[Pw_forum] espresso 2.1.4 and ifc8.1.024

Axel Kohlmeyer axel.kohlmeyer at theochem.ruhr-uni-bochum.de
Fri Jul 1 19:05:42 CEST 2005


On Fri, 1 Jul 2005, Patricia Paredes wrote:

PP> Hi, I was using until now pwscf-v2.0. This week I decided to upgrade to
PP> espresso. Also, I installed the latest version of intel compiler
PP> (ifc8.1.024) .For the compilation of pwscf-v2.0 we used ifc7, but
PP> espresso with ifc7 gives us error 0 0.  In the system we had installed
PP> cluster mkl 7.0. I changed the make.sys to include more optimization
PP> (-O3, -tpp7). 

patricia,

you have to be careful here. higher optimization levels can 
create problems sometimes, since the compiler may rearrange
the code a little bit, and that may not always be optimal
for numerical codes. it may slow down and sometimes even
produce incorrect results. however this very much depends 
on the compiler/hardware combination, the specific code and 
the problem. on pc-style hardware i found that using,
' -O2 -unroll -pc64' is a quite fast and secure choice.
PP> 

PP> The compilation went OK, so I began testing with and old input file. But
PP> the new pw.x runs slower than the old version (pwscf.v2.0).
PP> 
PP> 
PP> For example: I obtained for following results for a SCH3 radical in a
PP> box:
PP> 

can you please post this example input, so we can 
try to reproduce your problem. some flag's names 
or default settings may have changed since v2.0.

PP> 
PP> <excerpt>(single node, P4 HT, 3GHz, running Fedora and mpi6.0) 

ok, now i have a bunch of questions, since based the provided
information, there can be quite a few number of reasons for
the slowdown you are seeing:

do you have hyper-threading enabled or disabled?

are you running in parallel or serial?
if the former, does the slowdown only happen in parallel?

do you see the same problem with some of the integrated
(and smaller) test examples?

did you run the two jobs on the exact same machine with
the same communication library and hardware?

PP> 
PP> first scf iteration 
PP> 
PP> pwscf v2.0         47.55 secs
PP> 
PP> espresso v2.1.4   418.15 secs
PP> 
PP> 
PP> time for the fist ionic step
PP> 
PP> pwscf v2.0        358.88 secs
PP> 
PP> espresso v2.1.4   2103.36 secs
PP> 
PP> 
PP> Also, I tried with a periodic system with metallic atoms withe SCH3l
PP> adsorbed on the surface, running on 6 nodes:
PP> 
PP> the new executable was slower as in the case of the radical,  and the
PP> results were completely different. Since the default criterion for ionic
PP> convergence is 
PP> 
PP> setting by default in 1e-4 and 1e-3 and there are some changes in the
PP> value of the energy, the results are:
PP> 
PP> Pwscf-v2.0 converges in 12 iterations with a total force of 2 e-3.
PP> 
PP> espresso 2.1.4 has sligtly differents results (the total energy differs
PP> in the last 4 digits) and after iteration 30 something happens and got an
PP> error and the run stopped. 

what error?

 
PP> I'm aware of the problems with ifc as described in the archives and  in
PP> Axel Kohlmeyer web page, but I´m still in trouble.
PP> 
PP> 
PP> Did someone face the same problem?? Suppose the problem has to do with
PP> the compiler, but I don´t know at this stage how to solve it.

it need not be the compiler.

regards,
	axel.

PP> 
PP> Many thanks,
PP> 
PP> Patricia.
PP> 
PP> </excerpt>
PP> 
PP> <bold><underline>Prof. Dr. Patricia Paredes Olivera
PP> 
PP> </underline></bold>Facultad de Ciencias Químicas
PP> 
PP> Universidad Nacional de Córdoba
PP> 
PP> ph/fax: 054-0351-434-4972
PP> _______________________________________________
PP> Pw_forum mailing list
PP> Pw_forum at pwscf.org
PP> http://www.democritos.it/mailman/listinfo/pw_forum
PP> 
PP> 

-- 

=======================================================================
Dr. Axel Kohlmeyer   e-mail: axel.kohlmeyer at theochem.ruhr-uni-bochum.de
Lehrstuhl fuer Theoretische Chemie          Phone: ++49 (0)234/32-26673
Ruhr-Universitaet Bochum - NC 03/53         Fax:   ++49 (0)234/32-14045
D-44780 Bochum  http://www.theochem.ruhr-uni-bochum.de/~axel.kohlmeyer/
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.




More information about the users mailing list