DiscoPoP Empirical Autotuner

Executable

discopop_auto_tuner

Purpose

Identify the best configuration of parallel code which is achievable by applying a combination of the parallelization suggestion found by the DiscoPoP Explorer or optimizer. Internally, the DiscoPoP patch applicator is used to create different parallel codes, which will be compiled, executed, and evaluated based on the observed execution time and result validity. To find a beneficial configuration in a comparatively short amount of time, a greedy search is performed as described in the following.

• Initialize the “best_configuration” as the sequential code
• Focus on identified hotspots, followed by code regions which might be hotspots, and lastly coldspots
• Sort the loops in each category in descending order by their average runtime, thus focusing on “more important” loops wrt. execution time first
• For each loop, create a parallel version of the code for each suggestion applicable to this loop
• Compile the parallel code, execute, and verify it.
• Select the best parallel code and and save it as the new best_configuration
• Continue with the next loop, and successively hotspot category

Required input

• Parallel patterns in the form of a JSON file, created by the Explorer
• Detected hotspots in the form of a JSON file, created by the Hotspot detection
• Prepared patch files created by the DiscoPoP patch generator
• DP_COMPILE.sh, DP_EXECUTE.sh, and optionally DP_VALIDATE.sh scripts to compile, execute and validate the results of the created parallel code.

Output

The parallel code representing the identified best configuration will be stored in a copy of the project directory. Information on the configuration can be found in the included .discopop directory.

Limitations

Due to the empirical nature of the optimization approach described above, varying sets of input data might yield differing selected configurations. For this reason it is important, that the used input data (typically specified in DP_EXECUTE.sh) is representative for a production run of the software and large enough to allow for beneficial parallelizations.

Note

For a more detailed description of the available run-time arguments, please refer to the help string of the respective tool.

discopop_auto_tuner --help