Re: NAMD distribute

From: Robert Brunner (rbrunner_at_uiuc.edu)
Date: Thu Oct 13 2005 - 17:40:26 CDT

Another problem with using NAMD in a distributed system is that big
simulations are fairly communication-sensitive. NAMD uses spatial
decomposition, so the simulation space is divided up into a lot of
cubes. Each cube has to communicate with its neighboring cubes before
each time step can finish. That means each processor must communicate
with a number of other processors every few milliseconds, if you want
to get a fairly long simulation done in a timely manner. When your
neighbor-processors are on high-latency connections, the time to
complete each simulation step is limited by the time it takes to
communicate with your slow remote neighbors. That makes it difficult
to take advantage of distributed computing for a large single
simulation.

Distributing many small independent simulations out to a lot of
remote machines could be productive, although the security problems
previously mentioned must be considered.

Robert

On Oct 13, 2005, at 3:46 PM, Dan Strahs wrote:

>
> Thanks for the responses, Michael and Brian.
>
> So what I gather is that this would be doable if;
> 1) the communication between the grid and server is not significant
> and 2) if the calculations could be secured.
>
> These seem tractable issues for the most part, even for large
> simulations
> a la Schulten etc. Perhaps even load balancing could be partially
> parallelized and farmed out.
>
> Security will always be a problem in large-scale distributed
> computing due
> to systemic hacking, but small-scale? If the grid is geographically
> localized to a managable facility, such as a university or company,
> then
> security is enhanced by anonymity.
>
> Dan Strahs
>

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:41:15 CST