Re: [AMBER] multi-threading cpu's or no multi-threading

From: Jonathan Gough <jonathan.d.gough.gmail.com>
Date: Tue, 29 May 2012 22:25:33 -0400

I just set up an i7-920 machine with 12GB of RAM. and a tesla 1060 running
ubuntu 12.04. I'll give it a shot and see how it goes. i also set up a
hackentosh with a GTX 560, although i have been having issues compiling
amber on it...

On Tue, May 29, 2012 at 9:25 PM, Jason Swails <jason.swails.gmail.com>wrote:

> Hello,
>
> On Tue, May 29, 2012 at 3:40 PM, Jonathan Gough
> <jonathan.d.gough.gmail.com>wrote:
>
> > Dear all,
> >
> > I'm new to this and just setting up a new machine. My question is. if
> you
> > have a CPU that allows multi-threading (as most modern cpu's do) do you
> > have it enabled or disabled for running amber calculations?
> >
> > for example a quad core i7 will show 8 cpu's when in actuality it is only
> > really 4. The setting is changed in the bios (i know). But wondering if
> > and how it matters when setting a machine up to run amber. I know for QM
> > calculations (gamess) multi-threading doesn't help and can sometimes
> hurt,
> > and I always keep it turned off.
> >
>
> Try a couple of experiments: Look at the performance of 4 pmemd/sander
> threads on a simple benchmark with hyperthreading turned on and then again
> with hyperthreading turned off. Then, look at the performance difference
> of running 4 threads without hyperthreading vs. 8 threads with
> hyperthreading, and see how that change impacts performance.
>
> Performance is likely affected by several factors: what 'multithreading'
> scheme does your particular processor use? (Wikipedia suggests the
> existence of several and some are better than others at certain tasks)
> Does your OS effectively recognize the difference between physically
> different processing cores and virtual cores -- that is, will it ever
> (often?) assign 2 threads to the 2 virtual processors on the same core?
> The more it does this, the worse hyperthreading will be.
>
> Another downside of using hyperthreading cores vs. real, distinct cores is
> the on-chip cache. This memory is the fastest memory available to the
> processor, yet it's also available in the smallest amounts.
> Multi-threading cores must share this cache (and all other resources it
> has) with all of the threads running on that core, which could increase the
> number of delays and slowdowns to the point that overall timing actually
> suffers.
>
> All I've really done is provided excuses for why hyperthreading with Amber
> may perform poorly, but I don't think anyone can say for certain how your
> system will fare unless you actually try it out.
>
> HTH,
> Jason
>
> P.S. If you really want to see performance fly, get a non-super-overclocked
> GTX 580 ;)
>
> --
> Jason M. Swails
> Quantum Theory Project,
> University of Florida
> Ph.D. Candidate
> 352-392-4032
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue May 29 2012 - 19:30:03 PDT
Custom Search