On Saturday, June 08, 2013 12:28:04 PM Catalin Marinas wrote:
On Fri, Jun 07, 2013 at 07:08:47PM +0100, Preeti U Murthy wrote:
On 06/07/2013 08:21 PM, Catalin Marinas wrote:
I think you are missing Ingo's point. It's not about the scheduler complying with decisions made by various governors in the kernel (which may or may not have enough information) but rather the scheduler being in a better position for making such decisions.
My mail pointed out that I disagree with this design ("the scheduler being in a better position for making such decisions"). I think it should be a 2 way co-operation. I have elaborated below.
I agree with that.
Take the cpuidle example, it uses the load average of the CPUs, however this load average is currently controlled by the scheduler (load balance). Rather than using a load average that degrades over time and gradually putting the CPU into deeper sleep states, the scheduler could predict more accurately that a run-queue won't have any work over the next x ms and ask for a deeper sleep state from the beginning.
How will the scheduler know that there will not be work in the near future? How will the scheduler ask for a deeper sleep state?
My answer to the above two questions are, the scheduler cannot know how much work will come up. All it knows is the current load of the runqueues and the nature of the task (thanks to the PJT's metric). It can then match the task load to the cpu capacity and schedule the tasks on the appropriate cpus.
The scheduler can decide to load a single CPU or cluster and let the others idle. If the total CPU load can fit into a smaller number of CPUs it could as well tell cpuidle to go into deeper state from the beginning as it moved all the tasks elsewhere.
So why can't it do that today? What's the problem?
Regarding future work, neither cpuidle nor the scheduler know this but the scheduler would make a better prediction, for example by tracking task periodicity.
Well, basically, two pieces of information are needed to make target idle state selections: (1) when the CPU (core or package) is going to be used next time and (2) how much latency for going back to the non-idle state can be tolerated. While the scheduler knows (1) to some extent (arguably, it generally cannot predict when hardware interrupts are going to occur), I'm not really sure about (2).
As a consequence, it leaves certain cpus idle. The load of these cpus degrade. It is via this load that the scheduler asks for a deeper sleep state. Right here we have scheduler talking to the cpuidle governor.
So we agree that the scheduler _tells_ the cpuidle governor when to go idle (but not how deep).
It does indicate to cpuidle how deep it can go, however, by providing it with the information about when the CPU is going to be used next time (from the scheduler's perspective).
IOW, the scheduler drives the cpuidle decisions. Two problems: (1) the cpuidle does not get enough information from the scheduler (arguably this could be fixed)
OK, so what information is missing in your opinion?
and (2) the scheduler does not have any information about the idle states (power gating etc.) to make any informed decision on which/when CPUs should go idle.
That's correct, which is a drawback. However, on some systems it may never have that information (because hardware coordinates idle states in a way that is opaque to the OS - e.g. by autopromoting deeper states when idle for sufficiently long time) and on some systems that information may change over time (i.e. the availablility of specific idle states may depend on factors that aren't constant).
If you attempted to take all of the possible complications related to hardware designs in that area in the scheduler, you'd end up with completely unmaintainable piece of code.
As you said, it is a non-optimal one-way communication but the solution is not feedback loop from cpuidle into scheduler. It's like the scheduler managed by chance to get the CPU into a deeper sleep state and now you'd like the scheduler to get feedback form cpuidle and not disturb that CPU anymore. That's the closed loop I disagree with. Could the scheduler not make this informed decision before - it has this total load, let's get this CPU into deeper sleep state?
No, it couldn't in general, for the above reasons.
I don't see what the problem is with the cpuidle governor waiting for the load to degrade before putting that cpu to sleep. In my opinion, putting a cpu to deeper sleep states should happen gradually.
If we know in advance that the CPU can be put into idle state Cn, there is no reason to put it into anything shallower than that.
On the other hand, if the CPU is in Cn already and there is a possibility to put it into a deeper low-power state (which we didn't know about before), it may make sense to promote it into that state (if that's safe) or even wake it up and idle it again.
This means time will tell the governors what kinds of workloads are running on the system. If the cpu is idle for long, it probably means that the system is less loaded and it makes sense to put the cpus to deeper sleep states. Of course there could be sporadic bursts or quieting down of tasks, but these are corner cases.
It's nothing wrong with degrading given the information that cpuidle currently has. It's a heuristics that worked ok so far and may continue to do so. But see my comments above on why the scheduler could make more informed decisions.
We may not move all the power gating information to the scheduler but maybe find a way to abstract this by giving more hints via the CPU and cache topology. The cpuidle framework (it may not be much left of a governor) would then take hints about estimated idle time and invoke the low-level driver about the right C state.
Overall, it looks like it'd be better to split the governor "layer" between the scheduler and the idle driver with a well defined interface between them. That interface needs to be general enough to be independent of the underlying hardware.
We need to determine what kinds of information should be passed both ways and how to represent it.
Of course, you could export more scheduler information to cpuidle, various hooks (task wakeup etc.) but then we have another framework, cpufreq. It also decides the CPU parameters (frequency) based on the load controlled by the scheduler. Can cpufreq decide whether it's better to keep the CPU at higher frequency so that it gets to idle quicker and therefore deeper sleep states? I don't think it has enough information because there are at least three deciding factors (cpufreq, cpuidle and scheduler's load balancing) which are not unified.
Why not? When the cpu load is high, cpu frequency governor knows it has to boost the frequency of that CPU. The task gets over quickly, the CPU goes idle. Then the cpuidle governor kicks in to put the CPU to deeper sleep state gradually.
The cpufreq governor boosts the frequency enough to cover the load, which means reducing the idle time. It does not know whether it is better to boost the frequency twice as high so that it gets to idle quicker. You can change the governor's policy but does it have any information from cpuidle?
Well, it may get that information directly from the hardware. Actually, intel_pstate does that, but intel_pstate is the governor and the scaling driver combined.
Meanwhile the scheduler should ensure that the tasks are retained on that CPU,whose frequency is boosted and should not load balance it, so that they can get over quickly. This I think is what is missing. Again this comes down to the scheduler taking feedback from the CPU frequency governors which is not currently happening.
Same loop again. The cpu load goes high because (a) there is more work, possibly triggered by external events, and (b) the scheduler decided to balance the CPUs in a certain way. As for cpuidle above, the scheduler has direct influence on the cpufreq decisions. How would the scheduler know which CPU not to balance against? Are CPUs in a cluster synchronous? Is it better do let other CPU idle or more efficient to run this cluster at half-speed?
Let's say there is an increase in the load, does the scheduler wait until cpufreq figures this out or tries to take the other CPUs out of idle? Who's making this decision? That's currently a potentially unstable loop.
Yes, it is and I don't think we currently have good answers here.
The results of many measurements seem to indicate that it generally is better to do the work as quickly as possible and then go idle again, but there are costs associated with going back and forth from idle to non-idle etc.
The main problem with cpufreq that I personally have is that the governors carry out their own sampling with pretty much arbitrary resolution that may lead to suboptimal decisions. It would be much better if the scheduler indicated when to *consider* the changing of CPU performance parameters (that may not be frequency alone and not even frequency at all in general), more or less the same way it tells cpuidle about idle CPUs, but I'm not sure if it should decide what performance points to run at.
I would repeat here that today we interface cpuidle/cpufrequency policies with scheduler but not the other way around. They do their bit when a cpu is busy/idle. However scheduler does not see that somebody else is taking instructions from it and comes back to give different instructions!
The key here is that cpuidle/cpufreq make their primary decision based on something controlled by the scheduler: the CPU load (via run-queue balancing). You would then like the scheduler take such decision back into account. It just looks like a closed loop, possibly 'unstable' .
Why? Why would you call a scheduler->cpuidle->cpufrequency interaction a closed loop and not the new_scheduler = scheduler+cpuidle+cpufrequency a closed loop? Here too the scheduler should be made well aware of the decisions it took in the past right?
It's more like:
scheduler -> cpuidle/cpufreq -> hardware operating point ^ | +--------------------------------------+
You can argue that you can make an adaptive loop that works fine but there are so many parameters that I don't see how it would work. The patches so far don't seem to address this. Small task packing, while useful, it's some heuristics just at the scheduler level.
I agree.
With a combined decision maker, you aim to reduce this separate decision process and feedback loop. Probably impossible to eliminate the loop completely because of hardware latencies, PLLs, CPU frequency not always the main factor, but you can make the loop more tolerant to instabilities.
Well, in theory. :-)
Another question to ask is whether or not the structure of our software reflects the underlying problem. I mean, on the one hand there is the scheduler that needs to optimally assign work items to computational units (hyperthreads, CPU cores, packages) and on the other hand there's hardware with different capabilities (idle states, performance points etc.). Arguably, the scheduler internals cannot cover all of the differences between all of the existing types of hardware Linux can run on, so there needs to be a layer of code providing an interface between the scheduler and the hardware. But that layer of code needs to be just *one*, so why do we have *two* different frameworks (cpuidle and cpufreq) that talk to the same hardware and kind of to the scheduler, but not to each other?
To me, the reason is history, and more precisely the fact that cpufreq had been there first, then came cpuidle and only then poeple started to realize that some scheduler tweaks may allow us to save energy without sacrificing too much performance. However, it looks like there's time to go back and see how we can integrate all that. And there's more, because we may need to take power budgets and thermal management into account as well (i.e. we may not be allowed to use full performance of the processors all the time because of some additional limitations) and the CPUs may be members of power domains, so what we can do with them may depend on the states of other devices.
So I think we either (a) come up with 'clearer' separation of responsibilities between scheduler and cpufreq/cpuidle
I agree with this. This is what I have been emphasizing, if we feel that the cpufrequency/ cpuidle subsystems are suboptimal in terms of the information that they use to make their decisions, let us improve them. But this will not yield us any improvement if the scheduler does not have enough information. And IMHO, the next fundamental information that the scheduler needs should come from cpufreq and cpuidle.
What kind of information? Your suggestion that the scheduler should avoid loading a CPU because it went idle is wrong IMHO. It went idle because the scheduler decided this in first instance.
Then we should move onto supplying scheduler information from the power domain topology, thermal factors, user policies.
I agree with this but at this point you get the scheduler to make more informed decisions about task placement. It can then give more precise hints to cpufreq/cpuidle like the predicted load and those frameworks could become dumber in time, just complying with the requested performance level (trying to break the loop above).
Well, there's nothing like "predicted load". At best, we may be able to make more or less educated guesses about it, so in my opinion it is better to use the information about what happened in the past for making decisions regarding the current settings and re-adjust them over time as we get more information.
So how much decision making regarding the idle state to put the given CPU into should be there in the scheduler? I believe the only information coming out of the scheduler regarding that should be "OK, this CPU is now idle and I'll need it in X nanoseconds from now" plus possibly a hint about the wakeup latency tolerance (but those hints may come from other places too). That said the decision *which* CPU should become idle at the moment very well may require some information about what options are available from the layer below (for example, "putting core X into idle for Y of time will save us Z energy" or something like that).
And what about performance scaling? Quite frankly, in my opinion that requires some more investigation, because there still are some open questions in that area. To start with we can just continue using the current heuristics, but perhaps with the scheduler calling the scaling "governor" when it sees fit instead of that "governor" running kind of in parallel with it.
or (b) come up with a unified load-balancing/cpufreq/cpuidle implementation as per Ingo's request. The latter is harder but, with a good design, has potentially a lot more benefits.
A possible implementation for (a) is to let the scheduler focus on performance load-balancing but control the balance ratio from a cpufreq governor (via things like arch_scale_freq_power() or something new). CPUfreq would not be concerned just with individual CPU load/frequency but also making a decision on how tasks are balanced between CPUs based on the overall load (e.g. four CPUs are enough for the current load, I can shut the other four off by telling the scheduler not to use them).
As for Ingo's preferred solution (b), a proposal forward could be to factor the load balancing out of kernel/sched/fair.c and provide an abstract interface (like load_class?) for easier extending or different policies (e.g. small task packing).
Let me elaborate on the patches that have been posted so far on the power awareness of the scheduler. When we say *power aware scheduler* what exactly do we want it to do?
In my opinion, we want it to *avoid touching idle cpus*, so as to keep them in that state longer and *keep more power domains idle*, so as to yield power savings with them turned off. The patches released so far are striving to do the latter. Correct me if I am wrong at this.
Don't take me wrong, task packing to keep more power domains idle is probably in the right direction but it may not address all issues. You realised this is not enough since you are now asking for the scheduler to take feedback from cpuidle. As I pointed out above, you try to create a loop which may or may not work, especially given the wide variety of hardware parameters.
Also feel free to point out any other expectation from the power aware scheduler if I am missing any.
If the patches so far are enough and solved all the problems, you are not missing any. Otherwise, please see my view above.
Please define clearly what the scheduler, cpufreq, cpuidle should be doing and what communication should happen between them.
If I have got Ingo's point right, the issues with them are that they are not taking a holistic approach to meet the said goal.
Probably because scheduler changes, cpufreq and cpuidle are all trying to address the same thing but independent of each other and possibly conflicting.
Keeping more power domains idle (by packing tasks) would sound much better if the scheduler has taken all aspects of doing such a thing into account, like
- How idle are the cpus, on the domain that it is packing
- Can they go to turbo mode, because if they do,then we cant pack
tasks. We would need certain cpus in that domain idle. 3. Are the domains in which we pack tasks power gated? 4. Will there be significant performance drop by packing? Meaning do the tasks share cpu resources? If they do there will be severe contention.
So by this you add a lot more information about the power configuration into the scheduler, getting it to make more informed decisions about task scheduling. You may eventually reach a point where cpuidle governor doesn't have much to do (which may be a good thing) and reach Ingo's goal.
That's why I suggested maybe starting to take the load balancing out of fair.c and make it easily extensible (my opinion, the scheduler guys may disagree). Then make it more aware of topology, power configuration so that it makes the right task placement decision. You then get it to tell cpufreq about the expected performance requirements (frequency decided by cpufreq) and cpuidle about how long it could be idle for (you detect a periodic task every 1ms, or you don't have any at all because they were migrated, the right C state being decided by the governor).
There is another angle to look at that as I said somewhere above.
What if we could integrate cpuidle with cpufreq so that there is one code layer representing what the hardware can do to the scheduler? What benefits can we get from that, if any?
Rafael