Re: improving parallel on mac cluster

From: yu275197_at_yorku.ca
Date: Tue Feb 10 2004 - 18:23:39 CST

michael,
you mean i could be running all three processes on one node? well, i'm
monitoring the processor usage on all the nodes and they correlate
properly...i can also see namd being run in the job list inside the sys.
monitoring app. my simulation also runs faster on two systems. its just that
the efficiency doesn't increase after that. I've also tried it on a dual
processor G5 (by naming the system twice in the nodelist) and it works
fine(even faster).
my question is: how do we know namd is dividing the work properly? Is it
possible to control this?

min

Quoting Michael Grabe <mgrabe_at_itsa.ucsf.edu>:

> min,
>
> This may be a really stupid question, but are you sure that you are
> spawning 1 process per processor? A while ago, I was accidentally
> starting all of the processes on the same node because of a nodelist
> problem, and i spent a few hours wondering why my efficiency stunk
> until looked more closely. It was not obvious at first.
>
> -Michael
>
>
>
> On Tuesday, February 10, 2004, at 02:47 PM, yu275197_at_yorku.ca wrote:
>
> > Hi Brian
> > You are right. I don't think it's my network because i loose
> > efficiency after
> > the third machine. Do you know of any other reason why my simulation
> > won't run
> > any faster. It's very frustruating that i can't take advatage of the
> > 100 plus
> > systems i have here.
> >
> > Cheers
> > Min
> >
> >
> >
> >
> >
> > Quoting Brian Bennion <brian_at_youkai.llnl.gov>:
> >> Hello Min,
> >>
> >> Without more information on the system size I would only be guessing.
> >> With that caveat. I don't think the ethernet connection is going to
> >> be a
> >> problem (yet) unless it is less than 10MBs which I doubt. I have a 16
> >> cpu
> >> cluster running on top of a 100MBs network and for the systems I run
> >> (~45,000 atoms) I lose effeciency with more than 8 cpus due to the
> >> network communications. I use 11A cutoffs with switching and margins
> >> set
> >> to zero.
> >>
> >> Brian
> >>
> >>
> >>
> >> On Mon, 9 Feb 2004 yu275197_at_yorku.ca wrote:
> >>
> >>>
> >>> I've managed to run my simulation in parallel on a couple of G5 mac
> >> computers
> >>> but i've noticed that my performance does not increase after three
> >>> nodes.
> > I
> >> am
> >>> guessing the possible reason for this to be:
> >>> 1. simulation is bottlenecked by the ethernet connection
> >>> 2. my molecule is not configured properly to take advantage of
> >>> mutiple
> >> sys...
> >>> i have my cutoff set to a large number(essentially no cutoff) but
> >>> if i
> >> change
> >>> it to smaller number with switching on the simulation seems to be
> >> distributed
> >>> between the systems better.
> >>> 3. patch size...i think this is related to reason #2. because number
> >>> of
> >>> patches namd creates seems to be dependent of the conf file and
> >>> simulation
> >>> speed seems to be dependent on the number of patches.
> >>>
> >>> could anyone provide some insight into my situation.
> >>>
> >>> Cheers
> >>> min
> >>>
> >>
> >> *****************************************************************
> >> **Brian Bennion, Ph.D. **
> >> **Computational and Systems Biology Division **
> >> **Biology and Biotechnology Research Program **
> >> **Lawrence Livermore National Laboratory **
> >> **P.O. Box 808, L-448 bennion1_at_llnl.gov **
> >> **7000 East Avenue phone: (925) 422-5722 **
> >> **Livermore, CA 94550 fax: (925) 424-6605 **
> >> *****************************************************************
> >>
> >>
> >
> >
> >
> ------------------------------------------------------------------------
> --------------------------
> Michael Grabe, Ph.D.
> Post-doctoral Fellow
> Howard Hughes Medical Institute
> University of California, San Francisco
> 533 Parnassus Ave.
> San Francisco, CA 94143
> mgrabe_at_itsa.ucsf.edu
> tel: ++ 415.476.6585
> http://itsa.ucsf.edu/~mgrabe
>
>
>
>

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:37:20 CST