Dear Shaofeng,

scaling depends in general by number of factors. Moreover, you are providing no information about your input system. This results in a too generic question which is hard to answer if not making other questions (to be generally avoided on mailing lists).

I reckon you registered a better scaling from 2 to 4 cores and from 4 to 8 cores as you are reporting the issue only about scaling from 8 to 16 cores.
By experience I would say that you are dumping into two possible scenarios:

1) You reached a scaling limits.
In general px.x scales well up to 16 processes except you are working with a really small system.

2) You are actually working on a two quad-core CPU system with hyper-threading enabled. The hyper-threading extends to 8 the number of virtual cores per CPU socket (your /proc/cpuinfo reports 16 cores). In most cases, high-performance software applications (such as QE) do not benefit of this feature scaling only up to the number of physical cores available. Possibly you can google the model of you processor (see /proc/cpuinfo) to verify the number of physical cores on each CPU socket.

Regards,

Ivan


On 15/09/2015 03:40, Shaofeng Wang wrote:
Dear all,
I got a server computer equipped with two 8 core cpus. I installed intel fortran compiler (12.1) and openmpi (1.6.5). No errors took place during installation. Then I compiled parallel quantum esspresso (5.2.0) and the program run correctly. However, I found pw.x runs at very similar speed with 8 cores or 16 cores. Could any expert show me how to improve the calculation speed?
The make.sys file is attached.
Thanks in advance.
Shaofeng
--------------------------------------
Shaofeng Wang, Ph.D of Geochemistry
Environmental Molecular Science Group
Institute of Applied Ecology, Chinese Academy of Sciences
Shenyang, 110016, China
[email protected]
www.iae.cas.cn



_______________________________________________
Pw_forum mailing list
[email protected]
http://pwscf.org/mailman/listinfo/pw_forum

_______________________________________________
Pw_forum mailing list
[email protected]
http://pwscf.org/mailman/listinfo/pw_forum

Reply via email to