Performance improvements for efficient electromagnetic particle-in-cell computation on 1000s of CPUs

The finite-difference time-domain technique for simulation of electromagnetic and low-density plasma phenomena is computationally expensive and can require tens of thousands of computer hours to produce one solution. Substantial gains can be made through memory streamlining (factors of 2.3x faster), efficient cache usage (factors of 3x improvement), and through better parallel design (improving scalability to four times the number of CPUs). These improvements are documented and tested across five different supercomputing hardware platforms for idealized problems designed to highlight the effect of the changes. Then, the cumulative effect of these changes are tested across the five different systems for a typical problem of interest, a relativistic magnetron, on 48 CPUs which shows a factor of two to seven reduction in run-time, or best case, from 21 h to only 3 h. Index Terms--Electromagnetic, finite-difference time-domain (FDTD), particle-in-cell (PIC).