Research in Multithreaded Dataflow and Hybrid Architectures

点击翻译按钮获取中文摘要

The overall objective of the research is the investigation of architectural features for cost-effective, mainstream computing based on multithreading. In this vein, this research will focus in two directions; investigation of multithreaded dataflow paradigm, and the adaptation of the results to hybrid dataflow-von Neumann architecture. The dataflow model is ideally suited for multithreaded architecture since it facilitates instruction level context switching and fine-grained parallelism. A clear understanding of the issues in supporting multiple threads in dataflow environment will permit us to adapt them to hybrid dataflow/control flow architectures. Hybrid systems present the most interesting opportunities in the area of multiprocessing - they directly address problems that will be faced by future superscalar processors, such as long memory latencies, context switching overheads, multiple active instruction streams, fast and efficient support for task synchronization. Initially, this project will investigate the use of cache memories for multihreaded dataflow architecture, and compiler techniques for improving the performance of such machines. This project will then study architectural support for multithreading including support for thread management, thread scheduling, communication and synchronization among threads. Simultaneous multithreading in the context of dataflow and hybrid architectures will then be investigated. Finally, this project will explore the suitability of the most significant architectural features for use with convectional RISC technology.