The topology of the Infiniband network connecting the compute nodes of CHiC is hierarchical. That implies different communication latency between different pairs of compute nodes. To minimize this latency one can request compute nodes from the batch system which are in one rack (a rack houses up to 38 compute nodes) and thus connected to the same Infiniband switch. This can be done with the following requests:
qsub -I -l nodes=4:ppn=2:chic1b
#PBS -l nodes=4:ppn=2:chic1b
Where chic1b is the rack.
To see which racks have enough free resources type: