Hi Accelrys community,
I was wondering if anyone has any experience / knowledge of the parallelisation algorithms used by Forcite Plus. We are currently seeing some rather disturbing behaviour in some of our calculations and are trying to get to the bottom of what is happening.
As some background we are trying to simulate the interaction of single gas molecules with a surface, using molecular dynamics to model the gas/surface collision and resulting dynamics. We are using 3D periodic cells in which most of the cell is empty space, so that the gas molecule does not interact with the translated copy of the surface.
In some cases when we run the simulation on 4 or 8 CPUs, we see virtually no parallelisation. However making the surface wider improves substantially the parallel performance. Our best guess to explain this behaviour so for is to assume that ForcitePlus is using a domain decomposition type algorithm. In cases where the surface is small, the code divides the cell volume with planes parallel to the surface, leaving all of the atoms in one domain. When the surface is larger the cell volume may get divided by planes normal to the surface, more evenly dividing the atoms amoung the available domains.
Based on peoples knowledge of ForcitePlus, is this scenario likely? Also if it is the case is there any way to influence the way the cell volume is divided, either by explicit options or by the way the surface is constructed?
Thanks,
Bradw.