PDI-Cluster

• Latest news
• Accounts
• Compute nodes
• Data storage
• Queuing system
• Contact

Latest news

20091125 cnode07 is taken offline due to memory problems.

20091118 The cluster is moved to its PDI location. It can only be reached from the PDI intranet. Its address is now cc.pdi-berlin.de (10.7.31.60). Logins have not been changed. Happy computing!

Important: The cluster will soon be moved to its new location at the PDI and placed in a water cooled rack, which is expected to arrive on September 23rd. Expect some downtime the following week. Dates will be given when available.

Important: The cluster will be switched of between 20.06.-22.06.2009, due to a planned power outage. Please plan in advance.

20090923: Planned arrival of the water cooled rack

20090821: Openmpi upgraded to version 1.3.3.

20090821: Abinit version 1.8.4 installed.

20090417: Upgrade to Lenny finished. Tests successfull. Happy computing!

20090416: Upgrade to Lenny nearly completed, remaining nodes cnode03 and cnode05 (job still running).

20090407: Upgrading to Lenny, still in progress so far without interruption of the cluster. Cnode26, cnode27 are done.

Accounts

Accounts are given on request.

Important: Password change with

yppasswd

Please do this after first login.

Compute nodes

The cluster consists of the following nodes:

Name	Processor	CPUs	Memory	System	Properties	Remarks
cnode01	2 Quadcore	8	32 Gb	Debian	Gb16	Server, compute node
cnode02	2 Quadcore	8	32 Gb	Debian	Gb32, amd64	compute nodes
cnode03, cnode04	2 Quadcore	8	16 Gb	Debian	Gb16, amd64	compute nodes
cnode05, cnode06	2 Quadcore	8	32 Gb	Debian	Gb32, amd64	compute nodes
cnode07, cnode08	2 Quadcore	8	16 Gb	Debian	Gb16, amd64	compute nodes
cnode09, cnode10	2 Quadcore	8	32 Gb	Debian	Gb32, amd64	compute nodes
cnode11, cnode12	2 Quadcore	8	16 Gb	Debian	Gb16, amd64	compute nodes
cnode13-cnode18	2 Quadcore	8	16 Gb	Debian	Gb16, amd64	compute nodes
cnode19-cnode24	2 Quadcore	8	32 Gb	Debian	Gb32, amd64	compute nodes
cnode25-cnode28	2 Quadcore	8	8 Gb	Debian	Gb8, i386	compute nodes

top

Data storage

Hard disk space (approximate values) is available for user data on the following locations:

Location	Space	Comment
/users/username/	50 Gb total	Home directory
/disks/cnode01/	50  Gb	Data
/disks/cnode02/
...	100 Gb each	Data
/disks/cnode24/
/disks/cnode25/
...	180 Gb each	Data
/disks/cnode28/

top

Queuing system

Batch jobs are managed by Torque, a free version of PBS and successor of OpenPBS. Jobs can be submitted from the server (cnode01) and will be distributed according to the requirements across the cluster.

Queues

Name	User	Nodes	No. Cpus	Comment
defq	All			Default queue
GaSb	1	2	16
abinit	1	2	16

How to

First prepare a queue script in the directory where the job should be run (replace a.out with the name of your executable and modify the number of processors):

 

Serial shell script

Parallel shell script

Then execute

 qsub queuescript          with queuescript being the name of your file

and confirm that the job is submitted succesfully by

 qstat

Options which can be set in the queuescript can also be given as command line arguments to qsub. One can get an overview over the nodes by typing

 pbsnodelist

More information is given by

 pbsnodes

After the job is finished there are two additional files in the submission directory.

queuescript.o{$jobnumber}    containing the output of the queuing script

and

queuescript.e{$jobnumber}    containing the error messages (should be empty)

In a case where you want to stop your job before it is finished you execute

qdel {$jobnumber}

For parallel jobs, the queuing system allocates the processors and passes the information to mpirun, see sample script above.

Note that PBS allocates nodes, not CPUs in the first place. You may ask for 8 CPUs with
-l nodes=1:ppn=8 (8 CPUs on 1 node) or with
-l nodes=8 (1 CPU on each of 8 nodes).
In the latter case, the spare processors may be allocated to other jobs. Therefore, the first version is preferred if you want to have full control over the resources (memory, network traffic).
top

Contact

Suggestions or questions?
Email me: Frank Grosse
top