[Pw_forum] I/O performance on BG/P systems

[Paolo Giannozzi]

On Mar 25, 2011, at 13:58 , Gabriele Sclauzero wrote:

> Before moving some complaints to the sysadmins, I would like to have
> some feedback from those of you who have used this kind of  
> architecture.
> Are you aware of similar kind of I/O issues? Is there any performance
> tuning that can be done on the software/compiler side?

I also noticed that reading/writing the CP restart file on a BG takes a
sizable amount of time, in the order of a few minutes for a typical  
system
suitable for BG. I had attributed this to the "gathering" of the  
distributed
pieces of wavefunctions that are subsequently written to file by a  
single
process, though; in your case, it is the I/O by each process that  
takes a
lot of time. You may want to try to reduce the I/O to the strict  
minimum:
work with one k-point per processor (use k-point parallelization if  
needed),
use disk_io='low' (or 'none'): the code should write only at the end  
(or never).

P.
---
Paolo Giannozzi, Dept of Chemistry&Physics&Environment,
Univ. Udine, via delle Scienze 208, 33100 Udine, Italy
Phone +39-0432-558216, fax +39-0432-558222