Implementing Radix Sort on Emu 1

This paper discusses the implementation of radix sort on Emu Solutions’ Emu 1 system. Sorting is an important building block for a large class of data-intensive applications. Radix sort, in particular, is a good candidate for sorting large sequences of fixed-length integer keys (for example, indexes in a database), because it presents lower computational complexity with respect to comparison based sorting algorithms when the length of the keys is small with respect to the dataset size. In general, implementing scalable sorting algorithms for large datasets is challenging, due to the large amount of memory operations with low data locality to move keys to their new positions. The Emu architecture is a memory-centric design based around the concept of lightweight threads that migrate across nodelets, a combination of memory and multithreaded processors aimed at providing a scalable large-memory system. We show how the Emu 1 design provides scalability in performance and size to our radix sort implementation.