Re: [patch] slub: default min_partial to at least highest cpus per node

[Pekka Enberg <penberg@cs.helsinki.fi>]

Hi David,

On Tue, 7 Apr 2009, Pekka Enberg wrote:
>>> I'd be just as happy with the following, although it would require changing
>>> MIN_PARTIAL to be greater than its default of 5 if a node supports more cpus
>>> for optimal performance (the old patch did that automatically up to
>>> MAX_PARTIAL).
>> Hmm but why not move ->min_partial to struct kmem_cache_node as I suggested
>> and make sure it's adjusted properly as with nr_cpus_node()?

David Rientjes wrote:
> Sure, that's also possible except we'd lose the ability to tune 
> min_partial with /sys/kernel/slab/cache/min_partial, unless it would then 
> change n->min_partial for each N_NORMAL_MEMORY node.  We lack an interface 
> to change the per-node min_partial.
> 
> If you think that's acceptable, I'd be just as satisfied with that 
> approach as long as all archs have valid cpu_to_node() mappings at the 
> time of CPU_UP_PREPARE.

Well, that doesn't change the current behavior, so sure, I think it's 
acceptable. And if the new defaults seem reasonable enough, we can 
probably get rid of the tunable altogether.

David Rientjes wrote:
> Aside: we're lacking in the documentation of these sysfs tunables such as 
> remote_node_defrag_ratio to begin with, the only way to figure out what it 
> does is by reading the code or making assumptions based on its name.  I'd 
> be happy to add some documentation but it'd be good to keep it separate 
> from Documentation/vm/slub.txt.

AFAICT, the Documentation/ABI directory is the right place for this kind 
of stuff.

			Pekka