MPI-BIP An implementation of MPI over Myrinet |
It is available for evaluation.
Details about the experiments |
We have measured the performance of ping-pong tests where one master sends some data to a slave, the latter receives it and sends back the same message (of same size) to the master.
The timing results have been obtained by taking the time between the beginning of the send call and the end of the receive call on the master and dividing it by 2 to obtain one-way latency and message bandwith.
The values in the table are the median times of 100 successive experiments.
These tests were run between 2 Pentiums Pro 200 Mhz.
Size of the messages (Bytes) | Bandwidth (MB/s) | Elapse Time one-way (us) |
---|---|---|
0 | 0.00 | 11.02 |
4 | 0.35 | 11.41 |
8 | 0.65 | 12.33 |
16 | 1.22 | 13.16 |
32 | 2.32 | 13.80 |
64 | 4.02 | 15.93 |
128 | 6.33 | 20.22 |
256 | 4.49 | 57.04 |
512 | 8.98 | 57.01 |
1024 | 15.78 | 64.89 |
2048 | 26.79 | 76.45 |
4096 | 41.12 | 99.62 |
8192 | 56.57 | 144.82 |
16384 | 43.24 | 378.94 |
32768 | 56.08 | 584.32 |
65536 | 59.05 | 1109.89 |
131072 | 63.85 | 2052.77 |
262144 | 65.00 | 4033.11 |
524288 | 65.57 | 7996.01 |
1048576 | 65.76 | 15944.64 |
2097152 | 66.12 | 31718.34 |
4194304 | 66.32 | 63242.74 |
8388608 | 66.33 | 126465.65 |
The NAS parallel benchmarks |
MPI-BIP on PPro 200 | IBM SP (66/WN) | Cray T3E-900 | SGI Origin 2000-195 | Sun Enterprise 4000 | ||||||
---|---|---|---|---|---|---|---|---|---|---|
Mop/s | Speedup | Mop/s | Speedup | Mop/s | Speedup | Mop/s | Speedup | Mop/s | Speedup | |
IS @ 4 proc | 2.44 | 4.5 | 2.2 | 3.1 | 3.2 | N/A | 2.1 | 3.8 | 1.6 | 3.8 |
IS @ 8 proc | 2.11 | 8.8 | 2.0 | 5.6 | 3.8 | N/A | 2.4 | 8.6 | 1.5 | 6.9 |
LU @ 4 proc | 23.68 | 6 | 59.0 | 3.7 | 67.6 | N/A | 96.8 | N/A | 36.9 | 4.1 |
LU @ 8 proc | 22.73 | 11.6 | 57.2 | 7.1 | 66.4 | N/A | 103.3 | N/A | 37.3 | 8.4 |
SP @ 4 proc | 10.10 | 3.4 | 42.1 | 3.6 | 43.0 | N/A | 60.3 | N/A | 25.0 | 3.7 |
![]() |
Last modified: Mon Jun 7 14:35:27 CEST 1999 © BIP team |