The user-level performance metrics ("micro-benchmarks") most often cited for cluster interconnect are:
For Myrinet-2000 with MX-2G software, the following performance is typical:
|Myrinet-2000 User-level Performance|
|MX or MPI latency||3.2µs (D-card NICs)
2.6µs (E- or F-card NICs)
|MX or MPI unidirectional data rate||247 MBytes/s (one-port NICs)
495 MBytes/s (two-port NICs)
|TCP/IP data rate (MX ethernet emulation)||1.98 Gbits/s (one-port NICs)
3.95 Gbits/s (two-port NICs)
For performance of Myri-10G under MX-10G software, refer to the MX-10G Performance web page.
There are other performance metrics that are directly or indirectly predictive of application performance.
The links below will take you to detailed performance measurements for different Myrinet application-programming interfaces (APIs).
Myrinet-2000 Hardware Latency
The latency measurements presented on these pages were, unless otherwise noted, between NICs whose ports are connected through short cables through a single switch.
The measured latency includes ~0.5µs total hardware latency in the circuitry between the ports of the Lanai-X chips in the two NICs. This hardware latency is introduced by circuitry that converts to and from the Fiber Physical layer, and by the switch itself. Networks for more than 512 and up to 8192 hosts with XBar32-based switches have hardware latencies up to ~1.1µs, but the hardware latency is still small compared with the software latency. The "time of flight" latency of a fiber cable is ~0.0065µs per meter, i.e., ~1.3µs for a maximal-length, 200m fiber cable, but only 0.03µs for a 5m cable.