Concurrent self-adjusting distributed tree networks

Abstract: SplayNets are a distributed generalization of the classic splay tree data structures. Given a set of communication requests and a network comprised of n nodes, such that any pair of nodes is capable of establishing a direct connection, the goal is to dynamically find a (locally routable) binary tree topology, which connects all nodes and optimizes the routing cost for that communication pattern, making local topology transformations (rotations) before each request is served. In this work we present a distributed and concurrent implementation of SplayNets. Analytical results show that our proposed algorithm prevents loops and deadlocks from occurring between concurrent rotations. We compute the total amortized average cost of a splay request in number of rounds and number of time-slots and as a function of the empirical entropies of source and destination nodes of the splay requests.