Skip to content

Latest commit

 

History

History
196 lines (158 loc) · 10.5 KB

README.md

File metadata and controls

196 lines (158 loc) · 10.5 KB

What is ytopt?

ytopt is a machine-learning-based search software package that consists of sampling a small number of input parameter configurations, evaluating them, and progressively fitting a surrogate model over the input-output space until exhausting the user-defined time or the maximum number of evaluations. The package is built based on Bayesian Optimization that solves any optimization problem and is especially useful when the objective function is difficult to evaluate. It provides an interface that deals with unconstrained and constrained problems. The software is designed to operate in the manager-worker computational paradigm, where one manager node fits the surrogate model and generates promising input configurations and worker nodes perform the computationally expensive evaluations and return the outputs to the manager node. The asynchronous aspect of the search allows the search to avoid waiting for all the evaluation results before proceeding to the next iteration. As soon as an evaluation is finished, the data is used to retrain the surrogate model, which is then used to bias the search toward the promising configurations.

Directory structure

docs/	
    Sphinx documentation files
test/
    scipts for running benchmark problems in the problems directory
ytopt/	
    scripts that contain the search implementations  
ytopt/benchmark/	
    a set of problems the user can use to compare our different search algorithms or as examples to build their own problems

Install instructions

The autotuning framework requires the following components: ConfigSpace, CConfigSpace (optional), scikit-optimize, autotune, and ytopt.

  • We recommend creating isolated Python environments on your local machine using conda, for example:
conda create --name ytune python=3.7
conda activate ytune
  • Create a directory for ytopt tutorial as follows:
mkdir ytopt
cd ytopt

git clone https://github.com/ytopt-team/ConfigSpace.git
cd ConfigSpace
pip install -e .
cd ..

git clone https://github.com/ytopt-team/scikit-optimize.git
cd scikit-optimize
pip install -e .
cd ..

git clone -b version1 https://github.com/ytopt-team/autotune.git
cd autotune
pip install -e . 
cd ..

git clone -b main https://github.com/ytopt-team/ytopt.git
cd ytopt
pip install -e .
  • If you encounter installtion error, install psutil, setproctitle, mpich, mpi4py first as follows:
conda install -c conda-forge psutil
conda install -c conda-forge setproctitle
conda install -c conda-forge mpich
conda install -c conda-forge mpi4py
pip install -e .

  • [Optinal] Install CConfigSpace:

    • Prerequisites: autotools and the gsl

      • Ubuntu

        sudo apt-get install autoconf automake libtool libgsl-dev

      • MacOS

        brew install autoconf automake libtool gsl
    • Build and Install the library and python bindings: the configure command can take an optional --prefix= parameter to specify a different install path than the default one (/usr/local). Depending on the chosen location you may need elevated previleges to run make install. 
      git clone git@github.com:argonne-lcf/CCS.git
cd CCS
./autogen.sh
mkdir build
cd build
../configure
make
make install
cd ../bindings/python
pip install parglare==0.12.0
pip install -e .
    • Setup environment: in order for the python binding to find the CConfigSpace library, the path to the library install location (/usr/local/lib by default) must be appended to the LD_LIBRARY_PATH environment variable on Linux, while on MacOS the DYLD_LIBRARY_PATH environment variable serves the same purpose. Alternatively the LIBCCONFIGSPACE_SO_ environment variable can be made to point to the installed libcconfigspace.so file on Linux or to the installed libcconfigspace.dylib on MacOS.

  • [Optinal] Install Online tuning:

    • Online tuning with transfer learning interface is built on Synthetic Data Vault (SDV):
    • Install SDV: 
      cd ytopt
pip install -e .[online]

Tutorials

Who is responsible?

The core ytopt team is at Argonne National Laboratory:

The convolution-2d tutorial (source and python scripts) is contributed by:

Publications

  • J. Koo, P. Balaprakash, M. Kruse, X. Wu, P. Hovland, and M. Hall, "Customized Monte Carlo Tree Search for LLVM/Polly's Composable Loop Optimization Transformations," in Proceedings of 12th IEEE International Workshop on Performance Modeling, Benchmarking and Simulation of High Performance Computer Systems (PMBS21), pages 82–93, 2021. DOI: 10.1109/PMBS54543.2021.00015
  • X. Wu, M. Kruse, P. Balaprakash, H. Finkel, P. Hovland, V. Taylor, and M. Hall, "Autotuning PolyBench benchmarks with LLVM Clang/Polly loop optimization pragmas using Bayesian optimization (extended version)," Concurrency and Computation. Practice and Experience, vol. 11, 2021. ISSN 1532-0626 DOI: 10.1002/cpe.6683
  • X. Wu, M. Kruse, P. Balaprakash, H. Finkel, P. Hovland, V. Taylor, and M. Hall, "Autotuning PolyBench Benchmarks with LLVM Clang/Polly Loop Optimization Pragmas Using Bayesian Optimization," in Proceedings of 11th IEEE International Workshop on Performance Modeling, Benchmarking and Simulation of High Performance Computer Systems (PMBS20), pages 61–70, 2020. DOI: 10.1109/PMBS51919.2020.00012
  • P. Balaprakash, J. Dongarra, T. Gamblin, M. Hall, J. K. Hollingsworth, B. Norris, and R. Vuduc, "Autotuning in High-Performance Computing Applications," Proceedings of the IEEE, vol. 106, no. 11, 2018. DOI: 10.1109/JPROC.2018.2841200
  • T. Nelson, A. Rivera, P. Balaprakash, M. Hall, P. Hovland, E. Jessup, and B. Norris, "Generating efficient tensor contractions for GPUs," in Proceedings of 44th International Conference on Parallel Processing, pages 969–978, 2015. DOI: 10.1109/ICPP.2015.106

Acknowledgements

  • YTune: Autotuning Compiler Technology for Cross-Architecture Transformation and Code Generation, U.S. Department of Energy Exascale Computing Project (2017--Present)
  • Scalable Data-Efficient Learning for Scientific Domains, U.S. Department of Energy 2018 Early Career Award funded by the Advanced Scientific Computing Research program within the DOE Office of Science (2018--Present)
  • PROTEAS-TUNE, U.S. Department of Energy ASCR Exascale Computing Project (2018--Present)

Copyright and license

TBD