From owner-nwchem-users  Sun Sep 30 21:08:06 2001
Received: (from majordom@localhost)
	by odyssey.emsl.pnl.gov (8.8.8+Sun/8.8.5) id UAA21866
	for nwchem-users-outgoing; Sun, 30 Sep 2001 20:52:43 -0700 (PDT)
From: Drchemp@aol.com
Date: Sun, 30 Sep 2001 23:52:34 -0400 (EDT)
Subject: Message Passing in Linux 2.4.2 Kernel
To: nwchem-users@emsl.pnl.gov
Message-id: <15b.1ca356a.28e94282@aol.com>
MIME-version: 1.0
X-Mailer: AOL 4.0 for Windows 95 sub 113
Content-type: text/plain; charset="US-ASCII"
Content-transfer-encoding: 7bit
Sender: owner-nwchem-users@emsl.pnl.gov
Precedence: bulk

Following an upgrade to 2.4.2 kernel in RH Linux 7.1 to eliminate NaNs, the 
ratio of   CPU time to Wall clock is grossly in favor of the wall clock. This 
reminded me of the TCP Performance Fix for Short Messages provided as a patch 
to ipv4 routines from Josip Loncaric at www.icase.edu which worked very 
favorably applied to the 2.2 kernel (which suffers the NaN's).  Now, however, 
according to Josip the 2.4 kernel has thoroughly changed the TCP stack, and 
I've recognized this in the form of my trouble of attempting to 'past' the 
patched 2.2 kernel related ipv4 routines into a re-compilation of the new 
2.4.2 kernel.  

I'm running NWChem 4.0 recompiled under a running 2.4.2 smp kernel on a pair 
of dual PIII Xeon 550 MHz boxes with fast ethernet and one GB of RAM per box. 
 That is, four processors.


In usr/src/linux-2.4/net there is reference to TUNABLE in RH 7.1 install, 
apparently relating to some optimization capability.  Does anyone have any 
suggestions on how to return the message passing to better efficiency, 
assuming I have not overlooked the real cause of the poor ratio of CPU to 
Wall.

Thank you,
Michael Oberlander