AW: Re: AW: Why the computational speed does not increase with the increase of GPUs?

From: Norman Geist (norman.geist_at_uni-greifswald.de)
Date: Fri Oct 23 2015 - 03:49:49 CDT

Seems your message got truncated?

 

Norman Geist

 

Von: owner-namd-l_at_ks.uiuc.edu [mailto:owner-namd-l_at_ks.uiuc.edu] Im Auftrag von sunyeping
Gesendet: Freitag, 23. Oktober 2015 09:58
An: namd-l <namd-l_at_ks.uiuc.edu>; Norman Geist <norman.geist_at_uni-greifswald.de>
Betreff: namd-l: Re: AW: namd-l: Why the computational speed does not increase with the increase of GPUs?

 

 

 

Yeping Sun

Institute of Microbiology, Chinese Academy of Sciences

 

------------------------------------------------------------------

From:Norman Geist <norman.geist_at_uni-greifswald.de <mailto:norman.geist_at_uni-greifswald.de> >

Time:2015 Oct 22 (Thu) 23:12

To:namd-l < <mailto:namd-l_at_ks.uiuc.edu> namd-l_at_ks.uiuc.edu>, 孙业平 < <mailto:sunyeping_at_aliyun.com> sunyeping_at_aliyun.com>

Subject:AW: namd-l: Why the computational speed does not increase with the increase of GPUs?

 

You system is quite small with 60.000 Atoms. You may get an improvement by trying in your namd script:

 

twoawayx yes

 

If that helps, try in addition:

 

twoawayy yes

 

If that helps, try in addition:

 

twoawayz yes

 

How many cores does this machine have?

 

Norman Geist

 

 

Thank you. But it doesn't work.

 

Von: owner-namd-l_at_ks.uiuc.edu <mailto:owner-namd-l_at_ks.uiuc.edu> [mailto:owner-namd-l_at_ks.uiuc.edu] Im Auftrag von sunyeping
Gesendet: Donnerstag, 22. Oktober 2015 05:18
An: namd-l < <mailto:namd-l_at_ks.uiuc.edu> namd-l_at_ks.uiuc.edu>
Betreff: namd-l: Why the computational speed does not increase with the increase of GPUs?

 

Dear all,

 

I find increasing the number of GPU can not obviously increase the computational speed with NAMD. I've tested the computational speed of NAMD with two K80 GPUs and four K80 GPUs. The simulation system are the same, which contains about 60 thousand atoms, and the NAMD version is NAMD_CVS-2015-10-21_Linux-x86_64-multicore-CUDA. With two K80 GPUs, the computaional speed was 0.0566 day/ns, which with four K80 GPUs, the speed was 0.0456 day/ns. So there was little improvement with doubling the number of divices. It seems that there are some defets in NAMD code. But how can improve this situation?

Yeping Sun

Institute of Microbiology, Chinese Academy of Sciences

 

 

This archive was generated by hypermail 2.1.6 : Thu Dec 31 2015 - 23:22:10 CST