Re: IBM SP build

From: Jim Phillips (jim_at_ks.uiuc.edu)
Date: Thu Nov 13 2003 - 14:57:20 CST

Hi,

On Sat, 8 Nov 2003, David Skinner wrote:

> In building NAMD 2.5 for our SP I encountered a few questions:
>
> 1) Is sfftw (compare to rfftw) now implemented in order to decrease
> message volume? If so I assume the impact on roundoff is deemed
> acceptable. If not is there a switch to make everything doubles instead
> of floats?

Yes, using single precision for PME cuts communcation for this part of the
algorithm in half. We have seen no observable differences in the
simulation output (same energies for hundreds of steps), since the
reciprocal PME forces are much smaller than the direct ones. There are
some accuracy plots at http://www.fftw.org/accuracy/ showing
single-precision FFT having a relative RMS error of around 1e-7, which
should be sufficient for MD. There isn't a flag to switch to double,
although this might not be that hard to implement.

> 2) Is there a way to instruct charm to run threads within a node? I.e.
> running 32 nodes and 16 tasks per node will charm generate 32 or 512 MPI
> tasks?

If you build an SMP port of charm then +ppn on the command line controls
the number of (worker) threads per node. I've tried this on NCSA's p690
and wasn't thrilled with the performance. Also, this is still somewhat
experimental so you might happen across some rare new bugs.

> 3) The AIX binary distributed is 32 bit. What is the setting of bmaxdata
> ? Is building in 64 bit mode possible/tested? This is related to the
> last question in terms of memory consumed by MPI buffers at high
> concurrency. I could try a full 64 bit compile, but thought I would
> check first.

This should be trivial, but it hasn't been tested (by me at least).

-Jim

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:37:08 CST