The problem

Nereus, a java-based technology, provides a low-cost supercomputer by aggregating existing desktop computer power. The particular details of the problem have been omitted, however the complexity of the problem is determined by N, the size of the parameter space, given to be 11^N. At present a standard desktop will perform the optimisation calculation in a number of days.

"NereusV has been developed for over 5 years in the Physics Department at Oxford University, and was released to international acclaim last year. With an increasing interest over that time we were pleased to be able to help Willis with this interesting adaptation of their algorithm to run over the distributed Nereus network. This project demonstrated how easy this process was to implement and deploy, delivering dramatic speed improvements to a key business problem. NereusV turns the background latent processing power in a typical modern office into a supercomputer for 100 times lower cost than conventional alternatives," said academic supervisor Rhys Newman, University of Oxford.

The approach

A parallel version of the optimisation code was written in Java. Then, Nereus, a java-based technology to provide a platform for parallel computation, was implemented. Nereus is a grid computing technology developed at Oxford University by Dr. Rhys Newman. Written entirely in Java, Nereus aims to provide a simple platform in which a cluster can built over a network, allowing users to donate their computers' idle time to perform calculations. Since the Nereus cluster can be built using the existing Willis network, comprising of standard desktop computers meaning very few additional resources are required to use Nereus.

The parameter space is broken up into chunks of 100 calculations. One should be aware of two issues when choosing the size of the job. Too small a size would result in the computation time being less than the time it takes a client to download the information required to process the job. Too large of a choice increases the risk of client machines switching off during a calculation.

Consider the test contract defined for N = 5 layers. The number of jobs and the maximum number of computers allowed to be used is [11^5/100] = 1611. The speed up associated with the addition of clients to the Nereus network can be observed and the graph illustrates the theoretical speed up expected and the actual speed up obtained from the experiments.

"Based at the Catastrophe Risk Financing Centre, Smith School of Enterprise and the Environment, University of Oxford, we are working closely with both the private and public sectors to provide multidisciplinary approaches for facilitating decision-making in an uncertain world and to develop sustainable risk management strategies. Through the KTN internship at Willis, we were able to combine techniques from mathematics and computational science to substantially improve the efficiency of risk analysis modelling in the reinsurance sector. The opportunity to work with Willis has been extremely influential with regard to how we disseminate and communicate the results of our research and we look forward to further collaboration through the Willis Research Network," said academic supervisor Patrick McSharry, University of Oxford.