From owner-chemistry@ccl.net Thu Feb  2 16:24:01 2006
From: "Alex. A. Granovsky gran#%#classic.chem.msu.su" <owner-chemistry-x-server.ccl.net>
To: CCL
Subject: CCL: PES vs. # of CPUs
Message-Id: <-30759-060202155956-5040-yktIAxP8K/EbrbL+qlxBpw-x-server.ccl.net>
X-Original-From: "Alex. A. Granovsky" <gran-*-classic.chem.msu.su>
Content-Transfer-Encoding: 7bit
Content-Type: text/plain;
	charset="koi8-r"
Date: Thu, 2 Feb 2006 23:59:35 +0300
MIME-Version: 1.0


Sent to CCL by: "Alex. A. Granovsky" [gran . classic.chem.msu.su]
Jan, 

your statement is actually not always true. 

    More precisely, this is true for single-point calculations. 
The problem is that almost all energy minimization algorithms are not 
numerically stable on the complicated PES having non-trivial structure, 
and initially very close "trajectories" of minimization completely diverges 
over time. In this case, it is almost normal situation to get different answers 
(local minima) even with single-processor runs using different binaries 
(e.g., produced by different compilers). 

Best regards,
Alex Granovsky


> Sent to CCL by: janl ~ speakeasy.net
> While very small differences (truncation errors) are allowed
> between single processor and parallel runs (they may be due to
> some rounding errors when data are being sent between processors
> or machines), they should be below 0.001 kcal/mol. The differences
> 1000x larger indicate one or more:
>   1) Bug in the program
>   2) Bug in the compilation or compiler
>   3) Bug in the MPICH libraries
>   4) Faults in configuration
>   5) Errors in communications, e.g., lost packets, or connectionless
>      protocol (use TCP for testing, rather then UDP if you have
>      a choice of configuration).
> 
> Jan Labanowski
> jkl]^[ccl.net
> 
> 
> > -----Original Message-----
> > From: Jozsef Csontos jozsefcsontos~~creighton.edu [mailto:owner-chemistry]^[ccl.net]
> > Sent: Wednesday, February 1, 2006 10:52 PM
> > To: 'Labanowski, Jan -
> > Subject: CCL: PES vs. # of CPUs
> > 
> > Sent to CCL by: Jozsef Csontos [jozsefcsontos]|[creighton.edu]
> > Hi,
> > 
> > my short question is: 
> > is it possible to get different optimized final structures using the
> > same input and architecture, OS, software (and so on), but different
> > number of processors?
> > 
> > My longer comments are:
> > I understand that we use approximations (in QM and in life too:) and
> > there might be error accumulations and our number representation depends
> > on the hardware etc. But, I'm quite sure the answer is no. Maybe I'm
> > wrong, so I tell you the story.
> > 
> > The story:
> > I have been using a commercial quantum package, which uses the MPI
> > libraries (MPICH) for parallelization. I was about to map the PES of
> > several tripeptides in water using DFT. (Of course, the molecules are
> > flexible and the PESs are flat and more more approximations) I'm not
> > totally surprised that the results are different on 1 node with 1
> > processor and 6 node with 6 processors. (In the first case there is no
> > MPI and in the second one there is.) Interestingly, the results are
> > different on 3 nodes with 6 processors (2CPUs/node) and on 6 nodes with
> > 6 processors (1CPU/node), in addition the calculation using the previous
> > 6 processor configuration (2CPUs/node) differs from the 4 CPUs config
> > (2CPUs/node) ones.
> > The energy differences between the structures are small about 1kcal/mol,
> > however, the rmsd is quite large 1.9 Angstrom considering heavy atoms
> > and one of these structures is the deepest minimum on the PES, so these
> > structures belong to the low energy region. I haven't checked all the
> > structures whether they are minima or other stationary points. I guess
> > this is not related to the question, but if all of them were minima my
> > question would be louder.
> > 
> > I'm wondering what your opinions are?
> > 
> > Best wishes,
> > 
> > Jozsef
> > 
> > -- 
> > Jozsef Csontos, Ph.D.
> > 
> > Department of Biomedical Sciences
> > Creighton University,
> > Omaha, NE> 
> 
> 
>