Thanks Dave. Please see my responses inline. On Wed, Jun 3, 2015 at 5:35 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote: > On Wed, Jun 03, 2015 at 04:18:20PM -0700, Shrinand Javadekar wrote: >> Here you go! > > Thanks! > >> /dev/mapper/35000c50062e6a12b-part2 /srv/node/r1 xfs >> rw,nosuid,nodev,noexec,noatime,nodiratime,attr2,nobarrier,inode64,logbufs=8,noquota >> 0 0 > ..... >> meta-data=/dev/mapper/35000c50062e6a7eb-part2 isize=256 agcount=64, agsize=11446344 blks >> = sectsz=512 attr=2 >> data = bsize=4096 blocks=732566016, imaxpct=5 >> = sunit=0 swidth=0 blks >> naming =version 2 bsize=4096 ascii-ci=0 >> log =internal bsize=4096 blocks=357698, version=2 >> = sectsz=512 sunit=0 blks, lazy-count=1 >> realtime =none extsz=4096 blocks=0, rtextents=0 > > Ok, so agcount=64 is unusual, especially for a single disk > filesystem. What was the reason for doing this? I read few articles that recommend using an increased number of AGs, especially when there are large disks. I can use the default # of AGs (4?) and try again. > >> - Workload causing the problem: >> >> Openstack Swift. This is what it's doing: >> >> 1. A path like /srv/node/r1/objects/1024/eef/tmp already exists. >> /srv/node/r1 is the mount point. >> 2. Creates a tmp file, say tmpfoo in the patch above. Path: >> /srv/node/r1/objects/1024/eef/tmp/tmpfoo. >> 3. Issues a 256KB write into this file. >> 4. Issues an fsync on the file. >> 5. Closes this file. >> 6. Creates another directory named "deadbeef" inside "eef" if it >> doesn't exist. Path /srv/node/r1/objects/1024/eef/deadbeef. >> 7. Moves file tmpfoo into the deadbeef directory using rename(). >> /srv/node/r1/objects/1023/eef/tmp/tmpfoo --> >> /srv/node/r1/objects/1024/eef/deadbeef/foo.data >> 8. Does a readdir on /srv/node/r1/objects/1024/eef/deadbeef/ >> 9. Iterates over all files obtained in #8 above. Usually #8 gives only one file. > > Oh. We've already discussed this problem in a previous thread: > > http://oss.sgi.com/archives/xfs/2015-04/msg00256.html Yes, we touched upon this earlier and found that all files were getting created in the same AG. We fixed that by and my current testing includes that fix. Earlier the tmp file was /srv/node/r1/tmp. By moving it further down the filesystem hierarchy to /srv/node/r1/objects/1024/eef/tmp, we make sure there are several tmp directories. I'm told, the ideal solution using O_TMP and linkat() will have to be rolled out later when there is support for that in python. > > Next time, please make sure you start with a reference to previous > discussions on the same topic. Apologies, I will! > > Specifically, that discussion touched on problems your workload > induces in metadata layout and locality: > > http://oss.sgi.com/archives/xfs/2015-04/msg00300.html > > And you are using agcount=64 on these machines, so that's going to > cause you all sorts of locality problems, which will translate into > seek bound IO performance.... > >> - IOStat and vmstat output >> (attached) > > I am assuming these are 1 second samples, based on your 18s fast/12s > slow description earlier. Yes, these are 1 seconds samples. > > The vmstat shows fast writeback at 150-200MB/s, with no idle time, > anything up to 200 processes in running or blocked state and 20-30% > iowait, followed by idle CPU time with maybe 10 running/blocked > processes, writeback at 15-20MB/s with 70% idle time and 30% iowait. > > IOWs, the workload is cyclic - lots of incoming data with lots of > throughput, followed by zero incoming data processing on only small > amounts of writeback. My understanding is that the workload is either a) waiting for issued IOs to complete. b) not able to issue more IOs because XFS is busy flushing the journal entries. Is this not true? > > The iostat shows that when the system is running at 150MB/s, the IO > service time is ~7ms (running ~130 IOPS per drive) and the average > IO size is around 170kB, with a request queue depth of 20-30 IOs. > Device utilisation is 100%, so throughput is seek bound. > > When the system is mostly idle, the throughput is essentially > running a random 4k IO write workload - 180 IOPS, request size 4k, > service time 5ms, request queue depth ~140, average wait ~800ms, > device utilisation 100%. Again, seek bound, the only difference is > the IO size. Again, my understanding was that this idle time is because of XFS is busy writing metadata from the journal to the final locations on disk. > > The vmstat information implies that front end application processing > is stopping for some period of time, but it does not indicate why it > is doing so. When the disks are doing 4k writeback, can you grab > the output of 'echo w > /proc/sysrq-trigger' from dmesg and post the > output? That will tell us if the front end processing is blocked on > the filesystem at all... Aah.. ok. Will do and get back to you soon. > >> - Trace cmd report >> Too big to attach. Here's a link: >> https://www.dropbox.com/s/3xxe2chsv4fsrv8/trace_report.txt.zip?dl=0 > -Shri _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
