Hi, On Thu 03-08-17 22:39:51, Wang Shilong wrote: > Please send me patches, we could test and response you! So I finally have something which isn't obviously wrong (it survives basic testing and gives me improvements for some workloads). I have pushed out the patches to: git://git.kernel.org/pub/scm/linux/kernel/git/jack/linux-fs.git quota_scaling I'd be happy if you can share your results with my patches. I have not yet figured out a safe way to reduce the contention on dq_lock during update of on-disk structure when lot of processes bang single dquot. I have experimental patch but it didn't bring any benefit in my testing - I'll rebase it on top of other patches I have send it to you for some testing. Honza > On Thu, Aug 3, 2017 at 10:36 PM, Jan Kara <jack@xxxxxxx> wrote: > > Hello! > > > > On Thu 03-08-17 19:31:04, Wang Shilong wrote: > >> We DDN is investigating the same issue! > >> > >> Some comments comes: > >> > >> On Thu, Aug 3, 2017 at 1:52 AM, Andrew Perepechko <anserper@xxxxxxxxx> wrote: > >> >> On Tue 01-08-17 15:02:42, Jan Kara wrote: > >> >> > Hi Andrew, > >> >> > > >> >> I've been experimenting with this today but this idea didn't bring any > >> >> benefit in my testing. Was your setup with multiple users or a single user? > >> >> Could you give some testing to my patches to see whether they bring some > >> >> benefit to you? > >> >> > >> >> Honza > >> > > >> > Hi Jan! > >> > > >> > My setup was with a single user. Unfortunately, it may take some time before > >> > I can try a patched kernel other than RHEL6 or RHEL7 with the same test, > >> > we have a lot of dependencies on these kernels. > >> > > >> > The actual test we ran was mdtest. > >> > > >> > By the way, we had 15+% performance improvement in creates from the > >> > change that was discussed earlier in this thread: > >> > > >> > EXT4_SB(dquot->dq_sb)->s_qf_names[GRPQUOTA]) { > >> > + if (test_bit(DQ_MOD_B, &dquot->dq_flags)) > >> > + return 0; > >> > >> I don't think this is right, as far as i understand, journal quota need go > >> together with quota space change update inside same transaction, this will > >> break consistency if power off or RO happen. > >> > >> Here is some ideas that i have thought: > >> > >> 1) switch dqio_mutex to a read/write lock, especially, i think most of > >> time journal quota updates is in-place update, that means we don't need > >> change quota tree in memory, firstly try read lock, retry with write lock if > >> there is real tree change. > >> > >> 2)another is similar idea of Andrew's walkaround, but we need make correct > >> fix, maintain dirty list for per transaction, and gurantee quota updates are > >> flushed when commit transaction, this might be complex, i am not very > >> familiar with JBD2 codes. > >> > >> It will be really nice if we could fix this regression, as we see 20% performace > >> regression. > > > > So I have couple of patches: > > > > 1) I convert dqio_mutex do rw semaphore and use it in exclusive mode only > > when quota tree is going to change. We also use dq_lock to serialize writes > > of dquot - you cannot have two writes happening in parallel as that could > > result in stale data being on disk. This patch brings benefit when there > > are multiple users - now they don't contend on common lock. It shows > > advantage in my testing so I plan to merge these patches. When the > > contention is on a structure for single user this change however doesn't > > bring much (the performance change is in statistical noise in my testing). > > > > 2) I have patches to remove some contention on dq_list_lock by not using > > dirty list for tracking dquots in ext4 (and thus avoid dq_list_lock > > completely in quota modification path). This does not bring measurable > > benefit in my testing even on ramdisk but lockstat data for dq_list_lock > > looks much better after this - it seems lock contention just shifted to > > dq_data_lock - I'll try to address that as well and see whether I'll be > > able to measure some advantage. > > > > 3) I have patches to convert dquot dirty bit to sequence counter so that > > in commit_dqblk() we can check whether dquot state we wanted to write is > > already on disk. Note that this is different from Andrew's approach in that > > we do wait for dquot to be actually written before returning. We just don't > > repeat the write unnecessarily. However this didn't bring any measurable > > benefit in my testing so unless I'll be able to confirm it benefits some > > workloads I won't merge this change. > > > > If you can experiment with your workloads, I can send you patches. I'd be > > keen on having some performance data from real setups... > > > > Honza > > > >> > >> Thanks, > >> Shilong > >> > >> > dquot_mark_dquot_dirty(dquot); > >> > return ext4_write_dquot(dquot); > >> > > >> > The idea was that if we know that some thread is somewhere between > >> > mark_dirty and clear_dirty, then we can avoid blocking on dqio_mutex, > >> > since that thread will update the ondisk dquot for us. > >> > > > -- > > Jan Kara <jack@xxxxxxxx> > > SUSE Labs, CR -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR