The Optimization Of Accelerator Toolbox With Parallelization Methods

Due to the development of high energy physics and synchrotron radiation science, theaccelerators are becoming larger and more complex. There are thousands of elements in modernaccelerators. To study beam dynamics or design an accelerator requires extensive calculations.Furthermore, magnet field errors contributing nonlinear effects cannot be given analytically, whichwill have a great effect on the performance of the accelerator. In order to optimize acceleratorperformance, particle tracking is needed. For example, when calculating the dynamic aperture,different set of errors of the magnetic field data decomposed as a Taylor series added to the mainfield. Repeatedly, by checking the field error distributions, to evaluate their effects on dynamicaperture shrinkages.In modern high-energy circular accelerators, strong focusing quadrupole structure iscommonly used. It will lead to large chromaticity. To correct the chromaticity to near zero strongsextupoles are used. However sextupoles will increase the nonlinear effect in the accelerator anddecrease the dynamic aperture. It will lead shorter beam lifetime and lower injection efficiency.Using particle tracking results to determine the optimized sextupole setting is heavy acceleratordeveloping jobs. For a large Accelerator, these calculations require lots of computer time. Particletracking code is an efficient tool for beam dynamic study. By improving its efficiency, reducingprogram running time, the research cycle can be greatly shortened, faster adjustments can be made.Using parallel computing methods can significantly improve the speed of the program.Parallel computing methods can be divided into two kinds: GPU parallel computing and multi-core CPU parallel computing. This paper analyzes the Accelerator Toolbox (AT) program’srunning processes and tests a GPU parallel computing based on CUDA, hybrid programming andparallel computing based on OpenMP-MPI. Then the OpenMP-MPI method is taken as thebasement to build a multi-core compute platform for accelerator physics calculation.Computing platform integrates message passing and shared memory features. The underlyinglayer uses shared memory (SMP) in computing. The top layer uses non-uniform memory access(NUMA) in computing. Leading a node levels paralleling. The process-level paralleling is handledby MPI and the thread-level part is handled by OpenMP. Theoretical and numerical analysis showsthat OpenMP-MPI hybrid parallel structure combines well with AT program, by avoiding theinternal communication. The improved load balancing improves the efficiency. The platform isextensible. Platform’s computing speed can be increased by adding compute nodes to achieve greater goals. The computing speed grows linearly with the number of computing nodes. In theend, this paper summarized the work, as well as further way to increase the computing speed.