[CPMD-list] cpmd352 on dual-PCs - shmem communication feasible?

[Francesco Filippone]

Dear Christian, dear all,

it is possible to build a shared memory device (--with-device=sh_mem)  
with mpich. If you need to run only on the dual-processor PC 
this should be enough. And it works. Nevertheless mpich 1.2.1 to 
1.2.3  have always given me a speedup of about 1.3, exactly like in your 
case
. 
The same speedup has been obtained on a dual-processor IBM RS/6000 43P 
model 260, so I guess the problem is in mpich implementation. Someone 
claims better performances can be achieved using both fortran and 
C compilers from the same vendor. I have Portland Group only for fortran 
on my PC, so I don't have direct experience in this sense on this 
architecture. On the dual-p RS/6000, compilers are both from IBM, but this 
does not seem to improve performances.
 
Someone else claims that lam-mpi implementation  (www.lam-mpi.org) is 
faster than mpich. I don't have any experience since same vendor compilers 
are "encouraged" in this case. (http://www.lam-mpi.org/6.5/install.php)

Anyway, mpich implementation seems to have some problems in handling 
SHMEM systems, since, on both  machines I used, the speedup of CPMD on 2 
processors is about 0.8-0.9 (!) when using the option SYMMETRY 0 for 
isolated molecules.

Regards,
F.
-- 
----------------------------------------------------------------------------
  Francesco Filippone                            | Tanto gentile e
  CNR - ISTITUTO DI STRUTTURA DELLA MATERIA      | tant'onesta pare
  v. Salaria Km 29,300 - C.P. 10                 | la donna mia quand'ella
  I 00016 - Monterotondo Stazione (RM)           | altrui saluta...
  Tel + 39 06 90672342 - Fax +39 06 90672316     |
  e-mail: <francesco.filippone at mlib.cnr.it>      |     Dante
----------------------------------------------------------------------------