Newsgroups: comp.parallel.pvm
From: schiotz@oersted (Jakob Schiotz)
Reply-To: schiotz@fysik.dtu.dk
Subject: Re: Max buffer size for incoming messages?
Organization: Physics Department, Techn. Univ. of Denmark
Date: 28 Sep 1994 15:01:38 GMT
Message-ID: <36c0gi$7i5@news.uni-c.dk>

weicheng jiang (wjiang@thud.cs.utk.edu) wrote:
: In article <369f8c$o5d@news.uni-c.dk>, schiotz@oersted (Jakob Schiotz) writes:
: |> I have a PVM program where the slaves are sending 26 large messages
: |> (~100 kB) to other slaves. In theory up to 24 such messages may arrive
: |> while the slave is busy processing the first. I have tried to increase
: |> the message fragment size to improve performance, but if I increase
: |> the fragment size to say 64k (or 128k) the slaves crash (but until the
: |> crash the program runs faster). Is there a maximum buffer size for
: |> incoming messages, so the buffer may overrun if many messages arrive
: |> from different sources at the same time?
: |> 
: |> I am running PVM 3.3.3 on a CM-5E parallel computer.
: |> 

: Did you get any error messages in the log file?

No, I only got a coredump, apparently of the program running on the
Partition Manager while a PVM process is running on the nodes. THis is
the program that comes with PVM, and it was not compiled with -g. Of
course I can recompile PVM and investigate the problem, I just hoped
that it was a known problem.

I tried to reproduce it today, but I couldn't. The machine was however
rather heavily loaded, it may have influenced timings.


--
Jakob Schiotz                    !  The noble art of losing face
Physics Department               !  May someday save the human race
Technical University of Denmark  !  And turn into eternal merit
DK-2800 Lyngby, Denmark          !  What weaker mind would call disgrace
schiotz@fysik.dtu.dk             !                  - Piet Hein
              ^^^ Note: new domain

