The latency numbers below are for blocking operations. Nonblocking operations tend to provide better results depending on the benchmark used.
|Network Protocol||Latency Put (us)||Latency Get (us)|
|Myrinet-GM(2.4GHz Pentium-4, Linux 2.4, Myrinet C card, GM 1.64)||12.8||17.8|
|Quadrics Elan-3(1GHz ia64,Linux 2.4.20)||4.71||6.42|
|Quadrics Elan-4(1.4GHz AMD Opteron,Linux 2.4)||1.80||2.66|
|Quadrics Elan-4(1.5GHz ia64,Linux 2.4.20)||2.45||4.56|
|Infiniband(1GHz ia64,Linux 2.4.20)||7.4||16.0|
Linux cluster with dual 2.4GHz Pentium-4 nodes, Myrinet-2000 (M3F-PCI64C-2 Myrinet interface) located at the State University of New York at Buffalo. It employs GM (1.6.4) and MPICH-GM libraries provided by Myricom.
Comparison of latency of ARMCI get (nonblocking get followed by wait) operation with GM.
Non-blocking (overlapping communication with computation): % overlap for increasing message sizes for MPI and ARMCI (direct and server based protocols.
Linux cluster with dual 1 GHz Itanium-2 nodes, Myrinet-2000 ( M3F-PCI64B-2 Myrinet interface) located at Pacific Northwest National Laboratory. It employs GM (1.6.4) and MPICH-GM libraries provided by Myricom.