how can find out computational requirements of EnergyPlus Run?

asked 2015-03-25 15:46:44 -0500

Ralph Muehleisen
501 ●8 http://www.anl.gov/

updated 2015-07-11 09:44:11 -0500

I need to give some rough "computational effort" numbers for looking at very detailed models for urban regions. What I need to estimate is how much CPU power is needed to model something like every building in a large urban area (New York, Chicago, LA, etc) at a very high time resolution (maybe even using the next generation energy modeling engine and going beyond EnergyPlus).

I can do some simulations of prototypical buildings and then use multipliers to get computational efforts. But, what I need to figure out is the number of floating point and integer operations needed to run a model.

Is there an output in some report in EnergyPlus or OpenStudio that gives the number of flops/IOPS? I can't find one.

Thanks,

edit retag flag offensive close merge delete

add a comment

2 Answers

Sort by » oldest newest most voted

answered 2015-03-27 20:32:00 -0500

Ralph Muehleisen
501 ●8 http://www.anl.gov/

For the rest of the world, I did some experiments and analysis of other data to get some rough estimates. On my laptop, a small building runs in just under a minute and a larger building in closer to six minutes. A benchmarking program run with the same background processes yields a speed of 60 GFlops.

Figuring out the total number of floating point operations and then working back how many time steps, I get that a small building is, to rough order, about 1 kFlop and a larger building is about 10 kFlop.

I also got similar numbers working back from some results from ORNL where they ran 525k simulations on 130k CPU in about 68 min on Titan and knowing about the CPUs that were used (Opteron 6274s)

Now I can get some estimates of how much CPU power is required to do coupled building/transportation/grid simulations in large urban areas. Working from information about the building stock in Chicago, it takes about 10MFlop to simulate a time step of a 1 mi^2 area. The traffic system only takes about 100 kFlop per time step to simulate the same area. I haven't got estimates for a detailed grid simulation but I'm a bit scared that it will be several orders of magnitude higher computational requirements.

edit flag offensive delete link

add a comment

answered 2015-03-25 16:51:30 -0500

__AmirRoth__

4436 ●5 ●17 http://bleedinggreenna...

Why do you need FLOPS/IOPS as opposed to just wall time? Are you running in a batch environment that matches you to machines based on this information, e.g., Condor?

Either way, I don't think this has been done yet but the most straightforward way to do it would be to write a wrapper for EnergyPlus that resets and then reads the appropriate performance counters. I could probably spin this up in a few hours. Should I? How much will you pay me? ;)