[Gluster-users] Freezing during heal

Kevin Lemonnier lemonnierk at ulrar.net
Mon May 2 09:05:10 UTC 2016 

Hi,

So after some testing, it is a lot better but I do still have some problems with 3.7.11.
When I reboot a server it seems to have some strange behaviour sometimes, but I need to test
that better.
Removing a server from the network, waiting for a while then adding it back and letting it heal
works perfectly, completly invisible for the user and that's perfect !

However when I add a brick, changing the replica count from 2 to 3, it starts a heal
and some VMs switch to read only. I have to power them off then on again to fix it,
clearly it's better than with 3.7.6 which froze the VM until the heal was complete,
but I would still like to understand why some of the VMs are switching to readonly.
Looks like it happens everytime I add a brick to increase the replica, I would like
to test adding a whole replica set at once but I just don't have the hardware for that.

Rebooting a node looks like it's making some VMs go read only too, but I need to test
that better. For some reason it looks like rebooting a brick or adding a brick is causing
I/O errors on some VM disks and not others, and I have to power them off and then on to fix it.
I can't just reboot them, I guess I have to actually re-open the file to trigger a heal ?

Any idea on how to prevent that ? It's a lot better than 3.7.6 'cause it can be fixed in a minute,
but that's still not great to explain to the clients.

Thanks


On Mon, Apr 25, 2016 at 02:01:09PM +0200, Kevin Lemonnier wrote:
> Hi,
> 
> So I'm trying that now.
> I installed 3.7.11 on two nodes and put a few VMs on it, same config
> as before but with 64MB shards and the heal algo to full. As expected,
> if I poweroff one of the nodes, everything is dead, which is fine.
> 
> Now I'm adding a third node, a big heal was started after the add-brick
> of everything (7000+ shards), and for now everything seems to be working
> fine on the VMs. Last time I tried adding a brick, all those VM died for
> the duration of the heal, so that's already pretty good.
> 
> I'm gonna let it finish to copy everything on the new nodes, then I'll try
> to simulate nodes going down to see if my original problem of freezing and
> low heal time is solved with this config.
> For reference, here is the volume info, if someone sees something I should change :
> 
> Volume Name: gluster
> Type: Replicate
> Volume ID: e4f01509-beaf-447d-821f-957cc5c20c0a
> Status: Started
> Number of Bricks: 1 x 3 = 3
> Transport-type: tcp
> Bricks:
> Brick1: ipvr2.client_name:/mnt/storage/gluster
> Brick2: ipvr3.client_name:/mnt/storage/gluster
> Brick3: ipvr50.client_name:/mnt/storage/gluster
> Options Reconfigured:
> cluster.quorum-type: auto
> cluster.server-quorum-type: server
> network.remote-dio: enable
> cluster.eager-lock: enable
> performance.quick-read: off
> performance.read-ahead: off
> performance.io-cache: off
> performance.stat-prefetch: off
> features.shard: on
> features.shard-block-size: 64MB
> cluster.data-self-heal-algorithm: full
> performance.readdir-ahead: on
> 
> 
> It starts at 2 and jumps to 50 because the first server is doing something else for now,
> and I use 50 to be the temporary third node. If everything goes well, I'll migrate the production
> on the cluster, re-install the first server and do a replace-brick, which I hope will work just as well
> as the add-brick I'm doing now. Last replace-brick also brought everything down, but I guess that was the
> joy of 3.7.6 :).
> 
> Thanks !
> 
> 
> On Mon, Apr 18, 2016 at 08:17:05PM +0530, Krutika Dhananjay wrote:
> > On Mon, Apr 18, 2016 at 8:02 PM, Kevin Lemonnier <lemonnierk at ulrar.net>
> > wrote:
> > 
> > > I will try migrating to 3.7.10, is it considered stable yet ?
> > >
> > 
> > Oops, just realized 3.7.10 had a regression. Then 3.7.11 it is. :)
> > 
> > 
> > > Should I change the self heal algorithm even if I move to 3.7.10, or is
> > > that not necessary ?
> > > Not sure what that change might do.
> > >
> > 
> > So the other algorithm is 'diff' which computes rolling checksum on chunks
> > of the src(es) and sink(s), compares them and heals upon mismatch. This is
> > known to consume lot of CPU. 'full' algo on the other hand simply copies
> > the src into sink in chunks. With sharding, it shouldn't be all that bad
> > copying a 256MB file (in your case) from src to sink. We've used double the
> > block size and had no issues reported.
> > 
> > So you could change self heal algo to full even in the upgraded cluster.
> > 
> > -Krutika
> > 
> > 
> > >
> > > Anyway, I'll try to create a 3.7.10 cluster in the week end slowly move
> > > the VMs on it then,
> > > Thanks a lot for your help,
> > >
> > > Regards
> > >
> > >
> > > On Mon, Apr 18, 2016 at 07:58:44PM +0530, Krutika Dhananjay wrote:
> > > > Hi,
> > > >
> > > > Yeah, so the fuse mount log didn't convey much information.
> > > >
> > > > So one of the reasons heal may have taken so long (and also consumed
> > > > resources) is because of a bug in self-heal where it would do heal from
> > > > both source bricks in 3-way replication. With such a bug, heal would take
> > > > twice the amount of time and consume resources both the times by the same
> > > > amount.
> > > >
> > > > This issue is fixed at http://review.gluster.org/#/c/14008/ and will be
> > > > available in 3.7.12.
> > > >
> > > > The other thing you could do is to set cluster.data-self-heal-algorithm
> > > to
> > > > 'full', for better heal performance and more regulated resource
> > > consumption
> > > > by the same.
> > > >  #gluster volume set <VOL> cluster.data-self-heal-algorithm full
> > > >
> > > > As far as sharding is concerned, some critical caching issues were fixed
> > > in
> > > > 3.7.7 and 3.7.8.
> > > > And my guess is that the vm crash/unbootable state could be because of
> > > this
> > > > issue, which exists in 3.7.6.
> > > >
> > > > 3.7.10 saw the introduction of throttled client side heals which also
> > > moves
> > > > such heals to the background, which is all the more helpful for
> > > preventing
> > > > starvation of vms during client heal.
> > > >
> > > > Considering these factors, I think it would be better if you upgraded
> > > your
> > > > machines to 3.7.10.
> > > >
> > > > Do let me know if migrating to 3.7.10 solves your issues.
> > > >
> > > > -Krutika
> > > >
> > > > On Mon, Apr 18, 2016 at 12:40 PM, Kevin Lemonnier <lemonnierk at ulrar.net>
> > > > wrote:
> > > >
> > > > > Yes, but as I was saying I don't believe KVM is using a mount point, I
> > > > > think it uses
> > > > > the API (
> > > > >
> > > http://www.gluster.org/community/documentation/index.php/Libgfapi_with_qemu_libvirt
> > > > > ).
> > > > > Might be mistaken ofcourse. Proxmox does have a mountpoint for
> > > > > conveniance, I'll attach those
> > > > > logs, hoping they contain the informations you need. They do seem to
> > > > > contain a lot of errors
> > > > > for the 15.
> > > > > For reference, there was a disconnect of the first brick (10.10.0.1) in
> > > > > the morning and then a successfull
> > > > > heal that caused about 40 minutes downtime of the VMs. Right after that
> > > > > heal finished (if my memory is
> > > > > correct it was about noon or close) the second brick (10.10.0.2)
> > > rebooted,
> > > > > and that's the one I disconnected
> > > > > to prevent the heal from causing another downtime.
> > > > > I reconnected it one at the end of the afternoon, hoping the heal
> > > would go
> > > > > well but everything went down
> > > > > like in the morning so I disconnected it again, and waited 11pm
> > > (23:00) to
> > > > > reconnect it and let it finish.
> > > > >
> > > > > Thanks for your help,
> > > > >
> > > > >
> > > > > On Mon, Apr 18, 2016 at 12:28:28PM +0530, Krutika Dhananjay wrote:
> > > > > > Sorry, I was referring to the glusterfs client logs.
> > > > > >
> > > > > > Assuming you are using FUSE mount, your log file will be in
> > > > > > /var/log/glusterfs/<hyphenated-mount-point-path>.log
> > > > > >
> > > > > > -Krutika
> > > > > >
> > > > > > On Sun, Apr 17, 2016 at 9:37 PM, Kevin Lemonnier <
> > > lemonnierk at ulrar.net>
> > > > > > wrote:
> > > > > >
> > > > > > > I believe Proxmox is just an interface to KVM that uses the lib,
> > > so if
> > > > > I'm
> > > > > > > not mistaken there isn't client logs ?
> > > > > > >
> > > > > > > It's not the first time I have the issue, it happens on every heal
> > > on
> > > > > the
> > > > > > > 2 clusters I have.
> > > > > > >
> > > > > > > I did let the heal finish that night and the VMs are working now,
> > > but
> > > > > it
> > > > > > > is pretty scarry for future crashes or brick replacement.
> > > > > > > Should I maybe lower the shard size ? Won't solve the fact that 2
> > > > > bricks
> > > > > > > on 3 aren't keeping the filesystem usable but might make the
> > > healing
> > > > > > > quicker right ?
> > > > > > >
> > > > > > > Thanks
> > > > > > >
> > > > > > > Le 17 avril 2016 17:56:37 GMT+02:00, Krutika Dhananjay <
> > > > > > > kdhananj at redhat.com> a écrit :
> > > > > > > >Could you share the client logs and information about the approx
> > > > > > > >time/day
> > > > > > > >when you saw this issue?
> > > > > > > >
> > > > > > > >-Krutika
> > > > > > > >
> > > > > > > >On Sat, Apr 16, 2016 at 12:57 AM, Kevin Lemonnier
> > > > > > > ><lemonnierk at ulrar.net>
> > > > > > > >wrote:
> > > > > > > >
> > > > > > > >> Hi,
> > > > > > > >>
> > > > > > > >> We have a small glusterFS 3.7.6 cluster with 3 nodes running
> > > with
> > > > > > > >proxmox
> > > > > > > >> VM's on it. I did set up the different recommended option like
> > > the
> > > > > > > >virt
> > > > > > > >> group, but
> > > > > > > >> by hand since it's on debian. The shards are 256MB, if that
> > > matters.
> > > > > > > >>
> > > > > > > >> This morning the second node crashed, and as it came back up
> > > started
> > > > > > > >a
> > > > > > > >> heal, but that basically froze all the VM's running on that
> > > volume.
> > > > > > > >Since
> > > > > > > >> we really really
> > > > > > > >> can't have 40 minutes down time in the middle of the day, I just
> > > > > > > >removed
> > > > > > > >> the node from the network and that stopped the heal, allowing
> > > the
> > > > > > > >VM's to
> > > > > > > >> access
> > > > > > > >> their disks again. The plan was to re-connecte the node in a
> > > couple
> > > > > > > >of
> > > > > > > >> hours to let it heal at night.
> > > > > > > >> But a VM crashed now, and it can't boot up again : seems to
> > > freez
> > > > > > > >trying
> > > > > > > >> to access the disks.
> > > > > > > >>
> > > > > > > >> Looking at the heal info for the volume, it has gone way up
> > > since
> > > > > > > >this
> > > > > > > >> morning, it looks like the VM's aren't writing to both nodes,
> > > just
> > > > > > > >the one
> > > > > > > >> they are on.
> > > > > > > >> It seems pretty bad, we have 2 nodes on 3 up, I would expect the
> > > > > > > >volume to
> > > > > > > >> work just fine since it has quorum. What am I missing ?
> > > > > > > >>
> > > > > > > >> It is still too early to start the heal, is there a way to
> > > start the
> > > > > > > >VM
> > > > > > > >> anyway right now ? I mean, it was running a moment ago so the
> > > data
> > > > > is
> > > > > > > >> there, it just needs
> > > > > > > >> to let the VM access it.
> > > > > > > >>
> > > > > > > >>
> > > > > > > >>
> > > > > > > >> Volume Name: vm-storage
> > > > > > > >> Type: Replicate
> > > > > > > >> Volume ID: a5b19324-f032-4136-aaac-5e9a4c88aaef
> > > > > > > >> Status: Started
> > > > > > > >> Number of Bricks: 1 x 3 = 3
> > > > > > > >> Transport-type: tcp
> > > > > > > >> Bricks:
> > > > > > > >> Brick1: first_node:/mnt/vg1-storage
> > > > > > > >> Brick2: second_node:/mnt/vg1-storage
> > > > > > > >> Brick3: third_node:/mnt/vg1-storage
> > > > > > > >> Options Reconfigured:
> > > > > > > >> cluster.quorum-type: auto
> > > > > > > >> cluster.server-quorum-type: server
> > > > > > > >> network.remote-dio: enable
> > > > > > > >> cluster.eager-lock: enable
> > > > > > > >> performance.readdir-ahead: on
> > > > > > > >> performance.quick-read: off
> > > > > > > >> performance.read-ahead: off
> > > > > > > >> performance.io-cache: off
> > > > > > > >> performance.stat-prefetch: off
> > > > > > > >> features.shard: on
> > > > > > > >> features.shard-block-size: 256MB
> > > > > > > >> cluster.server-quorum-ratio: 51%
> > > > > > > >>
> > > > > > > >>
> > > > > > > >> Thanks for your help
> > > > > > > >>
> > > > > > > >> --
> > > > > > > >> Kevin Lemonnier
> > > > > > > >> PGP Fingerprint : 89A5 2283 04A0 E6E9 0111
> > > > > > > >>
> > > > > > > >> _______________________________________________
> > > > > > > >> Gluster-users mailing list
> > > > > > > >> Gluster-users at gluster.org
> > > > > > > >> http://www.gluster.org/mailman/listinfo/gluster-users
> > > > > > > >>
> > > > > > >
> > > > > > > --
> > > > > > > Envoyé de mon appareil Android avec K-9 Mail. Veuillez excuser ma
> > > > > brièveté.
> > > > > > > _______________________________________________
> > > > > > > Gluster-users mailing list
> > > > > > > Gluster-users at gluster.org
> > > > > > > http://www.gluster.org/mailman/listinfo/gluster-users
> > > > > > >
> > > > >
> > > > > --
> > > > > Kevin Lemonnier
> > > > > PGP Fingerprint : 89A5 2283 04A0 E6E9 0111
> > > > >
> > > > > _______________________________________________
> > > > > Gluster-users mailing list
> > > > > Gluster-users at gluster.org
> > > > > http://www.gluster.org/mailman/listinfo/gluster-users
> > > > >
> > >
> > > --
> > > Kevin Lemonnier
> > > PGP Fingerprint : 89A5 2283 04A0 E6E9 0111
> > >
> > > _______________________________________________
> > > Gluster-users mailing list
> > > Gluster-users at gluster.org
> > > http://www.gluster.org/mailman/listinfo/gluster-users
> > >
> 
> -- 
> Kevin Lemonnier
> PGP Fingerprint : 89A5 2283 04A0 E6E9 0111



> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://www.gluster.org/mailman/listinfo/gluster-users


-- 
Kevin Lemonnier
PGP Fingerprint : 89A5 2283 04A0 E6E9 0111
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: Digital signature
URL: <http://www.gluster.org/pipermail/gluster-users/attachments/20160502/b5af3d2d/attachment.sig>

More information about the Gluster-users mailing list