Variating calculation time

Submitted by OlegB on Tue, 11/24/2020 - 12:38

I'm running the same calculation (particles dropped in a box) on the same PC (4 core i7) using 4 threads in openmpi. Results are as expected the same, but calculation time varies drastically (up to 30%).
Commands are supplied from the same input file.
I'm using LIGGGHTS-PUBLIC 3.8.0

Anyone can explain why that could be?

Here are times with details from 4 different runs:
4089.4054
Pair time (%) = 2988.26 (73.0732)
Neigh time (%) = 50.9424 (1.24572)
Comm time (%) = 827.005 (20.2231)
Outpt time (%) = 11.3318 (0.277102)
Other time (%) = 211.866 (5.18085)

2829.6907
Pair time (%) = 2300.63 (81.3032)
Neigh time (%) = 43.5203 (1.53799)
Comm time (%) = 194.692 (6.88031)
Outpt time (%) = 7.83396 (0.276849)
Other time (%) = 283.016 (10.0017)

3401.7611
Pair time (%) = 2598.52 (76.3875)
Neigh time (%) = 47.2968 (1.39036)
Comm time (%) = 489.85 (14.3999)
Outpt time (%) = 7.9984 (0.235125)
Other time (%) = 258.095 (7.58711)

3510.7698
Pair time (%) = 2658.71 (75.7302)
Neigh time (%) = 47.0981 (1.34153)
Comm time (%) = 542.905 (15.464)
Outpt time (%) = 7.99947 (0.227855)
Other time (%) = 254.055 (7.23645)

AttachmentSize
Plain text icon in.gravitypacking.txt1.66 KB

OlegB | Tue, 11/24/2020 - 23:54

Yes, exactly. I run the same script and the execution time differs so much.
Is it possible the potential neighbors search mechanism can get different amount of atoms, so then more pairs are checked for collisions? Or some other reason that calculations are not exactly the same?
Thank you for any ideas to consider or possible checks to run to find the cause of that behavior.

richti83's picture

richti83 | Fri, 11/27/2020 - 22:11

This interesting fact has been carried out by my colleagues at engineerdo.com.
From a numerical point of view the same input script with the same processor distribution results in exactly the same neighbour and particle data, BUT the background tasks on GUI enabled machine makes a big difference in computational time.
See this blog post: https://www.engineerdo.com/2020/11/13/computing-on-a-budget/ at point "Avoid Background Tasks!".
From my personal experience the computanial time is nearly constant at profesionel managed HPC systems but vary on Linux Workstations and Windows installations.
Maybe cooling is another issue, as modern cpus throttles clock frequency when becoming to hot.
And as we know, clock frequency rules all :-).
Best,
Christian.

I'm not an associate of DCS GmbH and not a core developer of LIGGGHTS®
ResearchGate | Contact