On 2017年08月15日 18:36, Jesper Dangaard Brouer wrote: > On Tue, 15 Aug 2017 16:45:34 +0800 > Kemi Wang <kemi.wang@xxxxxxxxx> wrote: > >> Each page allocation updates a set of per-zone statistics with a call to >> zone_statistics(). As discussed in 2017 MM submit, these are a substantial > ^^^^^^ should be "summit" Hi, Jesper Thanks for reporting this issue and providing the benchmark to test raw performance of page allocation. It is really quite helpful to figure out the root cause. >> source of overhead in the page allocator and are very rarely consumed. This >> significant overhead in cache bouncing caused by zone counters (NUMA >> associated counters) update in parallel in multi-threaded page allocation >> (pointed out by Dave Hansen). > > Hi Kemi > > Thanks a lot for following up on this work. A link to the MM summit slides: > http://people.netfilter.org/hawk/presentations/MM-summit2017/MM-summit2017-JesperBrouer.pdf > Thanks for adding the link here. I should have done that in this cover letter. >> To mitigate this overhead, this patchset separates NUMA statistics from >> zone statistics framework, and update NUMA counter threshold to a fixed >> size of 32765, as a small threshold greatly increases the update frequency >> of the global counter from local per cpu counter (suggested by Ying Huang). >> The rationality is that these statistics counters don't need to be read >> often, unlike other VM counters, so it's not a problem to use a large >> threshold and make readers more expensive. >> >> With this patchset, we see 26.6% drop of CPU cycles(537-->394, see below) >> for per single page allocation and reclaim on Jesper's page_bench03 >> benchmark. Meanwhile, this patchset keeps the same style of virtual memory >> statistics with little end-user-visible effects (see the first patch for >> details), except that the number of NUMA items in each cpu >> (vm_numa_stat_diff[]) is added to zone->vm_numa_stat[] when a user *reads* >> the value of NUMA counter to eliminate deviation. > > I'm very happy to see that you found my kernel module for benchmarking useful :-) > >> I did an experiment of single page allocation and reclaim concurrently >> using Jesper's page_bench03 benchmark on a 2-Socket Broadwell-based server >> (88 processors with 126G memory) with different size of threshold of pcp >> counter. >> >> Benchmark provided by Jesper D Broucer(increase loop times to 10000000): > ^^^^^^^ > You mis-spelled my last name, it is "Brouer". > Dear Jesper, I am so sorry about it, please forgive me :) >> https://github.com/netoptimizer/prototype-kernel/tree/master/kernel/mm/bench >> >> Threshold CPU cycles Throughput(88 threads) >> 32 799 241760478 >> 64 640 301628829 >> 125 537 358906028 <==> system by default >> 256 468 412397590 >> 512 428 450550704 >> 4096 399 482520943 >> 20000 394 489009617 >> 30000 395 488017817 >> 32765 394(-26.6%) 488932078(+36.2%) <==> with this patchset >> N/A 342(-36.3%) 562900157(+56.8%) <==> disable zone_statistics >> >> Kemi Wang (2): >> mm: Change the call sites of numa statistics items >> mm: Update NUMA counter threshold size >> >> drivers/base/node.c | 22 ++++--- >> include/linux/mmzone.h | 25 +++++--- >> include/linux/vmstat.h | 33 ++++++++++ >> mm/page_alloc.c | 10 +-- >> mm/vmstat.c | 162 +++++++++++++++++++++++++++++++++++++++++++++++-- >> 5 files changed, 227 insertions(+), 25 deletions(-) >> > > > -- 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/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
