Automated Software Synthesis for Streaming Applications on Embedded Manycore Processors.

摘要

Stream applications are characterized by the requirement to process a virtually infinite sequence of data items. They appear in many areas including communication, networking, multimedia and cryptography. Embedded manycore systems, currently in the range of hundreds of cores, have shown a tremendous potential in achieving high throughput and low power consumption for such applications.;The focus of this dissertation is on automated synthesis of parallel software for stream applications on embedded manycore systems. Automated software synthesis significantly reduces the development and debug time. The vision is to enable seamless and efficient transformation from a higher-order specification of the stream application (e.g., dataflow graph) to parallel software code (e.g., multiple .C files) for a given target manycore system. This automated process involves many steps that are being actively researched, including workload estimation of tasks (actors) in the dataflow graph, allocation of tasks to processors, scheduling of tasks for execution on the processors, binding of processors to physical cores on the chip, binding of communications to physical channels on the chip, generation of the parallel software code, backend code optimization and estimation of throughput.;This dissertation improves on the state-of-the-art by making the following contributions. First, a versatile task allocation algorithm for pipelined execution is proposed that is provably-efficient and can be configured to target platforms with different underlying architectures. Second, a throughput estimation method is introduced that has acceptable accuracy, high scalability with respect to the number of cores, and a high degree of freedom in targeting platforms with different underlying onchip networks. Third, a task scheduling algorithm is proposed, based on iteration overlapping techniques, which explores the tradeoff between throughput and memory requirements for manycore platforms with and without FIFO-based onchip communication channels. Finally, to increase the scalability of application throughput with respect to the number of cores, a malleable dataflow specification model is proposed.