Experiments for Algorithm Engineering

Abstract

Hoare introduced and analyzed the Quicksort algorithm in the early 1960s 6. By the early 1970s, a highly tuned version of that algorithm was implemented in the Unix system’s main memory sort function, qsort. For two decades, that code performed admirably.

In the early 1990s, Alan Wilks and Rick Becker once again used that old, reliable program, and were stunned by the results. A run that should have taken a few minutes was cancelled after hours. They studied their input data and eventually produced an eight-line program that showed that their run would have taken weeks. They enclosed that program in a superb bug report that they e-mailed to me (details are described by Bentley [1]). They had stumbled across a problem that Hoare had foreseen: selection of the partitioning element. An implementation cleverness that had stood the test of two decades worth of time had finally failed catastrophically.