[Pw_forum] parallel pw.x memory requirements

[Paolo Giannozzi]

On Jan 6, 2007, at 14:36 , Alexander Shaposhnikov wrote:

> [...] what i have noticed, the amount of memory used per CPU
> does not decrease more than 2 times. For example, running on
> 8-node SMP machine with 8 threads consume 4 times more
> memory comparing with one only thread ( with 1 pool, procs/pool=8 )

> Is this expected behavior? Can somebody clarify this?

there is a "scalable" memory:

-  arrays containing quantities on either the real-space grid
    or the plane-wave basis set. These are distributed across
    processors (exception: in some cases, a single array as large
    as the entire real-space grid is used).

and a "nonscalable" memory:

-  most remaining arrays. These not distributed but replicated.
    (or they exist only on the master process). Note that arrays
    with dimensions Nstates*Nstates (still) belong to this category
- the code itself, libraries, etc

Right now the amount of "nonscalable" memory is not so small,
say > 200Mb, so the code will take a sizable amount of memory
no matter how many processors you use. Work is in progress
to reduce it to the strict minimum, in view of usage of massively
parallel machines with not-so-much memory per processor.
Meanwhile you can reduce the values of fixed dimensions (in
Modules/parameters.f90) from their default value, especially
those: ntypx, ndmx, nbrx .

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