[Bugs] [Bug 1467614] Gluster read/write performance improvements on NVMe backend

[bugzilla at redhat.com]

https://bugzilla.redhat.com/show_bug.cgi?id=1467614



--- Comment #39 from Manoj Pillai <mpillai at redhat.com> ---
(In reply to Manoj Pillai from comment #36)
> I think the investigation here would
> benefit from being able to separately control no. of io-threads on the
> client-side and brick i.e. from a fix for bz #1499649.

In the meantime, I'd suggest keeping io-thread-count=4. Here's why...

Result from 4 client, 1 brick test with io-thread-count=2:
read: IOPS=26.1k, BW=102Mi (107M)(6144MiB/60198msec)

And output from top on the brick:
<quote>
  PID USER      PR  NI    VIRT    RES    SHR S %CPU %MEM     TIME+ COMMAND
 3900 root      20   0 1319128  30236   4968 R 99.2  0.1   1:13.16 glusteriotw+
 3345 root      20   0 1319128  30236   4968 R 98.2  0.1   1:56.45 glusteriotw+
 2500 root      20   0 1319128  30236   4968 S 36.0  0.1   1:20.12 glusterrpcs+
</quote>

So the brick can only deliver about 26k IOPs with this setting. That is close
to the client-side limit of 23k IOPs. If we use io-thread-count=2, there is a
risk that we'll discard valuable client-side enhancements thinking that they
are not giving any benefits when in fact they might be. Also, we should
re-evaluate any enhancements we have discarded in this way :).

-- 
You are receiving this mail because:
You are on the CC list for the bug.
Unsubscribe from this bug https://bugzilla.redhat.com/token.cgi?t=DAl9egwqih&a=cc_unsubscribe