next up previous contents
Next: Visualisation Up: Computational Issues Previous: Disk space   Contents


Commodity computing

Although it is possible to perform micromagnetic simulations on a local workstation, throughput is limited. Having only one CPU means only one simulation can be effectively performed at one time. When more CPUs are available to use then more simulations can be performed simultaneously, making parameter dependence studies such as phase diagrams practical.

By configuring several powerful local workstations with MPI (Snir et al., 1995), it is possible to perform simulations normally impractical; code which can take advantage of MPI environments such as magpar is capable of using the total available memory of those MPI-enabled machines effectively as one contiguous block -- this allows larger simulations to be performed.

Condor (Litzkow et al., 1988) provides another mechanism for distributing and computing smaller problems. Whereas high-performance commodity computing systems such as Beowulf require dedicated compute resources, Condor is designed to take advantage of the CPU cycles left idle on ``normal'' workstations. Since these workstations are not dedicated, the jobs which run on them generally relinquish their resources when the owner of the workstation returns. Useful results can therefore only be acquired if the jobs which run via Condor are capable of completing in a short time.

Iridis is the University of Southampton's Linux-based clustered computational facility, consisting of several hundred Intel and AMD processors. Time is reserved in advance on this system and scheduling priority organised according to the size of the job in terms of CPU hours and node availability. As a dedicated compute resource it is designed to handle both batch and MPI jobs according to requirements, however competition for CPU cycles means that only a relatively small proportion of the total power can be used by an individual.

By using Iridis for extended studies with several variables, Condor for phase diagrams on smaller samples and optimised local workstation clusters for the most extreme situations, a varied cross-section of results from the different software can be acquired.


next up previous contents
Next: Visualisation Up: Computational Issues Previous: Disk space   Contents
Richard Boardman 2006-11-28