Re: How to optimize parameters in command for NAMD CUDA GPU MD calculations?

From: Aron Broom (broomsday_at_gmail.com)
Date: Thu Sep 12 2013 - 15:41:50 CDT

One thing first, since NAMD is still strongly dependent on CPU and GPU
communication bandwidth, I'm not sure how reliable a comparison is when
both the cards and the rest of the hardware is different.

But ignoring the above, in the above cases only a single CPU is used.
Since the calculation can't proceed until both the GPU and CPU steps have
completed, whichever is slower will be the bottleneck. So it could simply
be that even in the case of the tutorial the CPU was the limiting step, and
you're 1.87 factor mostly just illustrates the improvement in you i7 CPU
over whatever there was at the time.

You should test the different number of CPUs and see where your maximum
lies (are those 8 real cores or just 4? hyperthreading generally won't
help in this case). How you will compare that back to a GTX285 without
having the card I don't know.

On Thu, Sep 12, 2013 at 3:59 PM, Michael Shokhen
<michael.shokhen_at_biu.ac.il>wrote:

> Dear Namd Users,
>
>
> I wanted to estimate what is the computational advancement of GeForce
> GTX TITAN vs.
>
> GeForce GTX 285 in molecular dynamics (MD) simulations by NAMD software
> with implemented
>
> CUDA GPU.
>
> I have installed on my computer the
> NAMD_CVS-2013-06-14_Linux-x86_64-multicore-CUDA and
>
> vmd-1.9.1.bin.LINUXAMD64.opengl software packages.
>
>
> I used as a benchmark the “Membrane Proteins Tutorial” by Alex
> Aksimentiev et all where
>
> GeForce GTX 285 was used in MD calculations:
>
> http://www.ks.uiuc.edu/Training/Tutorials/science/membrane/mem-tutorial.pdf
>
>
> Following the instructions in Chapter 3 “Running a Simulation of KcsA ”
> page 34.
>
> I have ran in a terminal window the NAMD software by the command modified
> for CUDA GPU:
>
> namd2 +idlepoll kcsa_popcwimineq-01.conf > kcsa_popcwimineq-01.log
>
>
> The acceleration coefficient over the tutorial variant is only 1.87.
>
> See more details below in the LOG files for comparison.
>
> I would appreciate if somebody could advise me what command (what
> additional parameters)
>
> should I use for more optimal work of my computer hardware in order to get
> faster MD calculations on CUDA GPU. Unfortunately, I failed to find answer
> in internet.
>
>
> Thank you,
>
> Michael
>
>
>
> My workstation hardware:
>
> 8 core Intel(R) Core(TM) i7-3820 CPU @ 3.60GHz
>
> cpu MHz : 1200.000
>
> cache size : 10240 KB
>
> 1 GeForce GTX TITAN
>
> *
> *
>
> System:* *
> Ubuntu 12.04.2 LTS (GNU/Linux 3.2.0-45-generic x86_64)
>
> NVIDIA-Linux-x86_64-319.32
>
>
>
> Information from log files:
>
>
> My variant:
>
> Info: Running on 1 processors, 1 nodes, 1 physical nodes.
>
> Info: CPU topology information available.
>
> Info: Charm++/Converse parallel runtime startup completed at 0.00349903 s
>
> Did not find +devices i,j,k,... argument, using all
>
> Pe 0 physical rank 0 binding to CUDA device 0 on quant-lnx: 'GeForce GTX
> TITAN' Mem: 6143MB Rev: 3.5
>
> Info: 6.47266 MB of memory in use based on /proc/self/stat
>
> Info: Configuration file is kcsa_popcwimineq-01.conf
>
>
> WallClock: 14751.504883 CPUTime: 14751.504883 Memory: 182.707031 MB
>
> Program finished.
>
>
>
> Tutorial variant:
>
> Info: Running on 1 processors, 1 nodes, 1 physical nodes.
>
> Info: CPU topology information available.
>
> Info: Charm++/Converse parallel runtime startup completed at 0.0157971 s
>
> Pe 0 physical rank 0 binding to CUDA device 0 on home-lnx: 'GeForce GTX
> 285' Mem: 1023MB Rev: 1.3
>
> Info: 6.59766 MB of memory in use based on /proc/self/stat
>
> Info: Configuration file is kcsa_popcwimineq-01.conf
>
>
> WallClock: 27552.734375 CPUTime: 27552.734375 Memory: 129.031250 MB
>
> Program finished.
>
>
> *****************************
> Michael Shokhen, PhD
> Associate Professor
> Department of Chemistry
> Bar Ilan University,
> Ramat Gan, 52900
> Israel
> email: shokhen_at_mail.biu.ac.il
>
>
>
>
>

-- 
Aron Broom M.Sc
PhD Student
Department of Chemistry
University of Waterloo

This archive was generated by hypermail 2.1.6 : Tue Dec 31 2013 - 23:23:43 CST