On the parallelisation of MCMC by speculative chain execution
Byrd, Jonathan M. R., Jarvis, Stephen A. and Bhalerao, Abhir (2010) On the parallelisation of MCMC by speculative chain execution. In: Workshop on Multitheaded Architectures and Applications (MTAAP), Atlanta, Georgia, USA, 19-23 April 2010. Published in: 2010 IEEE International Symposium on Parallel & Distributed Processing, Workshops and Phd Forum (IPDPSW) pp. 1-8.
Restricted to Repository staff only
Official URL: http://dx.doi.org/10.1109/IPDPSW.2010.5470689
The increasing availability of multi-core and multi- processor architectures provides new opportunities for improving the performance of many computer simulations. Markov Chain Monte Carlo (MCMC) simulations are widely used for approximate counting problems, Bayesian inference and as a means for estimating very high-dimensional integrals. As such MCMC has had a wide variety of applications in fields including computational biology and physics, financial econometrics, machine learning and image processing.
One method for improving the performance of Markov Chain Monte Carlo simulations is to use SMP machines to perform ‘speculative moves’, reducing the runtime whilst producing statistically identical results to conventional sequential implementations. In this paper we examine the circumstances under which the original speculative moves method performs poorly, and consider how some of the situations can be addressed by refining the implementation. We extend the technique to perform Markov Chains speculatively, expanding the range of algorithms that maybe be accelerated by speculative execution to those with non-uniform move processing times. By simulating program runs we can predict the theoretical reduction in runtime that may be achieved by this technique. We compare how efficiently different architectures perform in using this method, and present experiments that demonstrate a runtime reduction of up to 35-42% where using conventional speculative moves would result in execution as slow, if not slower, than sequential processing.
|Item Type:||Conference Item (Paper)|
|Subjects:||Q Science > QA Mathematics > QA76 Electronic computers. Computer science. Computer software|
|Divisions:||Faculty of Science > Computer Science|
|Journal or Publication Title:||2010 IEEE International Symposium on Parallel & Distributed Processing, Workshops and Phd Forum (IPDPSW)|
|Official Date:||April 2010|
|Page Range:||pp. 1-8|
|Status:||Not Peer Reviewed|
|Conference Paper Type:||Paper|
|Title of Event:||Workshop on Multitheaded Architectures and Applications (MTAAP)|
|Type of Event:||Workshop|
|Location of Event:||Atlanta, Georgia, USA|
|Date(s) of Event:||19-23 April 2010|
Actions (login required)