Messages in this thread | | | Date | Tue, 27 Feb 2024 19:21:30 -0700 | Subject | Re: [PATCHSET v2 0/2] Split iowait into two states | From | Jens Axboe <> |
| |
On 2/27/24 2:06 PM, Jens Axboe wrote: > I haven't been able to properly benchmark patch 1, as the atomics are > noise in any workloads that approximate normality. I can certainly > concoct a synthetic test case if folks are interested. My gut says that > we're trading 3 fast path atomics for none, and with the 4th case > _probably_ being way less likely. There we grab the rq lock.
OK, so on Chris's suggestion, I tried his schbench to exercise the scheduling side. It's very futex intensive, so I hacked up futex to set iowait state when sleeping. I also added simple accounting to that path so I knew how many times it ran. A run of:
/schbench -m 60 -t 10 -p 8
on a 2 socket Intel(R) Xeon(R) Platinum 8458P with 176 threads, there's no regression in performance and try_to_wake_up() locking the rq of the task being scheduled in from another CPU doesn't seem to register much. On the previous run, I saw 2.21% there and now it's 2.36%. But it was also a better performing run, which may have lead to the increase.
Each run takes 30 seconds, and during that time I see around 290-310M hits of that path, or about ~10M/sec. Without modifying futex to use iowait, we obviously rarely hit it. About 200 times for a run, which makes sense as we're not really doing IO.
Anyway, just some data on this. If I leave the futex/pipe iowait in and run the same test, I see no discernable difference in profiles. In fact, the highest cost across the tests is bringing in the task->in_iowait cacheline.
-- Jens Axboe
| |