SIMULATION STUDY OF DYNAMIC LOAD BALANCING FOR PROCESSOR SHARING SERVERS WITH FINITE CAPACITY UNDER GENERALIZED HALFIN-WHITT REGIMES

摘要

Defining efficient decentralized load balancing schemes exhibiting low memory and communication costs is an important ongoing topic. In particular, characterizing critical regimes where a system optimizes resource usage is uncharted territory. We consider here dynamic balancing schemes in a set of processor sharing servers with finite capacity. Guided by recent results for insensitive load balancing schemes, we applied a modeling and simulation strategy to characterize systematically and extensively several classes of balancing policies under various statistical conditions. We found that there is a class of efficient policies for which a common critical regime can be identified and interpreted as a generalization of the Halfin-Whitt-Jagerman regime for one-server systems. We also study the gap between full and partial information systems, and analyze the performance sensitivity to jobs' size distribution. This study is motivated by the network architecture in the ATLAS experiment at CERN, where load balancing plays a key role.