Switchable Scheduling for Runtime Adaptation of Optimization

International audienceParallel applications used to be executed alone until their termination on partitions of supercomputers: a very static environment for very static applications. The recent shift to multicore architectures for desktop and embedded systems as well as the emergence of cloud computing is raising the problem of the impact of the execution context on performance. The number of criteria to take into account for that pur-pose is significant: architecture, system, workload, dynamic parameters, etc. Finding the best optimization for every context at compile time is clearly out of reach. Dynamic optimization is the natural solution, but it is often costly in execution time and may offset the optimization it is en-abling. In this paper, we present a static-dynamic compiler optimization technique that generates loop-based programs with dynamic auto-tuning capabilities with very low overhead. Our strategy introduces switchable scheduling, a family of program transformations that allows to switch between optimized versions while always processing useful computation. We present both the technique to generate self-adaptive programs based on switchable scheduling and experimental evidence of their ability to sustain high-performance in a dynamic environment