Improving Asynchronous Partial Overlay

Since its creation, the Asynchronous Partial Overlay (APO) protocol has received a great deal of attention because of its non-traditional approach to solving Distributed Constraint Satisfaction Problems (DCSPs). Its introduction led investigators to question the very definition of the word "distributed" and has subsequently inspired the community to create improved metrics for parallel computation, enhanced testing procedures, and most importantly new DCSP algorithms. These advances have raised concerns about APO's parallel efficiency by showing that, in some cases, APO performs very poorly compared to protocols such as Asynchronous Forward Checking, Conflict-directed Back jumping (AFC-CBJ). In addition, APO's soundness and completeness were brought into question when it was discovered that, under certain conditions, stale state information could cause the protocol's distributed locking mechanism to fail. This work revisits APO by reengineering the protocol to simplify it and increase its parallelism while ensuring its soundness and completeness. It also vastly improves the parallel efficiency of APO by replacing its central solver with a variant of the Forward Checking, Conflict-directed Back jumping (FC-CBJ) algorithm that is specifically tuned to complement the heuristic strategies used by APO to limit its centralization. This new version of APO is then evaluated against the AFC-CBJ protocol using random instances of both DCSPs and distributed 3-coloring problems. The end result is a protocol that is several orders of magnitude faster than the original APO, uses less messages, is more private, and outperforms the AFC-CBJ protocol in nearly every case tested.