From: Ajasja Ljubetič (ajasja.ljubetic_at_gmail.com)
Date: Wed Dec 07 2016 - 16:59:45 CST
Hi!
Probably if you increase the number of steps, the difference will decrease.
NAMD has some startup cost (parsing all the files etc...). What was the
absolute wall time? If it was less than about 100 s I would not trust the
results.
Best,
Ajasja
On 7 December 2016 at 23:02, Jan Fredin <jfredin_at_sgi.com> wrote:
> Hi,
>
> I am running the ApoA1 benchmark using 10,000 time steps and
> outputtiming=100 with all the other input from the standard benchmark
> apoa1.namd file. I’m running on a single node of Xeon Haswell with P100
> GPUs. I am interested in the ns/day and thought 10,000 time steps would be
> enough to have the WallClock time be usable in the ns/day calculation. I
> found that taking the reciprocal of the average of the 6 Benchmark time
> lines day/ns gave 26.3 ns/day but the straight calculation of ns simulated
> (0.01 ns) and WallClock time (XX sec / 86400 sec/day) gave only 14.3
> ns/day.
>
>
>
> What work contributing to the WallClock time is not being included in the
> Benchmark time lines?
>
> Under what circumstances do the 2 methods of calculating ns/day converge
> to the same value?
>
>
>
> Thanks,
>
> Jan
>
> --
> Dr. Jan Fredin
> SGI, Sr. Applications Engineer - Lead
> Austin, TX
> 651-683-7708 <(651)%20683-7708>
> jfredin_at_sgi.com
>
>
>
This archive was generated by hypermail 2.1.6 : Tue Dec 27 2016 - 23:22:41 CST