From owner-nwchem-users  Fri Mar  3 11:17:30 2000
Received: (from majordom@localhost)
        by odyssey.emsl.pnl.gov (8.8.8+Sun/8.8.5) id KAA09924
        for nwchem-users-outgoing; Fri, 3 Mar 2000 10:58:40 -0800 (PST)
Content-return: allowed
Date: Fri, 03 Mar 2000 10:58:28 -0800
From: "Fann, George I" &LTgi.fann@pnl.gov>
Subject: RE: parallel job on linux cluster
To: Todd Raeker &LTraeker@umich.edu>
Cc: nwchem-users@emsl.pnl.gov
Message-id: &LT6DBB005AF9DDD31193030008C7A49DC833C180@pnlmse8.pnl.gov>
MIME-version: 1.0
Content-type: text/plain
Sender: owner-nwchem-users@emsl.pnl.gov
Precedence: bulk


Todd,

as a preliminary step you may want to do some simple network tests...e.g.
tcgmsg using large messages or even rcp on large files to see what the real
performance on your network and whether it is doing 100 base T or not
..we are of course that you are running tcgmsg
over the your 100 base T and not tcgmsg over mpi over 100 base T ( real slow ).


I don't know how your file system works nor how your files are mounted on your
cluster.
If you can tell us a little about that it would be great.

thanks,

George

        ----------
        From:   Ricky Kendall
        Sent:   Friday, March 3, 2000 10:54 AM
        To:     Todd Raeker
        Cc:     nwchem-users@emsl.pnl.gov
        Subject:        Re: parallel job on linux cluster

        <&LTFile: mp2.xls>><&LTFile: rickyk.vcf>>
        Todd,
            This seems like a problem with the network.  We have an
        Alpha Cluster and I enclose a excel spreadsheet with the data
        The time decreases and the speedup (walltime) gets to 50% at 8 nodes.
        The only thing I would offer is the timing variance in such a short
        job may be large enough to caue this.  Try something with a little
        more meat in it time wise.  Also as you can tell from the spreadsheet
        the CPU based scaling is quite good.  e.g., we need a better network:)
        Regards,
        Ricky


        Todd Raeker wrote:

        > Hi all,
        >
        >   I am testing a two node linux cluster on a 100 Mbits/sec standalone
        > switched network.  A small scf single point energy calculation on one
and
        > two nodes results in the CPU time going from 44 sec. to 22 sec.
        > respectively.  This is great but when I look at the wall time I get 4
4
sec.
        > for one node verses 49 sec for the two node calculation.  This really
        > surprises me as the local net is running at 100 Mbits/sec and
completely
        > isolated from other networks.  I would like to talk to anybody out
there
        > with experience running nwchem on linux clusters.  My hardware
consists of
        > two machines with a 600 MHz AMD Athlon in each with 128 MB memory,
3com905
        > NIC and 3com superstack 3300 at 100 Mbits/sec network running RedHat
6.0
        > Linux.  I hope the slow wall time is a mistake in my network
configuration
        > rather than an high amount of nwchem builtin communication between
nodes
        > which increases overhead.  Any advice or info would be appreciated.
        >
        > Todd.
        >
        > Dr. Todd Raeker
        > Coordinator of Computer Services         raeker@umich.edu
        > Department of Chemistry                     (734)647-2867
        > University of Michigan
        > Ann Arbor, MI 48109