On 02/03/2012 10:06 AM, Gregory Farnum wrote:
On Feb 3, 2012, at 8:18 AM, Jim Schutt<jasc...@sandia.gov> wrote:
On 02/02/2012 05:28 PM, Gregory Farnum wrote:
On Thu, Feb 2, 2012 at 12:22 PM, Jim Schutt<jasc...@sandia.gov> wrote:
I found 0 instances of "waiting for commit" in all my OSD logs for my last
run.
So I never waited on the journal?
Looks like it. Interesting.
So far I'm looking at two behaviours I've noticed that seem anomalous to
me.
One is that I instrumented ms_dispatch(), and I see it take
a half-second or more several hundred times, out of several
thousand messages. Is that expected?
How did you instrument it? If you wrapped the whole function it's
possible that those longer runs are actually chewing through several
messages that had to get waitlisted for some reason previously.
(That's the call to do_waiters().)
Yep, I wrapped the whole function, and also instrumented taking osd_lock
while I was there. About half the time that ms_dispatch() takes more than
0.5 seconds, taking osd_lock is responsible for the delay. There's two
dispatch threads, one for ops and one for rep_ops, right? So one's
waiting on the other?
There's just one main dispatcher; no split for the ops and rep_ops .
The reason for that "dispatch_running" is that if there are requests
waiting then the tick() function will run through them if the
messenger dispatch thread is currently idle.
But it is possible for the Messenger to try and dispatch, and for that
to be blocked while some amount of (usually trivial) work is being
done by a different thread, yes. I don't think we've ever observed it
being a problem for anything other than updating OSD maps, though...
Ah, OK.
I guess I was confused by my log output, e.g.:
D'oh. Sorry, you confused me with your reference to repops, which
aren't special-cased or anything. But there are two messengers on the
OSD, each with their own dispatch thread. One of those messengers is
for clients and one is for other OSDs.
And now that you point that out, I wonder if the problem is lack of
Cond signaling in ms_dispatch. I'm on my phone right now but I believe
there's a chunk of commented-out code (why commented instead of
deleted? I don't know) that we want to uncomment for reasons that will
become clear when you look at it. :)
Try that and see how many of your problems disappear?
So I cherry-picked Sage's commit 7641a0e171f onto the code
I've been running (1fe75ee6419 + some debug stuff), and saw
no obvious difference in behaviour.
I also tested Sage's suggestion of separating journals and
data, by putting two journal partitions on half my disks,
and two data partitions on the other half. I made the data
partitions relatively small (~200 GiB each on a 1 TiB drive)
to minimize the effect of inner vs. outer tracks.
That didn't seem to help either.
Still looking -- Jim
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html