Framed Projective Cone Scheduling: Latency vs. Context-Switching Tradeoff in Data Centers

Abstract

Queue-processor service and communication switches are often reconfigured in data centers to dynamically reallocate resources based on demand and maximize utilization. However, reconfiguration introduces overhead that will reduce the usable processor bandwidth if done too frequently. We introduce a cost framework to determine how frequently such reconfigurations should occur in order to optimally trade off between the cost of the reconfiguration (or context switching) overhead and the latency cost due to the delayed reconfiguration. A general framing algorithm is introduced to optimize dynamic processor allocation that limits processors to only be reallocated at the beginning of a new frame, but allows a class of functions of the historical backlog to be employed when selecting the new allocation. We show that the system throughput is not affected by framing, however, the job latency increases with the frame's span. The cost model and framed allocation algorithm are investigated to determine how to balance a tolerable increase in job latency for significant reduction of system overhead due to processor reconfiguration.