Messages in this thread | | | Date | Fri, 13 Jul 2001 15:43:05 -0700 | From | Mike Kravetz <> | Subject | Re: CPU affinity & IPI latency |
| |
On Fri, Jul 13, 2001 at 12:51:53PM -0700, David Lang wrote: > A real-world example of this issue. > > I was gzipping a large (~800MB) file on a dual athlon box. the gzip prcess > was bouncing back and forth between the two CPUs. I actually was able to > gzip faster by starting up setiathome to keep one CPU busy and force the > scheduler to keep the gzip on a single CPU (I ran things several times to > verify it was actually faster) > > David Lang
That does sound like the same behavior I was seeing with lat_ctx. Like I said in my previous note, the scheduler does try to take CPU affinity into account. reschedule_idle() does a pretty good job of determining what CPU a task should run on. In the case of lat_ctx (and I believe your use of gzip), the 'fast path' in reschedule_idle() is taken because the CPU associated with the awakened task is idle. However, before schedule() is run on the 'target' CPU, schedule() is run on another CPU and the task is scheduled there.
The root cause of this situation is the delay between the time reschedule_idle() determines what is the best CPU a task should run on, and the time schedule() is actually ran on that CPU.
I have toyed with the idea of 'temporarily binding' a task to a CPU for the duration of the delay between reschedule_idle() and schedule(). It would go something like this,
- Define a new field in the task structure 'saved_cpus_allowed'. With a little collapsing of existing fields, there is room to put this on the same cache line as 'cpus_allowed'. - In reschedule_idle() if we determine that the best CPU for a task is the CPU it is associated with (p->processor), then temporarily bind the task to that CPU. The task is temporarily bound to the CPU by overwriting the 'cpus_allowed' field such that the task can only be scheduled on the target CPU. Of course, the original value of 'cpus_allowed' is saved in 'saved_cpus_allowed'. - In schedule(), the loop which examines all tasks on the runqueue will restore the value of 'cpus_allowed'.
This would preserve the 'best CPU' decision made by reschedule_idle(). On the down side, 'temporarily bound' tasks could not be scheduled until schedule() is run on their associated CPUs. This could potentially waste CPU cycles, and delay context switches. In addition, it is quite possible that while a task is 'temporarily bound' the state of the system could change in such a way that the best CPU is no longer best.
There appears to be a classic tradeoff between CPU affinity and context switch time.
Comments?
-- Mike Kravetz mkravetz@sequent.com IBM Linux Technology Center - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |