Messages in this thread | | | Date | Fri, 1 Aug 2008 14:13:36 -0500 | From | Robin Holt <> | Subject | Re: [Patch] Scale pidhash_shift/pidhash_size up based on num_possible_cpus(). |
| |
On Fri, Aug 01, 2008 at 11:27:20AM -0700, Eric W. Biederman wrote: > Robin Holt <holt@sgi.com> writes: > > > On Thu, Jul 31, 2008 at 03:04:56PM -0700, Eric W. Biederman wrote: > >> Robin Holt <holt@sgi.com> writes: > >> > >> > Like so??? > >> > > >> > I have not tested this yet. > >> > >> Looks reasonable to me. > >> > >> In what circumstances was the lookup in the pid hash table with > >> long changes causing a performance slowdown?. We don't perform > >> a lot of lookups. > > > > It was initially detected while profiling 'ps' on a 2048p machine that > > had 13 kernel threads per cpu. We added a couple more device drivers > > which added additional threads. We then started a pthread-on-process > > MPI job which had 2048 ranks each with 4 threads (test-case from > > customer job). There were misc other processes out there which brought > > our task count up to approx 63k. Larger page size helped the problem > > (went from 16k to 64k). > > Large page size? Do you mean larger hash size? > > What were you measuring that showed improvement with the large hash size?
Oops, confusing details. That was a different problem we had been tracking.
Sorry for the confusion, Robin
| |