Particle Resort Algorithm

[William French]

Hello,

I am running simulations of metallic nanowires undergoing elongation in vacuum. I have noticed some interesting performance-related issues when I adjust the particle resort algorithm in HOOMD. Since the relative particle positions in the nanowire do not change much during the simulation, performing particle resorts often should not be necessary. However, when I do occasionally perform a resort, the performance takes a hit (~5-6%); and I don't mean the resort itself is taking a long time, but rather that the TPS well after the resort has occurred is lower than it was before the resort. I have attached a plot to illustrate what I mean. 

I do not have a great understanding of how the space-filling curve works in HOOMD, but I suspect there is some implementation detail that does not work well for my particular system. Note, there are large void spaces on all sides of the wire to prevent self-interaction. Perhaps the issue is somehow related to these large void spaces?

Another thing I've noticed, if I disable the sorter using:

sorter.disable()

the performance is worse (~5-6%) throughout the simulation than if I simply use an arbitrarily long (longer than the total simulation length) resort period.

Anyone (who understands the particle resorting algorithm in HOOMD better than I do!) have ideas about what might be causing these issues?

Thanks,

Will

[Joshua Anderson]

The void spaces would make the space filling curve a non-optimal solution. It does fill all of space, after all. Thus I can understand a performance hit from the sort when compared to your initial ordering (already cache-friendly). 

I don't have any immediate explanation as to how no sorting at all is slower in your case. You can run for a bit with profiling enabled (run(..., profile=True) and see where all the extra time is going. That will shed some light on this behavior.

I've only ever tested the sorter for fluid, mostly diffusive self-assembly simulations. There, disabling the sorter causes a 500% performance degradation.

- Josh

--
You received this message because you are subscribed to the Google Groups "hoomd-users" group.
To post to this group, send email to hoomd...@googlegroups.com.
To unsubscribe from this group, send email to hoomd-users...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/hoomd-users?hl=en.