[gmx-users] No DD grid found

[Mark Abraham]

On 28/03/2011 10:37 PM, André Ferreira wrote:
> Dear all,
>
> I am having some problems with some molecules that I am currently 
> simulating. I got simulation box that is constituted by 100 oligomer 
> molecules (190 atoms each), in a cubic box with 7 nm side (50K). Since 
> the molecule is quite long I am using a rcoulomb of 2.8 nm and rvdw of 
> 3.4 nm.

That does not seem like a good reason to choose massive values for 
cutoffs. What about the parameters used in your system makes you think 
you need to treat interactions accurately out to these distances? 
Proteins have physical dimensions larger than that, and nobody uses 
cutoffs greater than about 1.2nm for all-atom models.

> When I try to perform the simulation mdrun can't find DD grid for more 
> than 48 cores, which is a huge handicap. This is the first time that I 
> am working with longer molecules, and I never had this problem.

In order to obtain good parallel scaling, the DD implementation requires 
that the interaction partners of all atoms reside on a cell located on 
the same node or a cell located on an adjacent node. Thus, there is a 
minimum cell size. The size of your cutoffs creates a minimum cell size 
that is inconsistent with this principle when applied to your system.

Fundamentally, a given parallelization algorithm is useful only for a 
certain range of problem sizes. Trivially, you won't get useful speedup 
for 100 atoms on 1000 processors without a seriously specialized 
algorithm. The same algorithm would be terrible for 1000 atoms on 10 
processors. If you had around 100 times the number of atoms, then 
48-core DD would probably be fine. So you need to change a feature of 
your model physics, use fewer processors, or make your system bigger.

Mark