Messages in this thread | | | From | "Rafael J. Wysocki" <> | Date | Mon, 26 Mar 2018 11:35:48 +0200 | Subject | Re: [PATCH v3] cpuidle: poll_state: Add time limit to poll_idle() |
| |
On Mon, Mar 26, 2018 at 9:13 AM, Doug Smythies <dsmythies@telus.net> wrote: > On 2018.03.25 23:00 Doug Smythies wrote: >> On 2018.03.25 14:25 Rik van Riel wrote: >> On Sun, 2018-03-25 at 23:34 +0200, Rafael J. Wysocki wrote: >>> On Sunday, March 25, 2018 10:15:52 PM CEST Rik van Riel wrote: > > ...[snip]... > >>>>> >>>>> OK, I am still seeing a performance >>>>> degradation with the above, though >>>>> not throughout the entire workload. >>>>> >>>>> It appears that making the idle loop >>>>> do anything besides cpu_relax() for >>>>> a significant amount of time slows >>>>> things down. >>>> >>>> I see. >> >> I have no proof, but I do not see that as >> the problem. >> >> I think the issue is the overall exiting >> and then re-entering idle state 0 much >> more often, and the related overheads, where >> interrupts are disabled for short periods.
That is a very good point.
So far we have assumed that the performance impact was due to what happened in poll_idle(), but it very well may be due to how often poll_idle() is called and returns over a unit of time.
There may be workloads in which wakeups occur so often that idle states above 1 are (almost) never selected and in these workloads there will be increased overhead related to entering and exiting state 0 (polling) with the timeout. That is inevitable.
Rik, if your workload is one of these, you will see performance impact. However, you also should see a difference in power (see below).
>> My jury rigged way of trying to create similar >> conditions seems to always have the ISR return with >> the need_resched() flag set, so there is no difference >> in idle state 0 entries per unit time between kernel >> 4.16-rc6 and one with the poll fixes added. >> >> i.e. the difference between these numbers over some time: >> >> cat /sys/devices/system/cpu/cpu*/cpuidle/state0/usage >> >> Rik, I wonder if you see a difference with your real >> workflow? > > Using iperf, I was able to show a difference on my computer. > Another computer was used as the server, and my test computer > was the client. (the other way around didn't show a difference) > > With Kernel 4.16-rc6 I got about ~2000 idle state 0 entries > per minute and ~155 seconds residency. ~32 watts package power. > > With the poll stuff included I got ~46000 idle state 0 entries > per minute and ~53 seconds residency. ~20 watts package power.
OK, so that's 30%+ of a difference in power.
| |