Currently I am working on the model development, debugging and calibration of the ECO3D model.
Due to the model complexity, I have spent great efforts to get things done in an appropriate way. Throughout the process, I have also acquired much experiences in model developments. Here I want to share some of the most useful tips. However, I will not discuss the technical aspects such as program language unless inevitable. You may use OpenMP to speed up the program, but it appears that the debugger, such as the one I use, Totalview, is not quite friendly when debugging in multiple-thread. So I suggest you can prepare two versions of CMakeList files, so you can switch between for different purposes; Use conditional breakpoint in Totalview can save you some time when you are only interested in certain point/condition; Ecosystem models, or Earth system models, require a “spin-up” simulation, which usually takes a great length of time ranging from hours to days. So you don’t want to do it every time when you test your model. To save time as well as HPC resources, we can save the steady state system state variables into files so we can restart from them to avoid repeated running; If your simulation objects are independent with each other under certain assumption, for example, the shortwave radiation of one pixel is independent with nearby pixels, you can speed up the debugging by shutting down/mask out the other “useless” pixel. In this way, you can focus on the only pixel that gives you troubles; In other scenarios, such as the cascade/lateral flow simulation, there is no way to test the model without a complete simulation because pixels are communicating with each other. Even though we have to run all the pixels, that doesn’t mean we have to output them, and we can still output single pixel for debug purpose; Some debugging information is useful for us to diagnose the problem. Often we can output (print, cout, etc.) the information directly on the terminal because it is the default I/O. We can also output them to a CSV file, so we can check the time series trend easily using Microsoft Excel. Having a deep understanding of the trend definitely help you understand what is causing the changes; Even on the most powerful HPC which provides a debug queue, the walltime (less than an hour?) is not enough for debugging. So we have to use queue with long walltime to debug. Also, because GUI debug causes a lot of communications, you should use Remote desktop hosted on HPC (ThinLinc, etc.) instead of local X windows. You should run the debug as an interactive job for as long as the debugging requires; Use debug flag control the workflow efficiently, so you don’t have to recompile your program every time you change something; Before you draw any conclusion about the algorithm, double check the input file with care. Lots of time we made assumptions about input data that are not true. Sometimes you can manually modify/create some input for testing purpose; Be consistent with some logical controls, if you change some important control, remember to check all the related ones. If some variables are used across modules, you should be even more careful; Keep more than one copy of the workspace with the same file system structure to speed up some checking processes. Both ftp/sftp and GUI tools can be used for data transfer. If you have to transfer large amount of data, consider Globus. Let me know if you have any questions.