Jean-François Méhaut

Density functional theory calculation on many-cores hybrid central processing unit-graphic processing unit architectures

We present the implementation of a full electronic structure calculation code on a hybrid parallel architecture with graphic processing units (GPUs). This implementation is performed on a free software code based on Daubechies wavelets. Such code shows very good performances, systematic convergence properties, and an excellent efficiency on parallel computers. Our GPU-based acceleration fully preserves all these properties. In particular, the code is able to run on many cores which may or may not have a GPU associated, and thus on parallel and massive parallel hybrid machines. With double precision calculations, we may achieve considerable speedup, between a factor of 20 for some operations and a factor of 6 for the whole density functional theory code.

Wavelet-Based Density Functional Theory Calculation on Massively Parallel Hybrid Architectures

Publisher Summary This chapter presents an implementation of a full DFT code that can run on massively parallel hybrid CPU-GPU clusters. The implementation is based on the architecture of NVIDIA GPU cards of compute capability at least of type 1.3, which support double-precision floating-point numbers. An overview of the BigDFT code is provided in order to describe why and how the use of GPU can be useful for accelerating the code operations. The set of basis functions used to express the KS orbital is of key importance for the nature of the computational operations that have to be performed. In the BigDFT code, the KS wave functions are expressed on Daubechies wavelets. The latter is a set of localized, real-space-based orthogonal functions that allow for a systematic, multi-resolution description. These basis functions are centered on the grid points of a mesh that is placed around the atoms. The port of the principal sections of an electronic structure code over graphic processing units (GPUs) has been shown. Such GPU sections have been inserted in the complete code in order to have a production DFT code that is able to run in a multi-GPU environment. The DFT code has high systematic convergence properties, very good performances, and excellent efficiency on parallel computation. The data transfers between the CPU and the GPU can be optimized in such a way to allow that more than one CPU core is associated to the same card. This is optimal for modern hybrid supercomputer architectures in which the number of GPU cards is generally smaller than the number of CPU cores.