challenges.tex

\comment{
\begin{frame}[t]
\frametitle{Harnessing Parallelism: Challenges}
\framesubtitle{Trends in System Architecture}
    \begin{itemize}
        \item Data movement is increasingly expensive
            \begin{itemize}
                \item (relatively) slow memory
                \item (relatively) slow interconnects
            \end{itemize}
        \item Energy / Power considerations will exacerbate these issues
        \pause
        \item Compute resources are cheap
            \begin{itemize}
                \item Smaller footprint on silicon
                \item Lower energy share compared to memory + interconnect
            \end{itemize}
    \end{itemize}
    \pause
    \begin{block}{Implications}
       More available compute cycles for each incoming byte of data
       %\pause
       %\begin{itemize}
       %    \item Restructure code to extract all flops from incoming operands
       %    \item Avoid moving data. Instead recompute!
       %    \item Commmunication-avoiding algorithms?
       %\end{itemize}
    \end{block}
\end{frame}


\begin{frame}[shrink]
\frametitle{Harnessing Parallelism: Challenges}
\framesubtitle{Trends in System Architecture}
  \begin{columns}
    \begin{column}{0.60\textwidth}
      \begin{itemize}
      \item However, compute resources are not faster cores, but \textbf{more cores}
      \item Unprecedented levels of available concurrency
        \begin{itemize}
        \item IBM BG/Q
          \begin{itemize}
          \item `Sequoia': 1,572,864 cores
          \item `Mira': 786,432 cores
          \end{itemize}
        \item Cray
          \begin{itemize}
          \item XE6 `Bluewaters`: $>$ 380,000 cores
          \item XK6 `Titan': 299,008 cores
          \end{itemize}
        \item K Supercomputer: 705,024 cores
        \end{itemize}
      \end{itemize}
    \end{column}
    \begin{column}{0.40\textwidth}
      \includegraphics[width=1\textwidth]{figures/mira.jpg}
    \end{column}
  \end{columns}

    \pause
    \begin{block}{Implications}
        \begin{itemize}
            \item Each thread of execution has to:
                \begin{itemize}
                    \item operate on lesser data
                    \item wait relatively longer for remote data
                \end{itemize}
            \item Have to operate in \textbf{strong scaling} regime
            % Even if you don't do anything, an 8GB problem will have to run on
            % more cores a few years from now simply because there will be many
            % more cores for the same 8GB 
            % If a product has to stay ahead of the competition, it has to scale
            % the same problem to even more cores with time to run faster
        \end{itemize}
    \end{block}
\end{frame}


\begin{frame}[t]
\frametitle{Harnessing Parallelism: Challenges}
\framesubtitle{Trends in System Architecture}
    \begin{itemize}
        \item Growing heterogeneity: Each thread of execution has different abilities
            \begin{itemize}
                \item NUMA cores on each node
                \item Secondary threads in SMT can be less capable (IBM Power 7)
                \item Accelerators (NVIDIA / AMD GPUs, Intel MIC etc)
                \item Some cores run system daemons
                \item Combinations of above
                    \begin{itemize}
                        \item GPUs only on some nodes (BlueWaters: NCSA)
                        \item MIC + SMT nodes (Stampede: TACC)
                    \end{itemize}
            \end{itemize}
        \pause
        \item Each architecture expects different treatment
            \begin{itemize}
                \item GPUs expect evenly divided pieces
                \item SMT might expect uneven pieces (IBM Power7)
            \end{itemize}
    \end{itemize}
    \pause
    \begin{block}{Implications}
        \begin{itemize}
            \item Achieving load balance on such hardware is challenging
            \item Minimizing idle time is extremely difficult
        \end{itemize}
    \end{block}
\end{frame}


\begin{frame}[t]
\frametitle{Harnessing Parallelism: Challenges}
\framesubtitle{Trends in System Architecture}
  \begin{itemize}
  \item Extracting performance on tighter power / energy budget
  \item Hardware component failures / faults
  \end{itemize}
\end{frame}


\begin{frame}[fragile]
\frametitle{Harnessing Parallelism: Challenges}
\framesubtitle{Application characteristics}
  \begin{itemize}
  \item Complex physics in multiple, interacting modules
  \item Adaptive, spatial and temporal resolutions
  \item Need for faster solutions (not just larger problems)
  \end{itemize}
\end{frame}

\begin{frame}[t]
\frametitle{Observations}
\framesubtitle{Exascale Applications}
  \begin{itemize}
    \item Development of new models must be driven by the needs of exascale applications
    \begin{itemize}
      \item Multi-resolution
      \item Multi-module (multi-physics)
      \item Dynamic/adaptive: to handle application variation
      \item Adapt to a volatile computational environment
      \item Exploit heterogeneous architecture
      \item Deal with thermal and energy considerations
    \end{itemize}
    \pause
    \item So? Consequences:
    \begin{itemize}
      \item Must support automated resource management
      \item Must support interoperability and parallel composition
    \end{itemize}
  \end{itemize}
\end{frame}
}
\begin{frame}[t]
\frametitle{Harnessing Parallelism: Challenges}
\framesubtitle{Complex machines and Sophisticated Applications}
\begin{itemize}
\item Hardware Challenges
  \begin{itemize}
    \item Large machines (largest have about 2 million cores!)
    \item Multicore node
    \item Heterogeneity
    \item Upcoming: power/energy/thermal considerations, and component failures
  \end{itemize}
\item{Applications are getting more sophisticated}
  \begin{itemize}
    \item Dynamic and adpative refinements
    \item Multi-physics codes
    \item Need for strong scaling
  \end{itemize}
\end{itemize}
\end{frame}

% \begin{frame}[fragile]
% \frametitle{Harnessing Parallelism: Challenges}
% \framesubtitle{2D Jacobi Iterations: 5-point Stencil}
%    \begin{center} \includegraphics[width=0.6\textwidth]{figures/2DJacobi_NeighborComm.jpg} \end{center}
% \end{frame}


% \begin{frame}[fragile]
% \frametitle{Harnessing Parallelism: Challenges}
% \framesubtitle{2D Jacobi Iterations: No Overlap}
%     \lstinputlisting{code/jacobi2D-mpi.c}
% \end{frame}


% \begin{frame}[fragile, shrink]
% \frametitle{Harnessing Parallelism: Challenges}
% \framesubtitle{2D Jacobi Iterations: Some Overlap}
%     \lstinputlisting{code/jacobi2D-mpi-nonblocking.c}
% \end{frame}

\comment{
\begin{frame}[fragile]
\frametitle{Harnessing Parallelism: Challenges}
\framesubtitle{Composing Independent Parallel Modules}
  \frametitle{Harnessing Parallelism}
  \framesubtitle{Composing Independent Parallel Modules}
  \begin{itemize}
    \item Sequential blocks (SEBs) in two modules can be partitioned in space
  \end{itemize}
  \begin{center}
    \includegraphics[width=0.8\textwidth]{figures/spaceDivision.pdf}
  \end{center}
\end{frame}

\begin{frame}[fragile]
\frametitle{Harnessing Parallelism: Challenges}
\framesubtitle{Composing Independent Parallel Modules}
  \begin{itemize}
    \item Sequential blocks (SEBs) in two modules can be partitioned in time
  \end{itemize}
  \begin{center}
    \includegraphics[width=0.8\textwidth]{figures/timeDivision.pdf}
  \end{center}
\end{frame}
}