A streaming computation framework for the Cell Processor

Abstract

The Cell architecture is a heterogeneous, distributed-memory multicore architecture that features a novel on-chip communication network. Stream programs are particularly well-suited for execution on Cell.This thesis implements a runtime library on Cell specifically designed to support streaming applications and streaming language compilers. The runtime library abstracts the details of Cell's communication network and provides facilities that simplify the task of scheduling stream actors. The library is designed in the context of the StreamIt programming language. This library is used to implement a dynamic scheduling framework. The programmability of high-level schedulers with and without the library is analyzed. We observe that the library does not add significant overhead, provides a number of useful features for schedulers, and greatly simplifies the code required to implement schedulers.