Parallelization methods for implementation of magnetic induction tomography forward models in symmetric multiprocessor clusters

Abstract

This paper describes three parallelization methods applied to a Magnetic Induction Tomography forward modeller suitable for implementation on computer systems with symmetric multiprocessor (SMP) architecture. The parallelization methods included (i) splitting by different coils using a distributed memory approach, (ii) splitting by physical domain decomposition using both distributed and shared memory and (iii) splitting by both coil and physical domain using distributed and shared memory approaches respectively. All three approaches were implemented on an IBM SP supercomputer. Distributed memory parallelization by coil was the most efficient method due to low inter-processor communication requirements but is limited by the number of coils within the MIT system. Physical domain decomposition allowed a larger number of processors to be employed but the efficiency versus number of processors was found to drop at a significantly faster rate in comparison to coil parallelization. The third approach, splitting by both coil and physical domain also allowed a larger number of processors to be employed but was found to provide higher efficiency than for physical domain parallelization only. This hybrid approach appears to offer an effective parallelization method for implementation of MIT forward models on SMP clusters. The results of forward model computations for an 80*80*80 voxel simulation space using 1 to 128 processors using the three approaches are given. © Springer-Verlag 2007.