Adventures in parallel statistical computation

As technology progresses, the processors used for statistical computation are not getting faster: there are just more of them. For example, there are no 10-GHz CPUs available on the market today, but there are quad-core, 3.2-GHz processors (albeit at a high price!). To take full advantage of this fact, one can turn to parallel computing, although existing algorithms tend to be serial designs, and a great challellge exists in the transformation of these algorithms into ones which can take advantage of parallel computing. When this is possible, however, great increases in performance can be achieved, with the impost of (often greatly) increased complexity of implementation. This paper describes one user's foray into parallel processing, with the twin aims of introducing parallel programming concepts and advantages, as well as encouraging more researchers to invest the time in such undertakings. The research described here involved the application of approximate Bayesian computation (ABC) to a DNA sequence segmentation problem. ABC is a stochastic parameter estimation algorithm that readily lends itself to parallel programming. Some of the pitfalls that were encountered are discussed, as well as the benefits gained from the final, parallel, version of the programme