Messages in this thread Patch in this message | | | Date | Wed, 24 Jun 2015 16:40:04 +0530 | From | Srikar Dronamraju <> | Subject | Regression in perf bench numa convergence stats |
| |
perf bench numa mem with -c / -m options on v4.1 and latest tip arent showing correct convergence statistics. I ran git bisect between v4.0 and v4.1. I have included the patch that fixed the problem for me.
After bisect, git bisect visualize shows
From e1e455f4f4d35850c30235747620d0d078fe9f64 Mon Sep 17 00:00:00 2001 From: Vinson Lee <vlee@twitter.com> Date: Mon, 23 Mar 2015 12:09:16 -0700 Subject: [PATCH] perf tools: Work around lack of sched_getcpu in glibc < 2.6. MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit
This patch fixes this build error with glibc < 2.6.
CC util/cloexec.o cc1: warnings being treated as errors util/cloexec.c: In function _perf_flag_probe_: util/cloexec.c:24: error: implicit declaration of function _sched_getcpu_ util/cloexec.c:24: error: nested extern declaration of _sched_getcpu_ make: *** [util/cloexec.o] Error 1
Signed-off-by: Vinson Lee <vlee@twitter.com> Acked-by: Jiri Olsa <jolsa@kernel.org> Acked-by: Namhyung Kim <namhyung@kernel.org> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Yann Droneaud <ydroneaud@opteya.com> Cc: stable@vger.kernel.org # 3.18+ Link: http://lkml.kernel.org/r/1427137761-16119-1-git-send-email-vlee@twopensource.com Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
# git log --oneline e1e455f e1e455f perf tools: Work around lack of sched_getcpu in glibc < 2.6. 77cfe38 perf kmem: Print big numbers using thousands' group 929a6bb tools lib traceevent: Factor out allocating and processing args e6d7c91 perf probe: Fix to get ummapped symbol address on kernel 228f14f perf tools: Remove (null) value of "Sort order" for perf mem report 2c7da8c perf annotate: Allow annotation for decompressed kernel modules bc84f46 perf tools: Try to lookup kernel module map before creating one 907fb50 perf tools: Remove is_kmodule_extension function e746b3e perf tools: Remove compressed argument from is_kernel_module 8dee9ff perf tools: Use kmod_path__parse in is_kernel_module
To further verify if the problem is because of e1e455f commit, I did roll back to e1e455f and its parent 77cfe38. I see this problem on more than one system.
# rpm -qa | grep glibc-2 glibc-2.17-55.el7.x86_64
git reset --hard e1e455f
# Running 'numa/mem' benchmark:
# Running main, "perf bench numa numa-mem --no-data_rand_walk -p 1 -t 64 -G 0 -P 0 -T 32 -l 800 -zZ0c" # #
### # 64 tasks will execute (on 4 nodes, 64 CPUs): # 800x 0MB global shared mem operations # 800x 0MB process shared mem operations # 800x 32MB thread local mem operations ###
### # # Startup synchronization: ... threads initialized in 0.512908 seconds. # # 0.1% [0.0 mins] 0/0 0/0 0/0 0/0 [ 0/0 ] l: -1-0 ( 1) {0-0} # 0.6% [0.0 mins] 0/0 0/0 0/0 0/0 [ 0/0 ] l: -1-0 ( 1) {0-0} # 5.1% [0.0 mins] 0/0 0/0 0/0 0/0 [ 0/0 ] l: -1-0 ( 1) {0-0} # 9.6% [0.1 mins] 0/0 0/0 0/0 0/0 [ 0/0 ] l: -1-0 ( 1) {0-0} # 14.0% [0.1 mins] 0/0 0/0 0/0 0/0 [ 0/0 ] l: -1-0 ( 1) {0-0}
###
4.903 secs slowest (max) thread-runtime 4.873 secs fastest (min) thread-runtime 4.941 secs average thread-runtime 0.301 % difference between max/avg runtime 4.228 GB data processed, per thread 270.583 GB data processed, total 1.160 nsecs/byte/thread runtime 0.862 GB/sec/thread speed 55.193 GB/sec total speed
and its parent 77cfe38 # git reset --hard 77cfe38
# Running 'numa/mem' benchmark:
# Running main, "perf bench numa numa-mem --no-data_rand_walk -p 1 -t 64 -G 0 -P 0 -T 32 -l 800 -zZ0c" # #
### # 64 tasks will execute (on 4 nodes, 64 CPUs): # 800x 0MB global shared mem operations # 800x 0MB process shared mem operations # 800x 32MB thread local mem operations ###
### # # Startup synchronization: ... threads initialized in 0.421336 seconds. # # 0.4% [0.0 mins] 16/1 16/1 16/1 16/1 [ 0/4 ] l: 1-20 ( 19) [95.0%] {4-4} # 2.6% [0.0 mins] 17/1 15/1 16/1 16/1 [ 2/4 ] l: 3-37 ( 34) [91.9%] {4-4} # 7.1% [0.0 mins] 17/1 15/1 16/1 16/1 [ 2/4 ] l: 32-67 ( 35) [52.2%] {4-4} # 11.8% [0.1 mins] 17/1 15/1 16/1 16/1 [ 2/4 ] l: 65-103 ( 38) [36.9%] {4-4} # 15.9% [0.1 mins] 17/1 15/1 16/1 16/1 [ 2/4 ] l: 98-136 ( 38) [27.9%] {4-4}
###
4.970 secs slowest (max) thread-runtime 4.940 secs fastest (min) thread-runtime 4.980 secs average thread-runtime 0.300 % difference between max/avg runtime 4.237 GB data processed, per thread 271.187 GB data processed, total 1.173 nsecs/byte/thread runtime 0.853 GB/sec/thread speed 54.562 GB/sec total speed
Even reverting e1e455f on top of tip/master seems to avoid the problem. The below patch fixes the problem.
-- Thanks and Regards Srikar Dronamraju
---->8--------------------------------------------
From 88199ad8a3d6495080eaa016b87a612bc742b1c4 Mon Sep 17 00:00:00 2001 From: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Date: Wed, 24 Jun 2015 16:23:22 +0530 Subject: [PATCH] perf tools:Fix perf_bench to show proper convergence
With commit: e1e455f (perf tools: Work around lack of sched_getcpu in glibc < 2.6), perf_bench numa mem with -c or -m option is not able to correctly calculate convergence. With the above commit, sched_getcpu always seems to return -1. The intention of commit e1e455f was to add a sched_getcpu in glibc < 2.6. Hence keep the sched_getcpu definition under an ifdef.
This regression happened occurred between v4.0 and v4.1
Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> --- tools/perf/util/cloexec.c | 4 ++++ 1 file changed, 4 insertions(+)
diff --git a/tools/perf/util/cloexec.c b/tools/perf/util/cloexec.c index 85b5238..2babdda 100644 --- a/tools/perf/util/cloexec.c +++ b/tools/perf/util/cloexec.c @@ -7,11 +7,15 @@ static unsigned long flag = PERF_FLAG_FD_CLOEXEC; +#ifdef __GLIBC_PREREQ +#if !__GLIBC_PREREQ(2, 6) int __weak sched_getcpu(void) { errno = ENOSYS; return -1; } +#endif +#endif static int perf_flag_probe(void) { -- 1.8.3.1
| |