Distributed large-scale simulation of magnetic resonance imaging.

The concept and the implementation of a parallelized and spin-based simulator for magnetic resonance (MR) imaging is presented. The dynamics of magnetization are modeled using the Bloch equation covering arbitrary radiofrequency (RF) pulses, gradients, main-field inhomogeneity, and relaxation. A temporal decomposition of a given sequence is introduced, leading to basic sequence elements called atoms. A concept of spatial sampling of the object by spins is proposed, in the course of which Shannon's sampling theorem must be respected. In biomedical MR imaging, spins can be modeled as noninteracting entities, permitting an efficient parallelization of the simulation. The simulator ParSpin was implemented on a heterogeneous, interconnected cluster of workstations based on existing message passing libraries. The communication overhead has been kept moderately small. The aggregate computing performance of many processors enables the research into very complex problems (e.g., three-dimensional or steady-state MR experiments requiring up to 10(6) spins). Additionally, ParSpin allows a comprehensive visualization for educational purposes.