B72MSYCW5J?jkR=m0GrHC$E$Zf)B&TWg{O^ z9}Uu#49`>5y^8kA>We^K8iqdgsOdK9=LDGyxIDF=`%F-#^hc2Tk78&W;F zNqlIT_Ws=u_C}U8*?y(a)IlfC#Z4C?PNxBkQO0PnN=rJ>@w1ZQKG~ofT~=CJ3fQf7 z{QkbTw%&v^0D+Hq9Cq##Q4hZXG-^FVw^Dz>rhQbgoZs`g8-ldd7{yIb|5QcKQQ+gm z%o??;==#h^;A0wYE;-k@E@kABJov*oY?_FY8MurfpTk)cbFjt6zQFaFIk*{DSROXF z5ETI5`JHnP&V~b69nNt8vTJ5=LPYORoRD1sjU+o4zXarafn}2!WIN;o=!1k2qQrTB z4Y$6}zPMSnW(*UkX20de)}bz1w%rE?(9gy@&9bFa!8M5FZ~Tr-Qj8IbnQERjHtdHm zpUllqN`LE)H3=0KR9uvS(Mq6eql}MFImm=uj*r~Ti_pZ!!)Zuq&Sn10h;v4MCzd@j zODh#8Ua-ArVpQ=wo>C2Cf)C@4I&Ya_SaZU>finvCnX(qN9)L8^UidkFp!^}pKlKGD zl+SAVl@l#H46FJihkI8wH7m(7GD2oZ9ksr5?3ZZBuV{ETn;Gt;*~7T6<{Q7WhYG$r z#yQ|`IrLF}W{dYuE_1}KSxk(ah*&}XAP7@&2|`QfQw;^XPu(x}X4@l4C6^;@vGAs` zeEH5TM`0oHos#&(o@41_DksS!I}}F-_@2k3cppVgot##!#c<^FYlCG>~mmrMqE7w|ktcl{^2}^tfWqJKO9_(-qpCu2~ ziWydM4CPj$I~KZY_QnP>XU@6K&lxtxA%^fbkmUCV0O`2e`td+t@OGvj7!5O8j4@jU zFv?c|sUz;{qYLPTpH)j98`KwdQ}r~1^?k$30pfP*FmKG2?@?viD7z|)9m><@ z?nJ1EVx6A{J@ba-If6`$J&O{iTa>zP#RZ9)!k2?T%<`b(VlUz7blyGx;EE;?YQK;j zoW@Eb-44i9M0gkb&hZ^^8(#Dmo^-(!(tQ_o2-z|JnRCSJs>Vd#cj}ypqDW#iaT zG0f2T#nfT|;yt?D4?#-i;Q8`^K7}x^X?`yL{7|z&*qLcxh_VDKr~kqN{&@NQ2Ao*f zLG1tLG@w82h5k1p_|JRx{}3mk|3RFL9n4MsFXY7dFXY7j#+`s{{~K~*=Xixoe?m^I ze??9}w*QWt{&>It*EjzE+Q<2yecg=gO#kI||DX40YRCQ)Iz0mzrbUbhNz6Y4Cd83e z?+Fz=y{m$;C|O7X!9TYXa#!AwZljv#BhPi)zyx<+5{1kW^Q{vf!8iqbLzi zQy^(-QTe&!XbXF}W}}5aQjY7afqJv!8D5LGsvStU5mkjn+bPL1xhl4L->emR?~jv^ z2UkKvqj%+)<#D+({`#-zI$%A?X4!K%d&8TFzgE7;9-{(^u5ee+lLd@AK0nuZS{X9o zZeHrr3`g{NQHw#Qjaed0XqyAxYbfTjvmgB)g~Kdt0O|td5a1Q&CESjjK|Zj1_Y~n3 zdg=HpWI*49fF*ihe!yM*xq@dqS& zgGF!8|0gW^!^!_bi$H&D`2RDo1pR5luPBK4jhwxbB1U4+TbuvtX8$6$uWt5+&4~ZV z`dYlagRP;Gv6B|V>lX(Yl#Sh-wEtq@Z+$1PZ)(gSukY{%TK}-YplIx9>+E1?>_`my z$6jL@V}wok2+HHHciy4UHAW z1Q{fVonGl4FE4}8YXx5|I1<0bx#FJy?cay;tvZf?|4|13!@n!>TEM?+@z%*w=0=WM z#D5^Qw_71w=eNH2lL^q@X||C5D|5CreQoutt$#Ecz#wO5Y$IU!#;~>iQ5J)`h9&g1htH__yg0E?I zAdnHr`&!}G?>f9f^?znzVSAmT7+)WAyk41Kudgo0#`voCI;uv_hQ|NWV;1^{4(qG# zANsF)%x@;&%)Onj`B|A>&sX`Y$}4Jry}nj~iRE7=Ut22t`hf{5VNOOMh#ABPWPCkY zSeR*nAW9&RlKM|Q=;mN-0ss<&H~_CD{qrDZWnp1rAvPiYhYZBZ!uUEcZwIl>KV`3J z(5nOgRR#jG{9Oim4X@Wg_`4q4-}Qj(?60x)H$8Ut*Vy};j04E_IwXHBkAs=@f688a z?r(Y^7FNbrpZrVSR~hr)?Sfd?U*rC7dar}=_jZ6-nHm3!>=jb~Jue&UYi#`0E-U+= z`^w2d|8qk$>MdU;Q{}1J-E#d$G -- 2.39.5