A Survey on Performance Modelling and Optimization Techniques for SpMV on GPUs

Sparse Matrix is a matrix consisting of very few non-zero entries. Large sparse matrices are often used in engineering and scientific operations. Especially sparse-matrix vector multiplication is an important operation for solving linear system and partial differential equations. However, there is a possibility that even though the matrix is partitioned and stored appropriately, the performance achieved is not significant. Hence a need arises to address these issues. System proposes an integrated analytical and profile based performance modelling that accurately predicts the kernel execution time of various SpMV CUDA kernels and also that of a given target sparse-matrix. Based on this the designed optimal solution auto-selection algorithm automatically reports the SpMV optimal solution for a target sparse-matrix. System was evaluated on NVIDIA Tesla C2050 and significant results were obtained. Proposed system would like enhance the existing system by trying the same on different SpMV CUDA kernels as well look for optimization. Proposed system would also like to try and execute the same on multi-GPU kernels. Proposed system would also like to evaluate the existing system on other NVIDIA GPU such as the NVIDIA GeForce GT 750M card. This paper presents a survey of various performance modelling and optimization techniques for SpMV CUDA kernels on GPUs. It also presents a survey of the various SpMV CUDA kernel implementation techniques. Keywords—SpMV, GPU, CUDA, performance modelling, optimization.