Myrinet Overview: "Messages can be sent and received without system calls, resulting in measured message latencies at the MPI level with the 'E card' and 'F card' NICs of ~2.7�s with MX, and ~6.5�s with GM. Message latencies using the popular, low-cost, 'D card' NICs are ~0.8�s higher. Thanks to the OS-pass mode of operation and to protocol processing being offloaded to the firmware, the host-CPU overhead is minimal, ~0.15�s per message for sending or receiving.
For either GM or MX, and with single-port NICs, measured, user-level, one-way, data rates are ~1.98 Gb/s (248 MBytes/s), and two-way (summed bidirectional) data rates are ~3.92 Gb/s (490 MBytes/s). With the high-end, dual-port NICs, the firmware takes care of distributing and reassembling packet data across the two ports. The measured, user-level, one-way data rate of ~3.9 Gb/s (~490 GBytes/s) approaches the capacity of the two links. On hosts with good PCI-X throughput, the two-way (summed bidirectional) data rate approaches 6.4 Gb/s (800 MBytes/s)."
No comments:
Post a Comment