| Message ID | 1434711418-20429-1-git-send-email-aderumier@odiso.com |
|---|---|
| State | New |
| Headers | show |
On Fri, 06/19 12:56, Alexandre Derumier wrote: > This adds "--enable-jemalloc" and "--disable-jemalloc" to allow linking > to jemalloc memory allocator. > > We have already tcmalloc support, > but it seem to not working well with a lot of iothreads/disks. > > The main problem is that tcmalloc use a shared thread cache of 16MB > by default. > With more threads, this cache is shared, and some bad garbage collections > can occur if the cache is too low. > > It's possible to tcmalloc cache increase it with a env var: > TCMALLOC_MAX_TOTAL_THREAD_CACHE_BYTES=256MB > > With default 16MB, performances are really bad with more than 2 disks. > Increasing to 256MB, it's helping but still have problem with 16 disks/iothreads. > > Jemalloc don't have performance problem with default configuration. > > Here the benchmark results in iops of 1 qemu vm randread 4K iodepth=32, > with rbd block backend (librbd is doing a lot of memory allocation), > 1 iothread by disk > > glibc malloc > ------------ > > 1 disk 29052 > 2 disks 55878 > 4 disks 127899 > 8 disks 240566 > 15 disks 269976 > > jemalloc > -------- > > 1 disk 41278 > 2 disks 75781 > 4 disks 195351 > 8 disks 294241 > 15 disks 298199 > > tcmalloc 2.2.1 default 16M cache > -------------------------------- > > 1 disk 37911 > 2 disks 67698 > 4 disks 41076 > 8 disks 43312 > 15 disks 37569 > > tcmalloc : 256M cache > --------------------------- > > 1 disk 33914 > 2 disks 58839 > 4 disks 148205 > 8 disks 213298 > 15 disks 218383 > > Signed-off-by: Alexandre Derumier <aderumier@odiso.com> > --- > configure | 29 +++++++++++++++++++++++++++++ > 1 file changed, 29 insertions(+) > > diff --git a/configure b/configure > index 222694f..2fe1e05 100755 > --- a/configure > +++ b/configure > @@ -336,6 +336,7 @@ vhdx="" > quorum="" > numa="" > tcmalloc="no" > +jemalloc="no" > > # parse CC options first > for opt do > @@ -1147,6 +1148,10 @@ for opt do > ;; > --enable-tcmalloc) tcmalloc="yes" > ;; > + --disable-jemalloc) jemalloc="no" > + ;; > + --enable-jemalloc) jemalloc="yes" > + ;; > *) > echo "ERROR: unknown option $opt" > echo "Try '$0 --help' for more information" > @@ -1420,6 +1425,8 @@ Advanced options (experts only): > --enable-numa enable libnuma support > --disable-tcmalloc disable tcmalloc support > --enable-tcmalloc enable tcmalloc support > + --disable-jemalloc disable jemalloc support > + --enable-jemalloc enable jemalloc support > > NOTE: The object files are built at the place where configure is launched > EOF > @@ -3344,6 +3351,11 @@ EOF > fi > fi > > +if test "$tcmalloc" = "yes" && test "$jemalloc" = "yes" ; then > + echo "ERROR: tcmalloc && jemalloc can't be used at the same time" > + exit 1 > +fi > + > ########################################## > # tcmalloc probe > > @@ -3361,6 +3373,22 @@ EOF > fi > > ########################################## > +# jemalloc probe > + > +if test "$jemalloc" = "yes" ; then > + cat > $TMPC << EOF > +#include <stdlib.h> > +int main(void) { malloc(1); return 0; } > +EOF > + > + if compile_prog "" "-ljemalloc" ; then > + LIBS="-ljemalloc $LIBS" > + else > + feature_not_found "jemalloc" "install jemalloc devel" > + fi > +fi > + > +########################################## > # signalfd probe > signalfd="no" > cat > $TMPC << EOF > @@ -4499,6 +4527,7 @@ echo "snappy support $snappy" > echo "bzip2 support $bzip2" > echo "NUMA host support $numa" > echo "tcmalloc support $tcmalloc" > +echo "jemalloc support $jemalloc" > > if test "$sdl_too_old" = "yes"; then > echo "-> Your SDL version is too old - please upgrade to have SDL support" > -- > 2.1.4 > > Reviewed-by: Fam Zheng <famz@redhat.com>
On 19/06/2015 12:56, Alexandre Derumier wrote: > This adds "--enable-jemalloc" and "--disable-jemalloc" to allow linking > to jemalloc memory allocator. > > We have already tcmalloc support, > but it seem to not working well with a lot of iothreads/disks. > > The main problem is that tcmalloc use a shared thread cache of 16MB > by default. > With more threads, this cache is shared, and some bad garbage collections > can occur if the cache is too low. > > It's possible to tcmalloc cache increase it with a env var: > TCMALLOC_MAX_TOTAL_THREAD_CACHE_BYTES=256MB It's also possible to do MallocExtension_SetNumericProperty("tcmalloc.max_total_thread_cache_bytes", num_io_threads << 24); What is the peak memory usage of jemalloc and tcmalloc? Paolo > With default 16MB, performances are really bad with more than 2 disks. > Increasing to 256MB, it's helping but still have problem with 16 disks/iothreads. > > Jemalloc don't have performance problem with default configuration. > > Here the benchmark results in iops of 1 qemu vm randread 4K iodepth=32, > with rbd block backend (librbd is doing a lot of memory allocation), > 1 iothread by disk > > glibc malloc > ------------ > > 1 disk 29052 > 2 disks 55878 > 4 disks 127899 > 8 disks 240566 > 15 disks 269976 > > jemalloc > -------- > > 1 disk 41278 > 2 disks 75781 > 4 disks 195351 > 8 disks 294241 > 15 disks 298199 > > tcmalloc 2.2.1 default 16M cache > -------------------------------- > > 1 disk 37911 > 2 disks 67698 > 4 disks 41076 > 8 disks 43312 > 15 disks 37569 > > tcmalloc : 256M cache > --------------------------- > > 1 disk 33914 > 2 disks 58839 > 4 disks 148205 > 8 disks 213298 > 15 disks 218383 > > Signed-off-by: Alexandre Derumier <aderumier@odiso.com> > --- > configure | 29 +++++++++++++++++++++++++++++ > 1 file changed, 29 insertions(+) > > diff --git a/configure b/configure > index 222694f..2fe1e05 100755 > --- a/configure > +++ b/configure > @@ -336,6 +336,7 @@ vhdx="" > quorum="" > numa="" > tcmalloc="no" > +jemalloc="no" > > # parse CC options first > for opt do > @@ -1147,6 +1148,10 @@ for opt do > ;; > --enable-tcmalloc) tcmalloc="yes" > ;; > + --disable-jemalloc) jemalloc="no" > + ;; > + --enable-jemalloc) jemalloc="yes" > + ;; > *) > echo "ERROR: unknown option $opt" > echo "Try '$0 --help' for more information" > @@ -1420,6 +1425,8 @@ Advanced options (experts only): > --enable-numa enable libnuma support > --disable-tcmalloc disable tcmalloc support > --enable-tcmalloc enable tcmalloc support > + --disable-jemalloc disable jemalloc support > + --enable-jemalloc enable jemalloc support > > NOTE: The object files are built at the place where configure is launched > EOF > @@ -3344,6 +3351,11 @@ EOF > fi > fi > > +if test "$tcmalloc" = "yes" && test "$jemalloc" = "yes" ; then > + echo "ERROR: tcmalloc && jemalloc can't be used at the same time" > + exit 1 > +fi > + > ########################################## > # tcmalloc probe > > @@ -3361,6 +3373,22 @@ EOF > fi > > ########################################## > +# jemalloc probe > + > +if test "$jemalloc" = "yes" ; then > + cat > $TMPC << EOF > +#include <stdlib.h> > +int main(void) { malloc(1); return 0; } > +EOF > + > + if compile_prog "" "-ljemalloc" ; then > + LIBS="-ljemalloc $LIBS" > + else > + feature_not_found "jemalloc" "install jemalloc devel" > + fi > +fi > + > +########################################## > # signalfd probe > signalfd="no" > cat > $TMPC << EOF > @@ -4499,6 +4527,7 @@ echo "snappy support $snappy" > echo "bzip2 support $bzip2" > echo "NUMA host support $numa" > echo "tcmalloc support $tcmalloc" > +echo "jemalloc support $jemalloc" > > if test "$sdl_too_old" = "yes"; then > echo "-> Your SDL version is too old - please upgrade to have SDL support" >
Hi Paolo, >>What is the peak memory usage of jemalloc and tcmalloc? I'll try to use an heap profiling to see. Don't known if "perf" can give me the info easily without profiling ? ----- Mail original ----- De: "pbonzini" <pbonzini@redhat.com> À: "aderumier" <aderumier@odiso.com>, "qemu-devel" <qemu-devel@nongnu.org> Envoyé: Mardi 23 Juin 2015 09:57:19 Objet: Re: [PATCH] configure: Add support for jemalloc On 19/06/2015 12:56, Alexandre Derumier wrote: > This adds "--enable-jemalloc" and "--disable-jemalloc" to allow linking > to jemalloc memory allocator. > > We have already tcmalloc support, > but it seem to not working well with a lot of iothreads/disks. > > The main problem is that tcmalloc use a shared thread cache of 16MB > by default. > With more threads, this cache is shared, and some bad garbage collections > can occur if the cache is too low. > > It's possible to tcmalloc cache increase it with a env var: > TCMALLOC_MAX_TOTAL_THREAD_CACHE_BYTES=256MB It's also possible to do MallocExtension_SetNumericProperty("tcmalloc.max_total_thread_cache_bytes", num_io_threads << 24); What is the peak memory usage of jemalloc and tcmalloc? Paolo > With default 16MB, performances are really bad with more than 2 disks. > Increasing to 256MB, it's helping but still have problem with 16 disks/iothreads. > > Jemalloc don't have performance problem with default configuration. > > Here the benchmark results in iops of 1 qemu vm randread 4K iodepth=32, > with rbd block backend (librbd is doing a lot of memory allocation), > 1 iothread by disk > > glibc malloc > ------------ > > 1 disk 29052 > 2 disks 55878 > 4 disks 127899 > 8 disks 240566 > 15 disks 269976 > > jemalloc > -------- > > 1 disk 41278 > 2 disks 75781 > 4 disks 195351 > 8 disks 294241 > 15 disks 298199 > > tcmalloc 2.2.1 default 16M cache > -------------------------------- > > 1 disk 37911 > 2 disks 67698 > 4 disks 41076 > 8 disks 43312 > 15 disks 37569 > > tcmalloc : 256M cache > --------------------------- > > 1 disk 33914 > 2 disks 58839 > 4 disks 148205 > 8 disks 213298 > 15 disks 218383 > > Signed-off-by: Alexandre Derumier <aderumier@odiso.com> > --- > configure | 29 +++++++++++++++++++++++++++++ > 1 file changed, 29 insertions(+) > > diff --git a/configure b/configure > index 222694f..2fe1e05 100755 > --- a/configure > +++ b/configure > @@ -336,6 +336,7 @@ vhdx="" > quorum="" > numa="" > tcmalloc="no" > +jemalloc="no" > > # parse CC options first > for opt do > @@ -1147,6 +1148,10 @@ for opt do > ;; > --enable-tcmalloc) tcmalloc="yes" > ;; > + --disable-jemalloc) jemalloc="no" > + ;; > + --enable-jemalloc) jemalloc="yes" > + ;; > *) > echo "ERROR: unknown option $opt" > echo "Try '$0 --help' for more information" > @@ -1420,6 +1425,8 @@ Advanced options (experts only): > --enable-numa enable libnuma support > --disable-tcmalloc disable tcmalloc support > --enable-tcmalloc enable tcmalloc support > + --disable-jemalloc disable jemalloc support > + --enable-jemalloc enable jemalloc support > > NOTE: The object files are built at the place where configure is launched > EOF > @@ -3344,6 +3351,11 @@ EOF > fi > fi > > +if test "$tcmalloc" = "yes" && test "$jemalloc" = "yes" ; then > + echo "ERROR: tcmalloc && jemalloc can't be used at the same time" > + exit 1 > +fi > + > ########################################## > # tcmalloc probe > > @@ -3361,6 +3373,22 @@ EOF > fi > > ########################################## > +# jemalloc probe > + > +if test "$jemalloc" = "yes" ; then > + cat > $TMPC << EOF > +#include <stdlib.h> > +int main(void) { malloc(1); return 0; } > +EOF > + > + if compile_prog "" "-ljemalloc" ; then > + LIBS="-ljemalloc $LIBS" > + else > + feature_not_found "jemalloc" "install jemalloc devel" > + fi > +fi > + > +########################################## > # signalfd probe > signalfd="no" > cat > $TMPC << EOF > @@ -4499,6 +4527,7 @@ echo "snappy support $snappy" > echo "bzip2 support $bzip2" > echo "NUMA host support $numa" > echo "tcmalloc support $tcmalloc" > +echo "jemalloc support $jemalloc" > > if test "$sdl_too_old" = "yes"; then > echo "-> Your SDL version is too old - please upgrade to have SDL support" >
On 23/06/2015 10:22, Alexandre DERUMIER wrote: > > What is the peak memory usage of jemalloc and tcmalloc? > > I'll try to use an heap profiling to see. > Don't known if "perf" can give me the info easily without profiling ? You can try using or modifying this systemtap script: https://sourceware.org/systemtap/examples/process/proctop.stp Paolo
>>You can try using or modifying this systemtap script: Thanks, I'll check this. If it can help, I have done a small top of qemu libc vs jemalloc vs tcmalloc libc-2.19.so ------------- started ------- PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 5294 root 20 0 26.953g 596744 18468 S 1.0 0.2 0:16.81 qemu fio running (200kiops) ---------------------- PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 5294 root 20 0 30.373g 633300 18552 S 2053 0.2 9:37.03 qemu jemalloc -------- started -------- PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 1789 root 20 0 11.640g 609224 18812 S 1.0 0.2 0:21.03 qemu fio running (300kiops) ---------------------- PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 1789 root 20 0 12.868g 684924 19144 S 2689 0.3 4:57.96 qemu tcmalloc (16MB) --------------- started ------- PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 1789 root 20 0 11.640g 609224 18812 S 1.0 0.2 0:21.03 qemu fio running (65k iops) ------------------- PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 6900 root 20 0 10.925g 643624 18432 S 1125 0.2 4:18.51 qemu tcmalloc (256MB) --------------- started ------- PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 10640 root 20 0 10.117g 584524 18236 S 0.7 0.2 0:16.28 qemu fio running (200k iops) ----------------------- PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 10640 root 20 0 10.983g 685136 18984 S 2566 0.3 12:34.73 qemu ----- Mail original ----- De: "pbonzini" <pbonzini@redhat.com> À: "aderumier" <aderumier@odiso.com> Cc: "qemu-devel" <qemu-devel@nongnu.org> Envoyé: Mardi 23 Juin 2015 10:57:03 Objet: Re: [PATCH] configure: Add support for jemalloc On 23/06/2015 10:22, Alexandre DERUMIER wrote: > > What is the peak memory usage of jemalloc and tcmalloc? > > I'll try to use an heap profiling to see. > Don't known if "perf" can give me the info easily without profiling ? You can try using or modifying this systemtap script: https://sourceware.org/systemtap/examples/process/proctop.stp Paolo
On 19/06/2015 12:56, Alexandre Derumier wrote: > This adds "--enable-jemalloc" and "--disable-jemalloc" to allow linking > to jemalloc memory allocator. > > We have already tcmalloc support, > but it seem to not working well with a lot of iothreads/disks. > > The main problem is that tcmalloc use a shared thread cache of 16MB > by default. > With more threads, this cache is shared, and some bad garbage collections > can occur if the cache is too low. > > It's possible to tcmalloc cache increase it with a env var: > TCMALLOC_MAX_TOTAL_THREAD_CACHE_BYTES=256MB > > With default 16MB, performances are really bad with more than 2 disks. > Increasing to 256MB, it's helping but still have problem with 16 disks/iothreads. > > Jemalloc don't have performance problem with default configuration. > > Here the benchmark results in iops of 1 qemu vm randread 4K iodepth=32, > with rbd block backend (librbd is doing a lot of memory allocation), > 1 iothread by disk > > glibc malloc > ------------ > > 1 disk 29052 > 2 disks 55878 > 4 disks 127899 > 8 disks 240566 > 15 disks 269976 > > jemalloc > -------- > > 1 disk 41278 > 2 disks 75781 > 4 disks 195351 > 8 disks 294241 > 15 disks 298199 > > tcmalloc 2.2.1 default 16M cache > -------------------------------- > > 1 disk 37911 > 2 disks 67698 > 4 disks 41076 > 8 disks 43312 > 15 disks 37569 > > tcmalloc : 256M cache > --------------------------- > > 1 disk 33914 > 2 disks 58839 > 4 disks 148205 > 8 disks 213298 > 15 disks 218383 > > Signed-off-by: Alexandre Derumier <aderumier@odiso.com> Rebased and applied, thanks. Paolo > --- > configure | 29 +++++++++++++++++++++++++++++ > 1 file changed, 29 insertions(+) > > diff --git a/configure b/configure > index 222694f..2fe1e05 100755 > --- a/configure > +++ b/configure > @@ -336,6 +336,7 @@ vhdx="" > quorum="" > numa="" > tcmalloc="no" > +jemalloc="no" > > # parse CC options first > for opt do > @@ -1147,6 +1148,10 @@ for opt do > ;; > --enable-tcmalloc) tcmalloc="yes" > ;; > + --disable-jemalloc) jemalloc="no" > + ;; > + --enable-jemalloc) jemalloc="yes" > + ;; > *) > echo "ERROR: unknown option $opt" > echo "Try '$0 --help' for more information" > @@ -1420,6 +1425,8 @@ Advanced options (experts only): > --enable-numa enable libnuma support > --disable-tcmalloc disable tcmalloc support > --enable-tcmalloc enable tcmalloc support > + --disable-jemalloc disable jemalloc support > + --enable-jemalloc enable jemalloc support > > NOTE: The object files are built at the place where configure is launched > EOF > @@ -3344,6 +3351,11 @@ EOF > fi > fi > > +if test "$tcmalloc" = "yes" && test "$jemalloc" = "yes" ; then > + echo "ERROR: tcmalloc && jemalloc can't be used at the same time" > + exit 1 > +fi > + > ########################################## > # tcmalloc probe > > @@ -3361,6 +3373,22 @@ EOF > fi > > ########################################## > +# jemalloc probe > + > +if test "$jemalloc" = "yes" ; then > + cat > $TMPC << EOF > +#include <stdlib.h> > +int main(void) { malloc(1); return 0; } > +EOF > + > + if compile_prog "" "-ljemalloc" ; then > + LIBS="-ljemalloc $LIBS" > + else > + feature_not_found "jemalloc" "install jemalloc devel" > + fi > +fi > + > +########################################## > # signalfd probe > signalfd="no" > cat > $TMPC << EOF > @@ -4499,6 +4527,7 @@ echo "snappy support $snappy" > echo "bzip2 support $bzip2" > echo "NUMA host support $numa" > echo "tcmalloc support $tcmalloc" > +echo "jemalloc support $jemalloc" > > if test "$sdl_too_old" = "yes"; then > echo "-> Your SDL version is too old - please upgrade to have SDL support" >
diff --git a/configure b/configure index 222694f..2fe1e05 100755 --- a/configure +++ b/configure @@ -336,6 +336,7 @@ vhdx="" quorum="" numa="" tcmalloc="no" +jemalloc="no" # parse CC options first for opt do @@ -1147,6 +1148,10 @@ for opt do ;; --enable-tcmalloc) tcmalloc="yes" ;; + --disable-jemalloc) jemalloc="no" + ;; + --enable-jemalloc) jemalloc="yes" + ;; *) echo "ERROR: unknown option $opt" echo "Try '$0 --help' for more information" @@ -1420,6 +1425,8 @@ Advanced options (experts only): --enable-numa enable libnuma support --disable-tcmalloc disable tcmalloc support --enable-tcmalloc enable tcmalloc support + --disable-jemalloc disable jemalloc support + --enable-jemalloc enable jemalloc support NOTE: The object files are built at the place where configure is launched EOF @@ -3344,6 +3351,11 @@ EOF fi fi +if test "$tcmalloc" = "yes" && test "$jemalloc" = "yes" ; then + echo "ERROR: tcmalloc && jemalloc can't be used at the same time" + exit 1 +fi + ########################################## # tcmalloc probe @@ -3361,6 +3373,22 @@ EOF fi ########################################## +# jemalloc probe + +if test "$jemalloc" = "yes" ; then + cat > $TMPC << EOF +#include <stdlib.h> +int main(void) { malloc(1); return 0; } +EOF + + if compile_prog "" "-ljemalloc" ; then + LIBS="-ljemalloc $LIBS" + else + feature_not_found "jemalloc" "install jemalloc devel" + fi +fi + +########################################## # signalfd probe signalfd="no" cat > $TMPC << EOF @@ -4499,6 +4527,7 @@ echo "snappy support $snappy" echo "bzip2 support $bzip2" echo "NUMA host support $numa" echo "tcmalloc support $tcmalloc" +echo "jemalloc support $jemalloc" if test "$sdl_too_old" = "yes"; then echo "-> Your SDL version is too old - please upgrade to have SDL support"
This adds "--enable-jemalloc" and "--disable-jemalloc" to allow linking to jemalloc memory allocator. We have already tcmalloc support, but it seem to not working well with a lot of iothreads/disks. The main problem is that tcmalloc use a shared thread cache of 16MB by default. With more threads, this cache is shared, and some bad garbage collections can occur if the cache is too low. It's possible to tcmalloc cache increase it with a env var: TCMALLOC_MAX_TOTAL_THREAD_CACHE_BYTES=256MB With default 16MB, performances are really bad with more than 2 disks. Increasing to 256MB, it's helping but still have problem with 16 disks/iothreads. Jemalloc don't have performance problem with default configuration. Here the benchmark results in iops of 1 qemu vm randread 4K iodepth=32, with rbd block backend (librbd is doing a lot of memory allocation), 1 iothread by disk glibc malloc ------------ 1 disk 29052 2 disks 55878 4 disks 127899 8 disks 240566 15 disks 269976 jemalloc -------- 1 disk 41278 2 disks 75781 4 disks 195351 8 disks 294241 15 disks 298199 tcmalloc 2.2.1 default 16M cache -------------------------------- 1 disk 37911 2 disks 67698 4 disks 41076 8 disks 43312 15 disks 37569 tcmalloc : 256M cache --------------------------- 1 disk 33914 2 disks 58839 4 disks 148205 8 disks 213298 15 disks 218383 Signed-off-by: Alexandre Derumier <aderumier@odiso.com> --- configure | 29 +++++++++++++++++++++++++++++ 1 file changed, 29 insertions(+)