Towards Learning-Augmented Peer-to-Peer Networks: Self-Stabilizing Graph Linearization with Untrusted Advice
Distributed peer-to-peer systems are widely popular due to their decentralized nature, which ensures that no peer is critical for the functionality of the system. However, fully decentralized solutions are usually much harder to design, and tend to have a much higher overhead compared to centralized approaches, where the peers are connected to a powerful server. On the other hand, centralized approaches have a single point of failure. Thus, is there some way to combine their advantages without inheriting their disadvantages? To that end, we consider a supervised peer-to-peer approach where the peers can ask a potentially unreliable supervisor for advice. This is in line with the increasingly popular algorithmic paradigm called algorithms with predictions or learning-augmented algorithms, but we are the first to consider it in the context of peer-to-peer networks. Specifically, we design self-stabilizing algorithms for the fundamental problem of distributed graph linearization, where peers are supposed to recover the "sorted line" network from any initial network after a transient fault. With the help of the supervisor, peers can recover the sorted line network in $O(\log n)$ time, if the advice is correct; otherwise, the algorithm retains its original recovery time (i.e., without any supervisor). A crucial challenge that we overcome is to correctly compose multiple self-stabilizing algorithms, that is, one that processes and exploits the advice, and another that does not rely on the advice at all. Our key technical contributions combine ideas from the fields of overlay networks and proof-labeling schemes. Finally, we give a matching lower bound of $\Omega(\log n)$ for the recovery time of any algorithm if the advice can be corrupted, where $n$ is the network size.