This page is provided for historical and archival purposes only. While the seminar dates are correct, we offer no guarantee of informational accuracy or link validity. Contact information for the speakers, hosts and seminar committee are certainly out of date.
The imprecise computation technique has been proposed as a way to handle transient overload and to enhance fault tolerance of real-time systems. In a system based on this technique, each time-critical task, or a set of related tasks, is designed and implemented in such a way that it can be decomposed logically into a mandatory part and an optional part. When the mandatory part completes, the job produces an approximate result of acceptable quality. Under normal operating conditions, the system also completes the optional part, and the result of the task attains the desired quality. During overloads, however, the system may choose to leave the optional part unfinished at the expense of the quality of the result produced by the task. An imprecise computation is similar to an anytime computation, a term used in the AI literature.
Several workload models have been used to characterize the behavior of imprecise computations and to quantify the costs and benefits of the tradeoff between result quality and processing time requirements for different classes of applications and implementation methods. This talk will present a brief overview of these models and the scheduling algorithms based on them. To provide an environment for experimentation with imprecise computations in several application domains, including multimedia data transmissions and database queries and updates, we are implementing a system, called ICS (Imprecise Computation Server) system on the Mach Operating System. This system integrates the user-directed imprecision mechanism for storing and returning approximate results of computations with a system-directed mechanism for traditional checkpointing for fault tolerance and error recovery. The goal of this integration is reduce the costs in providing fault tolerance. The architecture of this system will be described.
If you would like to meet with Dr. Liu contact Michelle Agie at (8)8818 or send e-mail to maa@cs