Hi Michal, Thanks for the thorough reviews for these 3 patches! On Fri, May 15, 2020 at 03:41:25PM +0800, Michal Hocko wrote: > On Fri 08-05-20 15:25:17, Feng Tang wrote: > > When checking a performance change for will-it-scale scalability > > mmap test [1], we found very high lock contention for spinlock of > > percpu counter 'vm_committed_as': > > > > 94.14% 0.35% [kernel.kallsyms] [k] _raw_spin_lock_irqsave > > 48.21% _raw_spin_lock_irqsave;percpu_counter_add_batch;__vm_enough_memory;mmap_region;do_mmap; > > 45.91% _raw_spin_lock_irqsave;percpu_counter_add_batch;__do_munmap; > > > > Actually this heavy lock contention is not always necessary. The > > 'vm_committed_as' needs to be very precise when the strict > > OVERCOMMIT_NEVER policy is set, which requires a rather small batch > > number for the percpu counter. > > > > So lift the batch number to 16X for OVERCOMMIT_ALWAYS and > > OVERCOMMIT_GUESS policies, and add a sysctl handler to adjust it > > when the policy is reconfigured. > > Increasing the batch size for weaker overcommit modes makes sense. But > your patch is also tuning OVERCOMMIT_NEVER without any explanation why > that is still "small enough to be precise". Actually, it keeps the batch algorithm for "OVERCOMMIT_NEVER", but change the other 2 policies, which I should set it clear in the commit log. > > Benchmark with the same testcase in [1] shows 53% improvement on a > > 8C/16T desktop, and 2097%(20X) on a 4S/72C/144T server. And no change > > for some platforms, due to the test mmap size of the case is bigger > > than the batch number computed, though the patch will help mmap/munmap > > generally. > > > > [1] https://lkml.org/lkml/2020/3/5/57 > > Please do not use lkml.org links in the changelog. Use > http://lkml.kernel.org/r/$msg instead. Thanks, will keep that in mind for this and future patches. > > Signed-off-by: Feng Tang <feng.tang@xxxxxxxxx> > > s32 vm_committed_as_batch = 32; > > > > -static void __meminit mm_compute_batch(void) > > +void mm_compute_batch(void) > > { > > u64 memsized_batch; > > s32 nr = num_present_cpus(); > > s32 batch = max_t(s32, nr*2, 32); > > - > > - /* batch size set to 0.4% of (total memory/#cpus), or max int32 */ > > - memsized_batch = min_t(u64, (totalram_pages()/nr)/256, 0x7fffffff); > > + unsigned long ram_pages = totalram_pages(); > > + > > + /* > > + * For policy of OVERCOMMIT_NEVER, set batch size to 0.4% > > + * of (total memory/#cpus), and lift it to 6.25% for other > > + * policies to easy the possible lock contention for percpu_counter > > + * vm_committed_as, while the max limit is INT_MAX > > + */ > > + if (sysctl_overcommit_memory == OVERCOMMIT_NEVER) > > + memsized_batch = min_t(u64, ram_pages/nr/256, INT_MAX); > > + else > > + memsized_batch = min_t(u64, ram_pages/nr/16, INT_MAX); Also as you mentioned there are real-world work loads with big mmap size and multi-threading, can we lift it even further ? memsized_batch = min_t(u64, ram_pages/nr/4, INT_MAX) Thanks, Feng
