DC OPTIMIZATION UTILITIES

Container for various utility files to be used for the "Forward Tracking code speed optimization" service works.

Setup.

Using these utilities requires a coatjava installation. Refer to its repository for installation instructions.

You are encouraged to add /path/to/coatjava/bin to your PATH variable for ease of access.

The files to be used for benchmarking are:

• sidis_vz0.hipo: gemc simulation with 10.000 events for quick comparison. Stored at /work/clas12/users/devita/clas12validation/infiles/sidis/sidis_vz0.hipo.
• clas_005038.1231.hipo: data file with 70.000 events for serious resolution tests. Stored at /work/clas12/users/devita/clas12validation/infiles/clas_005038.1231.hipo.

Profiling.

To profile code and find HotSpots use the visualvm tool installed in the benchmarking computer on the data file. Remember to not run anything else while running reconstruction in this machine to not spoil results. A screenshot of profiling results at different dates is included in the prof_results directory for future reference.

Speed Testing.

To compare the code speed of different versions simply run reconstruction in the benchmarking computer on the data file. At the end of reconstruction a solid summary of per-engine running time is printed, which is good enough for comparison.

The Plan.

Optimization work consists of three fields. These are detailed in the following subsections.

Java versions profiling and comparison

• Adapt branch of offline reconstruction software to work with OpenJDK 17.
• Make sure that reconstruction results are the same using compareBanks.groovy.
• Profile and compare speed from both branches.
• See if there's any obvious improvement to be found by using new methods from Java 17.

Refactoring and optimization of DC tracking

• Refactor and optimize HotSpot methods:
  • cnuphys.magfield.FieldProbe.contains() (23.2%)
  • org.jlab.rec.dc.track.RungeKuttaDoca.RK4transport() (21.4%)
  • cnuphys.magield.RotatedCompositeProbe.field() (11.5%)
  • org.jlab.rec.dc.cluster.ClusterCleanerUtilities.ClusterSplitter() ( 6.9%)
  • org.jlab.rec.dc.cluster.ClusterCleanerUtilities.LRAmbiguityResolver() ( 6.3%)
  • cnuphys.magield.CompositeProbe.field() ( 5.8%)
• Make sure that reconstruction results are the same via compareBanks.groovy.
• Refactor the whole DC codebase and document.
• See if any further optimization can be done on the new HotSpot methods.

Apply algorithmical improvements to HotSpot methods in DC tracking

• Switch RK4 for AM4.
• Compare reconstruction results via compareBanks.groovy.
• Get a good understanding on how swimming works.
• Decide on how to and if to optimize swimming.
• Compare reconstruction results via compareBanks.groovy.
• Figure out potential improvements to clustering.

Making a Pull Request.

Before making a pull request back to the original repository, remember to delete test files:

• validation/advanced-tests/run-profiling-tests.sh After making the pull request, add them back to the repository.

Notes.

• Garbage collector settings used in CLARA are: -XX:+UnlockExperimentalVMOptions -XX:+EnableJVMCI -XX:+UseJVMCICompiler.