Minimum delay load-balancing via nonparametric regression and no-regret algorithms

In the current network scenario, where traffic is increasingly dynamic and resource demanding, Dynamic Load-Balancing (DLB) has been shown to be an excellent Traffic Engineering tool. In particular, we are interested in the problem of minimum delay load-balancing. That is to say, we assume that the queueing delay of a link is given by a function of its load. The objective is then to adjust the traffic distribution over paths so that, for the current traffic demand, the addition of these functions times the load is minimized. The contribution of our article is twofold. Firstly, we analyze the possibility of using so-called no-regret algorithms to perform the load balancing. As opposed to other distributed optimization algorithms (such as the classical gradient descent) the algorithm we discuss requires no fine-tuning of any speed-controlling parameter. Secondly, we present a framework that does not assume any particular model for the queueing delay function, and instead learns it from measurements. This way, the resulting mean delay of optimizing with this learnt function is an excellent approximation of the real minimum delay traffic distribution. The whole framework is illustrated by several packet and flow level simulations.