Optimizing scheduling for stream structured programming for StreamIt

Abstract

As straightforward increases in performance on general purpose CPUs slow down, the shift to application specific implementations and hardware has accelerated. This shift to towards specialization improves performance but often at the cost of developer productivity in learning these new tools. StreamIt is a Domain Specific Language developed to increase performance of streaming applications while being relatively user-friendly. While designed to be parallelized easily, the scheduling backend of the StreamIt compiler is not adapted to the heterogeneous and distributed nature of new accelerator hardware. This thesis details the design and development of a scheduler interface that enables hardware customized schedulers to be developed quickly. The scheduler interface allows for schedulers to take advantage of the unique compiler optimizations enabled by StreamIt’s structure. Two schedulers, one search based and another heuristic based, are built using this interface to schedule StreamIt workloads to optimize differing metrics such as throughput and latency. Our experiments evaluate the performance of these workloads, and details future direction for expanding the interface and scheduler designs that could take advantage of it.