Francisco Jerez <currojerez@xxxxxxxxxx> writes:
> Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> writes:
>
>> Quoting Francisco Jerez (2020-04-14 22:00:25)
>>> Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> writes:
>>>
>>> > Quoting Francisco Jerez (2020-04-14 20:39:48)
>>> >> Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> writes:
>>> >>
>>> >> > Quoting Chris Wilson (2020-04-14 17:14:23)
>>> >> >> Try to make RPS dramatically more responsive by shrinking the evaluation
>>> >> >> intervales by a factor of 100! The issue is as we now park the GPU
>>> >> >> rapidly upon idling, a short or bursty workload such as the composited
>>> >> >> desktop never sustains enough work to fill and complete an evaluation
>>> >> >> window. As such, the frequency we program remains stuck. This was first
>>> >> >> reported as once boosted, we never relinquished the boost [see commit
>>> >> >> 21abf0bf168d ("drm/i915/gt: Treat idling as a RPS downclock event")] but
>>> >> >> it equally applies in the order direction for bursty workloads that
>>> >> >> *need* low latency, like desktop animations.
>>> >> >>
>>> >> >> What we could try is preserve the incomplete EI history across idling,
>>> >> >> it is not clear whether that would be effective, nor whether the
>>> >> >> presumption of continuous workloads is accurate. A clearer path seems to
>>> >> >> treat it as symptomatic that we fail to handle bursty workload with the
>>> >> >> current EI, and seek to address that by shrinking the EI so the
>>> >> >> evaluations are run much more often.
>>> >> >>
>>> >> >> This will likely entail more frequent interrupts, and by the time we
>>> >> >> process the interrupt in the bottom half [from inside a worker], the
>>> >> >> workload on the GPU has changed. To address the changeable nature, in
>>> >> >> the previous patch we compared the previous complete EI with the
>>> >> >> interrupt request and only up/down clock if both agree. The impact of
>>> >> >> asking for, and presumably, receiving more interrupts is still to be
>>> >> >> determined and mitigations sought. The first idea is to differentiate
>>> >> >> between up/down responsivity and make upclocking more responsive than
>>> >> >> downlocking. This should both help thwart jitter on bursty workloads by
>>> >> >> making it easier to increase than it is to decrease frequencies, and
>>> >> >> reduce the number of interrupts we would need to process.
>>> >> >
>>> >> > Another worry I'd like to raise, is that by reducing the EI we risk
>>> >> > unstable evaluations. I'm not sure how accurate the HW is, and I worry
>>> >> > about borderline workloads (if that is possible) but mainly the worry is
>>> >> > how the HW is sampling.
>>> >> >
>>> >> > The other unmentioned unknown is the latency in reprogramming the
>>> >> > frequency. At what point does it start to become a significant factor?
>>> >> > I'm presuming the RPS evaluation itself is free, until it has to talk
>>> >> > across the chip to send an interrupt.
>>> >> > -Chris
>>> >>
>>> >> At least on ICL the problem which this patch and 21abf0bf168d were
>>> >> working around seems to have to do with RPS interrupt delivery being
>>> >> inadvertently blocked for extended periods of time. Looking at the GPU
>>> >> utilization and RPS events on a graph I could see the GPU being stuck at
>>> >> low frequency without any RPS interrupts firing, for a time interval
>>> >> orders of magnitude greater than the EI we're theoretically programming
>>> >> today. IOW it seems like the real problem isn't that our EIs are too
>>> >> long, but that we're missing a bunch of them.
>>> >>
>>> >> The solution I was suggesting for this on IRC during the last couple of
>>> >> days wouldn't have any of the drawbacks you mention above, I'll send it
>>> >> to this list in a moment if the general approach seems okay to you:
>>> >>
>>> >> https://github.com/curro/linux/commit/f7bc31402aa727a52d957e62d985c6dae6be4b86
>>> >
>>> > We were explicitly told to mask the interrupt generation at source
>>> > to conserve power. So I would hope for a statement as to whether that is
>>> > still a requirement from the HW architects; but I can't see why we would
>>> > not apply the mask and that this is just paper. If the observation about
>>> > forcewake tallies, would this not suggest that it is still conserving
>>> > power on icl?
>>> >
>>>
>>> Yeah, it's hard to see how disabling interrupt generation could save any
>>> additional power in a unit which is powered off -- At least on ICL where
>>> even the interrupt masking register is powered off...
>>>
>>> > I haven't looked at whether I see the same phenomenon as you [missing
>>> > interrupts on icl] locally, but I was expecting something like the bug
>>> > report since the observation that render times are less than EI was
>>> > causing the clocks to stay high. And I noticed your problem statement
>>> > and was hopeful for a link.
>>> >
>>>
>>> Right. In the workloads I was looking at last week the GPU would often
>>> be active for periods of time several times greater than the EI, and we
>>> would still fail to clock up.
>>>
>>> > They sound like two different problems. (Underclocking for animations is
>>> > not icl specific.)
>>> >
>>>
>>> Sure. But it seems like the underclocking problem has been greatly
>>> exacerbated by 21abf0bf168d, which may have been mitigating the same ICL
>>> problem I was looking at leading to RPS interrupt loss. Maybe
>>> 21abf0bf168d wouldn't be necessary with working RPS interrupt delivery?
>>> And without 21abf0bf168d platforms earlier than ICL wouldn't have as
>>> much of an underclocking problem either.
>>
>> 21abf0bf168d ("drm/i915/gt: Treat idling as a RPS downclock event")
>>
>> is necessary due to that we can set a boost frequency and then never run
>> the RPS worker due to short activity cycles. See
>> igt/i915_pm_rc6_residency for rc6-idle, the bg_load is essentially just
>>
>> for (;;) {
>> execbuf(&nop);
>> sleep(.1);
>> }
>>
>> Without 21abf0bf168d if you trigger a waitboost just before, it never
>> recovers and power utilisation is measurably higher.
>>
>
> I can believe that, but can you confirm that the problem isn't a symptom
> of the PMINTRMSK issue I was talking about earlier? Assuming it isn't,
> couldn't you do something like 21abf0bf168d which simply stores a
> timestamp at GPU parking time and delays the actual frequency ramp-down
> to GPU unpark, at which point it's performed *only* if the time spent
> idle goes over a down-clock threshold based on the current "power" mode
> (e.g. something like the value of GEN6_RP_DOWN_EI minus
> GEN6_RP_DOWN_THRESHOLD) -- So the behavior of frequency rampdown due to
> parking more closely matches what the RPS would have done if it had been
> active.
>
If you want to match the behavior of the RPS even more closely you could
use a variable delay kind of like the one I'm using for overload
tracking in:
https://github.com/curro/linux/commit/0669ce23148eafa00e15fc44b9d07f7919802af0
However using a fixed delay based on GEN6_RP_DOWN_THRESHOLD as I was
describing above might be sufficient to avoid the waitboost energy
efficiency problem you were talking about.
>>> >> That said it *might* be helpful to reduce the EIs we use right now in
>>> >> addition, but a factor of 100 seems over the top since that will cause
>>> >> the evaluation interval to be roughly two orders of magnitude shorter
>>> >> than the rendering time of a typical frame, which can lead to massive
>>> >> oscillations even in workloads that use a small fraction of the GPU time
>>> >> to render a single frame. Maybe we want something in between?
>>> >
>>> > Probably; as you can guess these were pulled out of nowhere based on the
>>> > observation that the frame lengths are much shorter than the current EI
>>> > and that in order for us to ramp up to maxclocks in a single frame of
>>> > animation would take about 4 samples per frame. Based on the reporter's
>>> > observations, we do have to ramp up very quickly for single frame of
>>> > rendering in order to hit the vblank, as we are ramping down afterwards.
>>> >
>>> > With a target of 4 samples within say 1ms, 160us isn't too far of the
>>> > mark. (We have to allow some extra time to catch up rendering.)
>>>
>>>
>>> How about we stop ramping down after the rendering of a single frame?
>>> It's not like we save any power by doing that, since the GPU seems to be
>>> forced to the minimum frequency for as long as it remains parked anyway.
>>> If the GPU remains idle long enough for the RPS utilization counters to
>>> drop below the threshold and qualify for a ramp-down the RPS should send
>>> us an interrupt, at which point we will ramp down the frequency.
>>
>> Because it demonstrably and quite dramatically reduces power consumption
>> for very light desktop workloads.
>>
>>> Unconditionally ramping down on parking seems to disturb the accuracy of
>>> that RPS feedback loop, which then needs to be worked around by reducing
>>> the averaging window of the RPS to a tiny fraction of the oscillation
>>> period of any typical GPU workload, which is going to prevent the RPS
>>> from seeing a whole oscillation period before it reacts, which is almost
>>> guaranteed to have a serious energy-efficiency cost.
>>
>> There is no feedback loop in these workloads. There are no completed RPS
>> workers, they are all cancelled if they were even scheduled. (Now we
>> might be tempted to look at the results if they scheduled and take that
>> into consideration instead of unconditionally downclocking.)
>>
>
> Processing any pending RPS work seems better IMHO than unconditionally
> assuming they indicate a clock-down.
>
>>> > As for steady state workloads, I'm optimistic the smoothing helps. (It's
>>> > harder to find steady state, unthrottled workloads!)
>>>
>>> I'm curious, how is that smoothing you do in PATCH 1 better than simply
>>> setting 2x the EIs? (Which would also mean half the interrupt processing
>>> overhead as in this series)
>>
>> I'm anticipating where the RPS worker may not run until the next jiffie
>> or two, by which point the iir is stale. But yes, there's only a subtle
>> difference between comparing the last 320us every 160us and comparing
>> the last 320us every 320us.
>
> Sounds like the benefit from smoothing is very slight, considering its
> accuracy and interrupt processing overhead costs.
>
>> -Chris
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
> https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Attachment:
signature.asc
Description: PGP signature
