Reducing Randomization in the Power of Two Choices Load Balancing Algorithm

This paper proposes a new version of the Power of Two Choices, SQ(d), load balancing algorithm that improves the performance of the classical model based on the power of two choices randomized load balancing. This model considers jobs that arrive to a dispatcher as a Poisson stream of rate λn, λ 1, at a set of n servers. Using the power of two choices, the dispatcher chooses for each job some d constant independently and uniformly from the n servers in a random way, and sends the job to the server with the fewest number of jobs. This algorithm offers advantage over the load balancing based on shortest queue discipline, because it offers a good performance, and reduces the overhead over the servers and over the communication network. In this paper, we propose a new version, Shortest Queue of d with Randomization and Round Robin Policies, SQ-RR(d), that combines randomization techniques and static local balancing based on round robin policy. In this new version the dispatcher chooses the d servers as follows: one is selected using round robin policy and the d - 1 servers are chosen independently and uniformly in a random way from the η servers. Then, the dispatcher sends the job to the server with the fewest number of jobs. We demonstrate with an analytical approximation of this approach, that this new version improves the performance obtained with the classical solution for d 2, and obtains similar results for d 2, included systems at 99 percent of capacity. Furthermore, we provide simulations that demonstrate the analytical approximation developed and show the behavior of this algorithm with realistic workloads based on Google datacenter traces.