On Tue, 18 Jun 2013 14:17:25 +0000 Christoph Lameter <cl@xxxxxxxxx> wrote: > On Tue, 18 Jun 2013, Pekka Enberg wrote: > > > The changelog is way too vague. Numbers? Anyone who would want to > > use this in real world scenarios, please speak up! > > Steve? Steve's out this morning so I'll take a stab at it. This was an RT request. When we switched over to SLUB we saw an immediate overall performance boost over SLAB but encountered some 249 microsecond latency spikes when testing on large systems (40-core/256GB RAM). Latency traces showed that our spikes were SLUB's cpu_partial processing in unfreeze_partials(). We hacked up a boot script that would traverse the /sys/kernel/slab tree and write a zero to all the 'cpu_partial' entries (turning them off) but asked Christoph if he had a way to configure cpu_partial processing out, since running the script at boot did not actually catch all instances of cpu_partial. I'm sure it would be better to actually do cpu_partial processing in small chunks to avoid latency spikes in latency sensitive applications but for the short-term it's just easier to turn it off. Clark
Attachment:
signature.asc
Description: PGP signature