Multi-GPU Accelerated Finite-difference Time-domain Solver in Open Computing Language

This paper presents results of an evaluation of a novel multi-GPU OpenCL FDTD solver. Multi-GPU implementations are necessary, not only to speed up computations but also to aggregate relatively small GPU memories. The advantage of this solver is its portability between hardware manufactured by different vendors and also between highly specialized and parallel computing architectures available on the market, i.e., GPUs and multi-core CPUs. In our implementation, the computational domain is decomposed along the slowest direction, and electromagnetic field boundary data is shared between neighboring subdomains allocated on different GPUs. The communication overhead between GPUs is proportional to the area of the boundary and represents the rate-limiting step of the method. Portability and efficiency of the developed code were tested on a double CPU machine supported by four GPUs. For the utilized hardware devices, the communication overhead can be hidden by computations for sufficiently large simulation domains, giving scaling efficiency higher than 90%. Porting the OpenCL code dedicated to GPUs directly to multi-core CPUs gives unsatisfactory performance due to the application of architecture specific features in the GPU code. OpenCL kernels of the related FDTD code were therefore optimized to improve performance on multi-core CPUs. Despite this, OpenCL FDTD implementation on CPUs was up to 1.5 times slower than on a commercially available FDTD solver developed in the OpenMP standard. Subsequently, this paper presents an application of the developed multi-GPU OpenCL FDTD code in solving a real-life problem of computational electromagnetics in order to demonstrate its performance and applicability.