ON A NEW CLASS OF LOAD BALANCING NETWORK PROTOCOLS

Abstract

In this paper we study a new class of generic, parametrized, locally load-sensing (LLS) network-routing protocols over simple graphs, Y. These protocols are Y-"local" in the sense that they transmit packets only between Y-adjacent vertices and LLS since they base their "routing decisions" dynamically on queue-sizes of their neighbors and their relative distance to the destination. In the system each vertex has specific data-queues indexed by its respective Y-neighbors. The state of a vertex then consists of the collection of queue-sizes. The data-transmission protocols are formally specified in the framework of sequential dynamical systems, which allows us to categorize and classify our experiments. We will investigate the following scenario: for fixed Y we assume a single source/destination pair to be given and a system-update then consists of the collection of local protocol updates according to some fixed permutation of the Y-vertices. We then iterate the system-updates and thereby obtain the time evolution of the queue-sizes of the vertices. We will present and discuss results on the evolution of the load, i.e. the total number of packets in the network, the throughput, i.e. the rate at which packets arrive at the destination, and study the dependence of the queue-size dynamics on various other system parameters. In particular, we will analyze update schedule dependency and the impact of queue-capacity on system stability. We will show that our protocols can adapt and dynamically utilize new routes in a fixed network.