Towards a Framework for Automating the Workflow for Building Machine Learning Based Performance Tuning
MetadataShow full metadata
Recent interest in machine learning-based methods have produced many sophisticated models for performance modeling and optimi:,ation. These models tend to be sensitive to architectural parameters and are most effective when trained on the target platform. Training of these models; however; is a fairly involved process and requires knowledge of statistics and machine learning that the end-users of such models may not possess. This paper presents a framework for automatically generating machine learning-based performance models. Leveraging existing open-source software; we provide a tool-chain that provides automated mechanisms for sample generation; dynamic feature extraction; feature selection; data labeling; validation and model selection. We describe the design of the framework and demonstrate its effectiveness by developing a learning heuristic for register allocation of GPU kernels. The results show the newly created models are accurate and can predict register caps that lead to substantial improvements in execution time without incurring a penalty in power consumption.