Skip to content

Commit

Permalink
results plots
Browse files Browse the repository at this point in the history
  • Loading branch information
@milindasf
milindasf committed Jul 31, 2023
1 parent 89b9d11 commit ec4110a
Show file tree
Hide file tree
Showing 8 changed files with 411 additions and 0 deletions.
307 changes: 307 additions & 0 deletions miniapps/samplesort/dat/all_ppop.txt
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,307 @@
==================
strong scaling 500M
==================

cupy only
gpus all lsort1 sp_comp scatter_map all2all lsort2
2 4.30421114E-01 1.15388136E-01 5.68857702E-04 1.60563430E-03 1.71892890E-01 1.40931080E-01
3 3.58948328E-01 7.79944195E-02 8.44727608E-04 2.05807459E-03 1.92571960E-01 8.54445202E-02
4 3.20432234E-01 5.93451607E-02 1.06588820E-03 2.50257170E-03 1.87090397E-01 7.03794646E-02

crosspy w thread
gpus all lsort1 sp_comp scatter_map all2all lsort2
2 3.49802462E-01 1.15576801E-01 8.04035703E-04,1.79865579E-03 1.13606636E-01 1.18005798E-01
3 3.02869053E-01 7.79574179E-02 1.09557731E-03,2.38543470E-03 1.41963770E-01 7.94578821E-02
4 2.52899075E-01 5.91583280E-02 1.49350250E-03,3.06496951E-03 1.28813135E-01 6.03589629E-02

crosspy w parla
gpus all lsort1 sp_comp scatter_map all2all lsort2
2 5.77890648E-01 1.26540012E-01 2.95795511E-03 1.20897191E-03 1.13726599E-01 3.13458907E-01
3 4.49403549E-01 9.18094030E-02 3.92947550E-03 1.85868589E-03 1.63366586E-01 1.66963117E-01
4 3.87961231E-01 9.08590368E-02 5.41076250E-03 2.70663939E-03 1.42294630E-01 1.26029552E-01

Namespace(n=500000000, gpu=1, warm_up=5, runs=10, check=0)
All -- 2.2915E-01
|Local sort 1 -- 2.2913E-01
2.29148551E-01,2.29129795E-01,

Namespace(n=500000000, gpu=2, warm_up=5, runs=10, check=0)
All -- 4.3042E-01
|Local sort 1 -- 1.1539E-01
|Splitter comp. -- 5.6886E-04
|Scatter map -- 1.6056E-03
|All to all -- 1.7189E-01
|Local Sort 2 -- 1.4093E-01
4.30421114E-01,1.15388136E-01,5.68857702E-04,1.60563430E-03,1.71892890E-01,1.40931080E-01,

Namespace(n=500000000, gpu=3, warm_up=5, runs=10, check=0)
All -- 3.5895E-01
|Local sort 1 -- 7.7994E-02
|Splitter comp. -- 8.4473E-04
|Scatter map -- 2.0581E-03
|All to all -- 1.9257E-01
|Local Sort 2 -- 8.5445E-02
3.58948328E-01,7.79944195E-02,8.44727608E-04,2.05807459E-03,1.92571960E-01,8.54445202E-02,

Namespace(n=500000000, gpu=4, warm_up=5, runs=10, check=0)
All -- 3.2043E-01
|Local sort 1 -- 5.9345E-02
|Splitter comp. -- 1.0659E-03
|Scatter map -- 2.5026E-03
|All to all -- 1.8709E-01
|Local Sort 2 -- 7.0379E-02
3.20432234E-01,5.93451607E-02,1.06588820E-03,2.50257170E-03,1.87090397E-01,7.03794646E-02,

crosspy w thread
2,3.49802462E-01,1.15576801E-01,8.04035703E-04,1.79865579E-03,1.13606636E-01,1.18005798E-01,
3,3.02869053E-01,7.79574179E-02,1.09557731E-03,2.38543470E-03,1.41963770E-01,7.94578821E-02
4,2.52899075E-01,5.91583280E-02,1.49350250E-03,3.06496951E-03,1.28813135E-01,6.03589629E-02

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=500000000, gpu=1, warm_up=10, runs=10, mode='crosspy', check=0)
All -- 2.3002E-01
|Local sort 1 -- 2.3002E-01
2.30022741E-01,2.30018934E-01,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=500000000, gpu=2, warm_up=10, runs=10, mode='crosspy', check=0)
All -- 3.4980E-01
|Local sort 1 -- 1.1558E-01
|Splitter comp. -- 8.0404E-04
|Scatter map -- 1.7987E-03
|All to all -- 1.1361E-01
|Local Sort 2 -- 1.1801E-01
3.49802462E-01,1.15576801E-01,8.04035703E-04,1.79865579E-03,1.13606636E-01,1.18005798E-01,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=500000000, gpu=3, warm_up=10, runs=10, mode='crosspy', check=0)
All -- 3.0287E-01
|Local sort 1 -- 7.7957E-02
|Splitter comp. -- 1.0956E-03
|Scatter map -- 2.3854E-03
|All to all -- 1.4196E-01
|Local Sort 2 -- 7.9458E-02
3.02869053E-01,7.79574179E-02,1.09557731E-03,2.38543470E-03,1.41963770E-01,7.94578821E-02,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=500000000, gpu=4, warm_up=10, runs=10, mode='crosspy', check=0)
All -- 2.5290E-01
|Local sort 1 -- 5.9158E-02
|Splitter comp. -- 1.4935E-03
|Scatter map -- 3.0650E-03
|All to all -- 1.2881E-01
|Local Sort 2 -- 6.0359E-02
2.52899075E-01,5.91583280E-02,1.49350250E-03,3.06496951E-03,1.28813135E-01,6.03589629E-02,

crosspy w parla
2,5.77890648E-01,1.26540012E-01,2.95795511E-03,1.20897191E-03,1.13726599E-01,3.13458907E-01
3,4.49403549E-01,9.18094030E-02,3.92947550E-03,1.85868589E-03,1.63366586E-01,1.66963117E-01
4,3.87961231E-01,9.08590368E-02,5.41076250E-03,2.70663939E-03,1.42294630E-01,1.26029552E-01

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=500000000, gpu=1, warm_up=10, runs=10, mode='parla', check=0)
All -- 3.5456E-01
|Local sort 1 -- 3.3340E-01
3.54564987E-01,3.33402366E-01,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=500000000, gpu=2, warm_up=10, runs=10, mode='parla', check=0)
All -- 5.7789E-01
|Local sort 1 -- 1.2654E-01
|Splitter comp. -- 2.9580E-03
|Scatter map -- 1.2090E-03
|All to all -- 1.1373E-01
|Local Sort 2 -- 3.1346E-01
5.77890648E-01,1.26540012E-01,2.95795511E-03,1.20897191E-03,1.13726599E-01,3.13458907E-01,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=500000000, gpu=3, warm_up=10, runs=10, mode='parla', check=0)
All -- 4.4940E-01
|Local sort 1 -- 9.1809E-02
|Splitter comp. -- 3.9295E-03
|Scatter map -- 1.8587E-03
|All to all -- 1.6337E-01
|Local Sort 2 -- 1.6696E-01
4.49403549E-01,9.18094030E-02,3.92947550E-03,1.85868589E-03,1.63366586E-01,1.66963117E-01,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=500000000, gpu=4, warm_up=10, runs=10, mode='parla', check=0)
All -- 3.8796E-01
|Local sort 1 -- 9.0859E-02
|Splitter comp. -- 5.4108E-03
|Scatter map -- 2.7066E-03
|All to all -- 1.4229E-01
|Local Sort 2 -- 1.2603E-01
3.87961231E-01,9.08590368E-02,5.41076250E-03,2.70663939E-03,1.42294630E-01,1.26029552E-01,

cupy only
Namespace(n=100000000, gpu=1, warm_up=10, runs=10, check=0)
All -- 4.7102E-02
|Local sort 1 -- 4.7078E-02
4.71015733E-02,4.70779705E-02,

Namespace(n=200000000, gpu=2, warm_up=10, runs=10, check=0)
All -- 1.4722E-01
|Local sort 1 -- 4.6989E-02
|Splitter comp. -- 6.0652E-04
|Scatter map -- 1.6897E-03
|All to all -- 4.9992E-02
|Local Sort 2 -- 4.7902E-02
1.47215191E-01,4.69888416E-02,6.06516900E-04,1.68972869E-03,4.99917173E-02,4.79017944E-02,

Namespace(n=300000000, gpu=3, warm_up=10, runs=10, check=0)
All -- 2.1428E-01
|Local sort 1 -- 4.7049E-02
|Splitter comp. -- 8.0220E-04
|Scatter map -- 2.0283E-03
|All to all -- 1.1491E-01
|Local Sort 2 -- 4.9459E-02
2.14276175E-01,4.70493626E-02,8.02202406E-04,2.02832319E-03,1.14905252E-01,4.94592368E-02,

Namespace(n=400000000, gpu=4, warm_up=10, runs=10, check=0)
All -- 2.9637E-01
|Local sort 1 -- 4.8819E-02
|Splitter comp. -- 1.0439E-03
|Scatter map -- 2.4726E-03
|All to all -- 1.9257E-01
|Local Sort 2 -- 5.1418E-02
2.96371046E-01,4.88189334E-02,1.04394690E-03,2.47261780E-03,1.92572438E-01,5.14181675E-02,


===============
Weak scaling 100M grain sz
===============
crosspy w thread
gpus all lsort1 sp_comp scatter_map all2all lsort2
2 1.44663454E-01 4.71876461E-02 7.70231290E-04 1.83267320E-03 4.66794407E-02 4.81850781E-02
3 1.86291886E-01 4.72853744E-02 1.10351570E-03 2.55180609E-03 8.69384002E-02 4.84042116E-02
4 2.04527210E-01 4.75241148E-02 1.43491250E-03 3.07298410E-03 1.03705396E-01 4.87815639E-02

crosspy w parla
gpus all lsort1 sp_comp scatter_map all2all lsort2
2 2.00014174E-01 6.33237771E-02 2.86097311E-03 1.26846921E-03 4.67461797E-02 6.56800670E-02
3 2.99150044E-01 8.18834063E-02 3.84188270E-03 1.82645641E-03 1.26442941E-01 6.49946585E-02
4 3.29275617E-01 7.85926338E-02 5.42359711E-03 2.83029869E-03 1.35809122E-01 8.43319376E-02

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=100000000, gpu=1, warm_up=10, runs=10, mode='crosspy', check=0)
All -- 4.7240E-02
|Local sort 1 -- 4.7237E-02
4.72403154E-02,4.72366318E-02,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=200000000, gpu=2, warm_up=10, runs=10, mode='crosspy', check=0)
All -- 1.4466E-01
|Local sort 1 -- 4.7188E-02
|Splitter comp. -- 7.7023E-04
|Scatter map -- 1.8327E-03
|All to all -- 4.6679E-02
|Local Sort 2 -- 4.8185E-02
1.44663454E-01,4.71876461E-02,7.70231290E-04,1.83267320E-03,4.66794407E-02,4.81850781E-02,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=300000000, gpu=3, warm_up=10, runs=10, mode='crosspy', check=0)
All -- 1.8629E-01
|Local sort 1 -- 4.7285E-02
|Splitter comp. -- 1.1035E-03
|Scatter map -- 2.5518E-03
|All to all -- 8.6938E-02
|Local Sort 2 -- 4.8404E-02
1.86291886E-01,4.72853744E-02,1.10351570E-03,2.55180609E-03,8.69384002E-02,4.84042116E-02,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=400000000, gpu=4, warm_up=10, runs=10, mode='crosspy', check=0)
All -- 2.0453E-01
|Local sort 1 -- 4.7524E-02
|Splitter comp. -- 1.4349E-03
|Scatter map -- 3.0730E-03
|All to all -- 1.0371E-01
|Local Sort 2 -- 4.8782E-02
2.04527210E-01,4.75241148E-02,1.43491250E-03,3.07298410E-03,1.03705396E-01,4.87815639E-02,

crosspy w parla
USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=100000000, gpu=1, warm_up=10, runs=10, mode='parla', check=0)
All -- 7.6050E-02
|Local sort 1 -- 5.5860E-02
7.60504303E-02,5.58600824E-02,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=200000000, gpu=2, warm_up=10, runs=10, mode='parla', check=0)
All -- 2.0001E-01
|Local sort 1 -- 6.3324E-02
|Splitter comp. -- 2.8610E-03
|Scatter map -- 1.2685E-03
|All to all -- 4.6746E-02
|Local Sort 2 -- 6.5680E-02
2.00014174E-01,6.33237771E-02,2.86097311E-03,1.26846921E-03,4.67461797E-02,6.56800670E-02,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=300000000, gpu=3, warm_up=10, runs=10, mode='parla', check=0)
All -- 2.9915E-01
|Local sort 1 -- 8.1883E-02
|Splitter comp. -- 3.8419E-03
|Scatter map -- 1.8265E-03
|All to all -- 1.2644E-01
|Local Sort 2 -- 6.4995E-02
2.99150044E-01,8.18834063E-02,3.84188270E-03,1.82645641E-03,1.26442941E-01,6.49946585E-02,

USE_PYTHON_RUNAHEAD: True
CUPY_ENABLED: True
PREINIT_THREADS: True
DEFAULT SYNC: 0
Namespace(n=400000000, gpu=4, warm_up=10, runs=10, mode='parla', check=0)
All -- 3.2928E-01
|Local sort 1 -- 7.8593E-02
|Splitter comp. -- 5.4236E-03
|Scatter map -- 2.8303E-03
|All to all -- 1.3581E-01
|Local Sort 2 -- 8.4332E-02
3.29275617E-01,7.85926338E-02,5.42359711E-03,2.83029869E-03,1.35809122E-01,8.43319376E-02,


42 changes: 42 additions & 0 deletions miniapps/samplesort/dat/plot_ss.tex
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,42 @@
\documentclass[preview]{standalone}
\usepackage{tikz,pgfplots}
\usepackage{pgfplotstable}
\usetikzlibrary{patterns}

\makeatletter
\newcommand\resetstackedplotsxa{
\makeatletter
\pgfplots@stacked@isfirstplottrue
\makeatother
\addplot [forget plot,draw=none] coordinates{(2, 0) (3, 0) (4, 0)};
}
\makeatother
\begin{document}
\begin{figure}
\centering
\begin{tikzpicture}
\tikzstyle{every node}=[font=\footnotesize]
\begin{axis}[
ybar stacked, bar width=0.35cm,
xlabel={number of GPUs $\rightarrow$},
ylabel={time (s) $\rightarrow$ },symbolic x coords={2, 3, 4},width=12cm,height=7cm,
xtick = data,
legend style={text=black, at={(0.48,1.3)}, anchor=north},legend columns=3,grid=major]
\addplot [fill=red!50, fill opacity=0.5] [bar shift=0.20cm] table[x={gpus}, y expr = \thisrow{lsort1} + \thisrow{lsort2}] {ss_crosspy.txt};
\addplot [fill=blue!50, fill opacity=0.5] [bar shift=0.20cm] table[x={gpus}, y expr = \thisrow{sp_comp} + \thisrow{scatter_map}] {ss_crosspy.txt};
\addplot [fill=orange!50, fill opacity=0.5] [bar shift=0.20cm] table[x={gpus}, y = {all2all}] {ss_crosspy.txt};

\resetstackedplotsxa

\addplot [fill=red!50, fill opacity=0.5, postaction={pattern=north east lines}] [bar shift=-0.20cm] table[x={gpus}, y expr = \thisrow{lsort1} + \thisrow{lsort1}] {ss_parla.txt};
\addplot [fill=blue!50, fill opacity=0.5, postaction={pattern=north east lines}] [bar shift=-0.20cm] table[x={gpus}, y expr = \thisrow{sp_comp} + \thisrow{scatter_map}] {ss_parla.txt};
\addplot [fill=orange!50, fill opacity=0.5, postaction={pattern=north east lines}] [bar shift=-0.20cm] table[x={gpus}, y = {all2all}] {ss_parla.txt};
\legend{local sort (xp + Threads), splitters (xp + Threads), alltoall(xp + Threads), local sort (xp + Parla), splitters (xp + Parla), alltoall(xp + Parla)};
\end{axis}
\end{tikzpicture}
\caption{Strong scaling with global problem size of 500M}
\end{figure}

\end{document}


42 changes: 42 additions & 0 deletions miniapps/samplesort/dat/plot_ws.tex
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,42 @@
\documentclass[preview]{standalone}
\usepackage{tikz,pgfplots}
\usepackage{pgfplotstable}
\usetikzlibrary{patterns}

\makeatletter
\newcommand\resetstackedplotsxa{
\makeatletter
\pgfplots@stacked@isfirstplottrue
\makeatother
\addplot [forget plot,draw=none] coordinates{(2, 0) (3, 0) (4, 0)};
}
\makeatother
\begin{document}
\begin{figure}
\centering
\begin{tikzpicture}
\tikzstyle{every node}=[font=\footnotesize]
\begin{axis}[
ybar stacked, bar width=0.35cm,
xlabel={number of GPUs $\rightarrow$},
ylabel={time (s) $\rightarrow$ },symbolic x coords={2, 3, 4},width=12cm,height=7cm,
xtick = data,
legend style={text=black, at={(0.48,1.3)}, anchor=north},legend columns=3,grid=major]
\addplot [fill=red!50, fill opacity=0.5] [bar shift=0.20cm] table[x={gpus}, y expr = \thisrow{lsort1} + \thisrow{lsort2}] {ws_crosspy.txt};
\addplot [fill=blue!50, fill opacity=0.5] [bar shift=0.20cm] table[x={gpus}, y expr = \thisrow{sp_comp} + \thisrow{scatter_map}] {ws_crosspy.txt};
\addplot [fill=orange!50, fill opacity=0.5] [bar shift=0.20cm] table[x={gpus}, y = {all2all}] {ws_crosspy.txt};

\resetstackedplotsxa

\addplot [fill=red!50, fill opacity=0.5, postaction={pattern=north east lines}] [bar shift=-0.20cm] table[x={gpus}, y expr = \thisrow{lsort1} + \thisrow{lsort1}] {ws_parla.txt};
\addplot [fill=blue!50, fill opacity=0.5, postaction={pattern=north east lines}] [bar shift=-0.20cm] table[x={gpus}, y expr = \thisrow{sp_comp} + \thisrow{scatter_map}] {ws_parla.txt};
\addplot [fill=orange!50, fill opacity=0.5, postaction={pattern=north east lines}] [bar shift=-0.20cm] table[x={gpus}, y = {all2all}] {ws_parla.txt};
\legend{local sort (xp + Threads), splitters (xp + Threads), alltoall(xp + Threads), local sort (xp + Parla), splitters (xp + Parla), alltoall(xp + Parla)};
\end{axis}
\end{tikzpicture}
\caption{weak scaling with 100M array entries per GPU}
\end{figure}

\end{document}


4 changes: 4 additions & 0 deletions miniapps/samplesort/dat/ss_crosspy.txt
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,4 @@
gpus all lsort1 sp_comp scatter_map all2all lsort2
2 3.49802462E-01 1.15576801E-01 8.04035703E-04 1.79865579E-03 1.13606636E-01 1.18005798E-01
3 3.02869053E-01 7.79574179E-02 1.09557731E-03 2.38543470E-03 1.41963770E-01 7.94578821E-02
4 2.52899075E-01 5.91583280E-02 1.49350250E-03 3.06496951E-03 1.28813135E-01 6.03589629E-02
Loading

0 comments on commit ec4110a

Please sign in to comment.