On Sun, Nov 17, 2024 at 10:56:21AM +0800, Yafang Shao wrote: > On Fri, Nov 15, 2024 at 9:41 PM Michal Koutný <mkoutny@xxxxxxxx> wrote: > > > The load balancer is malfunctioning due to the exclusion of IRQ time from > > > CPU utilization calculations. > > > > Could this be fixed by subtracting (global) IRQ time from (presumed > > total) system capacity that the balancer uses for its decisions? (i.e. > > without exact per-cgroup breakdown of IRQ time) > > The issue here is that the global IRQ time may include the interrupted > time of tasks outside the target cgroup. As a result, I don't believe > it's possible to find a reliable solution without modifying the > kernel. Since there is no relation between the interrupt and the interrupted task (and through that its cgroup) -- all time might or might not be part of your cgroup of interest. Consider it a random distribution if you will. What Michael suggests seems no less fair, and possible more fair than what you propose: \Sum cgroup = total - IRQ As opposed to what you propose: \Sum (cgroup + cgroup-IRQ) = total - remainder-IRQ Like I argued earlier, if you have two cgroups, one doing a while(1) loop (proxy for doing computation) and one cgroup doing heavy IO or networking, then per your accounting the computation cgroup will get a significant amount of IRQ time 'injected', even though it is effidently not of that group. Injecting 'half' of the interrupts in the computation group, and missing 'half' of the interrupts from the network group will get 'wrong' load-balance results too. I remain unconvinced that any of this makes sense.
