SYMPHONY: A SCALABLE HIGH PERFORMANCE ARCHITECTURE FRAMEWORK FOR MEDIA APPLICATIONS

摘要

Media applications demand "real-time" performance and execution of multiple programs simultaneously. These applications are based on input data sequences that are streams. There exists a large amount of dynamic parallelism in such applications that can be exploited to achieve high performance. We present in this paper an architecture framework called SYMPHONY, suitable for media applications. SYMPHONY comprises several processors in an array, where each processor is capable of executing multithreaded programs. SYMPHONY can be viewed as an asynchronous processor pipeline that exploits coarse grain parallelism. Each processor in SYMPHONY can execute multi-threaded code and exploits coarse grain parallelism in terms of multiple threads executing simultaneously and fine grain parallelism in terms of instruction level parallelism (ILP) that is exploited by current day super-scalar architectures. SYMPHONY incorporates mechanisms for fine grain communication between processors in the array. The communication is low latency, and most often overlapped with computation. We also categorize and characterize the applications that can be executed on SYMPHONY efficiently.