Hello, on NUMA enabled server equipped with 4 Intel E5-4610 v2 CPUs we observe following performance degradation: Runtime to run "lu.C.x" test from NAS Parallel Benchmarks after booting the kernel: real 1m57.834s user 113m51.520s Then we disable and re-enable one core: echo 0 > /sys/devices/system/cpu/cpu1/online echo 1 > /sys/devices/system/cpu/cpu1/online and rerun the same test. Runtime is now degraded (by 40% for user time and by 30% for the real (wall-clock) time) using all 64 cores real 2m47.746s user 160m46.109s The issue was first reported in "The Linux Scheduler: a Decade of Wasted Cores" paper http://www.ece.ubc.ca/~sasha/papers/eurosys16-final29.pdf https://github.com/jplozi/wastedcores/issues/1 How to reproduce the issue: A) Get benchmark and compile it: 1) wget http://www.nas.nasa.gov/assets/npb/NPB3.3.1.tar.gz 2) tar zxvf NPB3.3.1.tar.gz 3) cd ~/NPB3.3.1/NPB3.3-OMP/config/ 4) ln -sf NAS.samples/make.def.gcc_x86 make.def (assuming using gcc compiler) 5) ln -sf NAS.samples/suite.def.lu suite.def 6) cd ~/NPB3.3.1/NPB3.3-OMP 7) make suite 8) You should have now in directory ~/NPB3.3.1/NPB3.3-OMP/bin benchmarks lu.*. The binaries are alphabetically sorted by runtime with "lu.A.x" having the shortest runtime. B) Reproducing the issue Remark: we have done the tests with autogroup disabled sysctl -w kernel.sched_autogroup_enabled=0 to avoid following issue on 4.7 kernel: https://bugzilla.kernel.org/show_bug.cgi?id=120481 The test was conducted on NUMA server with 4 nodes and using the all 64 cores. 1) (time bin/lu.C.x) |& tee $(uname -r)_lu.C.x.log_before_reenable_kernel.sched_autogroup_enabled=0 2) disable and re-enable one core echo 0 > /sys/devices/system/cpu/cpu1/online echo 1 > /sys/devices/system/cpu/cpu1/online 3) (time bin/lu.C.x) |& tee $(uname -r)_lu.C.x.log_after_reenable_kernel.sched_autogroup_enabled=0 grep "real\|user" *lu.C* You will see significant difference in both real and user time. Regarding to the authors of the paper the root cause of the problem is a missing call to regenerate domains inside NUMA nodes after re-enabling CPU. The problem was introduced in 3.19 kernel. Kernel 4.7 performance has improved significantly over 4.6 kernel but the bug still exists in 4.7 kernel. The authors of paper has proposed a patch which applies to 4.1 kernel. Here is the link: https://github.com/jplozi/wastedcores/blob/master/patches/missing_sched_domains_linux_4.1.patch ===========For the completeness here are the results with 4.6 kernel=========== AFTER BOOT real 1m31.639s user 89m24.657s AFTER core has been disabled and re-enabled real 2m44.566s user 157m59.814s Please notice that with 4.6 kernel problem is much more visible than with 4.7 rc5 kernel. At the same time, 4.6 kernel delivers better performance after boot than 4.7 rc5 kernel which might indicate that another problem is in play. ================================================================= Hello, on NUMA enabled server equipped with 4 Intel E5-4610 v2 CPUs we observe following performance degradation: Runtime to run "lu.C.x" test from NAS Parallel Benchmarks after booting the kernel: real 1m57.834s user 113m51.520s Then we disable and re-enable one core: echo 0 > /sys/devices/system/cpu/cpu1/online echo 1 > /sys/devices/system/cpu/cpu1/online and rerun the same test. Runtime is now degraded (by 40% for user time and by 30% for the real (wall-clock) time) using all 64 cores real 2m47.746s user 160m46.109s The issue was first reported in "The Linux Scheduler: a Decade of Wasted Cores" paper http://www.ece.ubc.ca/~sasha/papers/eurosys16-final29.pdf https://github.com/jplozi/wastedcores/issues/1 How to reproduce the issue: A) Get benchmark and compile it: 1) wget http://www.nas.nasa.gov/assets/npb/NPB3.3.1.tar.gz 2) tar zxvf NPB3.3.1.tar.gz 3) cd ~/NPB3.3.1/NPB3.3-OMP/config/ 4) ln -sf NAS.samples/make.def.gcc_x86 make.def (assuming using gcc compiler) 5) ln -sf NAS.samples/suite.def.lu suite.def 6) cd ~/NPB3.3.1/NPB3.3-OMP 7) make suite 8) You should have now in directory ~/NPB3.3.1/NPB3.3-OMP/bin benchmarks lu.*. The binaries are alphabetically sorted by runtime with "lu.A.x" having the shortest runtime. B) Reproducing the issue (see also attached script) Remark: we have done the tests with autogroup disabled sysctl -w kernel.sched_autogroup_enabled=0 to avoid this issue on 4.7 kernel: https://bugzilla.kernel.org/show_bug.cgi?id=120481 The test was conducted on NUMA server with 4 nodes and using all available 64 cores. 1) (time bin/lu.C.x) |& tee $(uname -r)_lu.C.x.log_before_reenable_kernel.sched_autogroup_enabled=0 2) disable and re-enable one core echo 0 > /sys/devices/system/cpu/cpu1/online echo 1 > /sys/devices/system/cpu/cpu1/online 3) (time bin/lu.C.x) |& tee $(uname -r)_lu.C.x.log_after_reenable_kernel.sched_autogroup_enabled=0 grep "real\|user" *lu.C* You will see significant difference in both real and user time. Regarding to the authors of the paper the root cause of the problem is a missing call to regenerate domains inside NUMA nodes after re-enabling CPU. The problem was introduced in 3.19 kernel. The authors of paper has proposed a patch which applies to 4.1 kernel. Here is the link: https://github.com/jplozi/wastedcores/blob/master/patches/missing_sched_domains_linux_4.1.patch ===========For the completeness here are the results with 4.6 kernel=========== AFTER BOOT real 1m31.639s user 89m24.657s AFTER core has been disabled and re-enabled real 2m44.566s user 157m59.814s Please notice that 4.6 kernel problem is much more visible than with 4.7 rc5 kernel. At the same time, 4.6 kernel delivers much better performance after boot than 4.7 rc5 kernel which might indicate that another problem is in play. ================================================================= I have also tested kernel provided by Peter Zijlstra on Friday, June 24th which provides fix for https://bugzilla.kernel.org/show_bug.cgi?id=120481. It does not fix this issue and kernel right after boot performs worse than 4.6 kernel right after boot so we may in fact face two problems here. ========Results with 4.7.0-02548776ded1185e6e16ad0a475481e982741ee9 kernel===== git://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git sched/urgent $ git rev-parse HEAD 02548776ded1185e6e16ad0a475481e982741ee9 AFTER BOOT real 1m58.549s user 113m31.448s AFTER core has been disabled and re-enabled real 2m35.930s user 148m20.795s ================================================================= Thanks a lot Jirka
Created attachment 221421 [details] Script to reproduce the problem It assumes that working directory is ~/NPB3.3.1/NPB3.3-OMP and under bin directory there is "lu.C.x" binary.
It turns out that lu.C.x results show quite big variation and tests have to be repeated several times and mean value of real time has to be used to get reliable results. There is NO regression on following CPUs 4x Xeon(R) CPU E5-4610 v2 @ 2.30GHz 4x Xeon(R) CPU E5-2690 v3 @ 2.60GHz but there is regression (slow down by factor 6x) on AMD Opteron(TM) Processor 6272 Kernel 4.7.0-0.rc7.git0.1.el7.x86_64 real_time to run ./lu.C.x benchmark (mean value out of 10 runs) Right after boot: 273 seconds After disabling and enabling a core: 1702 seconds!
WORKLOG ========================================================================== Jiri Olsa is working on a patch. This is the latest update from Jiri Olsa: ========================================================================== Jirka, Peter and Jean-Pierre reported performance drop on some cpus after making cpu offline and online again. The reason is the kernel logic that falls back to SMT level topology if more than one node is detected within CPU package. During the system boot this logic cuts out the DIE topology level and numa code adds NUMA level on top of this. After the boot if you reboot make the cpu offline and online again, this logic resets the SMT level topology, removing whole NUMA level stuff. Ensuring the SMT topology fallback happens only once during the boot so the NUMA topology level is kept once it's built. This problem is one of the issues reported in the wastedcores article [1]. There's similar patch to this issue attached to the article [2]. [1] https://github.com/jplozi/wastedcores [2] https://github.com/jplozi/wastedcores/blob/master/patches/missing_sched_domains_linux_4.1.patch Cc: Jean-Pierre Lozi <jplozi@unice.fr> Cc: Jirka Hladky <jhladky@redhat.com> Cc: Petr Surý <psury@redhat.com> Signed-off-by: Jiri Olsa <jolsa@kernel.org> --- arch/x86/kernel/smpboot.c | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-) diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c index 2a6e84a30a54..f2a769b2b3fe 100644 --- a/arch/x86/kernel/smpboot.c +++ b/arch/x86/kernel/smpboot.c @@ -487,7 +487,17 @@ static struct sched_domain_topology_level numa_inside_package_topology[] = { */ static void primarily_use_numa_for_topology(void) { - set_sched_topology(numa_inside_package_topology); + static bool once; + + /* + * We need to run it only during boot, once we are + * here due to getting cpu online again we have already + * NUMA topology setup done. + */ + if (!once) { + set_sched_topology(numa_inside_package_topology); + once = true; + } } void set_cpu_sibling_map(int cpu)
WORKLOG fixed by following upstream post: http://marc.info/?l=linux-kernel&m=147337376415456&w=2