Questions about the performance and architecture of the pscom layer compared to MPICH

[PaulinaEster]

Good afternoon,

I am using the ParaStation MPI (psmpi) library with the pscom communication layer and conducting a performance comparison against the original MPICH version. My reference for this study is the seminal paper:

WARSCHKO, T. M.; BLUM, J. M.; TICHY, W. F. ParaStation: Efficient parallel computing by clustering workstations: Design and evaluation. Journal of Systems Architecture, v. 44, n. 4, p. 241–260, 1998.

The paper shows that ParaStation achieves significantly lower latencies (on the order of microseconds) by removing the operating system from the critical communication path and providing an efficient emulation of the Unix socket interface. Based on this, I have the following questions:

• Does the TCP/IP protocol emulation still exist in current versions?
• Is there any configuration flag or runtime parameter that optimizes pscom performance?
• Is specific network hardware required to achieve the latency results described in the original paper?

Thank you.
@jensh

[spickartz]

Dear Paulina,

Thanks for your interest in the ParaStation MPI communication stack!

The feature you are referring to is no longer included as things have developed. Nowadays, TCP/IP is mainly used for development purpose, e.g., when there is no high-performance network available. Regarding runtime optimization of pscom, there are several parameters. Once built and installed, you can run pscom_info -cd to see the options available on your system.

Hope this helps!