|
Running Jobs on the Argonne BG/L System (BGL)
Access to BGL compute nodes is provided via the cqsub command.
cqsub -t <time> -n <nodecount> -c <#processors> -m <mode> <exe> [arg1,arg2,...]
Where:
<time> is in minutes (required)
<nodecount> is the number of nodes to use (required)
<#processors> use this if the number of nodes != number of procs
(optional, default is <nodecount>)
<mode> is one of 'co' or 'vn' (optional, default is 'co')
<exe> is the full path name to the mpi executable
[arg1,arg2,...] are any arguments to the mpi executable
The partition will be automatically selected from the available ones.
To check the queue:
cqstat
The Argonne BlueGene/L system (BG/L) has 1024 compute nodes that are
divided into partitions of the following sizes (the
partition size it the number of compute nodes, the number
of processors is double that because each compute node has two processors).
Please note that
when you use a partition size smaller than 512, your code cannot use the
full Torus network. Accessing the Torus requires the
midplanes links (only accessible from 512 node and larger partitions) because
the torus is formed by interleaving midplanes to avoid the long cabling
required at the end of a long succession of midplanes.
Not using the Torus network will most likely cause the performance to be
very poor. In addition, 64-node and 256-node partitions have sub-optimal
tree network topologies, so file I/O performance may suffer to some degree on
these partition sizes.
| Desired Usage |
Partition Size |
# of processors |
| Development, Scaling |
32 |
64 |
| Development, Scaling |
64 |
128 |
| Development, Scaling |
128 |
256 |
| Development, Scaling |
256 |
512 |
| No Development |
512 |
1024 |
| No Development |
1024 |
2048 |
|