GLOW Cluster Benchmarks
MCNP.pvm-pvm based parallel Monte Carlo transport code on GLOW compute node (includes current and heat loads in serial and parallel with and without hyperthreading)
NetPIPE-tcp, pvm, MPI network performance
Network performance benchmark (testing between 2 nodes):
- Cisco 3750 Gbps switch
- Red Hat kernel 2.4.20
- Netpipe 2.4 (old version used to compare to earlier cluster)
| Connection | Latency (usecs) | Maxthroughput (Mbits/sec) |
|---|---|---|
| TCP | 63 | 737 |
| PVM | 65 | 482 |
| MPICH-1.2.4 | 77 | 374 |
| Condor MPICH-1.2.4 | 163 | 215 |
Note: Measurements are more variable than that seen with Fast ethernet. Also, in Condor MPICH environment, both cpus on our dual cpu machines are typically loaded. In the non-condor tests the other cpu is not loaded.
Plots
