Performance degradation of GM!
"geoge_zhang" <[email protected]>
| Newsgroups | gmane.network.myrinet.general |
|---|---|
| Message-ID | <3BC166AC.18622@mta6> |
Hi: It's a little too silent here, and may need some noise:-) We've tested the performance of GM-1.4 on our new cluster platform (linux2.4.4 + Myrinet2000 + 64bit 66HMz +4dual xeon SMP). The test results for one-way communication performance is about the same with the results reported in Myrinet website. But for two-way communication, the performance degrades greatly. For two-way communication between two pairs nodes in the same direction (two source processes on the same node communicate with other two processes residing on another node), each way communication achieves about 50% of the performance of one-way communication (bandwidth). Which can be explained by the physical link performance limits and each way communication shares 50% of the communication capacity of one link. But for two-way communication between two pairs nodes in reverse direction, each way achieves only about 80% of the performance of one-way communication. As is said that Myrinet supports full-duplex link communication, then what's the cause for the performance degradation of two-way reverse communication ? Any explain would be appreciated ! Best regards Geoge