Newsgroups: comp.parallel.pvm
From: paulh@tycho.wynde.com (Paul Halenda)
Subject: problems with large messages
Keywords: large messages
Organization: Wyndemere
Date: 23 Aug 1994 19:37:09 GMT
Message-ID: <33dj55$730@potogold.rmii.com>

Howdy,

I'm trying to incorporate usage of PVM in place of socket I/O in a set of
programs running on Sun Sparc boxes connected via ethernet, 10baseT.  The 
conversion is going well, except for one startup message that's about
2.4mb.  For some reason it takes a very long time for this message to
get transferred from the sender to the receiver, on the order of 3 
minutes.  It doesn't seem to matter if the two processes are running on
the same physical machine or not, so I don't think the ethernet is the 
bottleneck (and a rough calculation would show this message ought to
transfer in seconds across the ethernet anyway).

The one thing I've tried is breaking the message up into smaller chunks; 
when split into 24 100k pieces, receipt of the first chunk occurs
about 25 seconds after it was written.  The other pieces seem to be
received quickly thereafter.

Is anyone aware of a size limitation, or number of messages limitation in PVM 
that might account for this?  One puzzling thing I've noticed is if I run
a couple bare-bones test programs to send various sizes of messages, it seems
like the delay is substantially smaller.  Is it possible I'm running into 
swapping overhead or something like that, since the "production" code is fairly
large?

Any comments or suggestions greatly appreciated, either via post or email.

Thanks,

Paul Halenda

paulh@wynde.com



