[gmx-users] GPU gets faster with more molecules in system

[Mark Abraham]

On 25/01/2011 8:25 AM, Christian Mötzing wrote:
> Hi,
>
> I compiled mdrun-gpu and tried some waterbox systems with different
> atoms counts.
>
> atoms  | GPU    | CPU
> 2.400  | 1.015s | 774s
> 4.800  | 1.225s | 1.202s
> 9.600  | 1.142s | 1.353s
> 19.200 | 2.984s | 2.812s
>
> Why does the system with 9.600 atoms finish faster than the one with
> 4.800? I tripple checked the simualtions and even GROMACs tells me that
> the atom count in the system is like above. So I think no mistaken
> there. A diff of md.log only shows differences in output values for each
> step.
>
> Is there any explanation for this behaviour?

As a guess, the cost of overheads for molecular simulations tend to have 
a weaker dependence on system size than the cost of computation (or none 
at all). Only once the latter dominate the cost do you see scaling with 
system size.

I expect you'd see similar behaviour running systems with 64, 128, 256, 
512 atoms on 64 processors.

Mark