2022-10-28-talk-fwam/main.tex

% vim: set foldmethod=marker foldmarker=<<<,>>>:

\input{ccmbeamer}
%\usepackage{svg}
\usetikzlibrary{graphdrawing.trees}
\usepackage{overpic}

\definecolor{c1} {rgb}{0,0,0}
\definecolor{c2} {rgb}{0.1216,0.4706,0.7059}
\definecolor{c3} {rgb}{0.2000,0.6275,0.1725}
\definecolor{c4} {rgb}{0.9843,0.6039,0.6000}
\definecolor{c5} {rgb}{0.8902,0.1020,0.1098}
\definecolor{c6} {rgb}{0.9922,0.7490,0.4353}
\definecolor{c7} {rgb}{1.0000,0.4980,     0}
\definecolor{c8} {rgb}{0.4157,0.2392,0.6039}
\definecolor{c9} {rgb}{0.6941,0.3490,0.1569}
\definecolor{c10}{rgb}{0.6510,0.8078,0.8902}
\definecolor{c11}{rgb}{0.6980,0.8745,0.5412}
\definecolor{c12}{rgb}{0.7922,0.6980,0.8392}
\definecolor{c12}{rgb}{1.0000,1.0000,0.6000}

\usepackage{minted}
\usemintedstyle{vs}
%\usemintedstyle{borland}

%\usemintedstyle{emacs}
%\usemintedstyle{perldoc}
%\usemintedstyle{friendly}
%%\usemintedstyle{pastie}
%%\usemintedstyle{vim}


\usepackage{fontspec}
\usepackage[nott]{inconsolata}

\usepackage{booktabs}

%<<< title, author, institute
  \title
  [What every programmer should know about \\ high performance computing]
  {What every programmer should know about \\ high performance computing}
  \subtitle{(instruction-level parallelism, memory)}
  \author[Dhairya Malhotra]{Codes: \url{https://github.com/dmalhotra/fwam2022} \\
  \phantom{.}\\
  Dhairya~Malhotra}

  %\institute{Flatiron Institute\\ \mbox{}  \\  \pgfuseimage{FIbig} }
  %\institute{\pgfuseimage{FIbig} }
  \institute{\Large $F_\omega(\alpha+m)!$}

  \date[]{Oct 28, 2022}
%>>>
%<<< packages
  \usepackage{tikz}
  \usetikzlibrary{fit,shapes.geometric,arrows,calc,shapes,decorations.pathreplacing,patterns}
  \usepackage{pgfplots,pgfplotstable}
  \pgfplotsset{compat=1.17}

  \usepackage{mathtools}
  \usepackage{multirow}
  \usepackage{multimedia}
  \usepackage{media9}
  %\usepackage{movie15} %(obsolete)
  \usepackage{animate}
  \usepackage{fp}
  %\usepackage{enumitem}
  \usepackage{bm}

  \beamertemplateballitem % Numbered bullets

  \usepackage{xstring}
  \usepackage{mathtools}% Loads amsmath

  \usepackage{stmaryrd}

  \newcommand{\vcenteredinclude}[1]{\begingroup\setbox0=\hbox{{#1}}\parbox{\wd0}{\box0}\endgroup}

  %%------------------------------------------------------------------------------
  %%- Latin-abbreviations
  %%------------------------------------------------------------------------------

  \usepackage{expl3}
  \ExplSyntaxOn
  \newcommand\latinabbrev[1]{
    \peek_meaning:NTF . {% Same as \@ifnextchar
      #1\@}%
    { \peek_catcode:NTF a {% Check whether next char has same catcode as \'a, i.e., is a letter
        #1.\@ }%
      {#1.\@}}}
  \ExplSyntaxOff

  %Omit final dot from each def.

  \def\eg{\latinabbrev{e.g}}
  \def\etal{\latinabbrev{et al}}
  \def\etc{\latinabbrev{etc}}
  \def\ie{\latinabbrev{i.e}}

%>>>


\begin{document}
  \setbeamercovered{transparent}% Dim out "inactive" elements

  \begin{frame}[t]%<<< Title
    \titlepage
  \end{frame}%>>>

  \input{intro}

  \input{ilp}

  \input{mem}

\end{document}





%%\input{openmp}


% Examples:
% Instruction level: polynomial evaluation, simple expressions (AXPY)
% Compute bound: GEMM
% Memory bound: AXPY, Gauss-Sidel / Gauss-Jacobi
% Latency bound: sorting

% Ideas to demonstrate:
% Hyper threading


% TOOLS:
% godbold
% https://quick-bench.com/
% Profiling: https://hpc-wiki.info/hpc/Compiler_Sanitizers
% Debugging: -fsanitize=address

% profile! profile! profile!
%omp_get_wtime() / MPI_Wtime()



% htop

%NUMA:
% numactl -H
% export OMP_PLACES="{0},{1},{2},{3}"
% numactl -l myBinary // local memory for each thread






% Distributed memory
% cost model
% load balancing
% minimizing communication












%false sharing, caching,




% GEMM cube volume and surface area




% Programming languages: https://hpc-wiki.info/hpc/Programming_Languages





% NUMA: https://hpc-wiki.info/hpc/Binding/Pinning
% export OMP_PROC_BIND=close/spread
% memory copy; OMP_NUM_THREADS=8
% non-temporal writes


% single thread can saturate memory bandwidth.
% do not optimize single-threaded, it may not reflect parallel performance.




% Diagnosing performance issues: https://hpc-wiki.info/hpc/Performance_Patterns






% Runtime profiling: https://hpc-wiki.info/hpc/Runtime_profiling