Transforming a set of office computers to a computing cluster.
Pihlgren, Pontus, Lars Lindbom, Niclas Jonsson
Department of Pharmaceutical Biosciences, Division of Pharmacokinetics and Drug Therapy, Uppsala University, Sweden
Objectives: Population PK/PD model building and evaluation potentially involves methods that use a large number of independent executions of regression software such as NONMEM. These methods may demand substantial amounts of time if the regressions are carried out one after another. Parallel computing environments, such as computer clusters, can reduce this run time markedly by allowing for a simultaneous execution of the regressions. However, a cluster solution using dedicated servers may not be feasible for smaller work groups, where the need for computing power is not enough to motivate the investment. We therefore show how a networked set of office computers easily can be transformed, temporarily or permanently, into a cluster. We also show how a bootstrap using NONMEM can benefit from this type of cluster.
Methods: An office cluster requires three parts: the computers, a computer network and cluster software. Our approach assumed that the computers are in place and connected to the same network and that NONMEM was installed on one of the computers. The Linux distribution clusterKnoppix was chosen as the cluster software . A solution for automatic configuration of NONMEM based on an existing installation was created. In addition, Perl-speaks-NONMEM (PsN) was used to facilitate parallel execution of the bootstrap . The software was packaged on a bootable CD. A test-cluster consisting of 10 computers was set up and evaluated using a bootstrap of 200 samples.
Results and conclusions: The bootstrap shows an approximate gain in speed of up to nine times using the test-cluster, compared to running on a single computer. The bootable CD was easy to use to set up an ad-hoc office computer. cluster and is freely available.