When Power-of-d-Choices Meets Priority

Power-of-d-choices (Pod) is a popular load balancing strategy, which has received much attention from both academia and industry. However, much prior work on Pod has focused on uniform tasks without priorities. In reality, tasks may have different priorities according to their service sensitivity, pricing, or importance to guarantee the quality of service (QoS). In this work, we distinguish two types of priorities in Pod: scheduling and service priorities. We propose Pod-SSP, which is a Pod algorithm with Scheduling and Service Priorities. To better understand the impact of priorities on the performance of tasks, we consider two simple variants of Pod-SSP: Pod with SCheduling Priorities (Pod-SCP) and Pod with SErvice Priorities (Pod-SEP). Utilizing mean-field approximation, we systematically study the performance of these protocols in the large-system regime. Our theoretical and simulation results show that high-priority tasks can have a more than 3x better delay relative to a system running the original Pod algorithm, and meanwhile, low-priority tasks only slightly sacrifice their delay.