Load Balancing in Parallel Queues and Rank-based Diffusions

Consider a queuing system with $K$ parallel queues in which the server for each queue processes jobs at rate $n$ and the total arrival rate to the system is $nK-\upsilon \sqrt{n}$ where $\upsilon \in (0, \infty)$ and $n$ is large. We study rank-based routing policies in which $O(\sqrt{n})$ of the incoming jobs are routed to servers with probabilities depending on their ranked queue length and the remaining jobs are routed uniformly at random. A particular case, referred to as the marginal join-the-shortest-queue (MJSQ) policy, is one in which the $O(\sqrt{n})$ jobs are routed using the join-the-shortest-queue (JSQ) policy. Our first result provides a heavy traffic approximation theorem for such queuing systems. It turns out that, unlike the JSQ system, there is no state space collapse in the setting of MJSQ (and for the more general rank-based routing schemes) and one obtains a novel diffusion limit which is the constrained analogue of the well studied Atlas model (and other rank-based diffusions) arising from mathematical finance. Next, we prove an interchange of limits result which shows that the steady state of the queuing system is well approximated by that of the limiting diffusion, given explicitly in terms of product laws of Exponential random variables. Using these results, we compute the time asymptotic workload in the heavy traffic limit for the MJSQ system. We find the striking result that although in going from JSQ to MJSQ the communication cost is reduced by a factor of $\sqrt{n}$, the asymptotic workload increases only by a constant factor which can be made arbitrarily close to one by increasing a MJSQ parameter. When the system is overloaded ($\upsilon<0$) we show that although the $K$-dimensional MJSQ system is unstable, the steady state difference between the maximum and minimum queue lengths stays bounded in probability (in the heavy traffic parameter $n$).