In 2012, several new nodes with Intel i7 quad-core processors (with hyperthreading) have been added to the Deltares H4 cluster. These new nodes are significantly faster than the old ones, and are accessible via a separate queue: 'normal-i7'. Since these nodes have 4 cores, they are most efficient when dealing with parallel jobs, for instance XBeach in MPI mode.
Writing shell scripts
Submitting a job to the cluster is done as usual, by submitting a shell script. The easiest way to create these shell scripts is using the XBeach MATLAB toolbox (for more information, see Matlab Toolbox, make sure to use an up-to-date version). The command used is xb_write_sh_scripts.
To create scripts for the i7 nodes use the syntax:
| Code Block |
|---|
xb_write_sh_scripts('localpath', 'runpath', 'binary','XBeach_binarypath', 'mpitype', 'openmpi', 'nodes', nr_of_nodes, 'name', 'runname', 'queuetype', 'normal-i7'); |
When choosing the normal-i7 queue, the job is split into 4 separate processes per node (so if you select 2 nodes, the job will be split up into 8 processes, in order to use all the cores on every node). Currently, this only works for 'openmpi' and not for 'mpich2'.
Be aware that both runpath and binary need to be defined in such a way that the H4 cluster can find it (hence in a /u/user/... format instead of U:/user/...). Also, the XBeach binary specified needs to be compiled with MPI support (see Running XBeach at Deltares H4 cluster)
Running xb_write_sh_scripts results in two shell scripts (xbeach.sh and mpi.sh) created in 'localpath'.
Submitting jobs
To submit the job to the cluster, use putty, go to the folder containing the shell scripts. Submit the job by using the command: qsub -q normal-i7 xbeach.sh
By adding
| Code Block |
|---|
-q normal-i7 |
the job is sent to the normal-i7 queue instead of the default queue (which is named 'normal'). xbeach.sh in turn automatically submits mpi.sh to the 'normal-i7' queue. Usage of all nodes can be viewed at the H4 cluster staus page. If all went well, the nodes that your xbeach simulation runs on have a load greater than 1 (actual load > maximum load of 1 core, so multiple cores are in use).