PAT calibration project - performance and parallelization
Hi everyone,
I have been trying to use PAT to calibrate a building model to measured data. The building model was developed using project specifications, i have measured data for almost a full year, with 10 min granularity (timestep) for rooms temperature and HVAC energy consumption. I also have a AMY weather file for that specific location.
I am currently running PAT in Algorithmic mode using OpenStudio-server on AWS, with a node type "t2.large" and max of 4 nodes.
My problem is the run time per each simulation. Running the model locally using EnergyPlus takes round 30s while in PAT is taking almost 3 minutes.
I have done some testing and believe that the problem is the reporting measure "Time Series Objective Function" that i am using to calculate calibration metrics (CVRMSE). This measure is being used per each room with temperature readings (8 rooms) plus for total HVAC energy consumption.
I would like to know if there is a way to improve simulation time and also if it is possible to run simulations in parallel.
Thanks!
Sorry to be brief as I am traveling, but all the algorithms run in parallel. My recommendation is to choose a better instance type, with more cores, etc.
Thank you for your answer and sorry for the late reply (was on vacations). I have manage to run simulations in parallel by changing the eksctl cluster settings, increasing both the number of nodes and minimum nodes (ex. --nodes 2 \ --nodes-min 2).
Regarding the reporting measure the problem still exists, since is increasing a lot the simulation time. I have also noticed that having more run periods (i have data with some missing days) also was a huge impact. Could this be related to the SQL querys being repeated for every run period?
Also regarding the measure, it has a lot of cool features (like the graphic representation), but could there be a way of just reporting the CVRMSE with less computation power required? Thank you!
ah, its coming back to me. sorry its been 6 years. there is an argument 'verbose_messages'. This is very useful to get the output of the measure for debugging and getting the sql query right, but its a major performance hit. make sure that is set to FALSE when doing production runs. i think there's a few other arguments that are not needed for production runs, like 'find_avail' which you can also set to FALSE.
Thanks again for your answer. I was already setting all these arguments to FALSE, but the reporting measures are still taking around 40% of the total simulation time. Is there something i am missing? Another question i have is related to parallelization. Is it possible to have several simulations runs per each node? I have managed to increase the number of parallel simulations to 6 by using the following setting of openstudio-server (set worker_hpa.minReplicas=6), but this is just adding 6 instances (nodes) that have a CPU utilization of arround 50%.
sounds like you are doing the right things. If speed is important, then make sure you are only reporting the variables you need and at the reporting frequency you need in the model. I would also suggest posting an issue about it here and include the datapoint logs and ideally a reproducible example.