quota: dqio_mutex design
Hello!
We have a heavy metadata related workload (ext4, quota journalling) and profiling shows that there's significant dqio_mutex contention.
Quoted 1 line
From the quota code, it looks like every time
dqio_mutexis taken
it protects access to only one quota file.
Is it possible to split dqio_mutex for each of MAXQUOTAS so that e.g. 2 parallel dquot_commit()'s can be running for user and group quota update? Am I missing any dqio_mutex function that requires dqio_mutex to be monolithic?
Thank you,
Andrew
Jan Kara
reviewerHello!
Quoted 10 lines · Andrew Perepechko
We have a heavy metadata related workload (ext4, quota journalling) and profiling shows that there's significant
dqio_mutexcontention.From the quota code, it looks like every time
dqio_mutexis taken it protects access to only one quota file.Is it possible to split
dqio_mutexfor each of MAXQUOTAS so that e.g. 2 parallel dquot_commit()'s can be running for user and group quota update? Am I missing anydqio_mutexfunction that requiresdqio_mutexto be monolithic?
So we can certainly make dqio_mutex less heavy. Making it per-quota-type would OK but I suspect it will not bring a big benefit. What would likely be more noticeable is if we avoided dqio_mutex for updates of quota information - that should not be that hard to do since we update that in-place and so don't really need the serialization for anything substantial. However we will need some restructuring of the code to make such locking scheme possible in a clean way...
Honza
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Hello!
Jan, do you think it makes sense, as an improvement until the code restructuring, to exit immediately from ext4_mark_dquot_dirty() if dquot_mark_dquot_dirty() returns 1?
It seems that in this case we are guaranteed that some thread is somewhere in the middle of mark_dquot_dirty() and clear_dquot_dirty(), so it will update the quota file buffer with the latest dquot data.
That would improve a single user/group scenario like:
thread 1) processing dquot_commit()
thread 2) dirtied dquot and is waiting for dqio_mutex thread 3, 4, 5 ...) dirtied dquot and are waiting for dqio_mutex
If we exit immediately on dquot dirtying, threads 3, 4, 5, ... can let thread 2 update the buffer data and themselves may not block on the mutex.
Thank you,
Andrew
Quoted 23 lines
Hello!
Quoted 10 lines · Andrew Perepechko
We have a heavy metadata related workload (ext4, quota journalling) and profiling shows that there's significant
dqio_mutexcontention.From the quota code, it looks like every time
dqio_mutexis taken it protects access to only one quota file.Is it possible to split
dqio_mutexfor each of MAXQUOTAS so that e.g. 2 parallel dquot_commit()'s can be running for user and group quota update? Am I missing anydqio_mutexfunction that requiresdqio_mutexto be monolithic?So we can certainly make
dqio_mutexless heavy. Making it per-quota-type would OK but I suspect it will not bring a big benefit. What would likely be more noticeable is if we avoideddqio_mutexfor updates of quota information - that should not be that hard to do since we update that in-place and so don't really need the serialization for anything substantial. However we will need some restructuring of the code to make such locking scheme possible in a clean way...Honza
Jan Kara
reviewerHi,
Quoted 9 lines · Andrew Perepechko
Jan, do you think it makes sense, as an improvement until the code restructuring, to exit immediately from
ext4_mark_dquot_dirty()ifdquot_mark_dquot_dirty()returns 1?It seems that in this case we are guaranteed that some thread is somewhere in the middle of
mark_dquot_dirty()andclear_dquot_dirty(), so it will update the quota file buffer with the latest dquot data.
Well, it would mostly work, except if process A dirties dquot outside of transaction (e.g. dquot_set_dqblk()), it could happen that other updates of dquot inside a running transaction will end up relying on update of dquot buffer by process A and that may end only in the next transaction thus breaking the journalling guarantees.
Honza
Quoted 25 lines
Quoted 23 lines
Hello!
Quoted 10 lines · Andrew Perepechko
We have a heavy metadata related workload (ext4, quota journalling) and profiling shows that there's significant
dqio_mutexcontention.From the quota code, it looks like every time
dqio_mutexis taken it protects access to only one quota file.Is it possible to split
dqio_mutexfor each of MAXQUOTAS so that e.g. 2 parallel dquot_commit()'s can be running for user and group quota update? Am I missing anydqio_mutexfunction that requiresdqio_mutexto be monolithic?So we can certainly make
dqio_mutexless heavy. Making it per-quota-type would OK but I suspect it will not bring a big benefit. What would likely be more noticeable is if we avoideddqio_mutexfor updates of quota information - that should not be that hard to do since we update that in-place and so don't really need the serialization for anything substantial. However we will need some restructuring of the code to make such locking scheme possible in a clean way...Honza
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Jan Kara
reviewerQuoted 14 lines · Jan Kara
Hello!
Quoted 10 lines · Andrew Perepechko
We have a heavy metadata related workload (ext4, quota journalling) and profiling shows that there's significant
dqio_mutexcontention.From the quota code, it looks like every time
dqio_mutexis taken it protects access to only one quota file.Is it possible to split
dqio_mutexfor each of MAXQUOTAS so that e.g. 2 parallel dquot_commit()'s can be running for user and group quota update? Am I missing anydqio_mutexfunction that requiresdqio_mutexto be monolithic?
Quoted 7 lines
So we can certainly make
dqio_mutexless heavy. Making it per-quota-type would OK but I suspect it will not bring a big benefit. What would likely be more noticeable is if we avoideddqio_mutexfor updates of quota information - that should not be that hard to do since we update that in-place and so don't really need the serialization for anything substantial. However we will need some restructuring of the code to make such locking scheme possible in a clean way...
So I'm experimenting with some patches. However I have trouble creating a workload where quota updates would show significant overhead. Can you share which workload is problematic for you? Thanks!
Honza
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
A message that is not in the archive (sent off-list or not yet received): 4181747.CBilgxvOab@panda
Jan Kara
reviewerHi Andrew,
Quoted 3 lines · Andrew Perepechko
The original workload was 50 threads sequentially creating files, each
thread in its own directory, over a fast RAID array.
OK, I can reproduce this. Actually I can reproduce on normal SATA drive. Originally I've tried on ramdisk to simulate really fast drive but there dq_list_lock and dq_data_lock contention is much more visible and the contention on dqio_mutex is minimal (two orders of magnitude smaller). On SATA drive we spend ~45% of runtime contending on dqio_mutex when creating empty files.
The problem is that if it is single user that is creating all these files, it is not clear how we could do much better - all processes contend to update the same location on disk with quota information for that user and they have to be synchronized somehow. If there are more users, we could do better by splitting dqio_mutex on per-dquot basis (I have some preliminary patches for that).
One idea I have how we could make things faster is that instead of having dquot dirty flag, we would have a sequence counter. So currently dquot modification looks like:
update counters in dquotdquot_mark_dquot_dirty(dquot);dquot_commit(dquot)
mutex_lock(dqio_mutex);
if (!clear_dquot_dirty(dquot))
nothing to do -> bail
->commit_dqblk(dquot)
mutex_unlock(dqio_mutex);
When several processes race updating the same dquot, they very often all end up updating dquot on disk even though another process has already written dquot for them while they were waiting for dqio_sem - in my test above the ratio of commit_dqblk / dquot_commit calls was 59%. What we could do is that dquot_mark_dquot_dirty() would return "current sequence of dquot", dquot_commit() would then get sequence that is required to be written and if that is already written (we would also store in dquot latest written sequence), it would bail out doing nothing. This should cut down dqio_mutex hold times and thus wait times but I need to experiment and measure that...
Honza
Quoted 60 lines
Quoted 55 lines · Jan Kara
Quoted 35 lines
Hello!
Quoted 12 lines · Andrew Perepechko
We have a heavy metadata related workload (ext4, quota journalling)
and profiling shows that there's significant
dqio_mutexcontention.From the quota code, it looks like every time
dqio_mutexis takenit protects access to only one quota file.
Quoted 7 lines
Is it possible to split
dqio_mutexfor each of MAXQUOTAS so thate.g. 2 parallel dquot_commit()'s can be running for user and group
quota update? Am I missing any
dqio_mutexfunction that requires
dqio_mutexto be monolithic?So we can certainly make
dqio_mutexless heavy. Making it per-quota-typewould OK but I suspect it will not bring a big benefit. What would likely
be more noticeable is if we avoided
dqio_mutexfor updates of quotaQuoted 7 lines
information - that should not be that hard to do since we update that
in-place and so don't really need the serialization for anything
substantial. However we will need some restructuring of the code to make
such locking scheme possible in a clean way...
So I'm experimenting with some patches. However I have trouble creating
a workload where quota updates would show significant overhead. Can you
share which workload is problematic for you? Thanks!
Honza
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Jan Kara
reviewerQuoted 5 lines · Jan Kara
Hi Andrew,
Quoted 2 lines · Andrew Perepechko
The original workload was 50 threads sequentially creating files, each
Quoted 8 lines
Quoted 1 line
thread in its own directory, over a fast RAID array.
OK, I can reproduce this. Actually I can reproduce on normal SATA drive. Originally I've tried on ramdisk to simulate really fast drive but there
dq_list_lockanddq_data_lockcontention is much more visible and the contention ondqio_mutexis minimal (two orders of magnitude smaller). On SATA drive we spend ~45% of runtime contending ondqio_mutexwhen creating empty files.
So this was just me misinterpretting lockstat data (forgot to divide the wait time by number of processes) - then the result would be that each process waits only ~1% of its runtime for dqio_mutex.
Anyway, my patches show ~10% improvement in runtime when 50 different processes create empty files for 50 different users. As expected there's not measurable benefit when all processes create files for the same user.
Quoted 20 lines
The problem is that if it is single user that is creating all these files, it is not clear how we could do much better - all processes contend to update the same location on disk with quota information for that user and they have to be synchronized somehow. If there are more users, we could do better by splitting
dqio_mutexon per-dquot basis (I have some preliminary patches for that).One idea I have how we could make things faster is that instead of having dquot dirty flag, we would have a sequence counter. So currently dquot modification looks like:
update counters in dquot
dquot_mark_dquot_dirty(dquot);dquot_commit(dquot)
mutex_lock(dqio_mutex);
if (!clear_dquot_dirty(dquot))nothing to do -> bail->commit_dqblk(dquot)
mutex_unlock(dqio_mutex);
Quoted 10 lines
When several processes race updating the same dquot, they very often all end up updating dquot on disk even though another process has already written dquot for them while they were waiting for
dqio_sem- in my test above the ratio ofcommit_dqblk/dquot_commitcalls was 59%. What we could do is thatdquot_mark_dquot_dirty()would return "current sequence of dquot",dquot_commit()would then get sequence that is required to be written and if that is already written (we would also store in dquot latest written sequence), it would bail out doing nothing. This should cut downdqio_mutexhold times and thus wait times but I need to experiment and measure that...
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
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Quoted 59 lines · Jan Kara
Quoted 5 lines
Hi Andrew,
Quoted 2 lines · Andrew Perepechko
The original workload was 50 threads sequentially creating files, each
Quoted 8 lines
Quoted 1 line
thread in its own directory, over a fast RAID array.
OK, I can reproduce this. Actually I can reproduce on normal SATA drive. Originally I've tried on ramdisk to simulate really fast drive but there
dq_list_lockanddq_data_lockcontention is much more visible and the contention ondqio_mutexis minimal (two orders of magnitude smaller). On SATA drive we spend ~45% of runtime contending ondqio_mutexwhen creating empty files.So this was just me misinterpretting lockstat data (forgot to divide the wait time by number of processes) - then the result would be that each process waits only ~1% of its runtime for
dqio_mutex.Anyway, my patches show ~10% improvement in runtime when 50 different processes create empty files for 50 different users. As expected there's not measurable benefit when all processes create files for the same user.
Quoted 23 lines
The problem is that if it is single user that is creating all these files, it is not clear how we could do much better - all processes contend to update the same location on disk with quota information for that user and they have to be synchronized somehow. If there are more users, we could do better by splitting
dqio_mutexon per-dquot basis (I have some preliminary patches for that).One idea I have how we could make things faster is that instead of having dquot dirty flag, we would have a sequence counter. So currently dquot modification looks like:
update counters in dquot
dquot_mark_dquot_dirty(dquot);dquot_commit(dquot)mutex_lock(dqio_mutex); if (!clear_dquot_dirty(dquot))nothing to do -> bail->commit_dqblk(dquot) mutex_unlock(dqio_mutex);Quoted 12 lines
When several processes race updating the same dquot, they very often all end up updating dquot on disk even though another process has already written dquot for them while they were waiting for
dqio_sem- in my test above the ratio ofcommit_dqblk/dquot_commitcalls was 59%. What we could
do is thatdquot_mark_dquot_dirty()would return "current sequence of dquot",dquot_commit()would then get sequence that is required to be written and if that is already written (we would also store in dquot latest
written sequence), it would bail out doing nothing. This should cut downdqio_mutexhold times and thus wait times but I need to experiment and measure that...
Quoted 6 lines
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;
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.
I think, you also mentioned that some mark_dquot_dirty callers, such as do_set_dqblk, may not be running with an open transaction handle, so we cannot assume this optimization is atomic. However, we don't use do_set_dqblk and seem safe wrt journalling.
Thank you,
Andrew
Jan Kara
reviewerQuoted 38 lines · Andrew Perepechko
Quoted 14 lines · Jan Kara
Quoted 4 lines
When several processes race updating the same dquot, they very often all end up updating dquot on disk even though another process has already written dquot for them while they were waiting for
dqio_sem- in my test above the ratio ofcommit_dqblk/dquot_commitcalls was 59%. What weQuoted 8 lines
could
do is thatdquot_mark_dquot_dirty()would return "current sequence of dquot",dquot_commit()would then get sequence that is required to be written and if that is already written (we would also store in dquot latest
written sequence), it would bail out doing nothing. This should cut downdqio_mutexhold times and thus wait times but I need to experiment and measure that...Quoted 6 lines
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]) {(quoted diff) +2
-0+ if (test_bit(DQ_MOD_B, &dquot->dq_flags)) + return 0; dquot_mark_dquot_dirty(dquot); return ext4_write_dquot(dquot);
Quoted 8 lines
The idea was that if we know that some thread is somewhere between
mark_dirtyandclear_dirty, then we can avoid blocking ondqio_mutex, since that thread will update the ondisk dquot for us.I think, you also mentioned that some
mark_dquot_dirtycallers, such asdo_set_dqblk, may not be running with an open transaction handle, so we cannot assume this optimization is atomic. However, we don't usedo_set_dqblkand seem safe wrt journalling.
OK, thanks for info. I'm reluctant to make ext4_mark_dquot_dirty() return before quota data is actually copied to the transaction which is what probably brings you the benefit. I'll think about it some more.
Honza
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Hello Guys,
We DDN is investigating the same issue!
Some comments comes:
Quoted 22 lines · Andrew Perepechko
Quoted 9 lines · Jan Kara
Quoted 2 lines
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]) {
Review
Quoted 2 lines
(quoted diff) +2
-0+ 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.
Thanks,
Shilong
Quoted 7 lines
dquot_mark_dquot_dirty(dquot); return ext4_write_dquot(dquot);The idea was that if we know that some thread is somewhere between
mark_dirtyandclear_dirty, then we can avoid blocking ondqio_mutex, since that thread will update the ondisk dquot for us.
Quoted 5 lines
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.
Hello Wang!
There is no transaction change in this case because all callers of this function have open handles for the same transaction.
If you enter that DQ_MOD_B check, you are guaranteed to reference the SAME transaction as the thread that's in between of mark_dirty and clear_dirty.
Thank you,
Andrew
Hi,
Quoted 6 lines · Andrew Perepechko
Quoted 5 lines
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.
Quoted 9 lines
Hello Wang!
There is no transaction change in this case because all callers of this function have open handles for the same transaction.
If you enter that
DQ_MOD_Bcheck, you are guaranteed to reference the SAME transaction as the thread that's in between ofmark_dirtyandclear_dirty.
This change mean if this dquot is dirty we skip, this
won't work because in this way, quota update is only kept in vfs dquota memory and newer update is not wrote to journal file and not wrapped into transaction too.
This is not what journal quota means to do.
Thanks,
Shilong
Quoted 2 lines
Thank you,
Andrew
Quoted 5 lines
This change mean if this dquot is dirty we skip, this
won't work because in this way, quota update is only kept in vfs dquota memory and newer update is not wrote to journal file and not wrapped into transaction too.
That's not true.
As I explained earlier, having DQ_MOD_B set at this point means another thread is going to write dquot but hasn't yet started doing so. This thread does not care whether it updates the ondisk dquot with its own data or with fresher data which came from another thread. In-core dquot has no indication of whose data in contains.
As I also explained earlier, the update cannot happen in the context of another transaction because thread A which sees DQ_MOD_B set and thread B which is running dquot_commit() both have journal handles to the same transaction. There's only one running transaction at a time and thread B does not switch to another transaction.
Please read the code carefully.
Quoted 9 lines
This is not what journal quota means to do.
Thanks,
ShilongQuoted 2 lines
Thank you,
Andrew
Let me put it this way:
Under file creation from different threads, ext4 will generate a series of dquot updates (incore and then ondisk, through journal):
dquot update1
dquot update2
dquot update3
...
dquot updateN
Either with my patch or without it, ondisk dquot update through journal may miss dquot update1, dquot update2, ... dquot update{N-1}.
You can easily see that from the code of dquot_commit():
int dquot_commit(struct dquot *dquot)
{
int ret = 0;
struct quota_info *dqopt = sb_dqopt(dquot->dq_sb);
mutex_lock(&dqopt->dqio_mutex);
spin_lock(&dq_list_lock);
if (!clear_dquot_dirty(dquot)) {
spin_unlock(&dq_list_lock);
goto out_sem;
}...
}
If actual dquot_commit() wrote dquot update N, the threads commiting updates 1 through N-1 will exit immediately once they get dqio_mutex since the dquot will NOT be dirty.
My patch only avoids blocking on dqio_mutex when we know for sure that another will NECESSARILY write the needed or a FRESHER dquot ondisk.
Quoted 29 lines
Quoted 4 lines
This change mean if this dquot is dirty we skip, this
won't work because in this way, quota update is only kept in vfs dquota memory and newer update is not wrote to journal file and not wrapped into transaction too.That's not true.
As I explained earlier, having
DQ_MOD_Bset at this point means another thread is going to write dquot but hasn't yet started doing so. This thread does not care whether it updates the ondisk dquot with its own data or with fresher data which came from another thread. In-core dquot has no indication of whose data in contains.As I also explained earlier, the update cannot happen in the context of another transaction because thread A which sees
DQ_MOD_Bset and thread B which is runningdquot_commit()both have journal handles to the same transaction. There's only one running transaction at a time and thread B does not switch to another transaction.Please read the code carefully.
Quoted 8 lines
This is not what journal quota means to do.
Thanks,
ShilongQuoted 2 lines
Thank you,
Andrew
Jan Kara
reviewerQuoted 31 lines · Andrew Perepechko
Let me put it this way:
Under file creation from different threads, ext4 will generate a series of dquot updates (incore and then ondisk, through journal):
dquot update1
dquot update2
dquot update3
...
dquot updateNEither with my patch or without it, ondisk dquot update through journal may miss dquot update1, dquot update2, ... dquot update{N-1}.
You can easily see that from the code of
dquot_commit():int
dquot_commit(struct dquot *dquot){ int ret = 0; struct quota_info *dqopt = sb_dqopt(dquot->dq_sb); mutex_lock(&dqopt->dqio_mutex); spin_lock(&dq_list_lock); if (!clear_dquot_dirty(dquot)) { spin_unlock(&dq_list_lock); goto out_sem; }...
}
Quoted 6 lines
If actual
dquot_commit()wrote dquot update N, the threads commiting updates 1 through N-1 will exit immediately once they getdqio_mutexsince the dquot will NOT be dirty.My patch only avoids blocking on
dqio_mutexwhen we know for sure that another will NECESSARILY write the needed or a FRESHER dquot ondisk.
Yeah, I agree with Andrew. What they did is *almost* safe for ext4. The only moment when it is not safe is when someone calls mark_dquot_dirty() outside of a scope of a transaction which happens when doing Q_SETQUOTA quotactl.
Another things which is subtle with Andrew's approach is that process modifying quota information can return and stop its handle before quota data gets copied to transaction buffer. This does not currently create any real problem since nobody is relying on that however it relies on intimate details of JBD2 transaction machinery and that could bite us in the future.
Honza
Quoted 31 lines
Quoted 29 lines
Quoted 4 lines
This change mean if this dquot is dirty we skip, this
won't work because in this way, quota update is only kept in vfs dquota memory and newer update is not wrote to journal file and not wrapped into transaction too.That's not true.
As I explained earlier, having
DQ_MOD_Bset at this point means another thread is going to write dquot but hasn't yet started doing so. This thread does not care whether it updates the ondisk dquot with its own data or with fresher data which came from another thread. In-core dquot has no indication of whose data in contains.As I also explained earlier, the update cannot happen in the context of another transaction because thread A which sees
DQ_MOD_Bset and thread B which is runningdquot_commit()both have journal handles to the same transaction. There's only one running transaction at a time and thread B does not switch to another transaction.Please read the code carefully.
Quoted 8 lines
This is not what journal quota means to do.
Thanks,
ShilongQuoted 2 lines
Thank you,
Andrew
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Jan Kara
reviewerHello!
Quoted 41 lines · Wang Shilong
We DDN is investigating the same issue!
Some comments comes:
Quoted 22 lines · Andrew Perepechko
Quoted 9 lines · Jan Kara
Quoted 2 lines
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]) {Review
Quoted 2 lines
(quoted diff) +2
-0+ 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_mutexto 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.
Quoted 7 lines
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
Quoted 11 lines
Thanks,
ShilongQuoted 7 lines
dquot_mark_dquot_dirty(dquot); return ext4_write_dquot(dquot);The idea was that if we know that some thread is somewhere between
mark_dirtyandclear_dirty, then we can avoid blocking ondqio_mutex, since that thread will update the ondisk dquot for us.
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Hello Jan,
Please send me patches, we could test and response you!
Thanks,
Shilong
Quoted 98 lines · Jan Kara
Hello!
Quoted 41 lines · Wang Shilong
We DDN is investigating the same issue!
Some comments comes:
Quoted 22 lines · Andrew Perepechko
Quoted 9 lines · Jan Kara
Quoted 2 lines
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]) {Review
Quoted 2 lines
(quoted diff) +2
-0+ 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_mutexto 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.Quoted 7 lines
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_mutexdo rw semaphore and use it in exclusive mode only when quota tree is going to change. We also usedq_lockto 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_lockby not using dirty list for tracking dquots in ext4 (and thus avoiddq_list_lockcompletely in quota modification path). This does not bring measurable benefit in my testing even on ramdisk but lockstat data fordq_list_locklooks much better after this - it seems lock contention just shifted todq_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
Quoted 11 lines
Thanks,
ShilongQuoted 7 lines
dquot_mark_dquot_dirty(dquot); return ext4_write_dquot(dquot);The idea was that if we know that some thread is somewhere between
mark_dirtyandclear_dirty, then we can avoid blocking ondqio_mutex, since that thread will update the ondisk dquot for us.--
Jan Kara <[email protected]>
SUSE Labs, CR
Jan Kara
reviewerHi,
Quoted 1 line · Wang Shilong
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
Quoted 99 lines · Jan Kara
Quoted 98 lines
Hello!
Quoted 41 lines · Wang Shilong
We DDN is investigating the same issue!
Some comments comes:
Quoted 22 lines · Andrew Perepechko
Quoted 9 lines · Jan Kara
Quoted 2 lines
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]) {Review
Quoted 2 lines
(quoted diff) +2
-0+ 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_mutexto 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.Quoted 7 lines
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_mutexdo rw semaphore and use it in exclusive mode only when quota tree is going to change. We also usedq_lockto 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_lockby not using dirty list for tracking dquots in ext4 (and thus avoiddq_list_lockcompletely in quota modification path). This does not bring measurable benefit in my testing even on ramdisk but lockstat data fordq_list_locklooks much better after this - it seems lock contention just shifted todq_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
Quoted 11 lines
Thanks,
ShilongQuoted 7 lines
dquot_mark_dquot_dirty(dquot); return ext4_write_dquot(dquot);The idea was that if we know that some thread is somewhere between
mark_dirtyandclear_dirty, then we can avoid blocking ondqio_mutex, since that thread will update the ondisk dquot for us.--
Jan Kara <[email protected]>
SUSE Labs, CR
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Hello Jan,
We have tested your patches, in generally, it helped in our case. Noticed,
our test case is only one user with many process create/remove file.
4.13.0-rc3 without any patches
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink
0 93,068 296,028 86,860 285,131 85,199 189,653 1 79,501 280,921 91,079 277,349 186,279 170,982 2 79,932 299,750 90,246 274,457 133,922 191,677 3 80,146 297,525 86,416 272,160 192,354 198,869
4.13.0-rc3/w Jan Kara patch
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink
0 73,057 311,217 74,898 286,120 81,217 288,138 ops/per second
1 78,872 312,471 76,470 277,033 77,014 288,057 2 79,170 291,440 76,174 283,525 73,686 283,526 3 79,941 309,168 78,493 277,331 78,751 281,377
4.13.0-rc3/with https://patchwork.ozlabs.org/patch/799014/
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink
0 100,319 322,746 87,480 302,579 84,569 218,969
1 728,424 299,808 312,766 293,471 219,198 199,389
2 729,410 300,930 315,590 289,664 218,283 197,871
3 727,555 298,797 316,837 289,108 213,095 213,458
4.13.0-rc3/w https://patchwork.ozlabs.org/patch/799014/ + Jan Kara patch
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink
0 100,312 324,871 87,076 267,303 86,258 288,137 1 707,524 298,892 361,963 252,493 421,919 282,492 2 707,792 298,162 363,450 264,923 397,723 283,675 3 707,420 302,552 354,013 266,638 421,537 281,763
In conclusion, your patches helped a lot for our testing, noticed, please ignored test0 running for creation, the first time testing will loaded inode cache in memory, we used test1-3 to compare.
With extra patch applied, your patches improved File creation(quota+project) 2X, File unlink 1.5X.
Thanks,
Shilong
________________________________________
From: Jan Kara [[email protected]]
Sent: Wednesday, August 09, 2017 0:06
To: Wang Shilong
Cc: Jan Kara; Andrew Perepechko; Shuichi Ihara; Wang Shilong; Li Xi; Ext4 Developers List; [email protected]
Subject: Re: quota: dqio_mutex design
Hi,
Quoted 1 line · Wang Shilong
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
Quoted 92 lines · Jan Kara
Quoted 91 lines
Hello!
Quoted 41 lines · Wang Shilong
We DDN is investigating the same issue!
Some comments comes:
Quoted 22 lines · Andrew Perepechko
Quoted 9 lines · Jan Kara
Quoted 2 lines
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]) {Review
Quoted 2 lines
(quoted diff) +2
-0+ 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_mutexto 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.Quoted 7 lines
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_mutexdo rw semaphore and use it in exclusive mode only when quota tree is going to change. We also usedq_lockto 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_lockby not using dirty list for tracking dquots in ext4 (and thus avoiddq_list_lockcompletely in quota modification path). This does not bring measurable benefit in my testing even on ramdisk but lockstat data fordq_list_locklooks much better after this - it seems lock contention just shifted todq_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
Quoted 7 lines
Thanks,
ShilongQuoted 3 lines
dquot_mark_dquot_dirty(dquot); return ext4_write_dquot(dquot);
Quoted 7 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines
The idea was that if we know that some thread is somewhere between
mark_dirtyandclear_dirty, then we can avoid blocking ondqio_mutex, since that thread will update the ondisk dquot for us.--
Jan Kara <[email protected]>
SUSE Labs, CR
--
Jan Kara <[email protected]>
SUSE Labs, CR
sorry, format did not look fine, please use attachment.
Quoted 176 lines · Wang Shilong
Hello Jan,
We have tested your patches, in generally, it helped in our case. Noticed,
our test case is only one user with many process create/remove file.4.13.0-rc3 without any patches
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 93,068 296,028 86,860 285,131 85,199 189,653 1 79,501 280,921 91,079 277,349 186,279 170,982 2 79,932 299,750 90,246 274,457 133,922 191,677 3 80,146 297,525 86,416 272,160 192,354 198,8694.13.0-rc3/w Jan Kara patch
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 73,057 311,217 74,898 286,120 81,217 288,138 ops/per second 1 78,872 312,471 76,470 277,033 77,014 288,057 2 79,170 291,440 76,174 283,525 73,686 283,526 3 79,941 309,168 78,493 277,331 78,751 281,3774.13.0-rc3/with https://patchwork.ozlabs.org/patch/799014/
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 100,319 322,746 87,480 302,579 84,569 218,969 1 728,424 299,808 312,766 293,471 219,198 199,389 2 729,410 300,930 315,590 289,664 218,283 197,871 3 727,555 298,797 316,837 289,108 213,095 213,4584.13.0-rc3/w https://patchwork.ozlabs.org/patch/799014/ + Jan Kara patch
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 100,312 324,871 87,076 267,303 86,258 288,137 1 707,524 298,892 361,963 252,493 421,919 282,492 2 707,792 298,162 363,450 264,923 397,723 283,675 3 707,420 302,552 354,013 266,638 421,537 281,763In conclusion, your patches helped a lot for our testing, noticed, please ignored test0 running for creation, the first time testing will loaded inode cache in memory, we used test1-3 to compare.
With extra patch applied, your patches improved File creation(quota+project) 2X, File unlink 1.5X.
Thanks,
Shilong________________________________________
From: Jan Kara [[email protected]]
Sent: Wednesday, August 09, 2017 0:06
To: Wang Shilong
Cc: Jan Kara; Andrew Perepechko; Shuichi Ihara; Wang Shilong; Li Xi; Ext4 Developers List; [email protected]
Subject: Re: quota:dqio_mutexdesignHi,
Quoted 1 line · Wang Shilong
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.gitquota_scalingI'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_lockduring 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
Quoted 92 lines · Jan Kara
Quoted 91 lines
Hello!
Quoted 41 lines · Wang Shilong
We DDN is investigating the same issue!
Some comments comes:
Quoted 22 lines · Andrew Perepechko
Quoted 9 lines · Jan Kara
Quoted 2 lines
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]) {Review
Quoted 2 lines
(quoted diff) +2
-0+ 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_mutexto 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.Quoted 7 lines
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_mutexdo rw semaphore and use it in exclusive mode only when quota tree is going to change. We also usedq_lockto 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_lockby not using dirty list for tracking dquots in ext4 (and thus avoiddq_list_lockcompletely in quota modification path). This does not bring measurable benefit in my testing even on ramdisk but lockstat data fordq_list_locklooks much better after this - it seems lock contention just shifted todq_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
Quoted 7 lines
Thanks,
ShilongQuoted 3 lines
dquot_mark_dquot_dirty(dquot); return ext4_write_dquot(dquot);Quoted 7 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines
The idea was that if we know that some thread is somewhere between
mark_dirtyandclear_dirty, then we can avoid blocking ondqio_mutex, since that thread will update the ondisk dquot for us.--
Jan Kara <[email protected]>
SUSE Labs, CR--
Jan Kara <[email protected]>
SUSE Labs, CR
Txt format attched.
BTW, Jan, it will be cool if you could point which patch help a lot for our test case, since there are a lot of patches there, we want to port some of patches to RHEL7.
Thanks,
Shilong
Quoted 170 lines · Wang Shilong
Hello Jan,
We have tested your patches, in generally, it helped in our case. Noticed,
our test case is only one user with many process create/remove file.4.13.0-rc3 without any patches
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 93,068 296,028 86,860 285,131 85,199 189,653 1 79,501 280,921 91,079 277,349 186,279 170,982 2 79,932 299,750 90,246 274,457 133,922 191,677 3 80,146 297,525 86,416 272,160 192,354 198,8694.13.0-rc3/w Jan Kara patch
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 73,057 311,217 74,898 286,120 81,217 288,138 ops/per second 1 78,872 312,471 76,470 277,033 77,014 288,057 2 79,170 291,440 76,174 283,525 73,686 283,526 3 79,941 309,168 78,493 277,331 78,751 281,3774.13.0-rc3/with https://patchwork.ozlabs.org/patch/799014/
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 100,319 322,746 87,480 302,579 84,569 218,969 1 728,424 299,808 312,766 293,471 219,198 199,389 2 729,410 300,930 315,590 289,664 218,283 197,871 3 727,555 298,797 316,837 289,108 213,095 213,4584.13.0-rc3/w https://patchwork.ozlabs.org/patch/799014/ + Jan Kara patch
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 100,312 324,871 87,076 267,303 86,258 288,137 1 707,524 298,892 361,963 252,493 421,919 282,492 2 707,792 298,162 363,450 264,923 397,723 283,675 3 707,420 302,552 354,013 266,638 421,537 281,763In conclusion, your patches helped a lot for our testing, noticed, please ignored test0 running for creation, the first time testing will loaded inode cache in memory, we used test1-3 to compare.
With extra patch applied, your patches improved File creation(quota+project) 2X, File unlink 1.5X.
Thanks,
Shilong________________________________________
From: Jan Kara [[email protected]]
Sent: Wednesday, August 09, 2017 0:06
To: Wang Shilong
Cc: Jan Kara; Andrew Perepechko; Shuichi Ihara; Wang Shilong; Li Xi; Ext4 Developers List; [email protected]
Subject: Re: quota:dqio_mutexdesignHi,
Quoted 1 line · Wang Shilong
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.gitquota_scalingI'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_lockduring 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
Quoted 96 lines · Jan Kara
Quoted 95 lines
Hello!
Quoted 41 lines · Wang Shilong
We DDN is investigating the same issue!
Some comments comes:
Quoted 22 lines · Andrew Perepechko
Quoted 9 lines · Jan Kara
Quoted 2 lines
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]) {Review
Quoted 2 lines
(quoted diff) +2
-0+ 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_mutexto 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.Quoted 7 lines
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_mutexdo rw semaphore and use it in exclusive mode only when quota tree is going to change. We also usedq_lockto 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_lockby not using dirty list for tracking dquots in ext4 (and thus avoiddq_list_lockcompletely in quota modification path). This does not bring measurable benefit in my testing even on ramdisk but lockstat data fordq_list_locklooks much better after this - it seems lock contention just shifted todq_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
Quoted 11 lines
Thanks,
ShilongQuoted 7 lines
dquot_mark_dquot_dirty(dquot); return ext4_write_dquot(dquot);The idea was that if we know that some thread is somewhere between
mark_dirtyandclear_dirty, then we can avoid blocking ondqio_mutex, since that thread will update the ondisk dquot for us.
Quoted 6 lines
Quoted 3 lines
Quoted 3 lines
--
Jan Kara <[email protected]>
SUSE Labs, CR--
Jan Kara <[email protected]>
SUSE Labs, CR
4.13.0-rc3 without any patches
no Quota -O quota -O quota,project
creation unlink creation unlink creation unlink
0 93,068 296,028 86,860 285,131 85,199 189,653 ops/per second 1 79,501 280,921 91,079 277,349 186,279 170,982 2 79,932 299,750 90,246 274,457 133,922 191,677 3 80,146 297,525 86,416 272,160 192,354 198,869
Jan Kara branch (quota_scaling)
no Quota -O quota -O quota,project
creation unlink creation unlink creation unlink 0 73,057 311,217 74,898 286,120 81,217 288,138 1 78,872 312,471 76,470 277,033 77,014 288,057 2 79,170 291,440 76,174 283,525 73,686 283,526 3 79,941 309,168 78,493 277,331 78,751 281,377
4.13.0-rc3 with v5 patch https://patchwork.ozlabs.org/patch/799014/
no Quota -O quota -O quota,project
creation unlink creation unlink creation unlink 0 100,319 322,746 87,480 302,579 84,569 218,969 1 728,424 299,808 312,766 293,471 219,198 199,389 2 729,410 300,930 315,590 289,664 218,283 197,871 3 727,555 298,797 316,837 289,108 213,095 213,458
Jan Kara branch (quota_scaling) with v5 patch https://patchwork.ozlabs.org/patch/799014/
no Quota -O quota -O quota,project
creation unlink creation unlink creation unlink
0 100,312 324,871 87,076 267,303 86,258 288,137 1 707,524 298,892 361,963 252,493 421,919 282,492 2 707,792 298,162 363,450 264,923 397,723 283,675 3 707,420 302,552 354,013 266,638 421,537 281,763
Jan Kara
reviewerHello,
Quoted 5 lines · Wang Shilong
Txt format attched.
BTW, Jan, it will be cool if you could point which patch help a lot for our test case, since there are a lot of patches there, we want to port some of patches to RHEL7.
Thanks for the test results! They are really interesting. Do you have any explanation why without any patches the '-O quota,project' runs for 'File Creation' are faster than runs without quota or any other runs in the test?
WRT which patches helped I don't have a good subset for you. In my testing each patch helped a bit. I expect in your setup the conversion of dqio_sem to rwsem and then to use dq_lock might not have that big impact. So you might try backporting patches from "quota: Fix possible corruption of dqi_flags" onward.
Honza
Quoted 209 lines · Wang Shilong
Quoted 170 lines
Hello Jan,
We have tested your patches, in generally, it helped in our case. Noticed,
our test case is only one user with many process create/remove file.4.13.0-rc3 without any patches
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 93,068 296,028 86,860 285,131 85,199 189,653 1 79,501 280,921 91,079 277,349 186,279 170,982 2 79,932 299,750 90,246 274,457 133,922 191,677 3 80,146 297,525 86,416 272,160 192,354 198,8694.13.0-rc3/w Jan Kara patch
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 73,057 311,217 74,898 286,120 81,217 288,138 ops/per second 1 78,872 312,471 76,470 277,033 77,014 288,057 2 79,170 291,440 76,174 283,525 73,686 283,526 3 79,941 309,168 78,493 277,331 78,751 281,3774.13.0-rc3/with https://patchwork.ozlabs.org/patch/799014/
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 100,319 322,746 87,480 302,579 84,569 218,969 1 728,424 299,808 312,766 293,471 219,198 199,389 2 729,410 300,930 315,590 289,664 218,283 197,871 3 727,555 298,797 316,837 289,108 213,095 213,4584.13.0-rc3/w https://patchwork.ozlabs.org/patch/799014/ + Jan Kara patch
no Quota -O quota' -O quota, project'
File Creation File Unlink File Creation File Unlink File Creation File Unlink0 100,312 324,871 87,076 267,303 86,258 288,137 1 707,524 298,892 361,963 252,493 421,919 282,492 2 707,792 298,162 363,450 264,923 397,723 283,675 3 707,420 302,552 354,013 266,638 421,537 281,763In conclusion, your patches helped a lot for our testing, noticed, please ignored test0 running for creation, the first time testing will loaded inode cache in memory, we used test1-3 to compare.
With extra patch applied, your patches improved File creation(quota+project) 2X, File unlink 1.5X.
Thanks,
Shilong________________________________________
From: Jan Kara [[email protected]]
Sent: Wednesday, August 09, 2017 0:06
To: Wang Shilong
Cc: Jan Kara; Andrew Perepechko; Shuichi Ihara; Wang Shilong; Li Xi; Ext4 Developers List; [email protected]
Subject: Re: quota:dqio_mutexdesignHi,
Quoted 1 line · Wang Shilong
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.gitquota_scalingI'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_lockduring 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
Quoted 96 lines · Jan Kara
Quoted 95 lines
Hello!
Quoted 41 lines · Wang Shilong
We DDN is investigating the same issue!
Some comments comes:
Quoted 22 lines · Andrew Perepechko
Quoted 9 lines · Jan Kara
Quoted 2 lines
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]) {Review
Quoted 2 lines
(quoted diff) +2
-0+ 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_mutexto 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.Quoted 7 lines
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_mutexdo rw semaphore and use it in exclusive mode only when quota tree is going to change. We also usedq_lockto 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_lockby not using dirty list for tracking dquots in ext4 (and thus avoiddq_list_lockcompletely in quota modification path). This does not bring measurable benefit in my testing even on ramdisk but lockstat data fordq_list_locklooks much better after this - it seems lock contention just shifted todq_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
Quoted 11 lines
Thanks,
ShilongQuoted 7 lines
dquot_mark_dquot_dirty(dquot); return ext4_write_dquot(dquot);The idea was that if we know that some thread is somewhere between
mark_dirtyandclear_dirty, then we can avoid blocking ondqio_mutex, since that thread will update the ondisk dquot for us.Quoted 6 lines
Quoted 3 lines
Quoted 3 lines
--
Jan Kara <[email protected]>
SUSE Labs, CR--
Jan Kara <[email protected]>
SUSE Labs, CR4.13.0-rc3 without any patches
no Quota -O quota -O quota,project
creation unlink creation unlink creation unlink0 93,068 296,028 86,860 285,131 85,199 189,653 ops/per second 1 79,501 280,921 91,079 277,349 186,279 170,982 2 79,932 299,750 90,246 274,457 133,922 191,677 3 80,146 297,525 86,416 272,160 192,354 198,869Jan Kara branch (
quota_scaling)
no Quota -O quota -O quota,projectcreation unlink creation unlink creation unlink 0 73,057 311,217 74,898 286,120 81,217 288,138 1 78,872 312,471 76,470 277,033 77,014 288,057 2 79,170 291,440 76,174 283,525 73,686 283,526 3 79,941 309,168 78,493 277,331 78,751 281,3774.13.0-rc3 with v5 patch https://patchwork.ozlabs.org/patch/799014/
no Quota -O quota -O quota,projectcreation unlink creation unlink creation unlink 0 100,319 322,746 87,480 302,579 84,569 218,969 1 728,424 299,808 312,766 293,471 219,198 199,389 2 729,410 300,930 315,590 289,664 218,283 197,871 3 727,555 298,797 316,837 289,108 213,095 213,458Jan Kara branch (
quota_scaling) with v5 patch https://patchwork.ozlabs.org/patch/799014/
no Quota -O quota -O quota,project
creation unlink creation unlink creation unlink0 100,312 324,871 87,076 267,303 86,258 288,137 1 707,524 298,892 361,963 252,493 421,919 282,492 2 707,792 298,162 363,450 264,923 397,723 283,675 3 707,420 302,552 354,013 266,638 421,537 281,763
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR