MPI+CUDA联合加速重力场反演的并行算法

Parallel Algorithm for Gravity Field Inversion Accelerated by MPI and CUDA

针对重力场解算过程中数据量巨大的问题,联合MPI(massage passing interface)与CUDA(compute unified device architecture)提出基于最小二乘法的重力场解算过程的并行加速算法。使用MPI完成复杂过程的任务分配,实现全局层面的并行加速;基于CUDA编写大规模矩阵相乘的并行加速程序,并针对不同类型的矩阵进行适配,同时联合MPI将法矩阵的计算过程进一步细分,实现对分进程内存峰值的压缩。在单机上完成30阶与120阶重力场仿真解算任务,结果表明,反演30阶重力场时加速比可达180;反演120阶重力场时,并行计算单次迭代仅耗时2 h,而串行模式下无法计算。

To address the problem of massive data computation in the process of gravity field calculation,we jointly complete a parallel acceleration algorithm for gravity field calculation based on the least squares method using MPI(massage passing interface)and CUDA(compute unified device architecture).MPI is used to complete the task allocation of complex processes and achieve global-level parallel acceleration.We write a parallel acceleration program for large-scale matrix multiplication based on CUDA and adapted for different types of matrices.At the same time,we use MPI to further subdivide the calculation process of the law matrix to compress the peak memory value within the sub-process.We complete the 30 and 120 order gravity field simulation calculation tasks on a single machine.The results show that when inverting the 30 order gravity field,the acceleration ratio can reach 180;when inverting the 120 order gravity field,parallel computing only takes 2 hours for a single iteration,while it cannot be calculated in serial mode.