AMBER Archive (2004)

Subject: RE: AMBER: question about parallel simulation

From: Ross Walker (ross_at_rosswalker.co.uk)
Date: Thu Dec 02 2004 - 16:37:12 CST


Hi Qiang,

> After the change of all bc_*** parameters. The code works
> fine for 2 nodes.

> But the time scale between
> 2 and 4 nodes are almost the same. The list of the time
> difference bellow
> shows that the problem is in
> the force adjustment calculation. Is there a way to overcome
> this problem?

This is a much much bigger problem to solve and is why programming in
parallel is so hard... First of all you should run some benchmarks of the
unmodified code on your system and plot the speeds. Then run your modified
code in exactly the same way an plot the speeds. The fact that your machines
are dual CPU is why it works fine on 2 cpus but slows down on 4. You are
either inadvertantly running all 4 threads on the same machine OR your back
plane gets swamped with communication between the two machines. Running the
unmodified code in the same way as you have run your modified code should
tell you if you are inadvertantly running all 4 threads on 1 machine.

As for the second case where communication is swamping your backplane this
is much more tricky. If it works fine with the unmodified code but not with
yours then it is mpost likely due to the fact that you made the arrays that
are broadcast MUCH MUCH bigger. This hurts you in 2 ways. First of all your
machines may not have enough memory to store the arrays in ram and so they
may be swapping. Secondly you are broadcasting much more information on each
step. Hence you should really make these arrays as small as you can get away
with. There is no point neelessly broadcasting data. Also, in order to get
more speedup you will need to be much cleverer, for example you could check
what has changed on each cpu and broadcast that data. This is significantly
more involved though and will take a considerable (read "months") of time.

So, take a look at the arrays you have and make them as efficient as
possible in what they store / require.

All the best
Ross

/\
\/
|\oss Walker

| Department of Molecular Biology TPC15 |
| The Scripps Research Institute |
| Tel:- +1 858 784 8889 | EMail:- ross_at_rosswalker.co.uk |
| http://www.rosswalker.co.uk/ | PGP Key available on request |
 

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu