On Thu 31-08-17 19:32:08, Wang Shilong wrote: > Hello, > > On Thu, Aug 31, 2017 at 5:48 PM, Jan Kara <jack@xxxxxxx> wrote: > > On Thu 31-08-17 17:09:26, Wang Shilong wrote: > > This is without your patch: > > > > no Quota quota quota,project > > 1 851,207 341,941 178,686 > > 2 850,368 342,233 191,755 > > 3 848,877 342,768 193,807 > > > > With your patchset: > > no quota quota quota,project > > 1 853,391 448,378 385,292 > > 2 851,379 448,375 407,716 > > 3 850,203 448,415 406,813 > > > > We still see creation regression here, but your patchset help a lot. > > OK, still quite some way to go... There will always be some cost associated > with quota bookkeeping especially since this has to be synchronized across > CPUs to allow for reliable limits checking. But 50% seems quite a bit. > > > We run tests with quota enabled, and collecting lock statics using 'perf > lock' > > Check IO/CPU overload! Actually, this doesn't look bad. Only 1626 contentions on s_inode_list_lock, and we totally waited only for 5ms over all CPUs. So I'm pretty confident contention on that lock is not costing us much. 'perf lock' seems to show only spinlocks while I think the real contention is going to be on sleeping locks. You should be able to see that using lockstat framework in the kernel (Documentation/locking/lockstat.txt). Honza > > Name acquired contended avg wait (ns) total wait (ns) max wait (ns) > min wait (ns) > > &(&s->s_inode_li... 16850 1626 3060 4976891 > 77915 1297 > &(&lru->node[i].... 8332 434 3506 1521648 > 317040 1227 > &(&journal->j_li... 3724 105 4784 502326 > 64196 1311 > sparse_irq_lock 3248 0 0 0 > 0 0 > &(&lock->wait_lo... 2636 144 10178 1465730 > 167087 1345 > &(&lock->wait_lo... 2184 81 11999 971986 > 272372 1369 > &(&zone->lock)->... 2010 51 2080 106127 > 4945 876 > &(&dentry->d_loc... 1945 0 0 0 > 0 0 > &(&dentry->d_loc... 1934 0 0 0 > 0 0 > &(&dentry->d_loc... 1933 0 0 0 > 0 0 > &(&dentry->d_loc... 1918 0 0 0 > 0 0 > &(&dentry->d_loc... 1915 0 0 0 > 0 0 > &(&dentry->d_loc... 1910 0 0 0 > 0 0 > &(&dentry->d_loc... 1906 0 0 0 > 0 0 > > > It looks we are still contending on @s_inode_list_lock which related to quota. > > > Thanks, > Shilong > > > Honza > > -- > Jan Kara <jack@xxxxxxxx> > SUSE Labs, CR > > -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
