SIParCS 2015- Srivatsa

Mudambi Srivatsa, University of Utah

Implementation of Radial Basis Function - Finite Difference (RBF-FD) Based PDE Solvers on Multi CPU-GPU Systems

(Slides) (Recorded Talk)

PDE solvers using Radial Basis Function - Finite Differences (RBF-FD) are gaining popularity as they work without the need for underlying meshes to structure nodes. RBF-FD method scales as O(N) with N being the total number of nodes in the domain and also offer high-order accuracy. The goal of the project is to implement RBF-FD based PDE solver on multi CPU-GPU systems and verify it using a two-dimensional advection PDE. As part of the project, single GPU, multi CPU and multi CPU-GPU variants are developed and their performance is evaluated for the maximum problem size of 160k nodes for the advection PDE. In the multi CPU-GPU variant, CUDA targets the GPU while Message Passing Interface (MPI) manages the communication and synchronization on the CPU. The performance is evaluated on commodity CPU-GPU hardware as well as on high performance cluster and the respective speed-ups, scalability results for different implementations.

Mentors: Raj Kumar and Rich Loft, CISL TDD and Natasha Flyer, CISL IMAGe