[Bugs] [Bug 1349953] thread CPU saturation limiting throughput on write workloads

bugzilla at redhat.com bugzilla at redhat.com
Thu Jun 30 13:29:37 UTC 2016


https://bugzilla.redhat.com/show_bug.cgi?id=1349953



--- Comment #15 from Pranith Kumar K <pkarampu at redhat.com> ---
(In reply to Xavier Hernandez from comment #14)
> Thanks Manoj for the testing. I was wrong. As Pranith says, writes need to
> be serialized in EC causing the queued writes to be resumed inside the epoll
> threads.
> 
> In this case I'm not sure why enabling client-io-threads improves
> performance so much. Maybe it allows fuse to process requests faster,
> causing the saturation of epoll-threads later...
> 
> In any case I think it would be interesting to move the execution of all
> these operations to other threads. Maybe using a thread pool or synctasks.
> 
> Another parameter that I tweak that gives a small extra boost in my tests is
> rpc.outstanding-rpc-limit. Default is 64. Using 128, 256 or 512 gives some
> more performance (probably at the cost of more CPU usage)
Xavi,
     In the latest releases(Even in 3.7.x) the outstanding-rpc-limit is
infinity i.e. value '0' for servers. It is enabled only for NFS servers.
These are the patches:
http://review.gluster.org/10267
http://review.gluster.org/10268

So I would say don't enable anything?

Pranith

-- 
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=K4PAOiMIUf&a=cc_unsubscribe


More information about the Bugs mailing list