-\title{Optimizing Energy Consumption with DVFS for Message \\
- Passing Applications \textcolor{blue}{with iterations} on \\
- Grid Architectures}
+\title{Optimizing the Energy Consumption \\
+of Message Passing Applications with Iterations \\
+Executed over Grids}
In this paper, a new online frequency selecting algorithm for grids, composed of heterogeneous clusters, is presented.
It selects the frequencies and tries to give the best
trade-off between energy saving and performance degradation, for each node
- computing the message passing application \textcolor{blue}{with iterations}.
+ computing the message passing application with iterations.
The algorithm has a small
overhead and works without training or profiling. It uses a new energy model
- for message passing applications \textcolor{blue}{with iterations} running on a grid.
+ for message passing applications with iterations running on a grid.
The proposed algorithm is evaluated on a real grid, the Grid'5000 platform, while
running the NAS parallel benchmarks. The experiments on 16 nodes, distributed on three clusters, show that it reduces on average the
energy consumption by \np[\%]{30} while the performance is on average only degraded
50.32 kilowatts.
Besides platform improvements, there are many software and hardware techniques
-to lower the energy consumption of these platforms, such as DVFS, scheduling \textcolor{blue}{and other techniques}.
+to lower the energy consumption of these platforms, such as DVFS, scheduling and other techniques.
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
trade-off between the energy reduction and performance degradation ratio. In
\cite{Our_first_paper} and \cite{pdsec2015}, a frequency selecting algorithm
was proposed to reduce the energy consumption of message passing
-applications \textcolor{blue}{with iterations} running over homogeneous and heterogeneous clusters respectively.
+applications with iterations running over homogeneous and heterogeneous clusters respectively.
The results of the experiments showed significant energy consumption
reductions. All the experimental results were conducted over the SimGrid
simulator \cite{SimGrid}, which offers easy tools to describe homogeneous and heterogeneous platforms, and to simulate the execution of message passing parallel
composed of heterogeneous clusters, is presented. It is applied to the NAS
parallel benchmarks and evaluated over a real testbed, the Grid'5000 platform
\cite{grid5000}. It selects for a grid platform running a message passing
- application \textcolor{blue}{with iterations} the vector of frequencies that simultaneously tries to
+ application with iterations the vector of frequencies that simultaneously tries to
offer the maximum energy reduction and minimum performance degradation
ratios. The algorithm has a very small overhead, works online and does not need
any training or profiling.
sequential, parallel or distributed architecture, homogeneous or heterogeneous
platform, synchronous or asynchronous application.
-In this paper, we are interested in reducing energy for message passing
- synchronous applications \textcolor{blue}{with iterations} running over heterogeneous grid platforms. Some
+In this paper, we are interested in reducing the energy consumption of message passing
+ synchronous applications with iterations running over heterogeneous grid platforms. Some
works have already been done for such platforms and they can be classified into
two types of heterogeneous platforms:
\begin{itemize}
following contributions :
\begin{enumerate}
\item two new energy and performance models for message passing
- synchronous applications \textcolor{blue}{with iterations} running over a heterogeneous grid platform. Both models
+ synchronous applications with iterations running over a heterogeneous grid platform. Both models
take into account communication and slack times. The models can predict the
required energy and the execution time of the application.
platforms. The algorithm has a very small overhead and does not need any
training nor profiling. It uses a new optimization function which
simultaneously maximizes the performance and minimizes the energy consumption
- of a message passing synchronous application \textcolor{blue}{with iterations}.
+ of a message passing synchronous application with iterations.
\end{enumerate}
\label{sec.exe}
\subsection{The execution time of message passing distributed
- applications \textcolor{blue}{with iterations} on a heterogeneous platform}
+ applications with iterations on a heterogeneous platform}
In this paper, we are interested in reducing the energy consumption of message
-passing distributed synchronous applications \textcolor{blue}{with iterations} running over
+passing distributed synchronous applications with iterations running over
heterogeneous grid platforms. A heterogeneous grid platform could be defined as a collection of
heterogeneous computing clusters interconnected via a long distance network which has lower bandwidth
and higher latency than the local networks of the clusters. Each computing cluster in the grid is composed of homogeneous nodes that are connected together via high speed network. Therefore, each cluster has different characteristics such as computing power (FLOPS), energy consumption, CPU's frequency range, network bandwidth and latency.
-The overall execution time of a distributed synchronous application \textcolor{blue}{with iterations}
+The overall execution time of a distributed synchronous application with iterations running
over a heterogeneous grid consists of the sum of the computation time and
the communication time for every iteration on a node.
-\textcolor{blue}{However, nodes from distinct clusters in a grid have different computing powers, thus
-while executing message passing \textcolor{blue}{with iterations} synchronous applications, fast nodes
+However, nodes from distinct clusters in a grid have different computing powers, thus
+while the application, 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. }
+periods are called idle or slack times.
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. \textcolor{blue}{For example, in Figure \ref{fig:heter} the task 1 is the slower task which has no slack time (not waits for the other nodes) and it is only has the communication times.}
+which has the highest computation time and almost no slack time. For example, in Figure \ref{fig:heter}, task 1 is the slower task and it does not have to wait for the other nodes to communicate with them because they all finish their computations before it.
\begin{figure}[!t]
\centering
\label{eq:s}
S = \frac{\Fmax}{\Fnew}
\end{equation}
-\textcolor{blue}{Where $\Fmax$ is the maximum frequency before applying DVFS and $\Fnew$ is the new frequency after applying DVFS.}
+where $\Fmax$ is the maximum frequency before applying any DVFS and $\Fnew$ is the new frequency after applying DVFS.
+
The execution time of a compute bound sequential program is linearly
proportional to the frequency scaling factor $S$. On the other hand, message
passing distributed applications consist of two parts: computation and
of these clusters, they may get different scaling factors represented by a scaling vector:
$(S_{11}, S_{12},\dots, S_{NM_i})$ where $S_{ij}$ is the scaling factor of processor $j$ in cluster $i$ . To
be able to predict the execution time of message passing synchronous
-applications \textcolor{blue}{with iterations} running over a heterogeneous grid, for different vectors of
+applications with iterations running over a heterogeneous grid, for different vectors of
scaling factors, the communication time and the computation time for all the
tasks must be measured during the first iteration before applying any DVFS
operation. Then the execution time for one iteration of the application with any
cluster $i$, $\TcpOld[ij]$ is the computation time of processor $j$ in the cluster $i$
and $\Tcm[hj]$ is the communication time of processor $j$ in the cluster $h$ during the
first iteration. The execution time for one iteration is equal to the sum of the maximum computation time for all nodes with the new scaling factors
-and \textcolor{blue}{the communication time of the slower node without slack time during one iteration.
-The slower node $h$ is the node that gives maximum execution time in all clusters befor scaling its frequency.}
+and the communication time of the slowest node without slack time during one iteration.
+ The slowest node $h$ is the node which takes the maximum execution time to execute an iteration before scaling down its frequency.
It means that only the communication time without any slack time is taken into account.
-Therefore, the execution time of the application \textcolor{blue}{with iterations} is equal to
+Therefore, the execution time of the application is equal to
the execution time of one iteration as in Equation (\ref{eq:perf}) multiplied by the
number of iterations of that application.
Both methods selects the frequencies that gives the best trade-off between
energy consumption reduction and performance for message passing
synchronous applications \textcolor{blue}{with iterations}. In this work we
-are interested in grids that are composed of heterogeneous clusters, \textcolor{blue}{where} the nodes
-have different characteristics such as dynamic power, static power, computation power,
+are interested in grids that are composed of heterogeneous clusters. The nodes from distinct clusters may have
+ different characteristics such as dynamic power, static power, computation power,
frequencies range, network latency and bandwidth.
Due to the heterogeneity of the processors, a vector of scaling factors should be selected
and it must give the best trade-off between energy consumption and performance.
\Require ~
\begin{description}
\item [{$N$}] number of clusters in the grid.
- \item [{$M$}] number of nodes in each cluster.
+ \item [{$M_i$}] number of nodes in each cluster.
\item[{$\Tcp[ij]$}] array of all computation times for all nodes during one iteration and with the highest frequency.
\item[{$\Tcm[ij]$}] array of all communication times for all nodes during one iteration and with the highest frequency.
\item[{$\Fmax[ij]$}] array of the maximum frequencies for all nodes.
of the slowest node and the computation time of the node $i$ as follows:
\begin{equation}
\label{eq:Scp}
- \Scp[ij] = \frac{ \mathop{\max\limits_{i=1,\dots N}}\limits_{j=1,\dots,M}(\Tcp[ij])} {\Tcp[ij]}
+ \Scp[ij] = \frac{ \mathop{\max\limits_{i=1,\dots N}}\limits_{j=1,\dots,M_i}(\Tcp[ij])} {\Tcp[ij]}
\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
+maximum frequency of node and its computed scaling factor as
follows:
\begin{equation}
\label{eq:Fint}
- F_{ij} = \frac{\Fmax[ij]}{\Scp[ij]},~{i=1,2,\dots,N},~{j=1,\dots,M}
+ F_{ij} = \frac{\Fmax[ij]}{\Scp[ij]},~{i=1,2,\dots,N},~{j=1,\dots,M_i}
\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
\subsection{Grid'5000 architecture and power consumption}
\label{sec.grid5000}
-Grid'5000~\cite{grid5000} is a large-scale testbed that consists of ten sites distributed all over metropolitan France and Luxembourg. All the sites are connected together via a special long distance network called RENATER,
+Grid'5000~\cite{grid5000} is a large-scale testbed that consists of ten sites distributed all over metropolitan France and Luxembourg. All the sites are connected together via a special long distance network called RENATER,
which is the French National Telecommunication Network for Technology.
Each site of the grid is composed of a few heterogeneous
computing clusters and each cluster contains many homogeneous nodes. In total,
Two types of local networks are used, Ethernet or Infiniband networks which have different characteristics in terms of bandwidth and latency.
Since Grid'5000 is dedicated to testing, contrary to production grids it allows a user to deploy its own customized operating system on all the booked nodes. The user could have root rights and thus apply DVFS operations while executing a distributed application. Moreover, the Grid'5000 testbed provides at some sites a power measurement tool to capture
-the power consumption for each node in those sites. The measured power is the overall consumed power by all the components of a node at a given instant, such as CPU, hard drive, main-board and memory. For more details refer to
+the power consumption for each node in those sites. The measured power is the overall consumed power by all the components of a node at a given instant. For more details refer to
\cite{Energy_measurement}. In order to correctly measure the CPU power of one core in a node $j$,
firstly, the power consumed by the node while being idle at instant $y$, noted as $\Pidle[jy]$, was measured. Then, the power was measured while running a single thread benchmark with no communication (no idle time) over the same node with its CPU scaled to the maximum available frequency. The latter power measured at time $x$ with maximum frequency for one core of node $j$ is noted $\Pmax[jx]$. The difference between the two measured power consumptions represents the
dynamic power consumption of that core with the maximum frequency, see Figure~\ref{fig:power_cons}.
the same block of operations several times, starting from the initial solution until reaching
the acceptable approximation of the exact solution.}
These applications have different computations and communications ratios and strategies which make them good testbed applications to evaluate the proposed algorithm and energy model.
-The benchmarks have seven different classes, S, W, A, B, C, D and E, that represent the size of the problem that the method solves. In this work, class D was used for all benchmarks in all the experiments presented in the next sections.
+The benchmarks have seven different classes, S, W, A, B, C, D and E, that represent the size of the problem that the method solves. In the next sections, the class D was used for all the benchmarks in all the experiments.
The main reason
for using these two scenarios is to evaluate the influence of long distance communications (higher latency) on the performance of the
scaling factors selection algorithm. Indeed, in the first scenario the computations to communications ratio
-is very low due to the higher communication times which reduce the effect of DVFS operations.
+is very low due to the higher communication times which reduces the effect of the DVFS operations.
The NAS parallel benchmarks are executed over
16 and 32 nodes for each scenario. The number of participating computing nodes from each cluster
& Graphene & Nancy & 5 \\ \cline{2-4}
& Griffon & Nancy & 6 \\
\hline
-\multirow{3}{*}{Tow sites / 32 nodes} & Taurus & Lyon & 10 \\ \cline{2-4}
+\multirow{3}{*}{Two sites / 32 nodes} & Taurus & Lyon & 10 \\ \cline{2-4}
& Graphene & Nancy & 10 \\ \cline{2-4}
& Griffon &Nancy & 12 \\
\hline
is exponentially related to the CPU's frequency value. On the other hand, the increase in the number of computing nodes can
increase the communication times and thus produces less energy saving depending on the
benchmarks being executed. The results of benchmarks CG, MG, BT and FT show more
-energy saving percentage in the one site scenario when executed over 16 nodes than over 32 nodes. LU and SP consume more energy with 16 nodes than 32 in one site because their computations to communications ratio is not affected by the increase of the number of local communications.
+energy saving percentage in the one site scenario when executed over 16 nodes than over 32 nodes. LU and SP consume more energy with 16 nodes than 32 nodes on one site because their computations to communications ratio is not affected by the increase of the number of local communications.
\begin{figure*}[!h]
\centering
Figure \ref{fig:per_d} presents the performance degradation percentages for all benchmarks over the two scenarios.
The performance degradation percentage for the benchmarks running on two sites with
16 or 32 nodes is on average equal to 8.3\% or 4.7\% respectively.
-For this scenario, the proposed scaling algorithm selects smaller frequencies for the executions with 32 nodes without significantly degrading their performance because the communication times are higher with 32 nodes which results in smaller computations to communications ratio. On the other hand, the performance degradation percentage for the benchmarks running on one site with
+For this scenario, the proposed scaling algorithm selects smaller frequencies for the executions with 32 nodes without significantly degrading their performance because the communication times are high with 32 nodes which results in smaller computations to communications ratio. On the other hand, the performance degradation percentage for the benchmarks running on one site with
16 or 32 nodes is on average equal to 3.2\% and 10.6\% respectively. In contrary to the two sites scenario, when the number of computing nodes is increased in the one site scenario, the performance degradation percentage is increased. Therefore, doubling the number of computing
nodes when the communications occur in high speed network does not decrease the computations to
communication ratio.
presented in Figures~\ref{fig:edp-eng}, \ref{fig:edp-perf} and \ref{fig:edp-dist} respectively.
As shown in these figures, the proposed frequencies selection algorithm, Maxdist, outperforms the EDP algorithm in terms of energy consumption reduction and performance for all of the benchmarks executed over the two scenarios.
-The proposed algorithm gives better results than the EDP method because it
-maximizes the energy saving and the performance at the same time.
+The proposed algorithm gives better results than the EDP method because the former selects the set of frequencies that
+gives the best tradeoff between energy saving and performance.
Moreover, the proposed scaling algorithm gives the same weight for these two metrics.
Whereas, the EDP algorithm gives sometimes negative trade-off values for some benchmarks in the two sites scenarios.
These negative trade-off values mean that the performance degradation percentage is higher than the energy saving percentage.
