AMBER Archive (2008)

Subject: Re: AMBER: MPI run problem

From: Robert Duke (rduke_at_email.unc.edu)
Date: Thu Apr 03 2008 - 15:57:33 CDT

Hello,
This is a limit imposed by Tom Darden on the pme code; it is in one sense
not firm at all; it is in another sense within a factor of two of getting
into trouble if you are on a symmetric multiprocessor where multiple
processors share memory and maybe each cpu does not have at least two
gigabytes, or running on 1 or 2 processors. Basically going from 256-->512
increases your fft memory requirements by 8; going from 256-->1024 increases
your fft memory requirements by 64. Now, with the distributed fft calc, you
divide back down, but you can still get into trouble if you go above a grid
size of about 512 (depending on processor count, system memory, etc.)
Also, it gets really expensive computationally to handle fft grids at
densities above 1 grid point per angstrom. I did increase the default max
to 512 in pmemd 10 because we are hitting some instances with really big
systems where a dimension is > 256 angstrom, though. Okay, so with all
these caveats, you can fix your copy of the code to allow a larger max grid
size by editing mdin_ewald_dat.fpp and recompiling. The critical value is:
integer, parameter :: gridhi = 512 (or whatever new value you want - it
is 256 in pmemd 9 I believe).
Regards - Bob Duke

----- Original Message -----
From: "David A. Case" <case_at_scripps.edu>
To: <amber_at_scripps.edu>
Sent: Thursday, April 03, 2008 4:38 PM
Subject: Re: AMBER: MPI run problem

> On Thu, Apr 03, 2008, Qi Yan wrote:
>>
>> | ERROR: nfft1 must be in the range of 6 to 256!
>
> Means what it says: in amber9 pmemd, the maximum grid dimesion is 256.
> You
> would either need a smaller system, a coarser fft grid, or updated code.
> Bob
> Duke can comment on how firm this limit is. I note that in Amber10
> (coming out
> soon!), the maximum grid dimension is increased to 512.
>
> ...dac
>
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber_at_scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu
>

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu