On Tue, May 19, 2015 at 8:36 PM, Shyam <srangana@xxxxxxxxxx> wrote:
On 05/19/2015 08:10 AM, Raghavendra G wrote:
After discussion with Vijaykumar mallikarjuna and other inputs in this
thread, we are proposing all quota tests to comply to following criteria:
* use dd always with oflag=append (to make sure there are no parallel
writes) and conv=fdatasync (to make sure errors, if any are delivered to
application. Turning off flush-behind is optional since fdatasync acts
as a barrier)
OR
* turn off write-behind in nfs client and glusterfs server.
What do you people think is a better test scenario?
Also, we don't have confirmation on the RCA that parallel writes are
indeed the culprits. We are trying to reproduce the issue locally.
@Shyam, it would be helpful if you can confirm the hypothesis :).
Ummm... I thought we acknowledge that quota checks are done during the WIND and updated during UNWIND, and we have io threads doing in flight IOs (as well as possible IOs in io threads queue) and we have 256K writes in the case mentioned. Put together, in my head this forms a good RCA that we write more than needed due to the in flight IOs on the brick. We need to control the in flight IOs as a resolution for this from the application.
In terms of actual proof, we would need to instrument the code and check. When you say it does not fail for you, does the file stop once quota is reached or is a random size greater than quota? Which itself may explain or point to the RCA.
The basic thing needed from an application is,
- Sync IOs, so that there aren't too many in flight IOs and the application waits for each IO to complete
- Based on tests below if we keep block size in dd lower and use oflag=sync we can achieve the same, if we use higher block sizes we cannot
Test results:
1) noac:
- NFS sends a COMMIT (internally translates to a flush) post each IO request (NFS WRITES are still with the UNSTABLE flag)
- Ensures prior IO is complete before next IO request is sent (due to waiting on the COMMIT)
- Fails if IO size is large, i.e in the test case being discussed I changed the dd line that was failing as "TEST ! dd if=/dev/zero of=$N0/$mydir/newfile_2 *bs=10M* count=1 conv=fdatasync" and this fails at times, as the writes here are sent as 256k chunks to the server and we still see the same behavior
- noac + performance.nfs.flush-behind: off + performance.flush-behind: off + performance.nfs.strict-write-ordering: on + performance.strict-write-ordering: on + performance.nfs.write-behind: off + performance.write-behind: off
- Still see similar failures, i.e at times 10MB file is created successfully in the modified dd command above
Overall, the switch works, but not always. If we are to use this variant then we need to announce that all quota tests using dd not try to go beyond the quota limit set in a single IO from dd.
2) oflag=sync:
- Exactly the same behavior as above.
3) Added all (and possibly the kitches sink) to the test case, as attached, and still see failures,
- Yes, I have made the test fail intentionally (of sorts) by using 3M per dd IO and 2 IOs to go beyond the quota limit.
- The intention is to demonstrate that we still get parallel IOs from NFS client
- The test would work if we reduce the block size per IO (reliably is a border condition here, and we need specific rules like block size and how many blocks before we state quota is exceeded etc.)
- The test would work if we just go beyond the quota, and then check a separate dd instance as being able to *not* exceed the quota. Which is why I put up that patch.
What next?
The only thing left out now is background quota accounting. Vijay mallikarjuna is working on making this behaviour configurable. The proposed behaviour of marker is:
1. If fd is opened with O_SYNC, make accounting foreground (write is unwound only after accounting is complete)
2. When soft-limit is crossed, enforcer instructs marker to make accounting foreground.
3. In all other cases, marker accounting is done in background wrt to fops.
Handling parallel writes is a bit more involved and hence will be sent out as a different patch. However our regression tests will be written using osync flags in dd there by avoiding parallel writes.
We are waiting for these changes to test whether the issue is fixed. Fingers crossed :).
regards,
Raghavendra.
On Tue, May 19, 2015 at 5:27 PM, Raghavendra G <raghavendra@xxxxxxxxxxx
<mailto:raghavendra@xxxxxxxxxxx>> wrote:
On Tue, May 19, 2015 at 4:26 PM, Jeff Darcy <jdarcy@xxxxxxxxxxGluster-devel@xxxxxxxxxxx <mailto:Gluster-devel@xxxxxxxxxxx><mailto:jdarcy@xxxxxxxxxx>> wrote:
> No, my suggestion was aimed at not having parallel writes. In this case quota
> won't even fail the writes with EDQUOT because of reasons explained above.
> Yes, we need to disable flush-behind along with this so that errors are
> delivered to application.
Would conv=sync help here? That should prevent any kind of
write parallelism.
An strace of dd shows that
* fdatasync is issued only once at the end of all writes when
conv=fdatasync
* for some strange reason no fsync or fdatasync is issued at all
when conv=sync
So, using conv=fdatasync in the test cannot prevent
write-parallelism induced by write-behind. Parallelism would've been
prevented only if dd had issued fdatasync after each write or opened
the file with O_SYNC.
If it doesn't, I'd say that's a true test failure somewhere in
our stack. A
similar possibility would be to invoke dd multiple times with
oflag=append.
Yes, appending writes curb parallelism (at least in glusterfs, but
not sure how nfs client behaves) and hence can be used as an
alternative solution.
On a slightly unrelated note flush-behind is immaterial in this test
since fdatasync is anyways acting as a barrier.
_______________________________________________
Gluster-devel mailing list
http://www.gluster.org/mailman/listinfo/gluster-devel
--
Raghavendra G
--
Raghavendra G
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://www.gluster.org/mailman/listinfo/gluster-devel
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://www.gluster.org/mailman/listinfo/gluster-devel
--
Raghavendra G
_______________________________________________ Gluster-devel mailing list Gluster-devel@xxxxxxxxxxx http://www.gluster.org/mailman/listinfo/gluster-devel
