[Gluster-users] Possible to put some sort of qos/rate-limit on glusterfsd?

Andrew Lau andrew at andrewklau.com
Sun Feb 2 09:18:10 UTC 2014


On Sun, Feb 2, 2014 at 7:05 PM, James <purpleidea at gmail.com> wrote:

> On Sun, Feb 2, 2014 at 12:46 AM, Andrew Lau <andrew at andrewklau.com> wrote:
> > Hi all,
> Hi,
>
> >
> > Sadly my google skills aren't finding me any results - is there an
> option to
> > limit the CPU usage and/or the disk IO intensity of glusterfsd.
> This is what cgroups are for. Looking forward to you setting this up
> and telling us how it goes and what the optimized values are. AFAIK,
> gluster doesn't have anything built in to limit this. I actually don't
> think that it makes sense for it to either.
>

Thanks for the suggestions on cgroups, I actually did not think about that.
I'll try that and report my findings.



>
> >
> > Example scenario, oVirt + gluster on the same host when it comes to
> adding
> > an extra host + replicated brick the original host with the brick goes
> crazy
> > with 500% cpu as it copies just under 1TB of data across to the new
> > replicated brick. Going crazy I mean everything else will hang, simple
> "ls"
> > command will take 30+ seconds.
> Not sure what the correct solution is here. Maybe other gluster dev's
> can comment. Although it's really not detailed and you comment is
> unclear, so i'd elaborate before expecting a good response.
>
>
Sorry for the poor choice of wordings - here's attempt 2:

Host 1 (1TB brick w/ no replica)
Host 2 (NEW - 1TB brick to join Host 1 as replica 2)

Top will report glusterfsd with up to 500% CPU usage (combined processes).
Causing the system to grind to a halt very quickly.



>  >
> > Limiting the network bandwidth to 200Mbps seems to solve this issue, I'm
> > quite sure this is a CPU issue rather than IO so I was wondering if
> there's
> > any possibility to limit this down so the nic's itself don't have to get
> > rate limited.
> I occasionally use shorewall+puppet-gluster to rate limit the network
> interfaces. Puppet-Gluster could do per interface and per port/gluster
> connection throttling.
>

Rate limiting seems to be only solution right now (until I take a look into
cgoups). It seems the CPU cannot handle the high cpu usage that glusterfsd
will create.


>
> >
> > Thanks,
> > Andrew
>
> HTH,
> James
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://supercolony.gluster.org/pipermail/gluster-users/attachments/20140202/8a222c2b/attachment.html>


More information about the Gluster-users mailing list