[complement-svn] SF.net SVN: complement: [1329] trunk/explore/perf/STL
Status: Pre-Alpha
Brought to you by:
complement
Menu
▾
▴
[complement-svn] SF.net SVN: complement: [1329] trunk/explore/perf/STL
|
From: <com...@us...> - 2006-10-16 21:05:28
|
Revision: 1329
http://svn.sourceforge.net/complement/?rev=1329&view=rev
Author: complement
Date: 2006-10-16 14:05:09 -0700 (Mon, 16 Oct 2006)
Log Message:
-----------
test with malloc in add-proxy added
Modified Paths:
--------------
trunk/explore/perf/STL/compare.tex
trunk/explore/perf/STL/string/add-proxy/Makefile
trunk/explore/perf/STL/string/test.sh
trunk/explore/perf/STL/string-MT/add-proxy/Makefile
trunk/explore/perf/STL/string-MT/test.sh
trunk/explore/perf/STL/strings.gplot
Added Paths:
-----------
trunk/explore/perf/STL/string/add-proxy/STLport-malloc/
trunk/explore/perf/STL/string/add-proxy/STLport-malloc/Makefile
trunk/explore/perf/STL/string-MT/add-proxy/STLport-malloc/
trunk/explore/perf/STL/string-MT/add-proxy/STLport-malloc/Makefile
Property Changed:
----------------
trunk/explore/perf/STL/
Property changes on: trunk/explore/perf/STL
___________________________________________________________________
Name: svn:ignore
- *.pcl
compare.dvi
compare.aux
compare.toc
compare.log
compare.ps
compare.pdf
alloc.eps
alloc.tex
strings-MT-t-libstd.eps
strings-MT-t-libstd.tex
strings-MT-t-stlp-m.eps
strings-MT-t-stlp-m.tex
strings-MT-t-stlp-n.eps
strings-MT-t-stlp-n.tex
strings-MT-t-stlp-r.eps
strings-MT-t-stlp-r.tex
str-rope.eps
str-rope.tex
str-rope-1.eps
str-rope-1.tex
strings-MT.eps
strings-MT.tex
strings-MT2.eps
strings-MT2.tex
strings.eps
strings.tex
streams.eps
streams.tex
str-rope-2.eps
str-rope-2.tex
intel-amd.eps
intel-amd.tex
build
+ *.pcl
compare.dvi
compare.aux
compare.toc
compare.log
compare.ps
compare.pdf
alloc.eps
alloc.tex
alloc-2_2_5.eps
alloc-2_2_5.tex
strings-MT-t-libstd.eps
strings-MT-t-libstd.tex
strings-MT-t-stlp-m.eps
strings-MT-t-stlp-m.tex
strings-MT-t-stlp-n.eps
strings-MT-t-stlp-n.tex
strings-MT-t-stlp-r.eps
strings-MT-t-stlp-r.tex
str-rope.eps
str-rope.tex
str-rope-1.eps
str-rope-1.tex
strings-MT.eps
strings-MT.tex
strings-MT2.eps
strings-MT2.tex
strings.eps
strings.tex
streams.eps
streams.tex
streams-1.eps
streams-1.tex
str-rope-2.eps
str-rope-2.tex
intel-amd.eps
intel-amd.tex
build
Modified: trunk/explore/perf/STL/compare.tex
===================================================================
--- trunk/explore/perf/STL/compare.tex 2006-10-16 20:16:12 UTC (rev 1328)
+++ trunk/explore/perf/STL/compare.tex 2006-10-16 21:05:09 UTC (rev 1329)
@@ -1,4 +1,4 @@
-% -*- LaTeX -*- Time-stamp: <06/10/16 19:00:51 ptr>
+% -*- LaTeX -*- Time-stamp: <06/10/17 00:43:24 ptr>
\documentclass[a4paper]{article}
@@ -49,12 +49,13 @@
In the tests was used following computers and operational envirinments:
\begin{enumerate}
\item Tyan Tiger motherboard, two $1.33$-GHz
- AMD$^{\mbox{\tiny\textregistered}}$ Athlon$^{\mbox{\tiny\textregistered}}$ XP 1500+ processors under Linux (kernel 2.6.16.26, \verb|glibc| 2{.}3{.}6);\label{AMD}
+ AMD$^{\mbox{\tiny\textregistered}}$ Athlon$^{\mbox{\tiny TM}}$ XP 1500+ processors under Linux (kernel 2{.}6{.}16{.}26, \verb|glibc| 2{.}3{.}6);\label{AMD}
+ \item Two $2.6$-GHz AMD$^{\mbox{\tiny\textregistered}}$ Opteron$^{\mbox{\tiny TM}}$ 252 processors under Linux (kernel 2{.}6{.}15, x86\_64, \verb|glibc| 2{.}3{.}6)\label{AMD64}
\end{enumerate}
\section{Compilers}
-In tests was used GNU gcc 3{.}4{.}4 or 4{.}1{.}1
+In tests was used GNU gcc 3{.}4{.}4 (on AMD64) or 4{.}1{.}1 (on AMD)
with appropriate
\verb|libstdc++| libraries (version~3).
@@ -137,7 +138,7 @@
\input strings.tex
\end{center}
\caption{STL strings implementations comparison. Single thread. $t$ is a wall time
- for tests.
+ for tests. On X-axis you see test's number (see text).
The implementation of strings in \STLport{} (non-COW)
drammaticaly faster than COW implementations (\libstd{} and
ropes in \STLport{}) in tests that modify strings.
@@ -232,6 +233,10 @@
For \libstd{} both \#8 and \#9 are the same test.
+The figure~\ref{STLport-gpp-strings} show that \STLport{} implementation
+give better results even without expression templates. But with
+expression templates the results are better more then twice.
+
\section{Role of Allocators}
Note: in the \verb|glibc| 2{.}3{.}6 enhancement of memory allocation
@@ -258,6 +263,14 @@
Wall time for tests. \label{STLport-alloc-strings}}
\end{figure}
+\begin{figure}
+ \begin{center}
+ \input alloc-2_2_5.tex
+ \end{center}
+ \caption{Role of memory allocator in strings implementation (\STLport), \texttt{glibc} 2{.}2{.}5.
+ Wall time for tests. \label{STLport-alloc-ii-ii-v-strings}}
+\end{figure}
+
Let's repeat tests from section~\ref{TestsDescr} for \STLport{} with different
allocators. We see (fig.~\ref{STLport-alloc-strings}) that all allocators
are good enough
@@ -266,10 +279,14 @@
%operations (test~1) the ``optimized'' \verb|node_alloc| has a tiny advantage
%(difference between best and worst allocators is $3\%$).
For search operations (test~2) the results as expected
-are the same (within measure of inaccuracy). The tests 4 and 5 show
+are the same (within measure of inaccuracy).
+
+The tests 4 and 5 show
that cost of \verb|malloc| system call---\verb|node_alloc| reuse once
allocated memory. But copy of string's content still present (compare with time of
-\libstd{} on fig.~\ref{STLport-gpp-strings}).
+\libstd{} on fig.~\ref{STLport-gpp-strings}). In the past (\verb|glibc| 2{.}2{.}5)
+the cost of \verb|malloc| was significant, but now it near zero (compare graphics
+\ref{STLport-alloc-strings} and \ref{STLport-alloc-ii-ii-v-strings}).
\section{Strings in Multithreaded Environment}
@@ -289,7 +306,7 @@
\input strings-MT.tex
\end{center}
\caption{Tests wall time (\libstd{} and \STLport{}
- strings) in MT environment (two threads).
+ strings) in multi-threaded environment (two threads).
\label{STLport-gpp-MT-strings}}
\end{figure}
@@ -298,7 +315,7 @@
\input strings-MT2.tex
\end{center}
\caption{Tests wall time (\libstd{} and \STLport{}
- strings) in MT environment (two threads) (same as fig.~\ref{STLport-gpp-MT-strings}, another scale).
+ strings) in multi-threaded environment (two threads) (same as fig.~\ref{STLport-gpp-MT-strings}, another scale).
\label{STLport-gpp-MT-strings2}}
\end{figure}
@@ -318,7 +335,7 @@
\end{center}
\caption{
Wall time, user time and system time per thread for
- \libstd{} in MT environment.
+ \libstd{} in multi-threaded environment.
\label{gpp-MT-strings-t}}
\end{figure}
@@ -342,7 +359,7 @@
\end{center}
\caption{
Wall time, user time and system time per thread for
- \STLport{} (node allocator) in MT environment.
+ \STLport{} (node allocator) in multi-threaded environment.
\label{STLport-MT-strings-t}}
\end{figure}
@@ -352,7 +369,7 @@
\end{center}
\caption{
Wall time, user time and system time per thread for
- \STLport{} (malloc allocator) in MT environment.
+ \STLport{} (malloc allocator) in multi-threaded environment.
\label{STLport-MT-strings-m}}
\end{figure}
@@ -362,7 +379,7 @@
\end{center}
\caption{
Wall time, user time and system time per thread for
- \STLport{} ropes in MT environment.
+ \STLport{} ropes in multi-threaded environment.
\label{STLport-MT-strings-r}}
\end{figure}
@@ -371,14 +388,14 @@
\texttt{node\_alloc}-based,
figure~\ref{STLport-MT-strings-m} show profile for tests with \texttt{malloc}-based allocators, and figure~\ref{STLport-MT-strings-r} for ropes.
-The surprise for me was that \texttt{malloc}-based variant
+The surprise for me was that \texttt{malloc}-based variants
win (in $1.5--2$ times faster) over \texttt{node\_alloc}-based variant in tests \#3,
\#4 and \#5.
This fact can be explained by usage of memory allocated chunks vector
in \texttt{node\_alloc}. This vector accessed from different threads
and such access should use thread synchronization primitives.
Compare figures~\ref{STLport-alloc-strings} and \ref{STLport-gpp-MT-strings}.
-The tests~1--2 has the same time in MT and non-MT context for
+The tests~1--2 has the same time in multi-threaded and single-threaded context for
\texttt{malloc}-based and \texttt{node\_alloc}-based allocators.
In test~3 \texttt{malloc}-based variant climb down, while
\texttt{node\_alloc}-based keep position.
@@ -516,7 +533,7 @@
significant better than C \verb|fwrite|). With raw write to string stream
(or in char buffer in case of C) the pure C variant is 3--4 times faster.
-As about Intel$^{\mbox{\tiny\textregistered}}$ vs. AMD$^{\mbox{\tiny\textregistered}}$, no comments---see figure~\ref{intel-amd} on page~\pageref{intel-amd}.
+% As about Intel$^{\mbox{\tiny\textregistered}}$ vs. AMD$^{\mbox{\tiny\textregistered}}$, no comments---see figure~\ref{intel-amd} on page~\pageref{intel-amd}.
\section{References}
Modified: trunk/explore/perf/STL/string/add-proxy/Makefile
===================================================================
--- trunk/explore/perf/STL/string/add-proxy/Makefile 2006-10-16 20:16:12 UTC (rev 1328)
+++ trunk/explore/perf/STL/string/add-proxy/Makefile 2006-10-16 21:05:09 UTC (rev 1329)
@@ -1,6 +1,6 @@
# -*- Makefile -*- Time-stamp: <05/12/27 00:39:55 ptr>
-DIRS = STLport-default STLport-add-str-proxy libstdc++
+DIRS = STLport-default STLport-malloc STLport-add-str-proxy libstdc++
all:
for d in $(DIRS); do \
Added: trunk/explore/perf/STL/string/add-proxy/STLport-malloc/Makefile
===================================================================
--- trunk/explore/perf/STL/string/add-proxy/STLport-malloc/Makefile (rev 0)
+++ trunk/explore/perf/STL/string/add-proxy/STLport-malloc/Makefile 2006-10-16 21:05:09 UTC (rev 1329)
@@ -0,0 +1,14 @@
+# -*- Makefile -*- Time-stamp: <05/12/27 00:41:26 ptr>
+
+SRCROOT := ../../../../..
+COMPILER_NAME := gcc
+
+ALL_TAGS := release-shared
+
+
+include ../Makefile.inc
+include ${SRCROOT}/Makefiles/top.mak
+
+DEFS += -D_STLP_USE_MALLOC
+
+release-shared: LDSEARCH = -L../../../build/lib/malloc -Wl,-rpath=../build/lib/malloc
Modified: trunk/explore/perf/STL/string/test.sh
===================================================================
--- trunk/explore/perf/STL/string/test.sh 2006-10-16 20:16:12 UTC (rev 1328)
+++ trunk/explore/perf/STL/string/test.sh 2006-10-16 21:05:09 UTC (rev 1329)
@@ -1,6 +1,6 @@
#!/bin/sh
#
-# Time-stamp: <06/10/16 18:25:07 ptr>
+# Time-stamp: <06/10/17 00:57:02 ptr>
#
BASEDIR=${PWD}/../../..
@@ -65,6 +65,9 @@
echo $j `../stat.awk s.log` >> STLport-malloc.dat
let j=j+1
done
+let j=j+1
+runtest add-proxy/STLport-malloc
+echo $j `../stat.awk s.log` >> STLport-malloc.dat
rm -f STLport-newalloc.dat
touch STLport-newalloc.dat
Modified: trunk/explore/perf/STL/string-MT/add-proxy/Makefile
===================================================================
--- trunk/explore/perf/STL/string-MT/add-proxy/Makefile 2006-10-16 20:16:12 UTC (rev 1328)
+++ trunk/explore/perf/STL/string-MT/add-proxy/Makefile 2006-10-16 21:05:09 UTC (rev 1329)
@@ -1,6 +1,6 @@
# -*- Makefile -*- Time-stamp: <05/12/27 00:39:55 ptr>
-DIRS = STLport-default STLport-add-str-proxy libstdc++
+DIRS = STLport-default STLport-malloc STLport-add-str-proxy libstdc++
all:
for d in $(DIRS); do \
Added: trunk/explore/perf/STL/string-MT/add-proxy/STLport-malloc/Makefile
===================================================================
--- trunk/explore/perf/STL/string-MT/add-proxy/STLport-malloc/Makefile (rev 0)
+++ trunk/explore/perf/STL/string-MT/add-proxy/STLport-malloc/Makefile 2006-10-16 21:05:09 UTC (rev 1329)
@@ -0,0 +1,14 @@
+# -*- Makefile -*- Time-stamp: <05/12/27 00:41:26 ptr>
+
+SRCROOT := ../../../../..
+COMPILER_NAME := gcc
+
+ALL_TAGS := release-shared
+
+
+include ../Makefile.inc
+include ${SRCROOT}/Makefiles/top.mak
+
+DEFS += -D_STLP_USE_MALLOC
+
+release-shared: LDSEARCH = -L../../../build/lib/malloc -Wl,-rpath=../build/lib/malloc
Modified: trunk/explore/perf/STL/string-MT/test.sh
===================================================================
--- trunk/explore/perf/STL/string-MT/test.sh 2006-10-16 20:16:12 UTC (rev 1328)
+++ trunk/explore/perf/STL/string-MT/test.sh 2006-10-16 21:05:09 UTC (rev 1329)
@@ -1,6 +1,6 @@
#!/bin/sh
#
-# Time-stamp: <06/10/16 21:52:27 ptr>
+# Time-stamp: <06/10/17 01:00:43 ptr>
#
BASEDIR=${PWD}/../../..
@@ -65,6 +65,9 @@
echo $j `../stat.awk s.log` >> STLport-malloc.dat
let j=j+1
done
+let j=j+1
+runtest add-proxy/STLport-malloc
+echo $j `../stat.awk s.log` >> STLport-malloc.dat
rm -f STLport-newalloc.dat
touch STLport-newalloc.dat
Modified: trunk/explore/perf/STL/strings.gplot
===================================================================
--- trunk/explore/perf/STL/strings.gplot 2006-10-16 20:16:12 UTC (rev 1328)
+++ trunk/explore/perf/STL/strings.gplot 2006-10-16 21:05:09 UTC (rev 1329)
@@ -55,7 +55,7 @@
#pause(-1)
-set output "alloc-2.2.5.eps"
+set output "alloc-2_2_5.eps"
plot [0.9:9.1] 'string/STLport-2.2.5.dat' using ($1):($6) notitle w l lt 1, \
'string/STLport-2.2.5.dat' using ($1):($6):($7) title "node alloc" w yerrorbars lt 1, \
@@ -68,7 +68,7 @@
set output "strings-MT.eps"
-plot [0.9:9.1][0:250] 'string-MT/STLport.dat' using ($1):($6) notitle w l lt 1, \
+plot [0.9:9.1] 'string-MT/STLport.dat' using ($1):($6) notitle w l lt 1, \
'string-MT/STLport.dat' using ($1):($6):($7) title "node alloc" w yerrorbars lt 1, \
'string-MT/libstd++.dat' using ($1):($6) notitle w l lt 2, \
'string-MT/libstd++.dat' using ($1):($6):($7) title "libstd++" w yerrorbars lt 2, \
@@ -81,7 +81,7 @@
set output "strings-MT2.eps"
-plot [0.9:9.1][0:50] 'string-MT/STLport.dat' using ($1):($6) notitle w l lt 1, \
+plot [0.9:9.1][0:25] 'string-MT/STLport.dat' using ($1):($6) notitle w l lt 1, \
'string-MT/STLport.dat' using ($1):($6):($7) title "node alloc" w yerrorbars lt 1, \
'string-MT/libstd++.dat' using ($1):($6) notitle w l lt 2, \
'string-MT/libstd++.dat' using ($1):($6):($7) title "libstd++" w yerrorbars lt 2, \
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
