From: Kenneth Huang (kennethneltharion_at_gmail.com)
Date: Mon Oct 12 2015 - 20:09:26 CDT
Hi,
I'm not an expert on the subject, but my understanding is that running any
MD or any sort of calculation with more than 1 CPU per GPU tends to cause a
bottleneck- each iteration has to pass from the GPU to the CPU to
communicate the information, which then has to go back to the GPU, which
compounds significantly the more CPUs you introduce. Whereas if you're just
running by 1 GPU/CPU, then you only need to communicate once as opposed to
16 times.
20 ns/day for a system of your size seems kind of slow, so I'd try it first
with just 1 CPU and 1 GPU, and see if you can't get some sort of speed up
that way, or just running solely on CPU. It also does depend on what type
of GPU you're using, since some have noticeable improvements over others.
Best,
Kenneth
On Mon, Oct 12, 2015 at 6:07 PM, jeevan gc <gcjeevanbdr_at_gmail.com> wrote:
> Dear all,
>
> In my protein membrane system of 121416 atoms,
> When I use 16 CPU +1 GPU, the total time is 3 days for 20ns for the regular MD.
> For the same system it takes 6 days to run 20ns accelerated MD.
>
> Why accelerated MD is slow in GPU? I am using NAMD2.10
>
> Thank you.
>
>
>
>
-- Ask yourselves, all of you, what power would hell have if those imprisoned here could not dream of heaven?
This archive was generated by hypermail 2.1.6 : Thu Dec 31 2015 - 23:22:06 CST