Newsgroups: comp.parallel.pvm
From: pikus@dolphin.physics.ucsb.edu (Fedor G. Pikus)
Subject: BIG problem with PVM on Paragon!
Organization: University of California, Santa Barbara
Date: 9 Dec 94 02:54:49 GMT
Message-ID: <pikus.786941689@dolphin.physics.ucsb.edu>

Hi,
I'm trying to port my PVM program to Paragon, after it has
been tested and ran on a cluster of workstations
and on Meiko. I have a problem, which seems hard to believe, but it's 
a fact: if I have Master-Slave program, and Master sends a MCast 
message to slaves (or uses Send to each slave, in a loop),
IT TAKES 1 SEC OF CPU TIME ON SLAVE NODE FOR EACH PVMFRECV!!!
If instead I use non-blocking receive and call to sleep in a loop,
it takes 6 sec of sleep for each message!!! I get this result
even if I have ONE SLAVE. 
It runs better on workstations with Ethernet than on Paragon!
Is there a way other than to rewrite the whole thing with NX 
calls?
Thanks
Fedor

