A Minicourse on Dynamic Multithreaded Algorithms

This tutorial teaches dynamic multithreaded algorithms using a Cilk-like [11, 8, 10] model. The material was taught in the MIT undergraduate class 6.046 Introduction to Algorithms as two 80-minute lectures. The style of the lecture notes follows that of the textbook by Cormen, Leiserson, Rivest, and Stein [7], but the pseudocode from that textbook has been “Cilkified” to allow it to describe multithreaded algorithms. The first lecture teaches the basics behind multithreading, including defining the measures of work and critical-path length. It culminates in the greedy scheduling theorem due to Graham and Brent [9, 6]. The second lecture shows how parallel applications, including matrix multiplication and sorting, can be analyzed using divide-and-conquer recurrences. 1 Dynamic multithreaded programming As multiprocessor systems have become increasingly available, interest has grown in parallel pro­ gramming. Multithreaded programming is a programming paradigm in which a single program is broken into multiple threads of control which interact to solve a single problem. These notes provide an introduction to the analysis of “dynamic” multithreaded algorithms, where threads can be created and destroyed as easily as an ordinary subroutine can be called and return. 1.1 Model Our model of dynamic multithreaded computation is based on the procedure abstraction found in virtually any programming language. As an example, the procedure FIB gives a multithreaded algorithm for computing the Fibonacci numbers:1 Support was provided in part by the Defense Advanced Research Projects Agency (DARPA) under Grant F3060297-1-0270, by the National Science Foundation under Grants EIA-9975036 and ACI-0324974, and by the SingaporeMIT Alliance. 1This algorithm is a terrible way to compute Fibonacci numbers, since it runs in exponential time when logarithmic methods are known [7, pp. 902–903], but it serves as a good didactic example.