MergeArray and scalable, relaxed, concurrent, mergeable priority queues

Abstract

The priority queue is a well-studied data structure which has prospered in the ever-growing field of distributed computing. However, in the asynchronous shared-memory model, one operation was left behind: merge. I present the MergeArray, a framework for implementing scalable, relaxed, concurrent, and mergeable objects, which exploits disjoint access parallelism by using an array of sequential objects and performs merges lazily, index-by-index. I use MergeArray to build a linearizable and scalable priority queue with lock-free merge and insert and a relaxed, deadlock-free remove-min with expected worst-case rank-error of O(plogp) for p threads under common assumptions. I show experimental evidence that supports this rank-error estimate in practice as well as increased performance and scalability on a relaxed Minimum Spanning Tree benchmark compared to SprayList, a cutting-edge relaxed priority queue.