%%%%%%%%%%%%%%%%%%%%
%% SLIDE 03 %%
%%%%%%%%%%%%%%%%%%%%
-\begin{frame}{Introduction and problem definition}
+\begin{frame}{Definition of parallel computing}
\section{\small {Introduction and Problem definition}}
\centering
\includegraphics[width=0.99\textwidth]{para.pdf}
\end{frame}
-
-
-
-
-
\begin{frame}{Execution of synchronous parallel tasks}
\vspace{-0.5 cm}
\end{minipage}%
\vspace{0.2cm}
\begin{minipage}{0.5\textwidth}
- \textcolor{blue}{2)} \small \bf \textcolor{black}{Increase the number of nodes.}
+ \textcolor{blue}{2)} \small \bf \textcolor{black}{Increase the number of computing
+ units.}
\textcolor{black}{The supercomputer Tianhe-2 has more than 3 million cores and consumes around 17.8 megawatts.}
\begin{itemize} \small \justifying
- \item Studying the effect of the scaling factor on the \textbf{energy consumption and performance } of parallel applications with iterations. \medskip
+ \item Studying the effect of the frequency scaling on the \textbf{energy consumption and performance } of parallel applications with iterations. \medskip
\item Discovering the \textbf{energy-performance trade-off relation} when changing the frequency of the processor.\medskip
- \item Proposing an algorithm for selecting the scaling factor that produces \textbf {the optimal trade-off} between the energy consumption and the performance. \medskip
+ \item Proposing an algorithm for selecting the scaling factor that produces \textbf {the good trade-off} between the energy consumption and the performance. \medskip
\item Comparing the proposed algorithm to existing methods.
+%%%%%%%%%%%%%%%%%%%%
+%% SLIDE 13 %%
+%%%%%%%%%%%%%%%%%%%%
+\begin{frame}{Performance evaluation of MPI programs}
+
+\small The frequency scaling factor is the ratio between the maximum and the new frequency, \textcolor{blue}{$S = \frac{F_{max}}{F_{new}}$}.
+ \vspace{5 mm}
+
+ \begin{femtoBlock}{}
+ \vspace{-5 mm}
+ \begin{block}{\small Execution time prediction model}
+ \centering{ $ \textcolor{red}{T_{new}} = \textcolor{blue}{T_{Max Comp Old} \cdot S + T_{{Min Comm Old}}}$}
+ \end{block}
+ \vspace{5 mm}
+ \centering{\includegraphics[width=.4\textwidth]{c1/cg_per}
+ \quad%
+ \includegraphics[width=.4\textwidth]{c1/lu_pre}}
+ \vspace{1 mm}
+
+ \small The maximum normalized error for CG=0.0073 \textbf{(the smallest)} and LU=0.031 \textbf{(the worst)}.
+ \end{femtoBlock}
+\end{frame}
+
+
+
+
+
+
+
+
+
+
%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Energy model for a homogeneous platform}
The power consumed by a processor is divided into two power metrics: the dynamic (\textcolor{red}{$P_d$}) and the static
- (\textcolor{red}{$P_s$}) power.
+ (\textcolor{red}{$P_s$}) powers.
\begin{equation}
\label{eq:pd}
\textcolor{red}{ P_d} = \textcolor{blue}{\alpha \cdot CL \cdot V^2 \cdot F}
\end{equation}
\scriptsize \underline{Where}: \\
- \scriptsize {\textcolor{blue}{$\alpha$}: switching activity \hspace{15 mm} \textcolor{blue}{$CL$}: load capacitance\\
- \textcolor{blue}{$V$}: the supply voltage \hspace{14 mm} \textcolor{blue}{$F$}: operational frequency}
+ \scriptsize {\textcolor{blue}{$\alpha$}: switching activity. \hspace{15 mm} \textcolor{blue}{$CL$}: load capacitance [F].\\
+ \textcolor{blue}{$V$}: the supply voltage [V]. \hspace{8 mm} \textcolor{blue}{$F$}: operational frequency [Hz].}
\begin{equation}
\label{eq:ps}
\small \textcolor{red}{P_s} = \textcolor{blue}{V \cdot N_{trans} \cdot K_{design} \cdot I_{Leak}}
\end{equation}
\underline{Where}:\\
- \scriptsize{ \textcolor{blue}{$V$}: the supply voltage. \hspace{28 mm} \textcolor{blue}{$N_{trans}$}: number of transistors. \\
- \textcolor{blue}{$K_{design}$}: design dependent parameter. \hspace{8 mm} \textcolor{blue}{$I_{leak}$}: technology dependent
- parameter.}
+ \scriptsize{ \textcolor{blue}{$V$}: the supply voltage [V]. \hspace{19 mm} \textcolor{blue}{$N_{trans}$}: number of transistors. \\
+ \textcolor{blue}{$K_{design}$}: design dependent parameter. \hspace{3 mm} \textcolor{blue}{$I_{leak}$}: technology dependent
+ parameter [A].}
+
- The frequency scaling factor is the ratio between the maximum and the new frequency, \textcolor{blue}{$S = \frac{F_{max}}{F_{new}}$}.
\end{frame}
+
+
%%%%%%%%%%%%%%%%%%%%
%% SLIDE 12 %%
%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Energy model for a homogeneous platform}
- \vspace{-0.77cm}
+ \vspace{-0.77cm}
\begin{figure}
\animategraphics[autopause,controls,scale=0.3,buttonsize=0.2cm]{10}{homo-model/a-}{0}{441}
%\includegraphics[width=0.6\textwidth]{homo-model/a-356}
\end{frame}
-
-
-%%%%%%%%%%%%%%%%%%%%
-%% SLIDE 13 %%
-%%%%%%%%%%%%%%%%%%%%
-\begin{frame}{Performance evaluation of MPI programs}
- \begin{femtoBlock}{}
- \vspace{-5 mm}
- \begin{block}{\small Execution time prediction model}
- \centering{ $ \textcolor{red}{T_{new}} = \textcolor{blue}{T_{Max Comp Old} \cdot S + T_{{Min Comm Old}}}$}
- \end{block}
- \vspace{10 mm}
- \centering{\includegraphics[width=.4\textwidth]{c1/cg_per}
- \quad%
- \includegraphics[width=.4\textwidth]{c1/lu_pre}}
- \vspace{5 mm}
-
- \small The maximum normalized error for CG=0.0073 \textbf{(the smallest)} and LU=0.031 \textbf{(the worst)}.
- \end{femtoBlock}
-\end{frame}
%%%%%%%%%%%%%%%%%%%%
%% SLIDE 17 %%
%%%%%%%%%%%%%%%%%%%%
-\begin{frame}{Experimental results }
+\begin{frame}{Experiment over SimGrid }
\begin{femtoBlock}{}
\begin{itemize}
\small
%% SLIDE 19 %%
%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Results comparison}
- \begin{block}{\small Rauber and Rünger's optimal scaling factor}
- $S_{opt} = \sqrt[3]{\frac{2}{N} \cdot \frac{P_{dyn}}{P_{static}} \cdot
- \left( 1 + \sum_{i=2}^{N} \frac{T_i^3}{T_1^3}\right) } $
- \end{block}
-
-
- \centering {
- %\includegraphics[width=.33\textwidth]{c1/c1.pdf}
- %\qquad
- %\includegraphics[width=.33\textwidth]{c1/c2.pdf}}
-
+ \small \textcolor{blue}{Rauber and Rünger's scaling factor \textcolor{black}{ \tiny \textsuperscript{2}}}
- \includegraphics[width=.55\textwidth]{c1/compare-c.pdf}}
+ \vspace{2 mm}
+
+ $S_{opt} = \sqrt[3]{\frac{2}{N} \cdot \frac{P_{dyn}}{P_{static}} \cdot
+ \left( 1 + \sum_{i=2}^{N} \frac{T_i^3}{T_1^3}\right) } $
+
+ \begin{center}
+ \includegraphics[width=.55\textwidth]{c1/compare-c.pdf}
+ \end{center}
+
+
+\vspace{-2 mm}
+ \tiny \textsuperscript{2} Thomas Rauber and Gudula Rünger. Analytical modeling and simulation of the energy consumption of independent tasks. In Proceedings of the Winter Simulation Conference, 2012.
\end{frame}
\item Studying the effect of the scaling factor $S$ on both the \textcolor{blue}{energy consumption and the performance} of
message passing iterative applications. \medskip
- \item Computing the vector of scaling factors ($S_1, S_2, ..., S_n$) producing \textcolor{blue} {the optimal trade-off} between
+ \item Computing the vector of scaling factors ($S_1, S_2, ..., S_n$) producing \textcolor{blue} {the good trade-off} between
the energy consumption and the performance.
\end{itemize}
%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Comparing the objective function to EDP}
- EDP is the products between the energy consumption and the delay.
+ EDP is the product between the energy consumption and the delay \tiny\textsuperscript{3}.
\vspace{-5 mm}
\begin{figure}[!t]
\centering
\end{figure}
+
+ \tiny \textsuperscript{3} Spiliopoulos et al, Green governors: A framework for continuously adaptive dvfs, in International Green Computing Conference and Workshops (IGCC), 2011.
\end{frame}
%\begin{frame}{Summary}
%\begin{itemize}
%%%%%%%%%%%%%%%%%%%%
%% SLIDE 46 %%
%%%%%%%%%%%%%%%%%%%%
-\begin{frame}{The scaling algorithm for Asynch. applications}
-\vspace{-0.1 mm}
-\centering
-\includegraphics[width=0.55\textwidth]{algo-hybrid.pdf}
-\end{frame}
+%\begin{frame}{The scaling algorithm for Asynch. applications}
+%\vspace{-0.1 mm}
+%\centering
+%\includegraphics[width=0.55\textwidth]{algo-hybrid.pdf}
+%\end{frame}
Science}, 2016.
\item Ahmed Fanfakh, Jean-Claude Charr, Raphaël Couturier, Arnaud Giersch. Energy Consumption Reduction for
- Asynchronous Message Passing Applications. \textit{Journal of Supercomputing}, 2016, (Submitted)
+ Asynchronous Message Passing Applications. \textit{Journal of Supercomputing}, 2016, (Accepted with minor revisions)
\end{enumerate}
\end{block}
\small \barrow The proposed algorithms for heterogeneous platforms should be applied to heterogeneous platforms composed of \textcolor{blue}{CPUs and GPUs}.
\small \barrow Comparing the results returned by the energy models to the values given by \textcolor{blue}{real instruments that measure the energy consumptions} of CPUs during the execution time.
+\small \barrow Considering the power consumed by the other devices in the node such as
+\textcolor{blue}{the memory and the hard drive} in the energy consumption model.
+
\end{itemize}
\end{frame}
