On Wed, Jun 03, 2015 at 04:18:20PM -0700, Shrinand Javadekar wrote:
> Here you go!
> /dev/mapper/35000c50062e6a12b-part2 /srv/node/r1 xfs
> 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?
> - 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:
> 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 -->
> 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
Oh. We've already discussed this problem in a previous thread:
Next time, please make sure you start with a reference to previous
discussions on the same topic.
Specifically, that discussion touched on problems your workload
induces in metadata layout and locality:
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
I am assuming these are 1 second samples, based on your 18s fast/12s
slow description earlier.
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.
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.
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...
> - Trace cmd report
> Too big to attach. Here's a link: