Applications and Extensions of the Imprecise-Computation Model
Wu-chang Feng · 1996
Real-time systems have become increasingly important in a growing number of application domains such as telecommunication networks and multimedia. The tasks in such a real-time system must produce logically correct results by their deadlines. The correctness of such a system depends not only on the logical correctness of the results but also on the tasks completing by their deadlines. Many of today''s real-time systems are built based on an estimate of the worst-case load level. When the load level of a real-time system temporarily exceeds its assumed worst-case load level, a transient overload occurs; some deadlines may be missed; and the performance of the system degrades. Under such an overload, the system must maintain an acceptable level of performance. The imprecise-computation technique is a way to deal with transient overloads. The technique is motivated by the fact that one can often trade off precision for timeliness. It prevents missed deadlines and provides graceful degradation during a transient overload. We illustrate its applicability by implementing an ATM congestion-control scheme. We also propose scheduling algorithms to handle real-time applications which contain tasks whose result quality is a function of both time and input quality.