| Message ID | 1462828814-32530-6-git-send-email-smuckle@linaro.org |
|---|---|
| State | New |
| Headers | show |
On Thu, May 19, 2016 at 01:44:36AM +0200, Rafael J. Wysocki wrote: > On Mon, May 9, 2016 at 11:20 PM, Steve Muckle <steve.muckle@linaro.org> wrote: > > The rate limit timestamp (last_freq_update_time) is currently advanced > > anytime schedutil re-evaluates the policy regardless of whether the CPU > > frequency is changed or not. This means that utilization updates which > > have no effect can cause much more significant utilization updates > > (which require a large increase or decrease in CPU frequency) to be > > delayed due to rate limiting. > > > > Instead only update the rate limiting timstamp when the requested > > target-supported frequency changes. The rate limit will now apply to > > the rate of CPU frequency changes rather than the rate of > > re-evaluations of the policy frequency. > > > > Signed-off-by: Steve Muckle <smuckle@linaro.org> > > I'm sort of divided here to be honest. It is true that this means we'll do more frequency re-evaluations, they will occur until an actual frequency change is requested. But the way it stands now, with a system's typical background activity there are so many minor events that it is very common for throttling to be in effect, causing major events to be ignored. > > > --- > > kernel/sched/cpufreq_schedutil.c | 3 +-- > > 1 file changed, 1 insertion(+), 2 deletions(-) > > > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c > > index e185075fcb5c..4d2907c8a142 100644 > > --- a/kernel/sched/cpufreq_schedutil.c > > +++ b/kernel/sched/cpufreq_schedutil.c > > @@ -117,12 +117,11 @@ static void sugov_update_commit(struct sugov_cpu *sg_cpu, int cpu, u64 time, > > struct sugov_policy *sg_policy = sg_cpu->sg_policy; > > struct cpufreq_policy *policy = sg_policy->policy; > > > > - sg_policy->last_freq_update_time = time; > > - > > if (sg_policy->next_freq == next_freq) { > > trace_cpu_frequency(policy->cur, cpu); > > You should at least rate limit the trace_cpu_frequency() thing here if > you don't want to advance the last update time I think, or you may > easily end up with the trace buffer flooded by irrelevant stuff. Going back to the reason this tracepoint exists, is it known why powertop thinks the CPU is idle when this tracepoint is removed? Maybe it's possible to get rid of this tracepoint altogether. Thanks for reviewing the series. thanks, Steve -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, May 20, 2016 at 02:24:19AM +0200, Rafael J. Wysocki wrote: > On Fri, May 20, 2016 at 1:34 AM, Steve Muckle <steve.muckle@linaro.org> wrote: > > On Thu, May 19, 2016 at 11:15:52PM +0200, Rafael J. Wysocki wrote: > >> But anyway this change again seems to be an optimization that might be > >> done later to me. > >> > >> I guess there are many things that might be optimized in schedutil, > >> but I'd prefer to address one item at a time, maybe going after the > >> ones that appear most relevant first? > > > > Calling the last two patches in this series optimizations is a stretch > > IMO. Issuing frequency change requests that result in the same > > target-supported frequency is clearly unnecessary and ends up delaying > > more urgent frequency changes, which I think is more like a bug. > > The [4/5] is pulling stuff where it doesn't belong. Simple as that. > Frequency tables don't belong in schedutil, so don't pull them in > there. > > If you want to do that cleanly, add a call to the driver that will > tell you what frequency would be selected by it if it were given a > particular target. > > I actually do agree with the direction of it and the [5/5], but I > don't like cutting corners. :-) Okay sure, makes sense and I'll work on a change to do that. > > > These patches are also needed in conjunction with the first three to address > > the remote wakeup delay. > > Well, does this mean that without the [4-5/5] the rest of the series > doesn't provide as much benefit as initially expected? At least in my testing the last two patches were required to show the improvement with the unit test I had developed. This is because all the minor system activity would keep pushing the rate limit out so when the remote cb came in it had no effect. So the last two patches aren't a hard requirement to theoretically solve the problem but in practice I don't think a benefit will be seen without them. > > Are there specific items you want to see addressed before these patches could go in? > > Do you mean in addition to what I already said in my comments? I was referring to the other possible optimizations/changes you mentioned in schedutil that might serve as cause to defer these patches. > > > I'm aware of the RT/DL support that needs improving, though > > that should be orthogonal. > > > > Also if it helps, I can provide a test case and/or traces to show the > > need for the last two patches. > > Yes, that will help. Okay I'll include a pointer to my unit test program and some trace data in my next post. thanks, Steve -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, May 20, 2016 at 02:37:17AM +0200, Rafael J. Wysocki wrote: > Also I think that it would be good to avoid walking the frequency > table twice in case we end up wanting to update the frequency after > all. With the [4/5] we'd do it once in get_next_freq() and then once > more in cpufreq_driver_fast_switch(), for example, and walking the > frequency table may be more expensive that doing the switch in the > first place. If a driver API is added to return the platform frequency associated with a target frequency, what do you think about requiring the fast_switch API to take a target-supported frequency? -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c index e185075fcb5c..4d2907c8a142 100644 --- a/kernel/sched/cpufreq_schedutil.c +++ b/kernel/sched/cpufreq_schedutil.c @@ -117,12 +117,11 @@ static void sugov_update_commit(struct sugov_cpu *sg_cpu, int cpu, u64 time, struct sugov_policy *sg_policy = sg_cpu->sg_policy; struct cpufreq_policy *policy = sg_policy->policy; - sg_policy->last_freq_update_time = time; - if (sg_policy->next_freq == next_freq) { trace_cpu_frequency(policy->cur, cpu); return; } + sg_policy->last_freq_update_time = time; sg_policy->next_freq = next_freq; if (sugov_queue_remote_callback(sg_policy, cpu))
The rate limit timestamp (last_freq_update_time) is currently advanced anytime schedutil re-evaluates the policy regardless of whether the CPU frequency is changed or not. This means that utilization updates which have no effect can cause much more significant utilization updates (which require a large increase or decrease in CPU frequency) to be delayed due to rate limiting. Instead only update the rate limiting timstamp when the requested target-supported frequency changes. The rate limit will now apply to the rate of CPU frequency changes rather than the rate of re-evaluations of the policy frequency. Signed-off-by: Steve Muckle <smuckle@linaro.org> --- kernel/sched/cpufreq_schedutil.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-) -- 2.4.10 -- To unsubscribe from this list: send the line "unsubscribe linux-pm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html