Data distribution on a multi-GPU node for TomoBayes CT reconstruction

Virtual conference (Covid)International audienceComputed tomography (CT) is an imaging technique that uses iterative algorithms to reconstruct the interior of a volume. Graphics Processing Units (GPUs) are currently the preffered technology for reconstruction in CT given their computational performance. In order to have a good reconstruction , the number of images needed is very large. Therefore, the collected data requires several Gigabytes of memory storing and can exceed the internal memory of current GPUs. In this paper, we present a strategy of data distribution between multiple GPUs. We provide experimental results showing that our memory-saving method increases the time of effective computing. We achieve a parallelization efficiency factor of 1 using 8 GPUs to reconstruct a volume of size 4 Gigabytes