Newsgroups: comp.parallel.pvm
From: scatter@scubed.com (Keith McLaughlin)
Subject: help: memory leak?
Keywords: master slave memory leak
Organization: S-CUBED Division of Maxwell Laboratories
Date: Fri, 24 May 1996 23:05:55 GMT
Message-ID: <1996May24.230555.22750@scubed.com>

Help!

We are running PVM3.3.11 on 16 SUN4 SUNOS4 computers to compute synthetic
seismograms using a wavenumber integration algorithm.
Each frequency is a separate task independent of all other frequencies. 
It is a single master, many slave paradigm.
We send each frequency out to the slaves with pvmfinitsend() and
wait for it to return with results. It then gets a new frequency.

We get the messages back from the slaves with non-blocking receive.

The master program and the local pvmd3 are growing and run
out of memory!  They look like each message is taking ~100K but
we have many messages (tasks) to send out and they get too big.

We do not use the pvmf[sr]buf() calls.
The manual says that pvmfinitsend() should re-use the buffers.
What is going wrong?

Thanks in advance

-- 
Keith L. McLaughlin, scatter@scubed.com
619-587-8436, FAX 619-755-0474
S-CUBED, Div. Maxwell Laboratories, POB 1620, La Jolla, CA 92038
URL: http://www.scubed.com/products/geop

