Efficient mapping of control applications into a given heterogeneous computing system
Abstract
An important issue in the use of distributed computing systems is the proper scheduling (or mapping) of the computational tasks into the distributed system. In control applications, the computational problem is usually modeled by a directed acyclic task graph DATG. The nodes in the DATG represent the tasks, and the arcs represent the precedence relations between the tasks. Control application require the execution of the prescribed tasks in the task graph in real-time, satisfying a specified time constraint, the completion time. The completion time is the finish time of the last task scheduled in the task graph. In many complex control applications, a single processing node cannot satisfy the real-time constraint. The problem is customarily solved using multiple interconnected processing nodes. For some applications, a sufficient solution was the design of a special purpose computer architecture tailored to the specific needs of the application. This approach limits the maximum usefulness of the computing system to just one application. A general approach is the use of a distributed system architecture, and the development of efficient task mapping algorithms. We adopt the second approach given a general, computing system architecture, and developing efficient mapping algorithms for the control problems on the given architecture. As in many applications the computational task is mixed, we further assume the processing system is heterogeneous. This approach has the advantage of maintaining the generality of the second approach, while allowing us to accommodate the advantage of special architecture on the task level. For example, if a given task requires parasitic functions (e.g., sines and cosines), or floating point arithmetic extensively, a special processing node that can execute such tasks at a fraction of the time in comparison to the other processing nodes can be incorporated into the system and will improve the overall performance of the computational system. In this thesis we address the processing of a deterministic set of task graphs on such a heterogeneous computing system.
Degree
Ph.D.
Advisors
Ahmad, Purdue University.
Subject Area
Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.