From: Jeff Comer (jeffcomer_at_gmail.com)
Date: Mon Jan 08 2018 - 10:04:52 CST
Just wanted to add my experience, in case it might be helpful to anyone.
Usually with for CPU-only jobs, there is a modest (~10%) improvement
if I run NAMD with two tasks per physical core. With GPUs, it seems to
vary depending on many factors, including system size, number of CPUs,
and the speed of the GPU. Usually with fast GPUs (1080 Tis), I see a
loss of performance with two tasks per physical core. The number of
tasks versus the number of GPUs seems to be an important issue—having
more than 16 tasks per GPU seems to be bad.
For example, with 6 physical cores (Xeon E5-1650 v2 @ 3.50GHz), a
GeForce GTX 1080 Ti, and a 14,303-atom system, I get the following
6 cores, 1 GPU: 199.6 ns/day
12 cores, 1 GPU: 142.8 ns/day
6 cores, 0 GPUs: 15.9 ns/day
12 cores, 0 GPUs: 17.8 ns/day
Note that it is necessary to use a new build of NAMD to get this
performance (Version Nightly Build), since the GPU performance has
improved since NAMD 2.12.
I also want to mention that I previously got improved performance by
using a smaller than standard cutoff (9 Å instead of 12 Å), but this
no longer gives improved performance with a 1080 Ti GPU and recent
CUDA builds of NAMD. This suggests that the performance bottleneck is
no longer direct non-bonded interactions.
Jeffrey Comer, PhD
Institute of Computational Comparative Medicine
Nanotechnology Innovation Center of Kansas State
Kansas State University
Office: P-213 Mosier Hall
On Mon, Jan 8, 2018 at 7:27 AM, Jérôme Hénin <jerome.henin_at_ibpc.fr> wrote:
> Hi Horacio,
> With this setup, do you gain anything from hyperthreading? What happens if
> you run just one task per physical core?
> On 8 January 2018 at 14:17, horacio poblete <elhoraki_at_gmail.com> wrote:
>> Hi Lewis.
>> I have a very similar setting-up than yours.
>> I have a 2 x E5-2683 v4, 16C. 2.1GHz (32 cores=64 threads total) plus 2
>> and 4 GTX 1080Ti.
>> I have been running on it 24/7 also, without any problem. The only issue
>> is that namd still yet been cpu dependent more than GPU.
>> example: With one of the servers running 2 calculation simultaneously (
>> each one: 32 threads +1GPU, ~45K atoms), I can reach ~100 ns/day but but
>> the performance of the GPUs reach only to the ~60% of the GPU.
>> on the other hand, in the other server, running 4 calculations
>> simultaniously (16 threads +1GPU, each) I can get ~60-for the 70 ns/day but
>> now I am using only the 38% of each CPU.
>> I dont know if the namd developers are working on improve it, I hope yes!
This archive was generated by hypermail 2.1.6 : Mon Sep 16 2019 - 23:19:21 CDT