ssic-linux-users

[SSI-users] CDSL's and typical SSI Linux cluster config

[Goetzman, D. <Dan...@bm...>]

Hello everyone,

 I have my OpenSSI cluster up and running on the Red Hat 7.3 distro. A
simple two node cluster, shared SCSI. Using CFS to share the filesystems and
HA-LVS for the virtual IP. To make things work I had to create the following
as CDSL's to allow each node to have a separate area;

 - /var/run
 - /var/log
 - /var/lock
 - /tmp

 Anyone have a list of other items that might need to be made a CDSL?
 I still have the /etc/sysconfig/network file mounted using the "mount
--bind" option, but maybe I should consider making this a CDSL as well?
 
 I followed the practice of creating a "{nodenum}" target in addition to the
"nodenum1", "nodenum2", targets so that I can boot the system up with a non
SSI Linux kernel and still have things working.

 Also, the standard start/stop scripts in /etc/init.d have interesting
behavior when /var/run is made unique via a CDSL. As the process PID's are
put in /proc and globally visible to each member of the cluster, the stop
action of the startup scripts can stop a process on ALL NODES of the
cluster. This is done by the "pidofproc" function in /etc/init.d/functions
(looking at Red Hat 7.3 here). It returns ALL pids found (on all nodes). IF
the start/stop script uses /var/run/<service>.pid then ONLY the pid for the
process on the node where the script is run will be killed (more like I
would have expected). All this kind of results in apparent random behavior.
Some start/stop scripts use a /var/run/.pid file, some do not. So, sometimes
the processes on just one node are stopped, other times they are stopped on
all nodes. And if the /var/run/.pid file is MIA for some reason, then the
process is killed on ALL nodes.

 I might consider a mod to the "pidofproc" function on my SSI cluster to
make things work more as I would expect.

 So, in summary...

 Did I miss any CDLS's I might need/want to create?
 Watch for the start/stop script and /var/run/.pid file interaction issue.

Thanks to all for the help in getting my first SSI Linux cluster up and
running.

Dan Goetzman
IS-PM&D Support
BMC Software, Houston Tx.

Re: [SSI-users] CDSL's and typical SSI Linux cluster config

[Brian J. W. <Bri...@hp...>]

Goetzman, Dan wrote:
 > To make things work I had to create the following
> as CDSL's to allow each node to have a separate area;
> 
>  - /var/run
>  - /var/log
>  - /var/lock
>  - /tmp

Is this needed for HA-LVS? I haven't played with it, so I don't know.

>  Anyone have a list of other items that might need to be made a CDSL?
>  I still have the /etc/sysconfig/network file mounted using the "mount
> --bind" option, but maybe I should consider making this a CDSL as well?

I think so. The --bind instructions dated back to time before we had CDSLs.

 > This is done by the "pidofproc" function in /etc/init.d/functions
> (looking at Red Hat 7.3 here). It returns ALL pids found (on all nodes). IF
> the start/stop script uses /var/run/<service>.pid then ONLY the pid for the
> process on the node where the script is run will be killed (more like I
> would have expected).

Returning all PIDs is more in line with the SSI philosophy. If the 
cluster is behaving like a single giant machine, then 
`/etc/init.d/xinetd stop` should kill all the xinetd daemons in the 
cluster. A new method to stop a service on a given node, however, would 
be useful. RC script handling is an aspect of OpenSSI that hasn't been 
fully designed. Input is more than welcome. :)

>  Did I miss any CDLS's I might need/want to create?

Dave added one for /var/lib/nfs/statd in OpenSSI 0.9.6 to support 
clusterized NFS client.

> Thanks to all for the help in getting my first SSI Linux cluster up and
> running.

No problem. I'm glad you like it.

Brian