From: "Nakajima, Jun" <[email protected]>
> We have found some performance regressions (e.g. SPECjbb) with the
> scheduler on a large IA-64 NUMA machine, and we are debugging it. On SMP
> machines, we haven't seen performance regressions.
I've run a flavor of AIM7 and kernbench on a variety of ia64-NUMA CPU counts,
ranging up to 128p (and my 2.6.4 + Piggin-scheduler kernel hangs during boot
at >=192p, vs. 2.6.4 + vanilla-scheduler booting and running at 512p), and the
Piggin scheduler is 10-15% slower at 64p, and even worse at 128p. I did,
however, produce superior performance (superior to the vanilla scheduler) with
AIM7 by increasing the sched_domain->busy_factor by 4x. I tried a few other
things, but nothing improved performance on these big systems better than a
simple increase of busy_factor.
John Hawkes