Doctor of Philosophy (Ph.D.)
As computing enters an era of heterogeneity and massive parallelism, it exhibits a distinct feature: the deepening non-uniform relations among the computing elements in both hardware and software. Besides traditional non-uniform memory accesses, much deeper non-uniformity shows in a processor, runtime, and application, exemplified by the asymmetric cache sharing, memory coalescing, and thread divergences on multicore and many-core processors. Being oblivious to the non-uniformity, current applications fail to tap into the full potential of modern computing devices.;My research presents a systematic exploration into the emerging property. It examines the existence of such a property in modern computing, its influence on computing efficiency, and the challenges for establishing a non-uniformity--aware paradigm. I propose several techniques to translate the property into efficiency, including data reorganization to eliminate non-coalesced accesses, asynchronous data transformations for locality enhancement and a controllable scheduling for exploiting non-uniformity among thread blocks. The experiments show much promise of these techniques in maximizing computing throughput, especially for programs with complex data access patterns.
© The Author
Wu, Bo, "Matching non-uniformity for program optimizations on heterogeneous many-core systems" (2014). Dissertations, Theses, and Masters Projects. Paper 1539624006.