Parallelizing a 3D finite difference MT inversion algorithm on a multicore PC using OpenMP

Technology for imaging geophysical properties of the Earth’s interior is increasingly demanding of computing resources as data set sizes increase and the fully three-dimensional (3D) nature of the problem is appreciated. One approach has been to develop distributed computing clusters (e.g., Hargrove et al., 2001), although these can require a substantial investment and facility footprint. An attractive alternative is to exploit multicore PC designs to which single-processor personal computers (PC’s) have evolved; this presents the prospect of parallel computing within an affordable, single-box, format. Here, we describe modifications made to a 3D magnetotelluric (MT) inversion program to allow it to run efficiently on a multicore desktop PC. Parallelization is accomplished using OpenMP, an easy to use application program interface developed for shared-memory platforms, such as multicore PC’s. Autoparallelization, while attractive, is unlikely to yield improvements as large as can be obtained by restructuring and explicit parallelization. We demonstrate that excellent scalability with increasing core number is achieved on important parts of the inverse problem, while particular processor architecture limits efficiency in other parts of the problem.