Autotuning Wavefront Applications for Multicore Multi-GPU Hybrid Architectures
Siddharth Mohanty, Murray Cole · 2014
Manual tuning of applications for heterogeneous parallel systems is tedious and complex. Optimizations are often not portable, and the whole process must be repeated when moving to a new system, or sometimes even to a different problem size. Pattern-based programming models provide structure which can assist in the creation of autotuners for such problems. We present a machine learning based auto-tuning framework which partitions the work created by applications which follow the wavefront pattern across systems comprising multicore CPUs and multiple GPU accelerators. The use of a pattern facilitates training on synthetically generated instances. Exhaustive search space exploration on real applications indicates that correct setting of the tuning factors leads to a maximum of 20x speedup over an optimized sequential baseline, with an average of 7.8x. Our machine learned heuristics obtain 98% of this speed-up, averaged across range of applications and architectures.