tech-kern archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: interesting skylake perf tidbit



Le 18/06/2018 à 14:29, maya%netbsd.org@localhost a écrit :
joerg called it stupid and said we should use monitor, he's probably
right. new arm also has a similar thing.

The thing is, there are several SPINLOCK_BACKOFF()s that we just can't
replace by monitor. For example because we want to measure the delay under
LOCKDEBUG.

(Perhaps a tunable read-mostly SPINLOCK_BACKOFF_* would be worthwhile?)

You mean, like, a configurable SPINLOCK_BACKOFF_MIN? Yes, that seems doable.

I guess we should do both; use "monitor" when possible, and in the places
that are still required to use "pause", use a lower BACKOFF_MIN (set at
boot time, depending on the cpu model) to compensate for the increased CPU
latency.

There will still be a loss in this case, because the cycle ratio is 14
(10c -> 140c), but BACKOFF_MIN equals 4.

By the way I'm wondering if KabyLake and later generations have the same
increased latency, if they do we definitely need to fix that.

I guess you have a Skylake, is there any seeable effect?

Also on the same topic: SPINLOCK_RUN_HOOK seems to be unused?

Yes, please remove.


Home | Main Index | Thread Index | Old Index