Semi-Automatic Performance Optimization of HPC Kernels

Steven Masnada 1, 2
1 CORSE - Compiler Optimization and Run-time Systems
Inria Grenoble - Rhône-Alpes, LIG - Laboratoire d'Informatique de Grenoble
2 POLARIS - Performance analysis and optimization of LARge Infrastructures and Systems
Inria Grenoble - Rhône-Alpes, LIG - Laboratoire d'Informatique de Grenoble
Abstract : High Performance Computing platforms are made to answer the need of huge computing power, however, taking advantage of their power is difficult as they are complex machines and each platform has a unique set of characteristics. Thus, the developer must program them with care and write specialized code. Tools exist to help the developer in this tricky task to generate optimized versions of an appli- cation. Finding high performing versions is the main concern because the search space can be huge (e.g GCC has about 500 compilation flags) and an exhaustive search is prohibitive. Hence, auto-tuning considers this as a mathematical opti- mization problem. To the best of our knowledge most auto-tuning frameworks mostly resort to generic optimization techniques combined to fully automatic ex- plorations. However, this approach excludes the user from the optimization pro- cess. Hence, it is difficult to know if further improvement can be made and the quality evaluation of the solution is complicated. To answer this problem we pro- pose a semi-automatic approach that gives power back to the user. This approach is based on linear regression techniques to predict the computation kernel perfor- mances. More precisely we used both least square regression and quantile regres- sion. It is also combined to techniques inspired from design of experiments which aim to reduce the experimental cost. We evaluated our approach using the case of a Laplacian kernel and compared it with other classical optimization techniques used in the auto-tuning literature. Our method gives very good results by finding almost every time near-optimal solutions. We provide an in depth analysis of the reason why our approach is much more effective than previously proposed one.
Complete list of metadatas

Cited literature [17 references]  Display  Hide  Download

https://hal.inria.fr/hal-01579422
Contributor : Steven Quinito Masnada <>
Submitted on : Thursday, August 31, 2017 - 10:30:29 AM
Last modification on : Friday, October 25, 2019 - 2:04:48 AM
Long-term archiving on : Friday, December 1, 2017 - 5:43:00 PM

Identifiers

  • HAL Id : hal-01579422, version 1

Citation

Steven Masnada. Semi-Automatic Performance Optimization of HPC Kernels. Performance [cs.PF]. 2016. ⟨hal-01579422⟩

Share

Metrics

Record views

225

Files downloads

107