[Myrinet] weird latency
Andrew Gaul <[email protected]>
| Newsgroups | gmane.network.myrinet.general |
|---|---|
| Message-ID | <[email protected]> |
Hi all,
I am measuring the latency between two hosts with Myrinet adapters and
have some curious results(*). I am using GM 1.5.1 on 33MHz 32bit
machines and am getting 65-70 us round-trip for 1000 byte packets using
a ping utility I wrote(**). There are >70 us outliers, which I expect,
but I also see <65 us outliers, clustered together in predictable
intervals (~13 ms). What could be causing these? I am running Linux
2.4.18 and have scheduled the process with the soft real-time scheduler.
I am measuring times by reading the cycle counter available on Pentium
and newer processors.
Also, I am working on a project that requires using Myrinet from
RTLinux; is has anyone done something similar? I have read a few
papers(***) written by Anthony Skjellum and his group, although their
code may not be available.
(*) http://gaul.org/images/myri.png
(**) http://gaul.org/files/myri-ping.c
(***) "A Synchronized Real-Time Linux Based Myrinet Cluster for
Deterministic High Performance Computer and MPI/RT" by Apte,
Chakravarthi, Padmanabhan, and Skjellum
Thank you for you time,
--
Andrew Gaul
[email protected]