A faster all parallel Mergesort algorithm for multicore processors

Sammendrag

The problem addressed in this paper is that we want to sort an integer array a[] of length n in parallel on a multi core machine with p cores using mergesort. Amdahl’s law tells us that the inherent sequential part of any algorithm will in the end dominate and limit the speedup we get from parallelisation. This paper introduces ParaMerge, an all parallel mergesort algorithm for use on an ordinary shared memory multi core machine that has just a few simple statements in its sequential part. The new algorithm is all parallel in the sense that by recursive decent it is two parallel in the top node, four parallel on the next level in the recursion, then eight parallel until we at least have started one thread for all the p cores. After parallelization, each thread then uses sequential recursion mergesort with a variant of insertion sort for sorting short subsections at the end. ParaMerge can be seen as an improvement over traditional parallelization of the mergesort algorithm where one follows the sequential algorithm and substitute recursive calls with the creation of parallel threads in the top of the recursion tree. This traditional parallel mergesort finally does a sequential merging of the two sorted halves of a[]. First at the next level it goes two-parallel, then four parallel on the next level, and so on. After parallelization my implementation of this traditional algorithm also use the same sequential mergesort and insertion sort algorithm as the ParaMerge algorithm in each thread. 
There are two main improvements in Paramerge: First the observation that merging can both be done from the start left to right picking the smallest elements of the two sections to be merged, and at the same time from the end of the same sections from right to left picking the largest elements. The second improvement is that the contract between a node and its two sub-nodes is changed. In a traditional parallelization a node is given a section of a[], sort this by merging two sorted halves it recursively receives from its own two sub nodes and returns its to its mother node. In Paramerge the two sub nodes each receive a full sorting from its two own sub nodes of the section itself got from its mother node (so this problem is already solved). Every node has a twin node. In parallel these two twin nodes then merge their two sorted sections, one from left and the other from right as described above. The two twin sub nodes have then sorted the whole section given to their common mother node. This goes also for the top node. We have thus raised the level of parallelization by a factor of two at each level of the top of the recursion tree. The ParaMerge algorithm also contains other improvements, such as a controlled sorting back and forth between a[] and a scratch area b[] of the same size such that the sorted result always ends up in a[] without any copy, and a special insertion sort that is central for achieving this copy-free feature. ParaMerge is compared with other published algorithms, and in only one case is one of the ‘new’ features in Paramerge found. This other algorithm is described and compared in some detail.
Finally, ParaMerge is empirically compared with three other algorithms sorting arrays of length n =10,20,…,50 m, and ..1000m when p=32. demonstrating that it is significantly faster than two other merge algorithms, the sequential and the traditional parallel algorithm, and Arrays.sort(), a sequential Quicksort algorithm from the Java library.