Newsgroups: comp.parallel.pvm
From: john@elroy.jpl.nasa.gov (John Wright)
Subject: Re: Limitations to size of buffer?
Organization: Image Analysis Systems Group, JPL
Date: 8 May 1995 14:32:27 GMT
Message-ID: <3ola1r$ehf@elroy.jpl.nasa.gov>

In article <UWE.95May5170948@aclx11.physik.rwth-aachen.de>,
 <uwe@aclx11.physik.rwth-aachen.de> wrote:
>I just do my first tests on pvm so please be patient.
>I just installed pvm 3.3.7 on a Linux-Cluster and everything seemed to
>work well. But now it looks like as if there was a limit to the size
>of the receive buffer. Sending more than 450 double values seems to be
>no problem, but somehow they are not unpacked correctly. Unpacking
>even leeds to a total crash of the pvm-machine. 
>The example programs do not to bother about buffer size. So the
>timing program coming with pvm sends very large arrays, but they are
>not unpacked by the corresponding slave program.
>
>Any hints or information?
>
>Thanks in advance
>
>Uwe
>

I had a similar sounding problem unpacking small arrays of doubles and
large arrays of bytes on a multiprocessor platform.  The workaround
was to change the architecture from SGIMP to SGI5 (multiprocessor
Silicon Graphics to uniprocessor SGI).  Apparently, the special
internal message passing functions in the MP implementation didn't
work right so forcing everything to go through the socket (or
wherever) made it work correctly with some slight loss of performance
(the processes on the MP platform were doing lots of comunication
with uniprocessor platforms anyway).

Since you used the word Cluster I'm guessing that you have an MP
platform.  Try to set the architecture to uniprocessor to see it
it works.  Just a suggestion from my admittedly limited experience :-)
Hope it helps.  P.S. I'm using version 3.3.3 so this problem may
already be taken care of.

John


-- 
john@tone.jpl.nasa.gov
---Signature under construction---
(I quail at the impossibility of capturing the 
essence of myself in four lines in an ASCII file)

