Hi Daniel,
> For the minimization step, is it possible to shorten the length of the
> simulation by decreasing the values for
You can reduce maxcyc which means your system will not minimize as well.
This is something you'd have to look at carefully to see what the GMAX is as
a function of the number of steps. To be honest minimization is quick
compared with MD because you do a lot less steps. Therefore I always do more
than I think is necessary since it doesn't hurt.
> maxcyc and nstlim? We are new to Amber and to MD and are therefore not
Now, nstlim is a MUCH MUCH more complex question and really needs an
understanding of statistical mechanics to appreciate why it is a more
complex question. Reducing nstlim will reduce the amount of sampling you do,
this will mean your results are not as well converged. You should always
check for convergence in your results by discarding some of the data and
seeing how the predicted values change. The amount of sampling needed is
VERY dependent on the system being simulated and the properties you want to
measure. Thus it is unfortunately not easy to tell you what to do with
regards to the length of your runs. This is something you'd have to look at
justifying yourself. Ultimately in a publication could you defend your
choice of run length?
> entirely sure about the effects of the various parameters on the
> simulation. Initially, we were using sander.MPI because this is the
> version of sander that was used in the example minimization step
> provided in tutorial 3. Also, I just found out that pmemd has not been
> compiled on our cluster. Thanks a bunch,
Yes, the tutorials all use Sander since it used to be somewhat complicated
(prior to AMBER 10) to compile PMEMD and most workshop or training machines
did not have pmemd installed. With AMBER 11 it is built as part of the main
build and therefore the recommendation is to use it if you can. PMEMD
supports a subset of the sander functionality, the exact differences are
detailed in the manual. Ultimately though it will read sander input and
right sander output the only difference being that it runs faster and scales
better.
Something I would suggest doing as well is running your simulation with
difference processor and node counts. Take a look at the mdinfo file after
about 3 to 4 minutes of running to get the performance numbers. Note running
more tasks is not always faster since once you go beyond the scaling limit
(which is simulation and hardware dependent) performance drops off
massively.
You also don't mention what your cluster is. If the interconnect is not a
good quality high speed interconnect, such as infiniband, but is instead
gigabit Ethernet for example then you will not be able to scale beyond a
single node. Running on multiple nodes in this case will give you
unbelievably bad performance, probably a lot worse than just running on one
node.
All the best
Ross
/\
\/
|\oss Walker
---------------------------------------------------------
| Assistant Research Professor |
| San Diego Supercomputer Center |
| Adjunct Assistant Professor |
| Dept. of Chemistry and Biochemistry |
| University of California San Diego |
| NVIDIA Fellow |
|
http://www.rosswalker.co.uk |
http://www.wmd-lab.org/ |
| Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
---------------------------------------------------------
Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu May 26 2011 - 10:30:03 PDT
