high performance computing on graphics processing units: hgpu.org

The Diagrammatic Determinantal Quantum Monte Carlo (DDQMC) algorithm [11, s. III] is used to solve quantum impurity models such as the Anderson model [13]. The topic of this dissertation is the efficient porting of an existing implementation of DDQMC to CUDA in order to use GPUs as accelerators. The main characteristics of quantum impurity models and Monte Carlo methods are introduced with regard to DDQMC, as well as NVIDIA’s GPU architecture and the CUDA programming model. There have been efforts to accelerate a similar algorithm targeting the linear algebra operations in the configuration updates (see [12]). However, within the Monte Carlo loop of the given DDQMC code the sampling of the two point Green’s function comprises the evaluation of multiple bilinears, which makes up for most of the compute time in the serial code and is well suited for concurrent execution on a GPU. The sampling is ported to CUDA and optimised yielding speedups of up to 103.5 times over the unaccelerated version of the code for physically relevant test cases.