Author

Paul Bone

Type

Ph.D. thesis

Institution

Department of Computing and Information Systems, The University of Melbourne, Australia

Date

December, 2012

Download

PDF

Slides

Completion Seminar Slides

Archive

University Library (unimelb.edu.au)

Source

LaTeX Source (github.com)

Abstract

Multicore computing is ubiquitous, so programmers need to write parallel programs to take advantage of the full power of modern computer systems. However, the most popular parallel programming methods are difficult and extremely error-prone. Most such errors are intermittent, which means they may be unnoticed until after a product has been shipped; they are also often very difficult to fix. This problem has been addressed by pure declarative languages that support explicit parallelism. However, this does nothing about another problem: it is often difficult for developers to find tasks that are worth parallelising. When they can be found, it is often too easy to create too much parallelism, such that the overheads of parallel execution overwhelm the benefits gained from the parallelism. Also, when parallel tasks depend on other parallel tasks, the dependencies may restrict the amount of parallelism available. This makes it even harder for programmers to estimate the benefit of parallel execution.

In this dissertation we describe our profile feedback directed automatic parallelisation system, which aims at solving this problem. We implemented this system for Mercury, a pure declarative logic programming language. We use information gathered from a profile collected from a sequential execution of a program to inform the compiler about how that program can be parallelised. Ours is, as far as we know, the first automatic parallelisation system that can estimate the parallelism available among any number of parallel tasks with any number of (non-cyclic) dependencies. This novel estimation algorithm is supplemented by an efficient exploration of the program's call graph, an analysis that calculates the cost of recursive calls (as this is not provided by the profiler), and an efficient search for the best parallelisation of N computations from among the 2^N-1 candidates.

We found that in some cases where our system parallelised a loop, spawning off virtually all of its iterations, the resulting programs exhibited excessive memory usage and poor performance. We therefore designed and implemented a novel program transformation that fixes this problem. Our transformation allows programs to gain large improvements in performance and in several cases, almost perfect linear speedups. The transformation also allows recursive calls within the parallelised code to take advantage of tail recursion.

Also presented in this dissertation are many changes that improve the performance of Mercury's parallel runtime system, as well as a proposal and partial implementation of a visualisation tool that assists developers with parallelising their programs, and helps researchers develop automatic parallelisation tools and improve the performance of the runtime system.

Overall, we have attacked and solved a number of issues that are critical to making automatic parallelism a realistic option for developers.