On Mon, 1 Aug 2011, Pekka Enberg wrote: > > More interesting than the perf report (which just shows kfree, > > kmem_cache_free, kmem_cache_alloc dominating) is the statistics that are > > exported by slub itself, it shows the "slab thrashing" issue that I > > described several times over the past few years. It's difficult to > > address because it's a result of slub's design. From the client side of > > 160 netperf TCP_RR threads for 60 seconds: > > > > cache alloc_fastpath alloc_slowpath > > kmalloc-256 10937512 (62.8%) 6490753 > > kmalloc-1024 17121172 (98.3%) 303547 > > kmalloc-4096 5526281 11910454 (68.3%) > > > > cache free_fastpath free_slowpath > > kmalloc-256 15469 17412798 (99.9%) > > kmalloc-1024 11604742 (66.6%) 5819973 > > kmalloc-4096 14848 17421902 (99.9%) > > > > With those stats, there's no way that slub will even be able to compete > > with slab because it's not optimized for the slowpath. > > Is the slowpath being hit more often with 160 vs 16 threads? Here's the same testing environment with CONFIG_SLUB_STATS for 16 threads instead of 160: cache alloc_fastpath alloc_slowpath kmalloc-256 4263275 (91.1%) 417445 kmalloc-1024 4636360 (99.1%) 42091 kmalloc-4096 2570312 (54.4%) 2155946 cache free_fastpath free_slowpath kmalloc-256 210115 4470604 (95.5%) kmalloc-1024 3579699 (76.5%) 1098764 kmalloc-4096 67616 4658678 (98.6%) Keep in mind that this is a default slub configuration, so kmalloc-256 has order-1 slabs and both kmalloc-1k and kmalloc-4k have order-3 slabs. If those were decreased, the free slowpath would become even worse, and if those were increased, the alloc slowpath would become even worse. I could probably get better numbers for 160 threads here if I let the free slowpath fall off the charts for kmalloc-256 and kmalloc-4k (which wouldn't be that bad, they're used 99.9% of the time) and make the alloc slowpath much easier to allocate order-0 slabs. It depends on how often we free to a partial slab, but it's a pointless exercise since users won't tune their slab allocator settings for specific caches or each workload. With regard to kmalloc-256 and kmalloc-4k on the 16 thread experiment, the lionshare of the allocations and free fastpath usage comes on the cpu taking the networking irq, whereas kmalloc-1k, the lionshare of free slowpath usage comes from that cpu. > As I said, > the problem you mentioned looks like a *scaling issue* to me which is > actually somewhat surprising. I knew that the slowpaths were slow but I > haven't seen this sort of data before. > Well, shoot, I wrote a patchset for it and presented similar data two years ago: https://lkml.org/lkml/2009/3/30/14 (back then, kmalloc-2k was part of the culprit and now it's kmalloc-4k). Although I agree that we don't want to rely on the heuristics that I created in that patchset for things like partial list ordering and it's probably not great to have an increment on a kmem_cache_cpu variable in the allocation fastpath, I still strongly advocate for some logic that only picks off a partial slab from while holding the per-node list_lock when it has a certain threshold of free objects, otherwise we keep pulling a partial slab that may have one object free and performance suffers. That logic is part of the patchset that I proposed back then and it helped performance, but that still comes at the cost of increased memory because we'd be allocating new slabs (and potentially order-3 as seen above) instead of utilizing sufficient partial slabs when the number of object allocations are low. I'm thinking this is part of the reason that Nick really advocated for optimizing for frees on remote cpus in slqb as a fundamental principle of the allocator's design. > I snipped the 'SLUB can never compete with SLAB' part because I'm > frankly more interested in raw data I can analyse myself. I'm hoping to > the per-CPU partial list patch queued for v3.2 soon and I'd be > interested to know how much I can expect that to help. > See my comment about having no doubt that you can improve performance of slub by throwing more memory in its direction, that is part of what the per-cpu partial list patchset does. Christoph posted it as an RFC and listed a few significant disadvantages to that approach, but I'm still happy to review it and see what can come of it. >From what I remember, though, each per-cpu partial list had a min_partial of half of what it currently is per-node. On my testing environment that I've been using here, they were stated to be two 16-core, 4 node systems for netperf client and server. kmalloc-256 currently has a min_partial of 8, and both kmalloc-1k and kmalloc-4k have min_partial of 10 for its current design of per-node partial lists, so that means we keep at minimum (absent kmem_cache_shrink() or reclaim) 8*4 kmalloc-256, 10*4 kmalloc-1k, and 10*4 kmalloc-4k empty slabs on the partial lists for later use on each of these systems. With the per-cpu partial lists the way I remember it, that would become 4*16 kmalloc-256, 5*16 kmalloc-1k, and 5*16 kmalloc-4k empty slabs on the partial lists. So now we've doubled the amount of memory we've reserved for the partial lists, so yeah, I'd expect better performance as a result of using (4*16 - 8*4) more order-1 slabs and 2 * (5*16 - 10*4) more order-3 slabs, about 700 pages for just those two caches systemwide. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>