[Linux-cluster] Known limits ?

[Patrick Caulfield]

Mathieu Avila wrote:
> Hello GFS team,
> 
> I'm trying to run a GFS filesystem on ~32 nodes, but there is a problem
> when i start the daemons.
> My nodes are called sam38 -> sam70.
> - I run ccsd on all nodes (using init script) at the same time, and it's
> ok.
> - I run cman on all nodes (using init script) at the same time, and it's
> ok. "cman_tool nodes" tell me alll nodes have rejoined the cluster.
> - I run fenced on all nodes, one by one every second, and it fails from
> sam57 to sam70.
> 
> From the last one that succeeds (sam56), i see :
> [root at sam56 ~]# cat /proc/cluster/services
> Service          Name                              GID LID State     Code
> Fence Domain:    "default"                           1   2 recover 4 -
> [21 19 9 8 7 1 2 3 4 6 11 13 16 23 26 27 28 32 33 25 20 14]
> 
> Then, at sam57 , i get in /var/log/messages:
> Jul 20 18:38:17 sam57 kernel: CMAN: got WAIT barrier not in phase 1
> TRANSITION.44 (2)
> 
> when trying to run fenced.
> 
> Then i /etc/init.d/fenced stop, and i get :
> Jul 20 18:47:44 sam57 fenced[28722]: process_events: service leave failed
> Jul 20 18:47:44 sam57 fenced: shutdown succeeded
> 
> 
> When i start it again:
> Jul 20 18:47:45 sam57 fenced[28964]: fence_domain_add: service set level
> failed
> 
> 
> After this step, i did stop everything on sam38 (fenced/ccsd/cman), to
> see whether getting one node out would let me get another one in, but i
> got this strange message on sam39:
> Jul 20 19:14:04 sam39 kernel: CMAN: node sam38 has been removed from the
> cluster : No response to messages
> Jul 20 19:14:12 sam39 kernel: SM: 00000001 process_recovery_barrier
> status=-104
> 
> 
> In a previous exprerience, running fenced all at the same time lead to a
> global cluster failure (everything did not respond): this is why i tried
> to run them one by one.
> 
> 
> All nodes are 64bits, i use the latest "cluster" code from the STABLE
> branch of the CVS. My configuration file is classical and works for a
> few nodes (tested many times on 5 nodes with no problem), except for this:
> <fence_daemon post_join_delay="30"></fence_daemon>
> 
> Why does fenced daemon fails to start where cman succeeded ? I thought
> it was just a service like any other, and was built on the top of CMAN.
> Also, what are the known limits of the cluster infrastructure, in terms
> of nodes ?
> 
> Do you have any advice on how to remove this problem ? particularly, at
> this point, does it change something to choose DLM instead of Gulm ?
> (no, i think, but i'd rather be sure)

You do seem to have hit a limit.

Personally I've only tested up to 31 nodes and recently someone posted to this
list with a similar problem on 38 nodes - however, when I looked at the logs
it actually seem to fall over at 32!

There's nothing hard-coded in cman to limit the number of nodes to that amount
 and I can't find anything obvious that should cause it to happen.

In the mean time all I suggest is that you use gulm for clusters with >= 32 nodes


-- 

patrick