Generative Code Specialisation for High-Performance Monte-Carlo Simulations

We address the tension between software generality and performance in the domain of scientific and financial simulations based on Monte-Carlo methods. To this end, we present a novel software architecture, centred around the concept of a specialising simulator generator, that combines and extends methods from generative programming, partial evaluation, runtime code generation, and dynamic code loading. The core tenet is that, given a fixed simulator configuration, a generator in a functional language can produce low-level code that is more highly optimised than a manually implemented generic simulator. We also introduce a skeleton, or template, capturing a wide range of Monte-Carlo methods and use it to explain how to design specialising simulator generators and how to generate parallelised simulators for multi-core and distributed-memory multiprocessors.

We evaluated the practical benefits and limitations of our approach by applying it to a highly relevant problem in computational chemistry. More precisely, we used a Markov-chain Monte-Carlo method for the study of advanced forms of polymerisation kinetics. The resulting implementation executes faster than all competing software products, while at the same time also being more general. The generative architecture allows us to cover a wider range of chemical reactions and to target a wider range of high-performance architectures (such as PC clusters and SMP multiprocessors).

We show that it is possible to outperform low-level languages with functional programming in domains with very stringent performance requirements if the domain also demands generality.