Cache Analysis of Non-uniform Distribution Sorting Algorithms

TL;DR

This paper analyzes the cache performance of distribution sorting algorithms with non-uniform key distributions, providing more accurate bounds and optimizing MSB radix sort for uniform floats, outperforming traditional methods.

Contribution

It offers a refined cache performance analysis for distribution sorting algorithms under non-uniform distributions and demonstrates improved sorting efficiency for uniform floats.

Findings

Tighter bounds for cache performance with non-uniform keys

Optimized MSB radix sort outperforms bucketsort and Quicksort for uniform floats

Enhanced understanding of cache behavior in distribution sorting algorithms

Abstract

We analyse the average-case cache performance of distribution sorting algorithms in the case when keys are independently but not necessarily uniformly distributed. The analysis is for both `in-place' and `out-of-place' distribution sorting algorithms and is more accurate than the analysis presented in \cite{RRESA00}. In particular, this new analysis yields tighter upper and lower bounds when the keys are drawn from a uniform distribution. We use this analysis to tune the performance of the integer sorting algorithm MSB radix sort when it is used to sort independent uniform floating-point numbers (floats). Our tuned MSB radix sort algorithm comfortably outperforms a cache-tuned implementations of bucketsort \cite{RR99} and Quicksort when sorting uniform floats from $[0, 1)$.