Abstract. We consider N single server infinite buffer queues with service rate β. Customers arrive at rate N α, choose L queues uniformly, and join the shortest. We study the processes t ∈ R + → R N t = (R N t (k)) k∈N for large N , where R N t (k) is the fraction of queues of length at least k at time t. Laws of large numbers (LLNs) are known, see Vvedenskaya et al. [15], Mitzenmacher [12] and Graham [5]. We consider certain Hilbert spaces with the weak topology. First, we prove a functional central limit theorem (CLT) under the a priori assumption that the initial data R N 0 satisfy the corresponding CLT. We use a compactness-uniqueness method, and the limit is characterized as an Ornstein-Uhlenbeck (OU) process. Then, we study the R N in equilibrium under the stability condition α < β, and prove a functional CLT with limit the OU process in equilibrium. We use ergodicity and justify the inversion of limits lim N →∞ lim t→∞ = lim t→∞ lim N →∞ by a compactnessuniqueness method. We deduce a posteriori the CLT for R N 0 under the invariant laws, an interesting result in its own right. The main tool for proving tightness of the implicitly defined invariant laws in the CLT scaling and ergodicity of the limit OU process is a global exponential stability result for the nonlinear dynamical system obtained in the functional LLN limit.
Key-words:Mean-field interaction, load balancing, resource pooling, ergodicity, non-equilibrium fluctuations, equilibrium fluctuations, birth and death processes, spectral gap, global exponential stability MSC2000: Primary: 60K35. Secondary: 60K25, 60B12, 60F05, 37C75, 37A30.
Introduction
PreliminariesWe consider a Markovian network constituted of N ≥ L ≥ 1 infinite buffer single server queues.Customers arrive at rate N α, are each allocated L distinct queues uniformly at random, and join the shortest, ties being resolved uniformly. Servers work at rate β. Arrivals, allocations, and services are independent. For L = 1 we have i.i.d. M α /M β /1/∞ queues. For L ≥ 2 the interaction structure depends only on sampling from the empirical measure of L-tuples of queue states: in statistical mechanics terminology, the system is in L-body mean-field interaction. We continue the large N study introduced by Vvedenskaya et al.