We want to investigate how to use GPU resources in the ATLAS analysis chain running on the CERN Grid. This required adapting existing code to run on GPUs, make the GPU resources visible to the grid monitor, and interfacing the submission agent so that jobs requesting GPU resources are brokered to the right cluster.
Currently analysis in ATLAS is done by dividing the input data into batches, where each batch is processed sequentially on a number of computation nodes. The computations done are in many cases well suited to run on GPUs, where many identical calculations are repeated for each input event. The possibility to use the GPU resources available could potentially increase the processing efficiency in the ATLAS chain.