Provably-efficient adaptive scheduling of multiprocessors for performance and energy

Abstract

Performance and energy are two crucial but conflicting objectives in the design of modern high-performance computing systems. This thesis considers efficient scheduling of parallel applications on multiprocessors to optimize performance and energy consumption. We mainly focus on a two-level framework and dynamic speed scaling with minimal or no assumption about the characteristics of the jobs, such as their release time, remaining work and parallelism profile. Both processor allocations and processor speeds (if possible) are periodically readjusted based on the estimates of a job's parallelism and the scheduling policy of the operating system.