From owner-nwchem-users@emsl.pnl.gov Wed Oct 27 16:22:15 2004
Received: from odyssey.emsl.pnl.gov (localhost [127.0.0.1])
	by odyssey.emsl.pnl.gov (8.12.10/8.12.10) with ESMTP id i9RNMF5J000165
	for <nwchem-users-outgoing@odyssey.emsl.pnl.gov>; Wed, 27 Oct 2004 16:22:15 -0700 (PDT)
Received: (from majordom@localhost)
	by odyssey.emsl.pnl.gov (8.12.10/8.12.10/Submit) id i9RNMFvU000164
	for nwchem-users-outgoing; Wed, 27 Oct 2004 16:22:15 -0700 (PDT)
Date: Wed, 27 Oct 2004 16:22:13 -0700
From: Edoardo Apra <edoardo.apra@pnl.gov>
Subject: Re: problems with parallel performance
In-reply-to: <DE1BFF2E-286C-11D9-A3B2-000393A15080@wsu.edu>
To: Kirk Peterson <kipeters@wsu.edu>
Cc: nwchem-users@emsl.pnl.gov
Message-id: <41802DA5.7040800@pnl.gov>
MIME-version: 1.0
Content-type: text/plain; charset=ISO-8859-1; format=flowed
Content-transfer-encoding: 7bit
X-Accept-Language: en-us, en
User-Agent: Mozilla Thunderbird 0.7.3 (X11/20040803)
References: <DE1BFF2E-286C-11D9-A3B2-000393A15080@wsu.edu>
Sender: owner-nwchem-users@emsl.pnl.gov
Precedence: bulk

Kirk
could you give more details about the nwchem binaries
your are using? More precisely how where these binaries compiled
(USE_MPI=? NWCHEM_TARGET=? ARMCI_NETWORK=?)

Edo

Kirk Peterson wrote:

> Hi,
>
> we've recently gotten NWChem (4.6) running on our Athlon Linux cluster
> (Red Hat, dual processor nodes), which uses myrinet for MPI
> communication. We observe two problems, which seem to be unrelated.
>
> The first is that when we run 2-way parallel across nodes, i.e., 1
> processor on each node, nwchem starts up 2 processes on each node for a
> total of 4 processes.  Each of these seems to accumulate essentially
> the same amounts of CPU cycles.  If I run a job 2-way on 1 node,
> everything looks normal.  I seem to remember a similar problem when I
> linked up Molpro to version 3.3.1 of the global array library.
>
> The 2nd problem is in regards to parallel performance.  I've only
> really tested this for one of the TCE coupled cluster test runs (H2O),
> but in general the parallel runs are much slower than a single
> processor run, e.g., 31 sec for 1 proc, 78 sec for 2 procs on 1 node,
> and 100 sec for 2 procs over 2 nodes (with the problems noted above). 
> I see similar behavior with just a simple MP2 calculation (Au-H2O), 9
> sec for 1 proc and 13 sec for 2 procs.  Certainly I would like to test
> this for more nodes, but not until the issue above is resolved.
>
> thanks in advance,
>
> Kirk
>
> PS - is it really true that i functions are not supported yet?
>

-- 
Edoardo Apra` - PNNL - P.O. Box 999, MS K8-91 - Richland, WA 99352
Tel +1-509-376-1280 Fax +1-509-376-0420