From: vfl@angewandte-geophysik.ruhr-uni-bochum.de (Tony Stoecker)
Newsgroups: comp.parallel.pvm
Subject: pvm_send slower than expected
Date: 28 Apr 1999 15:45:29 GMT
Organization: Ruhr-Universitaet Bochum, Rechenzentrum
Message-Id: <7g7aep$s1l$1@sunu789.rz.ruhr-uni-bochum.de>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Xref: ukc comp.parallel.pvm:8325


Hi all,

i am PVM-rookie. We startet our parallel-computing project with 
a 5 node PC-linux cluster connected with fast ethernet. The nodes 
are diskless running with Linux-Kernel-2.2.5 and netboot-0.8.  
We build some test programs to check the performance of the system. 
We changed the 'master.c slave.c' example from PVM-3.4 for our
purposes to have bigger data transfer (like our applications will have). 
Now the problem:
5 MB data returned from every node to the server needs approx. 4 seconds 
with pvm_pkdouble. => 25MB Data in 4 seconds. This is approx. 2 times
slower than we expected since we have a 100Mbit/s net. Now our first
idea was that our net configuration is bad. But then we saved the data
directly to disk from the slave program. That is the nodes saved the
data over NFS on their mounted disks on the server. The data transfer
time is now 2 seconds for all 25MB.
O.K. the server has fast disks (16-24MB/sec), but anyway this is a poor
solution and we believe that pvm_send should be faster than NFS since
RAM is still faster than the disk.

Is there anybody out there who knows what we are doing wrong?
Please email me. Thankyou very much.

cu,
Tony
 
--------------------------------------------------------- 
 Dipl.-Geophys. Tony Stoecker 
 Institut fuer Geophysik 
 Ruhr-Universitaet Bochum 
 44780 Bochum  
 Tel.: (0234) 700-3197 
 email: vfl@angewandte-geophysik.ruhr-uni-bochum.de
---------------------------------------------------------