= GCC cluster =

The GCC has its own 480 core cluster. The main workhorses are 10 servers each with
 * 48 cores
 * 256 GB RAM
 * 1 GBit management NIC
 * 10 GBit NIC for a dedicated fast IO connection to a 
 * 2 PB shared GPFS for storage

= For users =

== Login to the User Interface server ==

To submit jobs, check the status, test scripts, etc. you need to login on the user interface server a.k.a. cluster.gcc.rug.nl using SSH.
Please note that cluster.gcc.rug.nl is only available from within certain RUG/UMCG subnets. From outside you need a double hop; Firstly login to the proxy:
{{{
$> ssh [your_account]@proxy.gcc.rug.nl
}}}
followed by:
{{{
$> ssh [your_account]@cluster.gcc.rug.nl
}}}
If you are inside certain subnets of the RUG/UMCG network, you can skip the proxy and login to cluster.gcc.rug.nl directly.

== Available queues ==

In order to quickly test jobs you are allowed to run the directly on cluster.gcc.rug.nl outside the scheduler. Please think twice though before you hit enter: if you crash cluster.gcc.rug.nl others can no longer submit or monitor their jobs, which is pretty annoying. On the other hand it's not a disaster as the scheduler and execution daemons run on physically different servers and hence are not affected by a crash of cluster.gcc.rug.nl.

To test how your jobs perform on an execution node and get an idea of the typical resource requirements for your analysis you should submit a few jobs to the test queues first. The test queues run on a dedicated execution node, so in case your jobs make that server run out of disk space, out of memory or do other nasty things accidentally, it will not affect the production queues and ditto nodes.

Once you've tested your job scripts and are sure they will behave nice & perform well, you can submit jobs to the production queue named ''gcc''. In case you happen to be part of the gaf group and need to process high priority sequence data for the Genome Analysis Facility you can also use the ''gaf'' queue.

||**Queue**||**Job type**||**Limits**||
||test-short||debugging||10 minutes max. walltime per job; limited to a single test node / 48 cores||
||test-long||debugging||max 4 jobs running simultaneously per user; limited to half the test node / 24 cores||
||gcc||production - default prio||none||
||gaf||production - high prio||only available to users from the gaf group||

== Useful commands ==

Please refer to the Torque manuals for a complete overview. Some examples:

=== Submitting jobs: ===
Simple submit of job script to the default queue, which routes your job to the ''gcc'' production queue:
{{{
$> qsub myScript.sh
}}}
Submitting a job with a jobname different from the filename of the submitted script (default) and with a dependency on a previously submitted job.
This job will not start before the dependency has completed successfully:
{{{
$> qsub -N [nameOfYourJob] -W depend=afterok:[ID of a previously submitted job] myScript.sh
}}}
Instead of providing arguments to qsub on the commandline, you can also add them using the ''#PBS'' syntax as a special type of comments to your (bash) job script like this:
{{{
#!/bin/bash
#PBS -N jobName
#PBS -q test-short
#PBS -l nodes=1:ppn=2
#PBS -l walltime=00:06:00
#PBS -l mem=10mb
#PBS -e /some/path/to/your/testScript1.err
#PBS -o /some/path/to/your/testScript1.out

[Your actual work...]
}}}

=== Checking for the status of your jobs: ===
Default output for all users:
{{{
$> qstat
}}}
Long jobs names:
{{{
$> wqstat
}}}
Limit output to your own jobs
{{{
$> wqstat -u [your account]
}}}
Get "full" a.k.a detailed output for a specific job (you probably don't want that for all jobs....):
{{{
$> qstat -f [jobID]
}}}
Get other detailed status info for a specific job:
{{{
$> checkjob [jobID]
}}}

=== List jobs based on priority as in who is next in the queue: ===
{{{
$> diagnose -p
}}}

=== List available nodes: === 
{{{
$> pbsnodes
}}}

= For admins =


== Servers ==

||**Function**||**DNS**||**IP**||**Daemons**||**Comments**||
||User interface node||cluster.gcc.rug.nl||195.169.22.156||- (clients only)||Login node to submit and inspect jobs. [[BR]] Relatively powerful machine. [[BR]] Users can run code outside the scheduler for debugging purposes.||
||scheduler VM||scheduler01||195.169.22.214||''pbs_server''[[BR]]''maui''||Dedicated scheduler [[BR]] No user logins if this one is currently the production scheduler||
||scheduler VM||scheduler02||195.169.22.190||''pbs_server''[[BR]]''maui''||Dedicated scheduler [[BR]] No user logins if this one is currently the production scheduler||
||Execution node||targetgcc01||192.168.211.191||''pbs_mom''||Dedicated test node: only the ''test-short'' and ''test-long'' queues run on this node. [[BR]] Crashing the test node shall not affect production!.||
||Execution node||targetgcc02||192.168.211.192||''pbs_mom''||Redundant production node: only the default ''gcc'' and priority ''gaf'' queues run on this node.||
||Execution node||targetgcc03||192.168.211.193||''pbs_mom''||Redundant production node: only the default ''gcc'' and priority ''gaf'' queues run on this node.||
||Execution node||targetgcc04||192.168.211.194||''pbs_mom''||Redundant production node: only the default ''gcc'' and priority ''gaf'' queues run on this node.||
||Execution node||targetgcc05||192.168.211.195||''pbs_mom''||Redundant production node: only the default ''gcc'' and priority ''gaf'' queues run on this node.||
||Execution node||targetgcc06||192.168.211.196||''pbs_mom''||Redundant production node: only the default ''gcc'' and priority ''gaf'' queues run on this node.||
||Execution node||targetgcc07||192.168.211.197||''pbs_mom''||Redundant production node: only the default ''gcc'' and priority ''gaf'' queues run on this node.||
||Execution node||targetgcc08||192.168.211.198||''pbs_mom''||Redundant production node: only the default ''gcc'' and priority ''gaf'' queues run on this node.||
||Execution node||targetgcc09||192.168.211.199||''pbs_mom''||Redundant production node: only the default ''gcc'' and priority ''gaf'' queues run on this node.||
||Execution node||targetgcc10||192.168.211.200||''pbs_mom''||Redundant production node: only the default ''gcc'' and priority ''gaf'' queues run on this node.||

== PBS software / flavour ==

The current setup uses the resource manager **Torque 2.5.12** combined with the scheduler **Maui 3.3.1**.

=== Maui ===

Runs only on the schedulers with config files in $MAUI_HOME:
{{{
/usr/local/maui/
}}}

=== Torque ===

Torque clients are available on all servers.[[BR]]
Torque's pbs_server daemon runs only on the schedulers.[[BR]]
Torque's pbs_mom daemon runs only on the execution nodes where the real work is done.[[BR]]
Torque config files are installed in $TORQUE_HOME:
{{{
/var/spool/torque/
}}}

== Dual scheduler setup for seamless cluster upgrades ==

We use two schedulers: scheduler01 and scheduler02. These alternate as production and test scheduler. The production scheduler is hooked up to cluster.gcc.rug.nl and does not allow direct user logins. Hence you cannot submit jobs from the production scheduler, but only from cluster.gcc.rug.nl. The other is the test scheduler, which does not have a dedicated user interface machine and does allow direct user logins. You will need to login to the test scheduler in order to submit jobs. When it is time to upgrade software or tweak the !Torque/Maui configs:

 * We drain a few nodes: running jobs are allowed to finish, but no new ones will start.[[BR]]
   On the production scheduler as root:
{{{
$> qmgr -c 'set node targetgcc[0-9][0-9] state = offline'
}}}
 * Once ''idle'' move the drained nodes from the production to the test scheduler.[[BR]]
   Change the name of the scheduler in both these files on each node to be moved:
{{{
$TORQUE_HOME/server_name
$TORQUE_HOME/mom_priv/config
}}}
   On each execution node where the config changed run as root:
{{{
$> service pbs_mom restart
}}}
   On the test scheduler as root:
{{{
$> qmgr -c 'set node targetgcc[0-9][0-9] state = online'
}}}
 * Check the change in available execution nodes using:
{{{
$> pbsnodes
}}}
 * Test the new setup
 * Disable direct logins to the test scheduler
 * Enable direct logins to the production scheduler
 * Disable job submission from cluster.gcc.rug.nl on the production scheduler
 * Take cluster.gcc.rug.nl offline
 * Make cluster.gcc.rug.nl the user interface and submit host for the test scheduler
 * Take cluster.gcc.rug.nl back online: the test scheduler is now the new production scheduler and vice versa
 * Drain additional nodes and move them to the new production scheduler

== Installation details ==

=== Our current config files: ===
 * $TORQUE_HOME/mom_priv/[[attachment:gcc_pbs_mom.config.txt|config]]
 * $MAUI_HOME/[[attachment:gcc_maui.cfg.txt|maui.cfg]]
 * Other Torque settings can be loaded from a [[attachment:gcc_torque.txt|file]] using qmgr.[[BR]]
   To export/inspect the settings use:
{{{
$> qmgr -c 'p s'
}}}

=== Init scripts ===

Both the Torque and Maui source downloads contain a contrib folder with /etc/init.d/ scripts to start/stop the daemons:
 * /etc/init.d/[[attachment:suse.pbs_server|pbs_server]]
 * /etc/init.d/[[attachment:suse.maui|maui]]
 * /etc/init.d/[[attachment:suse.pbs_mom|pbs_mom]]
We use versions patched for:
* The location where the daemons are installed.
* The run levels at which the daemons should be started or stopped.
* Dependencies: GPFS is explicitly defined as service required for starting/stopping the Torque and Maui daemons.

To install:

On scheduler[01|02] as root:
{{{
$> cp suse.pbs_server /etc/init.d/pbs_server; chkconfig -a pbs_server; service pbs_server status
$> cp suse.maui       /etc/init.d/maui;       chkconfig -a maui;       service maui status
}}}

On targetgcc![01-10] as root:
{{{
$> cp suse.pbs_mom    /etc/init.d/pbs_mom;    chkconfig -a pbs_mom;    service pbs_mom status
}}}

=== wqstat ===

We have patched torque-2.5.12/src/cmds/[[attachment:qstat.c]] and recompiled the clients to create ''wqstat'', which reports long job names up to 40 characters as compared to the default 16. As normal user:
{{{
$> cd torque-2.5.12
$> ./configure --with-default-server=scheduler01 --disable-server --disable-mom --prefix=/some/other/location/cluster_clients/
$> make
$> make install
}}}
As root:
{{{
$> cp /some/other/location/cluster_clients/bin/qstat   /usr/local/bin/wqstat
}}}