On Tue, Oct 08, 2013 at 04:12:02PM +0900, Minchan Kim wrote: > On Tue, Oct 08, 2013 at 12:35:33AM -0400, KOSAKI Motohiro wrote: > > (10/7/13 11:07 PM), Minchan Kim wrote: > > >Hi KOSAKI, > > > > > >On Mon, Oct 07, 2013 at 10:51:18PM -0400, KOSAKI Motohiro wrote: > > >>>Maybe, int madvise5(addr, length, MADV_DONTNEED|MADV_LAZY|MADV_SIGBUS, > > >>> &purged, &ret); > > >>> > > >>>Another reason to make it hard is that madvise(2) is tight coupled with > > >>>with vmas split/merge. It needs mmap_sem's write-side lock and it hurt > > >>>anon-vrange test performance much heavily and userland might want to > > >>>make volatile range with small unit like "page size" so it's undesireable > > >>>to make it with vma. Then, we should filter out to avoid vma split/merge > > >>>in implementation if only MADV_LAZY case? Doable but it could make code > > >>>complicated and lost consistency with other variant of madvise. > > >> > > >>I haven't seen your performance test result. Could please point out URLs? > > > > > >https://lkml.org/lkml/2013/3/12/105 > > > > It's not comparison with and without vma merge. I'm interest how much benefit > > vmas operation avoiding have. > > I had an number but lost it so I should set up it in my KVM machine :( > And I needed old kernel 3.7.0 for testing vma-based approach. > > DRAM:2G, CPU : 12 > > kernel 3.7.0 > > jemalloc: 20527 records/s > jemalloc vma based approach : 5360 records/s > > vrange call made worse because every thread stuck with mmap_sem. > > kernel 3.11.0 > > jemalloc: 21176 records/s > jemalloc vroot tree approach: 103637 records/s > > It could enhance 5 times. And please keep in mind that vrange's user might want to small vrange like PAGE_SIZE. If we go with vma-based approach, we would consume memory with lots of vm_area_struct. -- Kind regards, Minchan Kim -- 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>