Histogram sort with sampling

Standard parallel sorting algorithms like sample sort rely on data partitioning techniques to distribute keys across processors. The sampling cost in sample sort for good load balance is prohibitive for massive clusters. We describe Histogram sort with sampling, an adaptation of the popular Histogram sort algorithm. We show that Histogram sort with sampling has sound theoretical guarantees and reduces the sample size requirements from O(p log N/epsilon^2) to O(k p sqrt[k]{log p/epsilon}) with k rounds of histogramming w.h.p.. Histogram sort with sampling is more efficient than Sample sort algorithms that achieve the same level of load balance, both in theory and practice, especially for massively parallel applications, scaling to tens of thousands of processors. We also show that an approximate but fairly accurate histogram can be obtained using a O( sqrt {p log N}/epsilon) sample on every processor. This can be used to speed up the histogramming step and can be of independent interest for answering general queries in large parallel processing systems. In our practical implementation, we exploit shared memory within nodes to improve the performance of our algorithm on large modern clusters