Generative Kernel Optimization (GKO)

GKO or Generative Kernel Optimization is a set of methods and tools for optimizing programming kernels to achieve high performance with a multi-stage method that includes performance evaluation, modeling, and automated optimization. For more details on the algorithmic methods used in GKO see below for a brief description and further reading sections.

Surrogate Performance Modeling

Based on the performance measurements from the past runs, GKO builds a performance surrogate model utilizing variants of Gaussian Process Regression. The specific configurations for initial runs are based on advanced sampling techniques that minimize the time spent in bootstrapping the surrogate. This may sometimes become a multi-stage processes that iteratively samples the performance configuration space, builds a model, and subsequently evaluates the resulting uncertainty to see if the stages have to be repeated to improve the precision of predictions before moving to the optimization stage. Many of the stages need advanced methods because of the high dimensionality and non-linear nature of the optimization and loss hyper-surfaces.

Installation

Please use the supplied requirements.txt to install the required packages from PyPI. After dependent packages were installed, the installation can be finished with setuptools using the provided pyproject.toml file.

Basic Usage

The simple invocation of the main entry point into the GKO methods relies on the proper settings residing in the configuration files:

python3 -m gko command [options]

python3 -m gko sample --level 5 --fidelity 0,12 --sample-count 25 --graph sample_graph.json --options sampleopts.json --infile old_samplesnorm.json --outfile new_samplesnorm.json

python3 -m gko convert --direction norm2orig --infile samplesnorm.json --outfile samplesorig.json

python3 -m gko model --samples samplesnorm.json --perf perf.json --options var/modopts.json --outfile model.pickle

python3 -m gko model --level 0:3 --fidelity 0 --fidelity 1 --fidelity 2 --samples samplesnorm.json --perf perf.json --graph model_graph.json --options var/modopts.json --infile old_model.pickle --outfile new_model.pickle

python3 -m gko model --samples samplesnorm.json --perf perfnorm.json --options var/modopts.json --outfile model.pickle

python3 -m gko predict --level 2 --fidelity 1 --m model.pickle --options var/modopts.json --param1 384 --param2 4 --param3 6 --param4 1

python3 -m gko search --level 2 --fidelity 0 --m model.pickle --model-opts var/modopts.json --search-opts searchopts.json --outfile candidnorm.json

python3 -m gko plot --level 2 --fidelity 1 --model model.pickle --options var/modopts.json --samples samplesnorm.json --perf perf.json --candidates candidnorm.json --outfile <image_name_template>

python3 -m gko merge --level 2 --fidelity 1 --samples samplesnorm.json --candidates candidnorm.json --perf perf.json --perf-candid <float,...>

Publications (Further Reading)

The major methods used by GKO include sample-efficient algorithms for sampling the configuration space that may have tens of dimensions, which makes basic methods inefficient in both time and exploration efficiency. The samples constitute a training set for building a surrogate performance model that uses versions of Gaussian Process Regression catered for the modeling circumstances. Subsequently, the model used in the optimization stage to find optimal configuration of parameters that can be used for obtaining high performance for the sampled system and the application kernel combination.

@InProceedings{sidlakhdar2022dgp,
author={Sid-Lakhdar, Wissam M.  and Aznaveh, Mohsen and Luszczek, Piotr and Dongarra, Jack},
title={Deep {Gaussian} process with multitask and transfer learning for performance optimization},
booktitle={2022 IEEE High Performance Extreme Computing Conference (HPEC)},
year={2022},
publisher={IEEE},
pages={1-7},
isbn={1665497866}
}

@InProceedings{luszczek2022sabath,
author={Luszczek, Piotr and Brown, Cade},
title={Surrogate {ML/AI} Model Benchmarking for {FAIR} Principles' Conformance},
booktitle={2022 IEEE High Performance Extreme Computing Conference (HPEC)},
year={2022},
publisher={IEEE},
pages={1-5},
isbn={1665497866}
}

@article{luszczek2023dgpautotune
,author={Luszczek, Piotr and Sid-Lakhdar, Wissam M. and Dongarra, Jack}
,title={Combining multitask and transfer learning with deep {Gaussian} processes for autotuning-based performance engineering}
,journal={The International Journal of High Performance Computing Applications}
,year=2023
,note={DOI:10.1177/10943420231166365}
}