Performance Analysis of Polymorphous Computing Architectures
Abstract
In general, polymorphous computing architectures are architectures that can be dynamically customized to various applications. This thesis is concerned with metrics, formulations, and algorithms for systematically synthesizing architectural configurations and software for such architectures, particularly in the domain of digital signal processing (DSP). In polymorphous system synthesis for DSP, one central aspect is managing trade-offs between latency and throughput, which are critical metrics for DSP applications. In Chapters 1-4 of the thesis, we develop a model for latency that is more appropriate than conventional models of latency for DSP system synthesis, and that takes into account central issues related to transient-state time, and we develop precise relationships between latency and throughput using this framework. Schedule post-processing strategies based on simulation and timing that reduce the latency and transient for a given throughput constraint are then presented, along with an approach based on graph-theoretic framework to streamline them, and their efficacy is substantiated with experimental results. Chapter 5-6 of the thesis then deal with the problem of efficient mapping of an application with stochastic execution times to a polymorphous computing architecture in accordance with the time-varying performance requirements for several metrics, which may include even non-trivial metrics such as the coupled latency/throughput metrics. A comprehensive model for system synthesis in this context is developed; results are developed on the complexity of system synthesis under this model; and preliminary algorithms that address the synthesis problem are presented, and evaluated experimentally.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 2001
- Accession Number
- ADA456425
Entities
People
- Sumit Lohani
Organizations
- University of Maryland