Above some old benchmark runs against swallowtail's infiniband switch.
Below new benchmarks with IPoIB. “ptile” is the number of jobs slots to allocate per node, so with 16 requested and ptile=1 sixteen individual hosts will be selected and all MPI traffic between hosts will go via the Voltaire infiniband switch.
| Amber | What | Host | N | JAC | Factor_IX | ptile= |
| 9+openmpi-1.2+intel-9 | Sander.MPI | n1 | 02 | 2m36s | 4m14s | |
| 9+openmpi-1.2+intel-9 | Sander.MPI | n1 | 04 | 1m23s | 2m14s | |
| 9+openmpi-1.2+intel-9 | Sander.MPI | n1 | 08 | 0m47s | 1m20s | |
| 9+openmpi-1.2+intel-9 | Sander.MPI | | 16 | 1m04s | 2m17s | 8 |
| 9+openmpi-1.2+intel-9 | Sander.MPI | | 16 | 0m42s | 2m07s | 1 |
| Amber | What | Host | N | 1g6r | ptile= |
| 9+openmpi-1.2+intel-9 | pmemd | n1 | 02 | 2m40s | |
| 9+openmpi-1.2+intel-9 | pmemd | n1 | 04 | 1m23s | |
| 9+openmpi-1.2+intel-9 | pmemd | n1 | 08 | 0m47s | |
| 9+openmpi-1.2+intel-9 | pmemd | | 16 | 0m43s | 8 |
| 9+openmpi-1.2+intel-9 | pmemd | | 16 | 0m39s | 1 |
Back