configure | 29 +++++++++++++++++++++++++++++ util/rcu.c | 6 ++++++ 2 files changed, 35 insertions(+)
Since there are some issues in memory alloc/free machenism
in glibc for little chunk memory, if Qemu frequently
alloc/free little chunk memory, the glibc doesn't alloc
little chunk memory from free list of glibc and still
allocate from OS, which make the heap size bigger and bigger.
This patch introduce malloc_trim(), which will free heap memory.
Below are test results from smaps file.
(1)without patch
55f0783e1000-55f07992a000 rw-p 00000000 00:00 0 [heap]
Size: 21796 kB
Rss: 14260 kB
Pss: 14260 kB
(2)with patch
55cc5fadf000-55cc61008000 rw-p 00000000 00:00 0 [heap]
Size: 21668 kB
Rss: 6940 kB
Pss: 6940 kB
Signed-off-by: Yang Zhong <yang.zhong@intel.com>
---
configure | 29 +++++++++++++++++++++++++++++
util/rcu.c | 6 ++++++
2 files changed, 35 insertions(+)
diff --git a/configure b/configure
index 0c6e757..6292ab0 100755
--- a/configure
+++ b/configure
@@ -426,6 +426,7 @@ vxhs=""
supported_cpu="no"
supported_os="no"
bogus_os="no"
+malloc_trim="yes"
# parse CC options first
for opt do
@@ -3857,6 +3858,30 @@ if test "$tcmalloc" = "yes" && test "$jemalloc" = "yes" ; then
exit 1
fi
+# Even if malloc_trim() is available, these non-libc memory allocators
+# do not support it.
+if test "$tcmalloc" = "yes" || test "$jemalloc" = "yes" ; then
+ if test "$malloc_trim" = "yes" ; then
+ echo "Disabling malloc_trim with non-libc memory allocator"
+ fi
+ malloc_trim="no"
+fi
+
+#######################################
+# malloc_trim
+
+if test "$malloc_trim" != "no" ; then
+ cat > $TMPC << EOF
+#include <malloc.h>
+int main(void) { malloc_trim(0); return 0; }
+EOF
+ if compile_prog "" "" ; then
+ malloc_trim="yes"
+ else
+ malloc_trim="no"
+ fi
+fi
+
##########################################
# tcmalloc probe
@@ -6012,6 +6037,10 @@ if test "$opengl" = "yes" ; then
fi
fi
+if test "$malloc_trim" = "yes" ; then
+ echo "CONFIG_MALLOC_TRIM=y" >> $config_host_mak
+fi
+
if test "$avx2_opt" = "yes" ; then
echo "CONFIG_AVX2_OPT=y" >> $config_host_mak
fi
diff --git a/util/rcu.c b/util/rcu.c
index ca5a63e..f403b77 100644
--- a/util/rcu.c
+++ b/util/rcu.c
@@ -32,6 +32,9 @@
#include "qemu/atomic.h"
#include "qemu/thread.h"
#include "qemu/main-loop.h"
+#if defined(CONFIG_MALLOC_TRIM)
+#include <malloc.h>
+#endif
/*
* Global grace period counter. Bit 0 is always one in rcu_gp_ctr.
@@ -272,6 +275,9 @@ static void *call_rcu_thread(void *opaque)
node->func(node);
}
qemu_mutex_unlock_iothread();
+#if defined(CONFIG_MALLOC_TRIM)
+ malloc_trim(4 * 1024 * 1024);
+#endif
}
abort();
}
--
1.9.1
On Fri, Nov 24, 2017 at 02:30:30PM +0800, Yang Zhong wrote: > diff --git a/configure b/configure > index 0c6e757..6292ab0 100755 > --- a/configure > +++ b/configure > @@ -426,6 +426,7 @@ vxhs="" > supported_cpu="no" > supported_os="no" > bogus_os="no" > +malloc_trim="yes" Looks pretty good, sorry I forgot to mention two things: Please add the --enable-malloc-trim/--disable-malloc-trim options so it's easy to build QEMU with or without this feature. For example, if someone is debugging a performance issue they may wish to rebuild with --disable-malloc-trim to confirm that trimming hasn't caused a regression. Also please change this line to malloc_trim="" so the "Disabling malloc_trim with non-libc memory allocator" error message is only printed when --enable-malloc-trim was explicitly given by the user. Otherwise the message is always printed when QEMU is built with jemalloc/tcmalloc - that's too noisy.
Hi, On 2017/11/24 14:30, Yang Zhong wrote: > Since there are some issues in memory alloc/free machenism > in glibc for little chunk memory, if Qemu frequently > alloc/free little chunk memory, the glibc doesn't alloc > little chunk memory from free list of glibc and still > allocate from OS, which make the heap size bigger and bigger. > > This patch introduce malloc_trim(), which will free heap memory. > > Below are test results from smaps file. > (1)without patch > 55f0783e1000-55f07992a000 rw-p 00000000 00:00 0 [heap] > Size: 21796 kB > Rss: 14260 kB > Pss: 14260 kB > > (2)with patch > 55cc5fadf000-55cc61008000 rw-p 00000000 00:00 0 [heap] > Size: 21668 kB > Rss: 6940 kB > Pss: 6940 kB > > Signed-off-by: Yang Zhong <yang.zhong@intel.com> > --- > configure | 29 +++++++++++++++++++++++++++++ > util/rcu.c | 6 ++++++ > 2 files changed, 35 insertions(+) > > diff --git a/configure b/configure > index 0c6e757..6292ab0 100755 > --- a/configure > +++ b/configure > @@ -426,6 +426,7 @@ vxhs="" > supported_cpu="no" > supported_os="no" > bogus_os="no" > +malloc_trim="yes" > > # parse CC options first > for opt do > @@ -3857,6 +3858,30 @@ if test "$tcmalloc" = "yes" && test "$jemalloc" = "yes" ; then > exit 1 > fi > > +# Even if malloc_trim() is available, these non-libc memory allocators > +# do not support it. > +if test "$tcmalloc" = "yes" || test "$jemalloc" = "yes" ; then > + if test "$malloc_trim" = "yes" ; then > + echo "Disabling malloc_trim with non-libc memory allocator" > + fi > + malloc_trim="no" > +fi > + > +####################################### > +# malloc_trim > + > +if test "$malloc_trim" != "no" ; then > + cat > $TMPC << EOF > +#include <malloc.h> > +int main(void) { malloc_trim(0); return 0; } > +EOF > + if compile_prog "" "" ; then > + malloc_trim="yes" > + else > + malloc_trim="no" > + fi > +fi > + > ########################################## > # tcmalloc probe > > @@ -6012,6 +6037,10 @@ if test "$opengl" = "yes" ; then > fi > fi > > +if test "$malloc_trim" = "yes" ; then > + echo "CONFIG_MALLOC_TRIM=y" >> $config_host_mak > +fi > + > if test "$avx2_opt" = "yes" ; then > echo "CONFIG_AVX2_OPT=y" >> $config_host_mak > fi > diff --git a/util/rcu.c b/util/rcu.c > index ca5a63e..f403b77 100644 > --- a/util/rcu.c > +++ b/util/rcu.c > @@ -32,6 +32,9 @@ > #include "qemu/atomic.h" > #include "qemu/thread.h" > #include "qemu/main-loop.h" > +#if defined(CONFIG_MALLOC_TRIM) > +#include <malloc.h> > +#endif > > /* > * Global grace period counter. Bit 0 is always one in rcu_gp_ctr. > @@ -272,6 +275,9 @@ static void *call_rcu_thread(void *opaque) > node->func(node); > } > qemu_mutex_unlock_iothread(); > +#if defined(CONFIG_MALLOC_TRIM) > + malloc_trim(4 * 1024 * 1024); > +#endif > } > abort(); > } > Looks like this patch introduces a performance regression. With this patch the time of booting a VM with 60 scsi disks on ARM64 is increased by 200+ seconds. Thanks, -- Shannon
On Sun, Nov 26, 2017 at 02:17:18PM +0800, Shannon Zhao wrote: > Hi, > > On 2017/11/24 14:30, Yang Zhong wrote: > > Since there are some issues in memory alloc/free machenism > > in glibc for little chunk memory, if Qemu frequently > > alloc/free little chunk memory, the glibc doesn't alloc > > little chunk memory from free list of glibc and still > > allocate from OS, which make the heap size bigger and bigger. > > > > This patch introduce malloc_trim(), which will free heap memory. > > > > Below are test results from smaps file. > > (1)without patch > > 55f0783e1000-55f07992a000 rw-p 00000000 00:00 0 [heap] > > Size: 21796 kB > > Rss: 14260 kB > > Pss: 14260 kB > > > > (2)with patch > > 55cc5fadf000-55cc61008000 rw-p 00000000 00:00 0 [heap] > > Size: 21668 kB > > Rss: 6940 kB > > Pss: 6940 kB > > > > Signed-off-by: Yang Zhong <yang.zhong@intel.com> > > --- > > configure | 29 +++++++++++++++++++++++++++++ > > util/rcu.c | 6 ++++++ > > 2 files changed, 35 insertions(+) > > > > diff --git a/configure b/configure > > index 0c6e757..6292ab0 100755 > > --- a/configure > > +++ b/configure > > @@ -426,6 +426,7 @@ vxhs="" > > supported_cpu="no" > > supported_os="no" > > bogus_os="no" > > +malloc_trim="yes" > > > > # parse CC options first > > for opt do > > @@ -3857,6 +3858,30 @@ if test "$tcmalloc" = "yes" && test "$jemalloc" = "yes" ; then > > exit 1 > > fi > > > > +# Even if malloc_trim() is available, these non-libc memory allocators > > +# do not support it. > > +if test "$tcmalloc" = "yes" || test "$jemalloc" = "yes" ; then > > + if test "$malloc_trim" = "yes" ; then > > + echo "Disabling malloc_trim with non-libc memory allocator" > > + fi > > + malloc_trim="no" > > +fi > > + > > +####################################### > > +# malloc_trim > > + > > +if test "$malloc_trim" != "no" ; then > > + cat > $TMPC << EOF > > +#include <malloc.h> > > +int main(void) { malloc_trim(0); return 0; } > > +EOF > > + if compile_prog "" "" ; then > > + malloc_trim="yes" > > + else > > + malloc_trim="no" > > + fi > > +fi > > + > > ########################################## > > # tcmalloc probe > > > > @@ -6012,6 +6037,10 @@ if test "$opengl" = "yes" ; then > > fi > > fi > > > > +if test "$malloc_trim" = "yes" ; then > > + echo "CONFIG_MALLOC_TRIM=y" >> $config_host_mak > > +fi > > + > > if test "$avx2_opt" = "yes" ; then > > echo "CONFIG_AVX2_OPT=y" >> $config_host_mak > > fi > > diff --git a/util/rcu.c b/util/rcu.c > > index ca5a63e..f403b77 100644 > > --- a/util/rcu.c > > +++ b/util/rcu.c > > @@ -32,6 +32,9 @@ > > #include "qemu/atomic.h" > > #include "qemu/thread.h" > > #include "qemu/main-loop.h" > > +#if defined(CONFIG_MALLOC_TRIM) > > +#include <malloc.h> > > +#endif > > > > /* > > * Global grace period counter. Bit 0 is always one in rcu_gp_ctr. > > @@ -272,6 +275,9 @@ static void *call_rcu_thread(void *opaque) > > node->func(node); > > } > > qemu_mutex_unlock_iothread(); > > +#if defined(CONFIG_MALLOC_TRIM) > > + malloc_trim(4 * 1024 * 1024); > > +#endif > > } > > abort(); > > } > > > > Looks like this patch introduces a performance regression. With this > patch the time of booting a VM with 60 scsi disks on ARM64 is increased > by 200+ seconds. > Hello Shannon, Thanks for your reply! As for your concerns, i did VM bootup compared tests, and results as below: #test command ./qemu-system-x86_64 -enable-kvm -cpu host -m 2G -smp cpus=4,cores=4,\ threads=1,sockets=1 -drive format=raw,\ file=test.img,index=0,media=disk -nographic #without patch root@intel-internal-corei7-64:~# systemd-analyze Startup finished in 4.979s (kernel) + 1.214s (userspace) = 6.193s root@intel-internal-corei7-64:~# systemd-analyze Startup finished in 4.922s (kernel) + 1.175s (userspace) = 6.097s root@intel-internal-corei7-64:~# systemd-analyze Startup finished in 4.990s (kernel) + 1.301s (userspace) = 6.291s root@intel-internal-corei7-64:~# systemd-analyze Startup finished in 5.063s (kernel) + 1.336s (userspace) = 6.400s root@intel-internal-corei7-64:~# systemd-analyze Startup finished in 4.820s (kernel) + 1.237s (userspace) = 6.057s avg: kernel 4.9548, userspace 1.2526 #with this patch root@intel-internal-corei7-64:~# systemd-analyze Startup finished in 5.099s (kernel) + 1.579s (userspace) = 6.679s root@intel-internal-corei7-64:~# systemd-analyze Startup finished in 5.003s (kernel) + 1.343s (userspace) = 6.347s root@intel-internal-corei7-64:~# systemd-analyze Startup finished in 4.853s (kernel) + 1.220s (userspace) = 6.074s root@intel-internal-corei7-64:~# systemd-analyze Startup finished in 4.836s (kernel) + 1.111s (userspace) = 5.948s root@intel-internal-corei7-64:~# systemd-analyze Startup finished in 4.917s (kernel) + 1.166s (userspace) = 6.083s avg: kernel 4.9416s, userspace: 1.2838 From above test results, there are almost not any performance regression on x86 platform. Sorry, there is not any ARM based platform in my hand, i can't give related datas. thanks! Regards, Yang > Thanks, > -- > Shannon
On 27/11/2017 04:06, Zhong Yang wrote: > #test command > ./qemu-system-x86_64 -enable-kvm -cpu host -m 2G -smp cpus=4,cores=4,\ > threads=1,sockets=1 -drive format=raw,\ > file=test.img,index=0,media=disk -nographic > > #without patch > root@intel-internal-corei7-64:~# systemd-analyze > Startup finished in 4.979s (kernel) + 1.214s (userspace) = 6.193s > > root@intel-internal-corei7-64:~# systemd-analyze > Startup finished in 4.922s (kernel) + 1.175s (userspace) = 6.097s > > root@intel-internal-corei7-64:~# systemd-analyze > Startup finished in 4.990s (kernel) + 1.301s (userspace) = 6.291s > > root@intel-internal-corei7-64:~# systemd-analyze > Startup finished in 5.063s (kernel) + 1.336s (userspace) = 6.400s > > root@intel-internal-corei7-64:~# systemd-analyze > Startup finished in 4.820s (kernel) + 1.237s (userspace) = 6.057s > > avg: kernel 4.9548, userspace 1.2526 > > > #with this patch > root@intel-internal-corei7-64:~# systemd-analyze > Startup finished in 5.099s (kernel) + 1.579s (userspace) = 6.679s > > root@intel-internal-corei7-64:~# systemd-analyze > Startup finished in 5.003s (kernel) + 1.343s (userspace) = 6.347s > > root@intel-internal-corei7-64:~# systemd-analyze > Startup finished in 4.853s (kernel) + 1.220s (userspace) = 6.074s > > root@intel-internal-corei7-64:~# systemd-analyze > Startup finished in 4.836s (kernel) + 1.111s (userspace) = 5.948s > > root@intel-internal-corei7-64:~# systemd-analyze > Startup finished in 4.917s (kernel) + 1.166s (userspace) = 6.083s > > avg: kernel 4.9416s, userspace: 1.2838 > > From above test results, there are almost not any performance regression > on x86 platform. Sorry, there is not any ARM based platform in my hand, > i can't give related datas. thanks! You are using only one disk, Shannon is using 200. That may make a difference, as PCI BAR setup in the guest becomes very expensive as you add more devices. Thanks, Paolo
© 2016 - 2024 Red Hat, Inc.