[Pw_forum] parallelization issue

[Giacomo Saielli]

Dear  Eyvaz and Paolo,
thank you very much for your comments.
In particular, Paolos' paper, Table IV report a time for cp step, T_i,  
of 3700 sec, 16 procs on SP3, system of 900 electrons, spin-unrestricted 
calculation. Then I guess that we can estimate roughly a half (1800 sec) 
for the same calculation spin-restricted, which would become about 
700-800 secs per time step if it were on SP4. Since I have about 600 
electrons instead of 900, this time would be reduced again to about 
400-500 or so? Then what I have found (1800secs/cp step) is about 4 
times larger than expected based on this very approximate estimate.
However I have used somewhat larger cutoffs than needed, as suggested by 
Paolo. Maybe I can improve the performance playing with these parameters.
Thank you again,
best wishes,
Giacomo

Paolo Giannozzi wrote:

>On Friday 15 October 2004 10:50, Giacomo Saielli wrote:
>
>  
>
>>I wonder if the comments of paragraph 7.4 of espresso manual v 2.1,
>>parallelization issue of pw.x, apply also for the cp.x module.
>>    
>>
>
>they mostly apply to cp.x as well. Of course there is no k-kpoint 
>parallelization in cp.x. I am not sure about the status of NEB image
>parallelization in cp.x.
>
>  
>
>>    ecutwfc = 30.0,
>>    ecutrho = 180.0,
>>    
>>
>
>25 and 100 Ry for C, N, H should be enough
>
>  
>
>>A test CPMD run with cp.x takes about 50 hours of total cpu 
>>(16 procs on IBM SP4 at CINECA, all PE of the same node) 
>>for 100 steps [...]. The question is if this numbers sounds
>>correct to you, for this system size
>>    
>>
>
>1800 s/time step seems to me way too large for a system like 
>yours. You can find reference CPU times on a SP3 here:
>http://www.nest.sns.it/~giannozz/Papers/JCP_57.pdf
>Divide by 2 or 3 to get an estimate for SP4.
>
>Paolo 
>
>  
>

-- 
Giacomo Saielli
Istituto per la Tecnologia delle Membrane del CNR, Sezione di Padova;
Via Marzolo, 1 - 35131 Padova, Italy; 
Tel: +39-049-8275279; Fax: -5239;
http://www.chfi.unipd.it/home/g.saielli/