Auto-Tuning for GPGPU Applications Using Performance and Energy Model

• Main Authors: Shih-Meng Teng, 鄧世孟
• Other Authors: Pao-Ann Hsiung
• Format: Others
• Language: en_US
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/44430574859489904355

碩士 === 國立中正大學 === 資訊工程研究所 === 101 === The graphic processing unit (GPU) is a popular accelerator for image processing systems because the algorithms are inherently and massively parallel and the workloads are divisible. GPU was further enhanced into General-Purpose GPU (GPGPU) which can be used for accelerating general applications such as scientific computations through parallelization. Several of the Top 500 list of supercomputers and the Green 500 list of power-efficient computers include millions of GPGPU devices. Distributing workload among the large number of cores in a GPGPU is currently still mostly a manual trial-and-error process. Programmers try out manually some configurations (distributions of workload) and might settle for a sub-optimal one leading to poor performance and/or high power consumption. The state-of-the-art methods for addressing this issue are mainly based on profiling of computation kernels. The work presented in this Thesis has two benefits. First, it proposes a model-based analytic approach for estimating performance and power consumption of kernels such that the estimated values have a high average fidelity. Second, an auto-tuning framework is proposed for automatically obtaining a near-optimal configuration for a kernel computation such that either of the following two optimizations can be performed: (a) the kernel's execution time is almost minimal while satisfying user-given upper bound on energy consumption, (b) the kernel's energy consumption is almost minimal while satisfying user-given upper bound on execution time. The proposed framework formulated the problem as an optimization and solved it using either simulated annealing (SA) or genetic algorithm (GA). Experiment results show that the fidelity of the proposed models for performance and energy consumption are 0.86 and 0.89, respectively. Further, the optimization algorithms result in a normalized optimality offset of 0.94\% and 0.79\% for SA and GA, respectively.