autosubmit.profiler

Important

If you only want to use the profiler built into the Autosubmit commands, simpler user-oriented guides are available for run, create and monitor.

The Autosubmit’s profiler

Autosubmit integrates a profiler that allows developers to easily measure the performance of entire functions or specific code fragments.

The profiler generates a comprehensive report with enough information to detect bottleneck functions during the execution of experiments, as well as information about the total memory consumed.

It mainly uses the cProfile library to make the report, so this module inherits its deterministic profiling and reasonable overhead features. However, it also limits profiling to a single thread, so please, do not use it on concurrent code. For memory profiling, it uses psutil.

Caution

This profiler was originally designed to be used in the autosubmit run command, so using it in other functions may produce unexpected results or errors. Now, its usage have been extended to autosubmit create and monitor.

The profiler instantiation requires an EXPID, and not all the functions in Autosubmit use it. This can be bypassed using another string, but keep in mind that there is no error handling in this case.

How to profile a function or a specific code fragment?

Depending on the Autosubmit function you want to profile, you must add a --profile argument to the parser. The autosubmit run, create and monitor subcommands already support it. It is recommended that the default value of this flag always be False, to ensure that the profiler does not interfere with the normal execution in an unwanted way. You will need to add something like this to your parser:

subparser.add_argument('-p', '--profile', action='store_true', default=False, required=False,
help='Prints performance parameters of the execution of this experiment.')

The function must receive the flag as argument to control the execution of the profiler. If the flag has value True, you should proceed as follows:

1. Instantiate a Profiler(EXPID) object. Specifying the EXPID is mandatory.

2. Run the profiler by calling the start() function of the instantiated object, at the beginning of the function or code fragment you want to evaluate. The measurement will start instantly.

3. Execute the stop() function of the profiler at the end of the function or code fragment to be evaluated. The process of taking measurements will stop instantly. The report will be generated automatically and the files will be stored in the <EXPID>/tmp/profile directory.

Important

Make sure, if necessary, that the call to stop() is always made, even if the Autosubmit code fails, in order to get the performance report.

The most relevant functions of the profiler, in detail

• The start() function: Starts taking measures, both of execution times thanks to cProfile, and memory thanks to psutil. It also manages errors to avoid illegal transitions between states.

• The stop() function: Same as the previous function, but terminating the taking of measurements. It will call the report function automatically.

• The _report() function: It is private, and its purpose is to generate the final performance report and storing it properly. It will print the report to the console output and log it at the same time. In addition, it will generate two files in the directory chosen when instantiating the Profiler object, a .txt file with the same report shown on screen, and a .prof file with the report generated by pstats. The .prof file can be manipulated with the appropriate tools. Our recommendation is to open it with SnakeViz, a graphical library that will interpret the data for you and display it in an interactive web interface.