A projection-based data partitioning method for distributed tomographic reconstruction

Tomography is a non-destructive technique for imaging the interior of a 3D object. We present an effcient data partitioning strategy for distributed tomographic reconstruction algorithms. Our novel partitioning method is a refinement of the previously published GRCB algorithm. Instead of taking as input a discrete set of lines corresponding to source-pixel pairs, the introduced algorithm works directly on the (coneshaped) projections. We introduce a geometric characterization of the communication volume, as well as a continuous model for load-balancing based on the varying line densities throughout the object volume. The resulting algorithm is orders of magnitude faster than the original algorithm while producing partitionings of similar quality. We introduce a novel communication data structure that can efficiently represent the communication metadata. An implementation on top of Bulk and the ASTRA toolbox is discussed. We provide experimental results of our method for various commonly used acquisition geometries. We achieve a speedup of 2:8x compared to ASTRA-MPI when using 32 GPUs to reconstruct an image for a circular-cone beam acquisition geometry.