Profiling large-scale lazy functional programs

Abstract

The LOLITA natural language processing system is an example of one of the ever increasing number of large-scale systems written entirely in a functional programming language. The system consists of over 50,000 lines of Haskell code and is able to perform a number of tasks such as semantic and pragmatic analysis of text, context scanning and query analysis. Such a system is more useful if the results are calculated in real-time, therefore the efficiency of such a system is paramount. For the past three years we have used profiling tools supplied with the Haskell compilers GHC and HBC to analyse and reason about our programming solutions and have achieved good results; however, our experience has shown that the profiling life-cycle is often too long to make a detailed analysis of a large system possible, and the profiling results are often misleading. A profiling system is developed which allows three types of functionality not previously found in a profiler for lazy functional programs. Firstly, the profiler is able to produce results based on an accurate method of cost inheritance. We have found that this reduces the possibility of the programmer obtaining misleading profiling results. Secondly, the programmer is able to explore the results after the execution of the program. This is done by selecting and deselecting parts of the program using a post-processor. This greatly reduces the analysis time as no further compilation, execution or profiling of the program is needed. Finally, the new profiling system allows the user to examine aspects of the run-time call structure of the program. This is useful in the analysis of the run-time behaviour of the program. Previous attempts at extending the results produced by a profiler in such a way have failed due to the exceptionally high overheads. Exploration of the overheads produced by the new profiling scheme show that typical overheads in profiling the LOLITA system are: a 10% increase in compilation time; a 7% increase in executable size and a 70% run-time overhead. These overheads mean a considerable saving in time in the detailed analysis of profiling a large, lazy functional program.