Job Scheduling for Maximal Throughput in Autonomic Computing Systems

Autonomic computing networks manage multiple tasks over a distributed network of resources. In this paper, we view an autonomic computing system as a network of queues, where classes of jobs/tasks are stored awaiting execution. At each point in time, local resources are allocated according to the backlog of waiting jobs. Service modes are selected corresponding to feasible configurations of computing (processors, CPU cycles, etc.), communication (slots, channels, etc.) and storage resources (shared buffers, memory places, etc.) We present a family of distributed algorithms which maximize the system throughput by dynamically choosing service modes in response to observed buffer backlogs. This class of policies, called projective cone scheduling algorithms, are related to maximum pressure policies in constrained queueing networks, and are shown to maintain stability under any arrival combination within the network capacity. They operate without knowledge of the arrival rates and require minimal information sharing between regions.