Running CASA: Timing Tests
A series of tests designed to find the optimal way to run CASA on the RTDC machines.
The tests run tclean on a mosaiced dataset (this includes the partition task for the parallel option). tclean is the CASA task that has the maximum gain from parallelization. Tests run using CASA 5.1.2.
Summary
- Massive time gains by running tclean in parallel. Even using 3 cores is ~3x faster than serial processing.
- Significant gains (~50%) from using machines with faster processors (actually a combination of CPU and disk performace).
- No significant gain to using > 7 cores.
- Better to use fewer, unshared cores, than more, shared cores.
| Data size | Disk type | Machine | Parallel? | # Cores | Other parameters | Time | Comments |
| 6.4G | NFS | RTDC8 | N | - | blockdev=6496 | 4h16m | Data read from rglinux13 |
| 6.4G | local | RTDC8 | N | - | blockdev=6496 | 3h50m | Local disks definitely better |
| 6.4G | local | RTDC8 | Y | 9 | blockdev=6496 | 45m | Repeated and it completed in 36m |
| 6.4G | local | RTDC8 | Y | 7 | blockdev=6496 | 42m | Repeated and it completed in 38m & 43m. Only a couple of minutes separate 7 and 9 cores. |
| 6.4G | local | RTDC8 | Y | 7 | blockdev=250 | 67m | Confirms that blockdev makes a significant difference |
| 6.4G | local | RTDC8 | Y | 5 | blockdev=6496 | 57m | Fewer cores = longer. Repeated and it completed in 47m. |
| 6.4G | local | RTDC8 | Y | 3 | blockdev=6496 | 85m | Fewer cores = longer. As expected. Repeated and it completed in 73m. |
| 6.4G | local | RTDC8 | N | - | blockdev=6496, mstransform[1, 256, 54] | 3h3m | Saving 20% on serial processing time using mstransform to trim |
| 6.4G | local | RTDC8 | Y | 7 | blockdev=6496, mstransform[1, 256, 54] | 39m | No time gains |
| 6.4G | NFS | RTDC8 | Y | 7 | blockdev=6496 | 56m | Data read from NFS mounted disk (rglinux13) & written locally. |
| 6.4G | NFS | RTDC8 | Y | 7 | blockdev=6496 | 78m | All work done on NFS mounted disk (rtdc9 mounted on rtdc8) |
| 6.4G | NFS | RTDC8 | Y | 7 | blockdev=6496 | 94m | All work done on NFS mounted disk (rglinux13 mounted on rtdc8) |
| 6.4G | NFS | RTDC8 | Y | 7 | blockdev=6496 | 58m | All work done on local RTDC8 disk that is mounted as NFS at mount point. |
| 6.4G | NFS | RTDC8 | Y | 7 | blockdev=6496 | 56m | Local disks on rtdc8 mounted as NFS |
| 6.4G | local | RGLINUX13 | Y | 7 | blockdev=6496, CPU speed 2.9GHz | 78m | |
| 13G | local | RTDC8 | Y | 5 | blockdev=6496 | 57m | Compare to 57 mins for half the data |
| 13G | local | RTDC8 | Y | 9 | blockdev=6496 | 54m | Compare to 42 mins for half the data |
| 13G | local | RTDC8 | Y | 13 | blockdev=6496 | 48m | |
| 13G | local | RTDC8 | Y | 15x2 | blockdev=6496 | 101m & 102m | Set two identical scripts running simultaneously. Better to run with fewer unshared cores, than to run with more cores that are shared. |
| 13G | local | RTDC8 | Y | 7x2 | blockdev=6496 | 79m & 77mm | |
| 6.4G | local | RTDC8 | Y | 5 | blockdev=6496 | 110m | Re-run with CASA 5.3.0 |
