From: "Jorgen Moth" <jorgen.moth@uni-c.dk>
Newsgroups: comp.parallel.mpi
Subject: latency and bandwidth on IBM SP2
Date: Tue, 8 Sep 1998 09:22:21 +0200
Organization: UNI-C
Message-Id: <6t2lve$oou$1@news.net.uni-c.dk>


Hi
I am measuring the latency (software latency) using non-blocking mpi_isend
on IBM SP2 with high-performance-switch. However my measurements (taken by
the built in rtc() real time clock routine from fortran) seems to indicate
that there is some copying going on from the user's buffer area to an
internal hidden buffer. I would expect an message-length-independent latency
(time from calling mpi_isend until control is returned to my fortran
program) round about 20-30 usec, however the time varies according to the
length of the message to be sent, i.e. several hundred useconds for 10 kB
messages. Does anybody (IBM?) want to commment?
Also, has anybody had experience or being using the basic driver routines
for the high-performance-switch (ccs0?).

Regards
Jorgen Moth
UNI-C, The Danish Computing Centre for Research and Education
jorgen.moth@uni-c.dk