Optimized Software-Based Hardening Strategies for Matrix Multiplication and Fast Fourier Transform
Zhijian Hui, Yangsheng Wang, Tao Qin, Honghui Tang, Haibin Wang · 2018
Nowadays, Graphics Processing Unit (GPU) has shown great potential in High-Performance Computing applications for its parallel computing structures, which can greatly accelerate the computing process. However, GPU reliability is critical in some applications like satellite or auto-driving. Thus, many researches have been carried out to improve its reliability using both hardware and software schemes. In this paper, we mainly focus on designing schemes to protect devices from soft errors. We analyze the performance of available algorithm-based fault tolerance (ABFT) schemes for two commonly used mathematical operations: Matrix Multiplication and Fast Fourier Transform (FFT). We developed optimized schemes to reduce the overhead of the available ones. The proposed schemes are based on two insights. First, in matrix multiplication, the scheme may classify some correct results as potential errors when there are multiple errors, introducing unnecessary overhead. Second, in FFTs, the overhead of the ABFT scheme depends on its size, so we can use butterfly module to detect errors and correct them. Finally, we use fault-injection simulation to evaluate the proposed schemes and in the simulation results they are proved to perform quite well in any error distributions.