Asymptotics of Insensitive Load Balancing and Blocking Phases

TL;DR

This paper analyzes the asymptotic behavior of insensitive load balancing schemes in large-scale systems, revealing different phases of blocking probability and providing generalized staffing rules for robust performance.

Contribution

It characterizes the asymptotic regimes and phases of blocking probability in insensitive load balancing schemes as the system scales, extending classical queueing results.

Findings

Identifies three deviation amplitudes in system response.

Derives phase transitions for blocking probability.

Generalizes the QED regime for multi-server systems.

Abstract

We address the problem of giving robust performance bounds based on the study of the asymptotic behavior of the insensitive load balancing schemes when the number of servers and the load scales jointly. These schemes have the desirable property that the stationary distribution of the resulting stochastic network depends on the distribution of job sizes only through its mean. It was shown that they give good estimates of performance indicators for systems with finite buffers, generalizing henceforth Erlang's formula whereas optimal policies are already theoretically and computationally out of reach for networks of moderate size. We study a single class of traffic acting on a symmetric set of processor sharing queues with finite buffers and we consider the case where the load scales with the number of servers. We characterize central limit theorems and large deviations, the response of symmetric systems under those schemes at different scales and show that three amplitudes of deviations can be identified. A central limit scaling takes place for a sub-critical load; for $\rho=1$, the number of free servers scales like $n^{ {\theta \over \theta+1}}$ ($\theta$ being the buffer depth and $n$ being the number of servers) and is of order 1 for super-critical loads. This further implies the existence of different phases for the blocking probability, Before a (refined) critical load $\rho_c(n)=1-a n^{- {\theta \over \theta+1}}$, the blocking is exponentially small and becomes of order $ n^{- {\theta \over \theta+1}}$ at $\rho_c(n)$. This generalizes the well-known Quality and Efficiency Driven (QED) regime or Halfin-Whitt regime for a one-dimensional queue, and leads to a generalized staffing rule for a given target blocking probability.