Parallel Performance Of "Synchrotron Radiation Workshop" Code: Partially Coherent Calculations For Storage Rings And Time-Dependent Calculations For Xfels

The "Synchrotron Radiation Workshop" (SRW) computer code is extensively used for the development of insertion devices (IDs) and X-ray beamlines at the National Synchrotron Light Source II and at other light source facilities. Among frequently used types of SRW calculations are the calculations of spontaneous emission from an ID in a storage ring, physical optics based simulations of propagation of this partially-coherent radiation through a beamline, and the simulations of propagation of 3D time-dependent radiation pulses through instruments of X-ray Free-Electron Lasers (XFELs). The two types of radiation propagation calculations are CPU-intensive, therefore for each of them parallel algorithms have been developed in SRW. For the storage ring related calculations, the parallel processing was implemented using the Message Passing Interface (MPI). For the XFEL calculations, a shared memory approach provided by the Open Multi-Processing (OpenMP) was adopted. The two parallelization methods, and their implementation in SRW, have different advantages and drawbacks: the MPI-parallelization of partially-coherent calculations for storage rings has a good scaling, but over-consumes memory, whereas the OpenMP-parallelization of time-dependent XFEL calculations is memory-efficient, but it can only scale within one multi-core server. We are reporting the results of the efficiency tests of these two types of parallel calculations, obtained for representative optical schemes. The tests were performed on an isolated server as well on a large computer cluster - the US DOE's NERSC scientific computing facility.