Le vendredi 12 avril 2013 à 19:45 +0200, Olivier Bonvalet a écrit :
> Le vendredi 12 avril 2013 à 10:04 -0500, Mark Nelson a écrit :
> > On 04/11/2013 07:25 PM, Ziemowit Pierzycki wrote:
> > > No, I'm not using RDMA in this configuration since this will eventually
> > > get deployed to production with 10G ethernet (yes RDMA is faster). I
> > > would prefer Ceph because it has a storage drive built into OpenNebula
> > > which my company is using and as you mentioned individual drives.
> > >
> > > I'm not sure what the problem is but it appears to me that one of the
> > > hosts may be holding up the rest... with Ceph if the performance of one
> > > of the hosts is much faster than others could this potentially slow down
> > > the cluster to this level?
> >
> > Definitely! Even 1 slow OSD can cause dramatic slow downs. This is
> > because we (by default) try to distribute data evenly to every OSD in
> > the cluster. If even 1 OSD is really slow, it will accumulate more and
> > more outstanding operations while all of the other OSDs complete their
> > requests. What will happen is that eventually you will have all of your
> > outstanding operations waiting on that slow OSD, and all of the other
> > OSDs will sit idle waiting for new requests.
> >
> > If you know that some OSDs are permanently slower than others, you can
> > re-weight them so that they receive fewer requests than the others which
> > can mitigate this, but that isn't always an optimal solution. Some
> > times a slow OSD can be a sign of other hardware problems too.
> >
> > Mark
> >
>
> and does response time of OSD are log somewhere, to identify that "weak
> link" ?
>
>
I think I found the answer with the admin socket :
ceph --admin-daemon /var/run/ceph/ceph-osd.14.asok perf dump
For example in the output I can see op_latency, op_w_latency,
op_r_latency, op_rw_latency, etc. Great.
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
