Operating System Support for Imprecise Computation

Recent research on imprecise computation (that is, flexible computation) at the University of Illinois has focused on the development of operating system support for imprecise computations. The scheduling algorithms and imprecise-computation environment resulting from this work enable the use of imprecise computation as a means to provide scheduling flexibility in real-time systems and to enhance their fault tolerance and allow graceful degradation [7,9]. A real-time system is one in which tasks have deadlines. A hard realtime system fails if any task does not produce its result by its deadline. Here, the term task loosely refers to a unit of work that is executed by the system to provide some service. FFT and multiple target tracking computations are examples of tasks, as are multimedia transmissions, displays, and operator actions. It is possible to design many real-time applications so that their tasks are flexible. The quality of service (QoS) provided by a flexible task improves with the amount of resources used by the task. For example, allocating a tracking task more processor time to complete a better but more complex tracking algorithm leads to a lower probability of producing false tracks. Research on imprecise computation in areas as diverse as digital control, multimedia, artificial intelligence, and databases have shown it to be a feasible and effective approach. We logically divide each imprecise task into two parts: a mandatory part and an optional part. This division is done in such a way that the system still performs acceptably as long as all the mandatory parts of the tasks are executed. Executing the optional part of a task improves the quality of the task’s output. If the optional part is run to completion, we say that the resulting output is precise. An anytime computation task is one that is entirely optional, while tasks in a system that do not allow premature termination of any task are entirely mandatory. As stated earlier, our aim is to develop a general framework for the use of imprecise computations in real-time systems. Our recent research has the following three thrusts: the design and development of ICE, an environment for developing imprecise systems, the development of scheduling and assignment algorithms that can deal with the propagation of errors among dependent tasks and the end-to-end nature of timing constraints, and the analysis and assessment of the effects of errors in the results of individual computations on the overall system behavior for several representative applications.