Re: [PATCH v2 0/4] Fix perf bench numa, futex and epoll to work with machines having #CPUs > 1K
From: Ian Rogers
Date: Wed Apr 06 2022 - 20:35:47 EST
On Wed, Apr 6, 2022 at 10:51 AM Athira Rajeev
> The perf benchmark for collections: numa, futex and epoll
> hits failure in system configuration with CPU's more than 1024.
> These benchmarks uses "sched_getaffinity" and "sched_setaffinity"
> in the code to work with affinity.
> Example snippet from numa benchmark:
> perf: bench/numa.c:302: bind_to_node: Assertion `!(ret)' failed.
> Aborted (core dumped)
> bind_to_node function uses "sched_getaffinity" to save the cpumask.
> This fails with EINVAL because the default mask size in glibc is 1024.
> Similarly in futex and epoll benchmark, uses sched_setaffinity during
> pthread_create with affinity. And since it returns EINVAL in such system
> configuration, benchmark doesn't run.
> To overcome this 1024 CPUs mask size limitation of cpu_set_t,
> change the mask size using the CPU_*_S macros ie, use CPU_ALLOC to
> allocate cpumask, CPU_ALLOC_SIZE for size, CPU_SET_S to set mask bit.
> Fix all the relevant places in the code to use mask size which is large
> enough to represent number of possible CPU's in the system.
> Fix parse_setup_cpu_list function in numa bench to check if input CPU
> is online before binding task to that CPU. This is to fix failures where,
> though CPU number is within max CPU, it could happen that CPU is offline.
> Here, sched_setaffinity will result in failure when using cpumask having
> that cpu bit set in the mask.
> Patch 1 and Patch 2 address fix for perf bench futex and perf bench
> epoll benchmark. Patch 3 and Patch 4 address fix in perf bench numa
> Athira Rajeev (4):
> tools/perf: Fix perf bench futex to correct usage of affinity for
> machines with #CPUs > 1K
> tools/perf: Fix perf bench epoll to correct usage of affinity for
> machines with #CPUs > 1K
> tools/perf: Fix perf numa bench to fix usage of affinity for machines
> with #CPUs > 1K
> tools/perf: Fix perf bench numa testcase to check if CPU used to bind
> task is online
> From v1 -> v2:
> Addressed review comment from Ian Rogers to do
> CPU_FREE in a cleaner way.
> Added Tested-by from Disha Goel
The whole set:
Acked-by: Ian Rogers <irogers@xxxxxxxxxx>
> tools/perf/bench/epoll-ctl.c | 25 ++++--
> tools/perf/bench/epoll-wait.c | 25 ++++--
> tools/perf/bench/futex-hash.c | 26 ++++--
> tools/perf/bench/futex-lock-pi.c | 21 +++--
> tools/perf/bench/futex-requeue.c | 21 +++--
> tools/perf/bench/futex-wake-parallel.c | 21 +++--
> tools/perf/bench/futex-wake.c | 22 ++++--
> tools/perf/bench/numa.c | 105 ++++++++++++++++++-------
> tools/perf/util/header.c | 43 ++++++++++
> tools/perf/util/header.h | 1 +
> 10 files changed, 242 insertions(+), 68 deletions(-)