Load sharing in soft real-time distributed systems

Abstract

In a soft real-time distributed computer system, a job submitted at a node in the system must complete execution within a specified deadline, otherwise it is considered lost. When a job cannot be serviced within its deadline at a node, it may still be possible to service this job within its deadline by transferring it to some other node in the system. We examine three load sharing policies and show that we can achieve a considerable improvement by using simple load sharing algorithms. We introduce a new variation of these algorithms and show that improved performance results can be obtained at low loads. Approximate analytical models are developed, solved numerically and performance comparisons are made. The performance results suggest that, over a wide range of system parameters, the performance of these algorithms is substantially better than that of no load sharing and is often close to that of perfect load sharing.