Mean Waiting Time in Large-Scale and Critically Loaded Power of d Load Balancing Systems

Mean field models are a popular tool used to analyse load balancing policies. In some exceptional cases the waiting time distribution of the mean field limit has an explicit form. In other cases it can be computed as the solution of a set of differential equations. In this paper we study the limit of the mean waiting time E[W-lambda] as the arrival rate lambda approaches 1 for a number of load balancing policies in a large-scale system of homogeneous servers which finish work at a constant rate equal to one and exponential job sizes with mean 1 (i.e. when the system gets close to instability). As E[W-lambda] diverges to infinity, we scale with - log(1 - lambda) and present a method to compute the limit lim(lambda -> 1-) -E[W.]/log(1 -lambda). We show that this limit has a surprisingly simple form for the load balancing algorithms considered. More specifically, we present a general result that holds for any policy for which the associated differential equation satisfies a list of assumptions. For the well-known LL(d) policy which assigns an incoming job to a server with the least work left among d randomly selected servers these assumptions are trivially verified. For this policy we prove the limit is given by 1/d-1. We further show that the LL(d, K) policy, which assigns batches of K jobs to the K least loaded servers among d randomly selected servers, satisfies the assumptions and the limit is equal to K/d-K. For a policy which applies LL(d(i)) with probability p(i), we show that the limit is given by 1/Sigma(i)p(i)d(i)-1. We further indicate that our main result can also be used for load balancers with redundancy or memory. In addition, we propose an alternate scaling - log(p(lambda)) instead of - log(1 - lambda), where p(lambda) is adapted to the policy at hand, such that lim(lambda -> 1-) -E[W-lambda]/log(1-lambda) = lim(lambda -> 1-) -E[W-lambda]/log(p(lambda)), where the limit lim(lambda -> 0+) -E[W-lambda]/log(p(lambda)) is well defined and non-zero (contrary to lim(lambda -> 0+)-E[W-lambda]/log(1 -lambda)). This allows to obtain relatively flat curves for -E[W-lambda]/log(lambda(p)) for lambda is an element of [0, 1] which indicates that the low and high load limits can be used as an approximation when. is close to one or zero. Our results rely on the earlier proven ansatz which asserts that for certain load balancing policies the workload distribution of any finite set of queues becomes independent of one another as the number of servers tends to infinity.