[Gluster-users] glusterfs, pacemaker and Filesystem RA

[Uwe Weiss]

Hello List
due to a mistake my post from yesterday has been cut.  This is why I try to
send my post again and open it as new thread. I hope it will work this time.

<---- Original postet mail starts here ---->

Hello Marcel, hello Samuel,

sorry for my late answer, but I was away for two months and for that I could
continue my tests last week.

First of all thank you for your patch of the Filesystem RA. It works like a
charm but I have some little remarks. 

What I found out is that the test of the filesystem access with
OCF_CHECK_LEVEL is not working with glusterfs.

If I use the nvpair OCF_CHECK_LEVEL with a value of 10 I get an err_message
with the content: 
' 192.168.51.1:/gl_vol0 is not a block device, monitor 10 is noop'

If I use the nvpair OCF_CHECK_LEVEL with a value of 20 I get an err_message
with the content:
 ' ERROR: dd said: dd: opening
`/virtfs0/.Filesystem_status/res_glusterfs_sp0:0_0_vmhost1': Invalid
argument'
After that the resource is trying to restart permanently

Unfortunately I am not familiar enough with scripting to fix it by myself
and to contribute it.

Another item I would like to discuss is a bit more general. 
As Samuel pointed out the Filesystem RA (with native client) needs the
gluster node it connects to (by using the device attribute of the
Filsesystem RA) up and running only on startup of the client. 

After that the native client detects by itself if a gluster node is gone or
not. This is correct so far but in my setup this could be a SPOF.

I would like to build a cluster of three machines (A,B,C) and start a
Filesystem RA clone on all three clusternodes. Each of that nodes is a
glusterfs server offering a glusterfs replicated share  and also a glusterfs
client which mounts that share (from server A initially).  If all three
servers are up there is no problem. 

Even if one of the servers is going down everything will work fine.

But If the node the clients are connected to on startup (Server A) crashes
and I have afterwards a need to reboot one of the remaining servers (B or C)
, this server is not able to reconnect as client  cause node A is still
down.

>From my point of view it could be a solution to give some nvpairs (e.g.
glusterhost1=IPorName1ofNodeA:/glustervolume,
glusterhost2=IPorName2ofNodeB:/glustervolume, ...
glusterhostN=IPorNameNofNodeN:/glustervolume). 

These pairs could be pinged and/or tested before the Filesystem RA tries to
connect to them. In case that one of these nodes is not reachable or does
not respond to the connection attempt  the RA could try a connection with
the next nvpair.

Background:

I would like to build a openais/pacemaker cluster consisting of three nodes.

On each node should run a gluster server providing a replicated glusterfs
share, a glusterfs client (Filesystem RA clone) connected to this share and
one or more KVM-VMs 

Due to load reason the VMs should be distributed over the cluster. In case
of crash of one of these servers the affected VM shall fail over to the
remaining nodes.

I hope I was able to explain my concerns and you or anybody else could give
me a hint to solve my problem.

Thx in advance
Uwe



-----Ursprüngliche Nachricht-----
Von: gluster-users-bounces at gluster.org
[mailto:gluster-users-bounces at gluster.org] Im Auftrag von Marcel Pennewiß
Gesendet: Montag, 18. Juli 2011 14:00
An: gluster-users at gluster.org
Betreff: Re: [Gluster-users] glusterfs and pacemaker

On Monday 18 July 2011 13:26:00 samuel wrote:
> I don't know from which version on but, if you use the native client 
> for mounting the volumes, it's only required to have the IP active in 
> the mount moment. After that, the native client will transparently 
> manage node's failure.

ACK, that's why we use this shared IP (e.g. for backup issues via nfs).
AFAIR glusterFS retrieves Volfile (via shared IP) and connects to the nodes.

Marcel
_______________________________________________
Gluster-users mailing list
Gluster-users at gluster.org
http://gluster.org/cgi-bin/mailman/listinfo/gluster-users