On Fri, Aug 09, 2002 at 09:59:38AM -0700, Gordon Messmer wrote: > Note that the CPU affinity feature in question relates to IRQ's, not > applications.
Linux has long had irq affinity. Ingo Molnar implemented that in 2.1.x; see /proc/irq/ process cpu affinity is a different and orthogonal feature. [rml's toolset implement both irqset and taskset.] > Binding applications to a specific CPU doesn't make much > sense. It would be very rare for a pair of applications to be balanced > in CPU use enough that you'd want to bind one set to the first CPU and > the other set to a second set in a dual processor system. Not if processor affinity means cache-affinity. [And with NUMA and soon the x86-64 SUMO architecture, it will mean memory efficiency, also.] It depends on the workload and the scheduler. I have a mostly compute-bound trading application that ping-pongs rapidly between processors on Solaris 8 unless we bind it. I don't know whether Solaris has tunables to fix this, but it doesn't matter, as we are in the process of moving everything over to Linux. There are several potential disadvantages to binding, of course, both in terms of system performance, and physically, from unequal heating of the processors. It may pay to rotate the affinity at regular intervals so all of the CPU's get exercised more-or-less equally. I'll soon see how well Ingo's O(1) scheduler does on our workload ... Regards, Bill Rugolsky -- redhat-list mailing list unsubscribe mailto:[EMAIL PROTECTED]?subject=unsubscribe https://listman.redhat.com/mailman/listinfo/redhat-list
