Re: How to best use GPUs

From: Itamar Kass (itamar.kass_at_gmail.com)
Date: Thu Jun 25 2020 - 12:20:37 CDT

Shalom Josh and Giacomo,

Thanks for the quick reply. I am using NAMD2.13
(Linux-x86_64-multicore-CUDA). NAMD recognised 4 TITAN Xp cards.

I am now running benchmarking with different numbers of CPU to see if this
effect.

Cheers,
Itamar

On Thu, 25 Jun 2020 at 19:11, Josh Vermaas <joshua.vermaas_at_gmail.com> wrote:

> Something else that I've done without paying attention is to run a
> CPU-only NAMD binary and then wonder why GPUs weren't helping performance.
> What build of NAMD does the top of your log report?
>
> -Josh
>
> On Thu, Jun 25, 2020 at 6:53 AM Giacomo Fiorin <giacomo.fiorin_at_gmail.com>
> wrote:
>
>> Hi, regardless of GPU use, there is nothing to gain (and often something
>> to lose) in running more threads than physical cores available. Using all
>> the "hardware" threads (which are 2x the number of cores in most CPUs) only
>> makes sense for codes that spend a lot of time waiting for data to be
>> loaded into the CPU, and are essentially processing a multitude of
>> *independent* tasks. This is not the case for NAMD. Also, try to
>> further decrease the number of threads, which are certainly competing for
>> CPU-GPU bandwidth.
>>
>> It may be helpful to know that the 3.0 alpha version may be more
>> efficient in the single-GPU setup by running on it exclusively:
>> https://www.ks.uiuc.edu/Research/namd/3.0alpha/
>>
>> Giacomo
>>
>> On Thu, Jun 25, 2020 at 6:08 AM Itamar Kass <itamar.kass_at_gmail.com>
>> wrote:
>>
>>> Hi,
>>>
>>> I am using NAMD 2.13 on a single Linux machine with 32CPU (64 threads)
>>> and 4 NVIDIA TITAN Xp cards. I run a protein solvated in TIP3 water (total
>>> system size ~220K).
>>>
>>> I have tried to run with few flags:
>>> namd2 +p64 +setcpuaffinity +devices 0,1,2,3 MD004.conf
>>> namd2 +p64 +setcpuaffinity +idlepoll +devices 0,1,2,3, MD004.conf
>>> namd2 +p64 +idlepoll +devices 0,1,2,3, MD004.conf
>>>
>>> But in all cases, it seems that the GPUs do not contribute anything to
>>> the run (I look at the GPU load while running and I compared the run time
>>> to a run without GPUs).
>>>
>>> In addition, I set:
>>> *outputEnergies 5000*
>>>
>>> Thanks in advance,
>>> Itamar
>>>
>>>
>>> --
>>> -----
>>> Itamar Kass, PhD.
>>>
>>> https://fromthekeyboardofastructuralbiologist.wordpress.com
>>>
>>
>>
>> --
>> Giacomo Fiorin
>> Associate Professor of Research, Temple University, Philadelphia, PA
>> Research collaborator, National Institutes of Health, Bethesda, MD
>> http://goo.gl/Q3TBQU
>> https://github.com/giacomofiorin
>>
>

-- 
-----
Itamar Kass, PhD.
https://fromthekeyboardofastructuralbiologist.wordpress.com

This archive was generated by hypermail 2.1.6 : Thu Dec 31 2020 - 23:17:13 CST