Distributed CPU multi-core implementation of SIRT with vectorized matrix kernel for micro-CT

We describe an implementation of SIRT for execution using a cluster of multi-core PCs. Algorithmic techniques are presented for reducing the size and computational cost of a reconstruction including near-optimal relaxation, scalar preconditioning, orthogonalized ordered subsets, and data-driven focus of attention. Implementation wise, a scheme is outlined which provides each core mutex-free access to its local shared memory while also balancing the workload across the cluster, and the system matrix is computed on-the-fly using vectorized code. Experimental results show the efficacy of the approach. (orig.)