Skip to content

pbbs benchmarks for OpenCL runtime for CPU parallel-for partitioner implementation in Eigen runtime

License

Notifications You must be signed in to change notification settings

blonded04/composable-parallel-scheduler-pbbsbench

 
 

Repository files navigation

pbbsbench

The Problem Based Benchmark Suite (PBBS) is a collection of over 20 benchmarks defined in terms of their IO characteristics. They are designed to make it possible to compare different algorithms, or implementations in different programming languages. The emphasis is on parallel algorithms. Our default implementations are for shared-memory multicore machines, although we would be delighted if others implemented versions on GPUs or distributed machines and made comparisons.

A list of the benchmarks can be found on here.

Information on the organization and on how to run PBBS can be found on here.

And here is a paper outlining the benchmarks:

The problem-based benchmark suite (PBBS), V2
Daniel Anderson, Guy E. Blelloch, Laxman Dhulipala, Magdalen Dobson, and Yihan Sun
ACM SIGPLAN Symposium on Principles & Practice of Parallel Programming (PPoPP), 2022

How to run

It is proposed to use run_benches.py script.

  • Currently 3 backends are supported: OMP, TBB and Eigen
    • OMP and Eigen are used similarly: provide --omp and/or --eigen options respectively
    • To use TBB, you'll need to have it built somewhere. Provide --tbb-path path/to/tbb and --tbb flags. It is expected that path/to/tbb has include/ directory on the top level and .so somewhere within
  • Sometimes checks work weirdly (and always long) so you might want to add --nocheck flag
  • If you don't need the full measurement, use --small flag, then input sizes will be reduced. (Though, few benches will not work properly with small inputs)
  • Script will generate a directory with an ugly name for logs for each backend. If you want to provide your own directory instead, add --dir logs/dir
    • There is a separate script that turns logs into .jsons

Example:

./run_benches.py --tbb-path=$HOME/local --tbb --dir tmp_res/ --small
for file in tmp_res/*.txt ; do cat $file | ./gather_stats.py > $file.json ; done

After that in graphs.ipynb provide your logs directory (tmp_res/ in example) and run all the cells. There are a few things you might need to tweak:

  • filter_exec function: it's responsible for choosing backends which results will be displayed
  • target_execs will highlight results for provided backends
  • base_exec is a baseline backend. It also has to deal with highlighting
  • curr_target will filter out all graphs where curr_target backend is absent

About

pbbs benchmarks for OpenCL runtime for CPU parallel-for partitioner implementation in Eigen runtime

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C++ 68.3%
  • Jupyter Notebook 12.6%
  • C 12.1%
  • Python 2.3%
  • HTML 1.3%
  • Shell 1.3%
  • Other 2.1%