Messages in this thread | | | Date | Tue, 14 Sep 2010 17:02:39 -0700 (PDT) | From | David Rientjes <> | Subject | Re: [PATCH v2 2/2] SLUB: Mark merged slab caches in /proc/slabinfo |
| |
On Wed, 15 Sep 2010, Pekka Enberg wrote:
> > > In my not-so-humble opinion, either the merging needs to go away > > > entirely, or the misleading output needs to be fixed. > > Cache merging may have been advertised as a bigger performance improvement > > than it actually is, and I don't do it in my own slab allocator for other > > reasons, but it does lead to more effective memory use by reducing slab > > fragmentation. On one of my benchmarking servers, over 60% of caches are > > merged and /sys/kernel/slab/.../partial reports roughly the same percent > > of fewer total partial slabs over the system in comparison to > > slub_nomerge. > Last time I checked (and it's been a while), it did reduce _internal > fragmentation_ for the naive "memory used after boot" scenario. I don't think > I ever advertised it as a performance improvement. Dunno if somebody else did. >
I believe that cache merging was advertised as being both a small cacheline optimization and a method for reducing slab fragmentation by filling up partial slabs. Regardless, I measured the latter to be much more significant and found only a very slight improvement for netperf TCP_RR on machines with large cpu counts in comparison to slub_nomerge.
I don't believe that we need to remove cache merging and allocate more memory by default to be able to identify a particular cache using an egregious amount of slab when troubleshooting a problem. That amount of memory (probably 4-8MB) is significant without kmem_cache_shrink() since, using my example where 60% of caches were merged into others, the partial list won't shrink below min_partial slabs for those ~120 caches.
I prefer my suggestion of using the unique ids of each cache in /proc/slabinfo since it won't unfairly attribute slab to caches that didn't allocate it. Sure, /proc/slabinfo can't tell you which cache is at fault without booting with slub_nomerge, enabling tracing, using SLAB_TRACE, or kmemleak, but I think that's better than sacrificing the slab fragmentation by default.
| |