From: Norman Geist (norman.geist_at_uni-greifswald.de)
Date: Tue Oct 20 2015 - 03:53:51 CDT
You need to look at the walltime and forget about the pure CPU time.
Von: Puneet Singh [mailto:singh.punit1990_at_gmail.com]
Gesendet: Dienstag, 20. Oktober 2015 10:29
An: Norman Geist <norman.geist_at_uni-greifswald.de>
Betreff: Re: namd-l: Regarding significance of days per ns (days/ns) parameter in NAMD benchmark
Thankyou for a quick reply , this is a follow up question on your reply. I now understand that days taken per nanosecond simulation should be lower, but this gave rise to another confusion:
In my benchmark seconds per step are lower i.e. 0.078224 whereas anl.gov <http://anl.gov> 's 0.105658 s/step is higher.
Also my days/ns is 905.37 and their 1.2229 days/ns .
So as per my understanding with s/step my results seem better where as on days/ns anl.gov <http://anl.gov> 's result is good.
So here which parameter should be preferred for stating a clear winner in terms of performance? OR i should go with wall time instead?
Awaitiing your reply,
On 20 October 2015 at 11:59, Norman Geist <norman.geist_at_uni-greifswald.de <mailto:norman.geist_at_uni-greifswald.de> > wrote:
NAMD just calculates this value from the step times benchmarked at simulation start, cause it is a more human readable value than seconds per timestep. Lower is better as it is days needed per nanosecond simulated.
Von: owner-namd-l_at_ks.uiuc.edu <mailto: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 Puneet Singh
Gesendet: Dienstag, 20. Oktober 2015 08:01
An: namd-l_at_ks.uiuc.edu <mailto:namd-l_at_ks.uiuc.edu>
Betreff: namd-l: Regarding significance of days per ns (days/ns) parameter in NAMD benchmark
I wish to know that what is meant by the parameter days/ns & how it relates to namd performance?
my benchmark on Intel(R) Xeon(R) E5-2660 v2 ( NAMD V2.10 ApoA1 example - only cpu - no MIC/GPU):
16 CPUs 0.078224 s/step 905.37 days/ns 380.207 MB memory (42.40 seconds)
19 CPUs 0.0665057 s/step 769.742 days/ns 411.355 MB memory (36 seconds)
114 CPUs 0.203145 s/step 2351.21 days/ns 1976.23 MB memory (91.94 seconds)
Benchmark@https://wiki.anl.gov/cnm/HPC/Applications/namd ( ApoA1 example):
16 CPUs 0.105658 s/step 1.2229 days/ns 50.3477 MB memory (NAMD V2.7)
16 CPUs 0.103924 s/step 1.20282 days/ns 225.961 MB memory(NAMD V2.9)
comparing these i get a sense that with increasing version of NAMD , memory requirement grows, & maybe s/step (500 nsteps in my case , anl.gov <http://anl.gov> nsteps =? ) are better on my benchmark,
but i am confused with days/ns , so if anyone can explain the significance of days/ns it would be very kind of you. Further please mention that the higher value is better or lower value is better wrt performance ?
Eagerly awaiting your replies,
This archive was generated by hypermail 2.1.6 : Tue Dec 27 2016 - 23:21:25 CST