Project GitLab Link : https://gitlab.com/hepcedar/rivet/ https://gitlab.com/hepcedar/yoda/
Project Proposal Link : doc
- Profiling of Rivet/Yoda
- Optimization for Performance Improvement
- Thread Safety
- We did the through profiling of the Rivet and YODA
- Vtune, ARMMap, cache grind, Gprof were used for profiling, out of which Vtune was most helpful. the profiling mostly contained the non IO performance of Rivet
- Profiling Methodology and Initial Report is available at https://gitlab.com/Aditya_Addy/rivet-profiling
- Profiling Methodology and the code used can be found in the repo.
- Helped in improving the Rivet non I/O performance by over 60% by removiong the Deep Copy of HepMC Event
- The major CPU sink was coming from a line in
Event.cc
file where a deep copy of event was going on, solving it gave us a major performance gain
- Implemented the loop realignment for Invisible particle and PID optimization for finding charged particle
- To find the invisible particle we were doing the heavy operation first, similarly In finding charge of the particle we hardcoded some of the particle which is very frequent. Both combined gave us a performance boost of around 5%
- Implemented the realigned of the if-else in Histo1D::fill method
- This method is called very frequently for multiweighted event analysis, even for as small as 10 analysis the figure goes to millions, so realigning the xMax, xMin gave us performance improvement of couple of percent
- Implemented the One level Cache in YODA to Improve the BinSearcher efficiency, couple of percent gain in Performance
- A one level cache is added to YODA BinSearcher, this significantly increases the efficiency of BinSearcher in case of non-uniform binning
- [ Almost Done ] Implemented the basics of Thread Safety by making the ProjectionHandler the member of AnalysisHandler and lazy initialization of projection
- ProjectionHandler was global so in Rivet crashes in case of multi threaded program, to solve this issue we have added ProjectionHandler to each Analysishandler and pass it to the base class
ProjectionApplier
of Analysis and Projection, Implementation using raw pointers is done however use of unique pointers would be more efficient.
- Issue Assigned
- Important Commits
- Rivet/master
- Rivet/master
- Rivet/thread-safety
- Rivet/thread-safety
- Yoda/binSearcher-caching
- Yoda/binSearcher-caching
- Merge Request
I want to thank Google, CERN-HSF for creating such oppurtunity, the mentors especially Andy Buckley without whom this project couldn't have came this far, Andy helped me at every step and guided me throughout the program, due to the mentors like Andy this GSoC community is flourshing this much. Lastly I want to thank Holger Schulz for helping me out in the profiling methodology. Overall this was a great learning experience and I wish to contribute to the open source community in future.