Parallelizing using SMP techniques requires the processors to be connected to a shared memory via a common bus. Normally Linux clusters have a distributed memory system. With regards message passing (I've only used MPI) there would really not be much computational benefit. Because each voxel is analysed separately, the number of messages passed would be very few. It would be possible to separate each dataset into 20 pieces and then recombine it at the end, but I don't think there is much point. Data is normally transferred across a slower network than messages and so this could end up costing time rather than gaining. If you had twenty subjects, it would be quicker to just analyse one on each node.