[Pw_forum] relaxation problem on IBM bluegene machine

Paolo Giannozzi giannozz at democritos.it
Wed Jan 7 09:11:20 CET 2009


pushpa raghani wrote:

> When I use c(2x3) cell (with U and/or without U), everything goes 
> fine. When I use c(2x4) cell without U (just spin polarized 
> calculation), then also there is no such problem [...].
> The problem arises when I do LDA+U for c(2x4) cell.

then it might also be a new version of an old problem. Excerpt from
http://www.quantum-espresso.org/wiki/index.php/Developer_Manual#Parallelization:
---
Beware: replicated calculations may either be performed independently on 
each processor, or performed on one processor and broadcast to all 
others. The first approach requires less programming, but it is unsafe: 
in principle all processors should yield exactly the same results, if 
they work on the same data, but sometimes they don't (depending on the 
machine, compiler, and libraries). Even a tiny difference in the last 
significant digit can eventually cause serious trouble if allowed to 
build up, especially when a replicated check is performed (in which case 
the code may hang if the check yields different results on different 
processors). Never assume that the value of a variable produced by 
replicated calculations is exactly the same on all processors: when in 
doubt, broadcast the value calculated on a specific processor (the root 
processor) to all others.
---
Guess what: IBM bluegene is one of those machine where different
processors may yield slightly different results starting with the
same data. Some time ago it took me a week to figure out why the
parallel diagonalization was hanging on a BG and not on other machines.

>     what happens if you change options (try to keep
>     data into memory instead of saving them to file), etc. etc.. 
> 
> This could work but how to do that?

option "disk_io", in namelist &control

In any case: try the latest stable version first, and maybe also
the cvs version (a snapshot is available from the download page)

Paolo
-- 
Paolo Giannozzi, Democritos and University of Udine, Italy



More information about the users mailing list