Impact of Restart Write on Wall Time

[mkwunder]

I’m doing a bit of an experiment with FDS. I am trying to see how much of an impact writing an FDS restart file has on wall time.

I am running a variety of runs, starting with no restarts, 2 restarts (every 2 seconds), 4 restarts (every 5 seconds.), and so on until 64 restarts (every 0.3125 seconds). [like a defacto strong scaling study]

However, my initial results show that runs with minimal restarts, say every 2.5 or every 1.25 seconds are significantly slower compared to the runs where a restart occurs every .6 or .3 seconds.

This could be caused by a particularly slow node I am using; however, I wanted to check if FDS even registers restart file writes at such short intervals. Perhaps because the interval is so tiny, it never writes a restart?

I can share more information if needed, and would greatly appreciate if anyone has any information they could share.

[drjfloyd]

If your timestep is greater than DT_RESTART you will skip some writes. i.e if DT_RESTART=1 and your timestep is 2 your writes will be at 2,4,6.

[mkwunder]

Interesting, I did not know that; although, my step size appears to be smaller than my dt_restart (adaptive step size being in the neighborhood of 0.379E-03).

Is there anyway to confirm how many restart files are written? If they are written?

[mcgratta]

RESTART files over-write each other. There is no counter for the number of writes.