\begin{abstract}
Dynamic Voltage Frequency Scaling (DVFS) can be applied to modern CPUs. This
technique is usually used to reduce the energy consumed by a CPU while
- computing. Indeed, power consumption by a processor at a given instant is
+ computing. Indeed, power consumption by a processor at a given time is
exponentially related to its frequency. Thus, decreasing the frequency
reduces the power consumed by the CPU. However, it can also significantly
affect the performance of the executed program if it is compute bound and if a
constructors have been regularly increasing the number of computing cores and
processors in supercomputers (for example in November 2013, according to the
TOP500 list~\cite{43}, the Tianhe-2 was the fastest supercomputer. It has more
-than 3 millions of cores and delivers more than \np[Tflop/s]{33} while consuming
+than 3 million of cores and delivers more than \np[Tflop/s]{33} while consuming
\np[kW]{17808}). This large increase in number of computing cores has led to
large energy consumption by these architectures. Moreover, the price of energy
is expected to continue its ascent according to the demand. For all these
-reasons energy reduction became an important topic in the high performance
-computing field. To tackle this problem, many researchers used DVFS (Dynamic
+reasons energy reduction has become an important topic in the high performance
+computing field. To tackle this problem, many researchers use DVFS (Dynamic
Voltage Frequency Scaling) operations which reduce dynamically the frequency and
voltage of cores and thus their energy consumption. Indeed, modern CPUs offer a
set of acceptable frequencies which are usually called gears, and the user or
the operating system can modify the frequency of the processor according to its
needs. However, DVFS also degrades the performance of computation. Therefore
-researchers try to reduce the frequency to minimum when processors are idle
+researchers try to reduce the frequency to the minimum when processors are idle
(waiting for data from other processors or communicating with other processors).
-Moreover, depending on their objectives they use heuristics to find the best
+Moreover, depending on their objectives, they use heuristics to find the best
scaling factor during the computation. If they aim for performance they choose
the best scaling factor that reduces the consumed energy while affecting as
little as possible the performance. On the other hand, if they aim for energy
reduction, the chosen scaling factor must produce the most energy efficient
execution without considering the degradation of the performance. It is
-important to notice that lowering the frequency to minimum value does not always
+important to notice that lowering the frequency to the minimum value does not always
give the most energy efficient execution due to energy leakage. The best
scaling factor might be chosen during execution (online) or during a
pre-execution phase. In this paper, we present an algorithm that selects a
objective of HPC systems is to execute as fast as possible the application.
Therefore, our algorithm selects the scaling factor online with very small
footprint. The proposed algorithm takes into account the communication times of
-the MPI program to choose the scaling factor. This algorithm has ability to
+the MPI program to choose the scaling factor. This algorithm has the ability to
predict both energy consumption and execution time over all available scaling
factors. The prediction achieved depends on some computing time information,
gathered at the beginning of the runtime. We apply this algorithm to seven MPI
The offline scaling factor selection methods are executed before the runtime of
the program. They return static scaling factor values to the processors
-participating in the execution of the parallel program. On one hand, the
+participating in the execution of the parallel program. On the one hand, the
scaling factor values could be computed based on information retrieved by
analyzing the code of the program and the computing system that will execute it.
In~\cite{40}, Azevedo et al. detect during compilation the dependency points
between tasks in a multi-task program. This information is then used to lower
the frequency of some processors in order to eliminate slack times. A slack
-time is the period of time during which a processor that have already finished
-its computation, have to wait for a set of processors to finish their
+time is the period of time during which a processor that has already finished
+its computation, has to wait for a set of processors to finish their
computations and send their results to the waiting processor in order to
continue its task that is dependent on the results of computations being
executed on other processors. Freeh et al. showed in~\cite{17} that the
communication times of MPI programs do not change when the frequency is scaled
down. On the other hand, some offline scaling factor selection methods use the
-information gathered from previous full or partial executions of the program. A
-part or the whole program is usually executed over all the available frequency
-gears and the the execution time and the energy consumed with each frequency
-gear are measured. Then an heuristic or an exact method uses the retrieved
+information gathered from previous full or partial executions of the program. The whole program or, a
+part of it, is usually executed over all the available frequency
+gears and the execution time and the energy consumed with each frequency
+gear are measured. Then a heuristic or an exact method uses the retrieved
information to compute the values of the scaling factor for the processors.
In~\cite{29}, Xie et al. use an exact exponential breadth-first search algorithm
to compute the scaling factor values that give the optimal energy reduction
while respecting a deadline for a sequential program. They also present a
linear heuristic that approximates the optimal solution. In~\cite{8} , Rountree
et al. use a linear programming algorithm, while in~\cite{38,34}, Cochran et
-al. use multi logistic regression algorithm for the same goal. The main
-drawback for these methods is that they all require executing a part or the
-whole program on all frequency gears for each new instance of the same program.
+al. use a multi-logistic regression algorithm for the same goal. The main
+drawback of these methods is that they all require executing the
+whole program or, a part of it, on all frequency gears for each new instance of the same program.
\subsection{Online scaling factor selection methods}
The online scaling factor selection methods are executed during the runtime of
the program. They are usually integrated into iterative programs where the same
block of instructions is executed many times. During the first few iterations,
-many informations are measured such as the execution time, the energy consumed
+a lot of information is measured such as the execution time, the energy consumed
using a multimeter, the slack times, \dots{} Then a method will exploit these
measurements to compute the scaling factor values for each processor. This
operation, measurements and computing new scaling factor, can be repeated as
gear and then compute the scaling factor values for the rest of the processors.
Although this model was built for parallel architectures, it can be adapted to
distributed architectures by taking into account the communications. The
-primary contribution of our paper is presenting a new online scaling factor
+primary contribution of our paper is to present a new online scaling factor
selection method which has the following characteristics:
\begin{enumerate}
\item It is based on Rauber and Rünger analytical model to predict the energy
\item It is well adapted to distributed architectures because it takes into
account the communication time.
\item It is well adapted to distributed applications with imbalanced tasks.
-\item It has very small footprint when compared to other methods
+\item It has a very small footprint when compared to other methods
(e.g.,~\cite{19}) and does not require profiling or training as
in~\cite{38,34}.
\end{enumerate}
\subsection{Parallel tasks execution on homogeneous platform}
-A homogeneous cluster consists of identical nodes in terms of hardware and
+A homogeneous cluster consists in identical nodes in terms of hardware and
software. Each node has its own memory and at least one processor which can be
a multi-core. The nodes are connected via a high bandwidth network. Tasks
executed on this model can be either synchronous or asynchronous. In this paper
\caption{Parallel tasks on homogeneous platform}
\label{fig:homo}
\end{figure*}
-Therefore, the execution time of a task consists of the computation time and the
+Therefore, the execution time of a task consists in the computation time and the
communication time. Moreover, the synchronous communications between tasks can
lead to slack times while tasks wait at the synchronization barrier for other
tasks to finish their tasks (see figure~(\ref{fig:h1})). The imbalanced
-communications happen when nodes have to send/receive different amount of data
-or they communicate with different number of nodes. Another source of slack
-times is the imbalanced computations. This happens when processing different
+communications happen when nodes have to send/receive different amounts of data
+or they communicate with different numbers of nodes. Other sources of slack
+times are imbalanced computations. This happens when processing different
amounts of data on each processor (see figure~(\ref{fig:h2})). In this case the
fastest tasks have to wait at the synchronization barrier for the slowest ones
to begin the next task. In both cases the overall execution time of the program
\subsection{Energy model for homogeneous platform}
-Many researchers~\cite{9,3,15,26} divide the power consumed by a processor to
+Many researchers~\cite{9,3,15,26} 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 on, the latter is only consumed during
computation times. The dynamic power $P_{dyn}$ is related to the switching
\end{equation}
where $T$ is the execution time of the program, $T_{Comp}$ is the computation
time and $T_{Comp} \leq T$. $T_{Comp}$ may be equal to $T$ if there is no
-communications, no slack times and no synchronizations.
+communication, no slack time and no synchronization.
DVFS is a process that is allowed in modern processors to reduce the dynamic
power by scaling down the voltage and frequency. Its main objective is to
S_i = S \cdot \frac{T_1}{T_i}
= \frac{F_\textit{max}}{F_\textit{new}} \cdot \frac{T_1}{T_i}
\end{equation}
-In this paper we depend on Rauber and Rünger energy model EQ~(\ref{eq:energy})
-for two reasons: (1) this model is used for any number of concurrent tasks, and
-(2) we compare our algorithm with Rauber and Rünger scaling factor selection
-method which is based on EQ~(\ref{eq:energy}). The optimal scaling factor is
-computed by minimizing the derivation for this equation which produces
+In this paper we use Rauber and Rünger's energy model, EQ~(\ref{eq:energy}), because it can be applied to homogeneous clusters if the communication time is taken in consideration. Moreover, we compare our algorithm with Rauber and Rünger's scaling factor selection
+method which uses the same energy model. In their method, the optimal scaling factor is
+computed by minimizing the derivation of EQ~(\ref{eq:energy}) which produces
EQ~(\ref{eq:sopt}).
\begin{equation}
\section{Performance evaluation of MPI programs}
\label{sec.mpip}
-The performance (execution time) of parallel synchronous MPI applications depend
+The performance (execution time) of parallel synchronous MPI applications depends
on the time of the slowest task as in figure~(\ref{fig:homo}). If there is no
communication and the application is not data bounded, the execution time of a
parallel program is linearly proportional to the operational frequency and any
DVFS operation for energy reduction increases the execution time of the parallel
program. Therefore, the scaling factor $S$ is linearly proportional to the
-execution time. However, in most of MPI applications the processes exchange
+execution time. However, in most MPI applications the processes exchange
data. During these communications the processors involved remain idle until the
-communications are finished. For that reason any change in the frequency has no
+communications are finished. For that reason, any change in the frequency has no
impact on the time of communication~\cite{17}. The communication time for a
task is the summation of periods of time that begin with an MPI call for sending
or receiving a message till the message is synchronously sent or received. To
be able to predict the execution time of MPI program, the communication time and
-the computation time for the slower task must be measured before scaling. These
+the computation time for the slowest task must be measured before scaling. These
times are used to predict the execution time for any MPI program as a function
of the new scaling factor as in EQ~(\ref{eq:tnew}).
\begin{equation}
In this paper, this prediction method is used to select the best scaling factor
for each processor as presented in the next section.
-\section{Performance to energy competition}
+\section{Performance and energy reduction trade-off}
\label{sec.compet}
-This section demonstrates our approach for choosing the optimal scaling factor.
-This factor gives maximum energy reduction taking into account the execution
-times for both computation and communication. The relation between the energy
-and the performance is nonlinear and complex, because the relation of the energy
-with scaling factor is nonlinear and with the performance it is linear
-see~\cite{17}. Moreover, they are not measured using the same metric. For
-solving this problem, we normalize the energy by calculating the ratio between
+This section presents our approach for choosing the optimal scaling factor.
+This factor gives maximum energy reduction while taking into account the execution
+times for both computation and communication. The relation between the performance
+and the energy is nonlinear and complex. Thus, unlike the relation between the performance and the scaling factor, the relation of energy with the scaling factor is nonlinear, for more details refer to~\cite{17}. Moreover, they are not measured using the same metric. To
+solve this problem, we normalize the energy by calculating the ratio between
the consumed energy with scaled frequency and the consumed energy without scaled
frequency:
\begin{multline}
P_\textit{dyn} \cdot \left(T_1+\sum_{i=2}^{N}\frac{T_i^3}{T_1^2}\right) +
P_\textit{static} \cdot T_1 \cdot N }
\end{multline}
-By the same way we can normalize the performance as follows:
+In the same way we can normalize the performance as follows:
\begin{equation}
\label{eq:pnorm}
P_\textit{Norm} = \frac{T_\textit{New}}{T_\textit{Old}}
\end{equation}
The second problem is that the optimization operation for both energy and
performance is not in the same direction. In other words, the normalized energy
-and the performance curves are not in the same direction see
+and the performance curves are not at the same direction see
figure~(\ref{fig:r2}). While the main goal is to optimize the energy and
performance in the same time. According to the equations~(\ref{eq:enorm})
and~(\ref{eq:pnorm}), the scaling factor $S$ reduce both the energy and the
performance simultaneously. But the main objective is to produce maximum energy
reduction with minimum performance reduction. Many researchers used different
strategies to solve this nonlinear problem for example see~\cite{19,42}, their
-methods add big overhead to the algorithm for selecting the suitable frequency.
-In this paper we present a method to find the optimal scaling factor $S$ for
-optimizing both energy and performance simultaneously without adding big
-overheads. Our solution for this problem is to make the optimization process
-have the same direction. Therefore, we inverse the equation of normalize
+methods add big overheads to the algorithm to select the suitable frequency.
+In this paper we present a method to find the optimal scaling factor $S$ to optimize both energy and performance simultaneously without adding a big
+overhead. Our solution for this problem is to make the optimization process
+for energy and performance follow the same direction. Therefore, we inverse the equation of the normalized
performance as follows:
\begin{equation}
\label{eq:pnorm_en}
According to this equation all the nodes may have the same frequency value if
they have balanced workloads, otherwise, they take different frequencies when
having imbalanced workloads. Thus, EQ~(\ref{eq:fi}) adapts the frequency of the
-CPU to the nodes' workloads to maintain performance.
+CPU to the nodes' workloads to maintain the performance of the program.
\section{Experimental results}
\label{sec.expe}
Our experiments are executed on the simulator SimGrid/SMPI v3.10. We configure
the simulator to use a homogeneous cluster with one core per node. The detailed
-characteristics of our platform file are shown in the
+characteristics of our platform file are shown in
table~(\ref{table:platform}). Each node in the cluster has 18 frequency values
from \np[GHz]{2.5} to \np[MHz]{800} with \np[MHz]{100} difference between each
two successive frequencies. The simulated network link is \np[GB]{1} Ethernet
\subsection{Performance prediction verification}
In this section we evaluate the precision of our performance prediction method
-based on EQ~(\ref{eq:tnew}) by applying it the NAS benchmarks. The NAS programs
-are executed with the class B option for comparing the real execution time with
+based on EQ~(\ref{eq:tnew}) by applying it to the NAS benchmarks. The NAS programs
+are executed with the class B option to compare the real execution time with
the predicted execution time. Each program runs offline with all available
scaling factors on 8 or 9 nodes (depending on the benchmark) to produce real
execution time values. These scaling factors are computed by dividing the
(EP, CG, MG, FT, BT, LU and SP) which were run with three classes (A, B and C).
For each instance the benchmarks were executed on a number of processors
proportional to the size of the class. Each class represents the problem size
-ascending from the class A to C. Additionally, depending on some speed up
+ascending from class A to C. Additionally, depending on some speed up
points for each class we run the classes A, B and C on 4, 8 or 9 and 16 nodes
respectively. Depending on EQ~(\ref{eq:energy}), we measure the energy
-consumption for all the NAS MPI programs while assuming the power dynamic with
+consumption for all the NAS MPI programs while assuming that the dynamic power with
the highest frequency is equal to \np[W]{20} and the power static is equal to
\np[W]{4} for all experiments. These power values were also used by Rauber and
Rünger in~\cite{3}. The results showed that the algorithm selected different
program as in the plots~(\ref{fig:nas}). These plots illustrate that there are
different distances between the normalized energy and the normalized inverted
performance curves, because there are different communication features for each
-benchmark. When there are little or not communications, the inverted
+benchmark. When there are little or no communications, the inverted
performance curve is very close to the energy curve. Then the distance between
the two curves is very small. This leads to small energy savings. The opposite
happens when there are a lot of communication, the distance between the two
optimize both the energy and the performance simultaneously for all NAS
benchmarks. In table~(\ref{table:factors results}), we record all optimal
scaling factors results for each benchmark running class C. These scaling
-factors give the maximum energy saving percent and the minimum performance
-degradation percent at the same time from all available scaling factors.
+factors give the maximum energy saving percentage and the minimum performance
+degradation percentage at the same time from all available scaling factors.
\begin{figure*}[t]
\centering
\includegraphics[width=.328\textwidth]{fig/ep}\hfill%
\label{table:factors results}
% is used to refer this table in the text
\end{table}
-As shown in the table~(\ref{table:factors results}), when the optimal scaling
-factor has big value we can gain more energy savings for example as in CG and
-FT. The opposite happens when the optimal scaling factor is small value as
-example BT and EP. Our algorithm selects big scaling factor value when the
+
+As shown in table~(\ref{table:factors results}), when the optimal scaling
+factor has a big value we can gain more energy savings as in CG and
+FT benchmarks. The opposite happens when the optimal scaling factor has a small value as in BT and EP benchmarks. Our algorithm selects a big scaling factor value when the
communication and the other slacks times are big and smaller ones in opposite
-cases. In EP there are no communications inside the iterations. This make our
-algorithm to selects smaller scaling factor values (inducing smaller energy
+cases. In EP there are no communication inside the iterations. This leads our
+algorithm to select smaller scaling factor values (inducing smaller energy
savings).
\subsection{Results comparison}
factor is not optimal for energy reduction, the results in these tables prove
that our algorithm returns the best scaling factor that satisfy our objective
method: the largest distance between energy reduction and performance
-degradation. Negative values in the energy-performance column mean that one of
+degradation. Figure~\ref{fig:compare} illustrates even better the distance between the energy reduction and performance degradation. The negative values mean that one of
the two objectives (energy or performance) have been degraded more than the
other. The positive trade-offs with the highest values lead to maximum energy
savings while keeping the performance degradation as low as possible. Our
algorithm always gives the highest positive energy to performance trade-offs
-while Rauber and Rünger method ($R_{E-P}$) gives in some time negative
+while Rauber and Rünger's method, ($R_{E-P}$), gives sometimes negative
trade-offs such as in BT and EP.
-%\begin{figure*}[t]
-% \centering
-% \includegraphics[width=.328\textwidth]{fig/compare_class_A}
-% \includegraphics[width=.328\textwidth]{fig/compare_class_B}
-% \includegraphics[width=.328\textwidth]{fig/compare_class_C}
-% \caption{Comparing our method to Rauber and Rünger methods}
-% \label{fig:compare}
-%\end{figure*}
+\begin{figure*}[t]
+ \centering
+ \includegraphics[width=.328\textwidth]{fig/compare_class_A}
+ \includegraphics[width=.328\textwidth]{fig/compare_class_B}
+ \includegraphics[width=.328\textwidth]{fig/compare_class_C}
+ \caption{Comparing our method to Rauber and Rünger's methods}
+ \label{fig:compare}
+\end{figure*}
\section{Conclusion}
\label{sec.concl}
the energy and the inverted performance curves. To evaluate this method, we
have applied it to the NAS benchmarks and it was compared to Rauber and Rünger
methods while being executed on the simulator SimGrid. The results showed that
-our method, outperforms Rauber and Rünger methods in terms of energy-performance
+our method, outperforms Rauber and Rünger's methods in terms of energy-performance
ratio.
In the near future, we would like to adapt this scaling factor selection method
