Learning-Augmented Distributed Directories*

We study distributed directory protocols for accessing shared objects in large-scale distributed systems under the recently proposed framework of learning-augmentation. Each shared object has an owner node that can modify its value. The ownership may change by moving the object from one node to another in response to move requests. The value of an object can be read by other nodes with lookup requests. The existing directory protocols were designed in the online model where both the arrival time of requests and the nodes issuing requests are not known a priori. We consider the learned-augmented framework that involves a priori knowledge on nodes that issue requests; the arrive time of requests as well as whether in fact predicted nodes issue those requests are unknown (i.e., the predictions may be error-prone). We design two distributed directory protocols, one tree-based and another cluster-based, and provide better guarantees that were known in the literature in the online model, when predictions are perfect (no prediction error). We additionally show that the guarantees degrade gracefully with prediction error but do not get worse than the guarantees in the online model even with maximum prediction error. To the best of our knowledge, this is the first study of distributed directory protocols under learning-augmented framework.