Newsgroups: comp.parallel.pvm
From: "Ralf - G. Au" <au_r@informatik.fh-hamburg.de>
Subject: Data loss in pvm_send/recv, or pvm_pk*/upk*
Organization: Fachhochschule Hamburg, FB Informatik
Date: Wed, 09 Oct 1996 16:06:12 +0200
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <325BB154.167E@informatik.fh-hamburg.de>

Hello again, PVM-Programmers all over the world !

My fellow student and I are _still_ porting a raytracer to PVM, and
again we're experiencing some trouble.

We are using PVM Version 3.3.11 on a cluster of DECStation 5000/200's
(PMAX architecture referring to pvmgetarch) running DEC Ultrix
V4.4 (Rev. 69).

We have implemented a farmer/worker model with the workers doing the
actual raytracing and the farmer coordinating them and doing all the
messy I/O stuff.

The workers transmit the raw picture data (4 float, 1 int per pixel) to
the farmer where it all gets stitched together, and saved to a file.

The problem is: the picture data is transferred line by line and almost
always half of every line gets lost, resulting in black pixels for the
rest of the line. (Sometimes more than half a line is transmitted,
though, but never the whole line). 
We think we have traced down the error down to the pvm_pk* - pvm_send -
pvm_recv - pkvm_upk* part, but now we are at the end of our wits.

Our question now is: is it possible (if so, how?) to get the address of
the message buffer used by the PVM system ? All library functions
dealing with buffers always only refer to a buffer id.
Having the address of a buffer would make it possible to check by hand
if the pvm_pk*/upk* did work correctly.

Also, has anyone experienced similar problems, and could provide us with
hints or even workarounds ?

Any input would be greatly appreciated.


Greetings from Germany

Ralf Au
--
Ralf Au <au_r@informatik.fh-hamburg.de>
Fachhochschule Hamburg, Germany
Fachbereich Informatik / Informatics Faculty

