Performance of co-scheduling on a network of workstations

Abstract

In a set of high performance workstations connected by a network, many workstations may be underutilized by their owners. While each workstation may be primarily responsible for executing its owner's tasks with the highest priority, the unused processing capacity may be made available to computationally intensive tasks submitted externally to the system. Static co-scheduling for such an environment has been considered previously (M.J. Atallah et al., 1991), where the goal was to maximize the speedup by partitioning the task among many workstations. The authors consider the problem from the system point of view, and develop a queuing model and efficient algorithms to minimize the mean response time. The results obtained show that significant improvements in the mean response time can be obtained through co-scheduling over that of the M/M/m system where each task would be assigned to a single workstation as a whole.<>