Just some more info for you. I'm glad to see there has been thought put into
some sort of "backgrounding" of the AFR. I saw "massive" across all bricks
when I started using AFR (with 750GB of user data, 270K files). I was doing
an RSYNC to update the glusterfs data from the master file server. So
anytime anyone wants to increase the number of relicas, and then go through
all the files (eg. using an rsync), there will be a lot of activity going
on.
I don't like the performance hit, and it is not "critical" that all the
replica happen right away -- in fact, it is extremely low priority. My
highest priority is low latency to my master files. It seems that the best
time to do the replica is not on the open(), but maybe some talk to a
"scheduler" xlator who is also looking at bandwidth availability.
I want to bring bandwidth into the picture, because making replicas across a
wan connection makes sense for disaster recovery. In the case of the bricks
out of sync with email files (where the replica data could chew up all
available wan bandwidth), maybe the AFR, and a separate "garbage cleanup" or
file syncing could be done by another process that is also working in the
background. I could forsee that with large datasets, and network outages,
that large datasets will be out of sync, and have to resync, that resyncing
is never "finished". Having a large performance hit or demand on the wan
bandwidth when a network outage is fixed would be horrible.
From: "Anand Avati" <avati@xxxxxxxxxxxxx>
To: "Gerry Reno" <greno@xxxxxxxxxxx>
CC: gluster-devel <gluster-devel@xxxxxxxxxx>
Subject: Re: self-heal behavior
Date: Wed, 4 Jul 2007 19:33:14 +0530
Gerry,
your question is appropriate, but the answer to 'when to resync' is not
very simple. when a brick which was brought down is brought up later, it
may
be a completely new (empty) brick. In that case starting to sync every file
would most likely be the wrong decision. (we should rather sync the file
which the user needs than some unused file). Even if we chose to sync files
without user accessing them it would be very sluggish too since it would be
intervening in other operations.
The current approach is to sync files on the next open() on it. This is
usually a good balance since, during open() if we were to sync a file, even
if it was a GB it would take 10-15 secs, and for normal files (in the order
of few MBs) it is almost not noticable. But if this were to happen together
for all files whether the user accessed them or not there would be a lot of
traffic and be very sluggish.
This approach of syncing on open() is what even other filesystems which
support redundancy do.
Detecting 'idle time' and beginning sync-up and pausing the sync-up when
user begins activity is a very tricky job, but that is definitely what we
aim at finally. It is not enough if AFR detects the client is free, because
the servers may be busy serving files to another client and syncing at that
time may not be the most apprpriate time. The following versions of AFR
will
have more options to tune 'when' to sync. Currently it is only at open().
We
plan to add options to make it sync on lookup() (happens on ls). Later
versions would have pro-active syncing (detecting that both server and
clients are idle etc).
thanks,
avati
2007/7/4, Gerry Reno <greno@xxxxxxxxxxx>:
I've been doing some testing of self-heal. Basically taking down one
brick and then copying some files to one of the client mounts, then
bringing the downed brick back up. What I see is that when I bring the
downed brick back up, no activity occurs. It's only when I start doing
something in one of the client mounts that something occurs to rebuild
the out-of-sync brick. My concern with this is that if I have four
applications on different client nodes (separate machines) using the
same data set (mounted on GlusterFS). The brick on one of these nodes
is out-of-sync, and it is not until some user is trying to use the
application that the brick starts to resync. This results in sluggish
performance to the user as all the data has to be brought over the
network from other bricks since the local brick is out-of-sync. Now
there may have been ten minutes of idle time prior to this user trying
to access the data but glusterfs did not make any use of this time to
rebuild the out-of-sync brick but rather waited until a user tried to
access data. To me, it appears that glusterfs should be making use of
such opportunity and this would diminish the overall impact to users of
the out-of-sync condition.
Regards,
Gerry
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxx
http://lists.nongnu.org/mailman/listinfo/gluster-devel
--
Anand V. Avati
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxx
http://lists.nongnu.org/mailman/listinfo/gluster-devel
_________________________________________________________________
http://newlivehotmail.com
