Srinivas Pandruvada writes: > On Mon, 2020-07-20 at 16:20 -0700, Francisco Jerez wrote: >> "Rafael J. Wysocki" writes: >> >> > On Fri, Jul 17, 2020 at 2:21 AM Francisco Jerez < >> > currojerez@riseup.net> wrote: >> > > "Rafael J. Wysocki" writes: >> > > > {...] > >> > Overall, so far, I'm seeing a claim that the CPU subsystem can be >> > made >> > use less energy and do as much work as before (which is what >> > improving >> > the energy-efficiency means in general) if the maximum frequency of >> > CPUs is limited in a clever way. >> > >> > I'm failing to see what that clever way is, though. >> Hopefully the clarifications above help some. > > To simplify: > > Suppose I called a function numpy.multiply() to multiply two big arrays > and thread is a pegged to a CPU. Let's say it is causing CPU to > finish the job in 10ms and it is using a P-State of 0x20. But the same > job could have been done in 10ms even if it was using P-state of 0x16. > So we are not energy efficient. To really know where is the bottle neck > there are numbers of perf counters, may be cache was the issue, we > could rather raise the uncore frequency a little. A simple APRF,MPERF > counters are not enough. Yes, that's right, APERF and MPERF aren't sufficient to identify every kind of possible bottleneck, some visibility of the utilization of other subsystems is necessary in addition -- Like e.g the instrumentation introduced in my series to detect a GPU bottleneck. A bottleneck condition in an IO device can be communicated to CPUFREQ by adjusting a PM QoS latency request (link [2] in my previous reply) that effectively gives the governor permission to rearrange CPU work arbitrarily within the specified time frame (which should be of the order of the natural latency of the IO device -- e.g. at least the rendering time of a frame for a GPU) in order to minimize energy usage. > or we characterize the workload at different P-states and set limits. > I think this is not you want to say for energy efficiency with your > changes. > > The way you are trying to improve "performance" is by caller (device > driver) to say how important my job at hand. Here device driver suppose > offload this calculations to some GPU and can wait up to 10 ms, you > want to tell CPU to be slow. But the p-state driver at a movement > observes that there is a chance of overshoot of latency, it will > immediately ask for higher P-state. So you want P-state limits based on > the latency requirements of the caller. Since caller has more knowledge > of latency requirement, this allows other devices sharing the power > budget to get more or less power, and improve overall energy efficiency > as the combined performance of system is improved. > Is this correct? Yes, pretty much.