Re: [UnifiedV4 00/16] The Unified slab allocator (V4)

[Pekka Enberg]

(Adding more people who've taken interest in slab performance in the
past to CC.)

On Tue, Oct 5, 2010 at 9:57 PM, Christoph Lameter <cl@xxxxxxxxx> wrote:
> V3->V4:
> - Lots of debugging
> - Performance optimizations (more would be good)...
> - Drop per slab locking in favor of per node locking for
>  partial lists (queuing implies freeing large amounts of objects
>  to per node lists of slab).
> - Implement object expiration via reclaim VM logic.
>
> The following is a release of an allocator based on SLAB
> and SLUB that integrates the best approaches from both allocators. The
> per cpu queuing is like in SLAB whereas much of the infrastructure
> comes from SLUB.
>
> After this patches SLUB will track the cpu cache contents
> like SLAB attemped to. There are a number of architectural differences:
>
> 1. SLUB accurately tracks cpu caches instead of assuming that there
>   is only a single cpu cache per node or system.
>
> 2. SLUB object expiration is tied into the page reclaim logic. There
>   is no periodic cache expiration.
>
> 3. SLUB caches are dynamically configurable via the sysfs filesystem.
>
> 4. There is no per slab page metadata structure to maintain (aside
>   from the object bitmap that usually fits into the page struct).
>
> 5. Has all the resiliency and diagnostic features of SLUB.
>
> The unified allocator is a merging of SLUB with some queuing concepts from
> SLAB and a new way of managing objects in the slabs using bitmaps. Memory
> wise this is slightly more inefficient than SLUB (due to the need to place
> large bitmaps --sized a few words--in some slab pages if there are more
> than BITS_PER_LONG objects in a slab) but in general does not increase space
> use too much.
>
> The SLAB scheme of not touching the object during management is adopted.
> The unified allocator can efficiently free and allocate cache cold objects
> without causing cache misses.
>
> Some numbers using tcp_rr on localhost
>
>
> Dell R910 128G RAM, 64 processors, 4 NUMA nodes
>
> threads unified         slub            slab
> 64      4141798         3729037         3884939
> 128     4146587         3890993         4105276
> 192     4003063         3876570         4110971
> 256     3928857         3942806         4099249
> 320     3922623         3969042         4093283
> 384     3827603         4002833         4108420
> 448     4140345         4027251         4118534
> 512     4163741         4050130         4122644
> 576     4175666         4099934         4149355
> 640     4190332         4142570         4175618
> 704     4198779         4173177         4193657
> 768     4662216         4200462         4222686

Are there any stability problems left? Have you tried other benchmarks
(e.g. hackbench, sysbench)? Can we merge the series in smaller
batches? For example, if we leave out the NUMA parts in the first
stage, do we expect to see performance regressions?
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/