Improving Communication Sensitive Parallel Radix Sort for Unbalanced Data

Abstract

Radix sort is an efficient method to sort integer keys on parallel computers. It is easy to parallelize and simple to implement. The main drawbacks of existing algorithms are load balancing problems and communication overhead. These problems are caused in data characteristics like data-skew and duplicates. There are several approaches how to parallelize the radix sort algorithm, which yield to reduce communication operations or to improve the load balance. If an algorithm has its focus on the optimization of the load balance then, its communication is inefficient. Otherwise, if the focus is on the communication minimization, then the algorithms are only efficient for well-distributed data. For the latter case, we will present an efficient improvement which helps to overcome the problems with unbalanced data characteristics. The suggested improvements are tested practically on a Linux-based SMP cluster.