SC17 Denver, CO

P76: A Compiler Agnostic and Architecture Aware Predictive Modeling Framework for Kernels


Authors: William Killian (University of Delaware), Ian Karlin (Lawrence Livermore National Laboratory), David Beckingsale (Lawrence Livermore National Laboratory), John Cavazos (University of Delaware)

Abstract: Multi-architecture machines make program characterization for modeling a regression outcome difficult. Determining where to offload compute-dense portions requires accurate prediction models for multiple architectures. To productively achieve portable performance across these diverse architectures, users are adopting portable programming models such as OpenMP and RAJA.

When adopted, portable models make traditional high-level source code analysis inadequate for program characterization. In this poster, we introduce a common microarchitecture instruction format (ComIL) for program characterization. ComIL is capable of representing programs in an architecture-aware compiler-agnostic manner. We evaluate feature extraction with ComIL by constructing multiple regression-objective models for performance (execution time) and correctness (maximum absolute error). These models perform better than the current state of the art -- achieving a mean error rate of only 4.7% when predicting execution time. We plan to extend this work to handle multiple architectures concurrently and evaluate with more representative physics kernels.

Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf


Poster Index