Simple in-place yet comparison-optimal Mergesort

Abstract

Mergesort is one of the few efficient sorting algorithms and, despite being the oldest one, often still the method of choice today. In contrast to some alternative algorithms, it always runs efficiently using O(n log n) element comparisons and usually works in a stable manner. Its only practical disadvantage is the need for a second array, and thus twice the amount of memory. This can be an impeding factor, especially when handling large amounts of data, where it is often either impractical or even impossible to fall back to slower or unstable sorting alternatives. Therefore, many attempts have been made to fix this problem by adapting Mergesort to work in place. While it is known that such algorithms exist, the previously published solutions are mostly not efficient, become unstable, and/or are very complex. This renders them practically useless for real-world applications. In this paper, we propose a novel in-place Mergesort algorithm that is stable by design. Albeit its running time of O(n log^2 n) is not quite optimal, it still works efficiently, both in theory using the optimal number of O(n log n) comparisons and in practice with low constants, while being easily comprehensible. The baseline for this new algorithm includes just two prerequisites: 1) an optimal array rotation; and 2) the co-ranking idea, published in 2014 and originally intended to parallelize Mergesort. Although it would certainly be possible to parallelize the presented algorithm, this paper focuses on the sequential aspect of this efficient, stable and in-place Mergesort algorithm. Additionally, we implemented our algorithm and present performance results measured on one of the largest shared memory systems currently available.