Messages in this thread | | | Date | Tue, 7 Apr 2009 15:01:45 -0700 (PDT) | From | David Rientjes <> | Subject | Re: [patch] slub: default min_partial to at least highest cpus per node |
| |
On Tue, 7 Apr 2009, Pekka Enberg wrote:
> Yeah, something more dynamic makes sense. But we probably need to do > ilog2(nr_cpu_ids) or something; otherwise we will have very long partial lists > on big iron machines (think 4096 cpus here). >
The suggestion was to replace MAX_PARTIAL with nr_cpu_ids, not as the default value for n->min_partial, which is ilog2(size). To exceed the current value of MAX_PARTIAL, which is 10, size would need to be 2K or greater.
I agree that the default n->min_partial would be better represented in terms of cpus than object size. ilog2(nr_cpu_ids) certainly seems appropriate and free partial slabs can always be reclaimed by utilizing /sys/kernel/slab/cache/shrink.
The 4096 cpu example is certainly realistic, but probably not UMA.
| |