LASLiN: A Learning-Augmented Peer-to-Peer Network
Abstract
We introduce a learning-augmented peer-to-peer (P2P) network design that leverages the predictions of traffic patterns to optimize the network's topology. While keeping formal guarantees on the standard P2P metrics (routing path length, maximum degree), we optimize the network in a demand-aware manner and minimize the path lengths weighted by the peer-to-peer communication demands. Our protocol is learning-augmented, meaning that each node receives an individual, possibly inaccurate prediction about the future traffic patterns, with the goal of improving the network's performances. We strike a trade-off between significantly improved performances when the predictions are correct (consistency) and polylogarithmic performances when the predictions are arbitrary (robustness). We have two main contributions. First, we consider the centralized setting and show that the problem of constructing an optimum static skip list network (SLN) is solvable in polynomial time and can be computed via dynamic programming. This problem is the natural demand-aware extension of the optimal skip list problem. Second, we introduce the Uniform P2P protocol which generalizes skip list networks (SLN) by relaxing the node's heights from discrete to continuous. We show that Uniform achieves state-of-the-art performances: logarithmic routing and maximum degree, both with high probability. We then use Uniform to build a learning-augmented P2P protocol in order to incorporate demand-awareness, leading to our main contribution, LASLiN. We prove that the performances of LASLiN are consistent with those of an optimum static SLN with correct predictions (given via our dynamic programming approach), and are at most a logarithmic factor off the state-of-the-art P2P protocols if the predictions are arbitrary wrong. For the special case of highly sparse demands, we show that LASLiN achieves improved performances.