On 03/03/12 05:24, Adrian Chadd wrote:

As I can see, my patch has nothing to do with the problem. My patch improves SMP load balancing, while in this case problem is different. In some cases, when not all CPUs are busy, my patch could mask the problem by using more CPUs, but not in this case when dnets consumes all available CPUs.

I still not feel very comfortable with ULE math, but as I understand, in both illustrated cases there is a conflict between clearly CPU-bound dnets threads, that consume all available CPU and never do voluntary context switches, and more or less interactive other threads. If other threads detected to be "interactive" in ULE terms, they should preempt dnets threads and everything will be fine. But "batch" (in ULE terms) threads never preempt each other, switching context only about 10 times per second, as hardcoded in sched_slice variable. Kernel build by definition consumes too much CPU time to be marked "interactive". exo-helper-1 thread in interact.out could potentially be marked "interactive", but possibly once it consumed some CPU to become "batch", it is difficult for it to get back, as waiting in a runq is not counted as sleep and each time it is getting running, it has some new work to do, so it remains "batch". May be if CPU time accounting was more precise it would work better (by accounting those short periods when threads really sleeps voluntary), but not with present sampled logic with 1ms granularity. As result, while dnets threads each time consume full 100ms time slices, other threads are starving, getting running only 10 times per second to voluntary switch out in just a few milliseconds.