Evaluation of Different Methods for Runtime Optimization of Longitudinal Phase Space Tomography

In order to measure the parameters, and therefore the quality, of the particle beam in the CERN synchrotrons, various methods are used for beam diagnostics. One of them is longitudinal phase space tomography, with which the distribution of particles in the phase space can be reconstructed from measured bunch profiles. Due to the large number of parameters and the high number of calculations, the entire process for a single reconstruction usually takes a few seconds. For some applications, a large number of reconstructions are required, so that the order of magnitude can quickly increase from a few seconds to several minutes or even hours. For others, fast reconstructions are desired, in which case several seconds is too long. This work compares different approaches to optimize the runtime of the longitudinal phase space tomography. Thereby, mainly approaches with GPUs are used as well as frameworks for optimizing Python code. The runtimes are compared with the current CPU implementation and a GPU based version is implemented on the basis of the findings. In the end, a faster application is created using CuPy and CUDA C++, reaching speedups between six and 19, depending on the GPU model, and reconstruction runtimes of below 300 milliseconds, while the whole application workflow takes less than two seconds.