GPU Architecture Aware Instruction Scheduling for Improving Soft-Error Reliability

Abstract

The demand for low-power and high-performance computing has been driving the semiconductor industry for decades. The semiconductor technology has been scaled down to satisfy these demands. At the same time, the semiconductor technology has faced severe reliability challenges like soft-error. Research has been conducted to improve the soft-error reliability of the GPU, which has been improved by using various methodologies such as redundancy methodologies. However, the GPU compiler has yet to be considered for improving the soft-error reliability of the GPU. In this paper, in order to improve the soft-error reliability of the GPU, we propose a novel GPU architecture aware compilation methodology. The proposed methodology jointly considers the parallel behavior of the GPU hardware and the applications, and minimizes the vulnerability of the GPU applications during instruction scheduling. In addition, the proposed methodology is able to complement any hardware based soft-error reliability improvement techniques. We compared our compilation methodology with the state-of-the-art soft-error reliability aware techniques and the performance aware instruction scheduling. We have injected the soft-errors during the experiments and have compared the number of correct executions that have no erroneous output. Our methodology requires less performance and power overhead than the state-of-the-art soft-error reliability methodologies in most cases. Compilation time overhead of our methodology is 8.13 seconds on average. The experimental results show that our methodology improves the soft-error reliability by 23 percent and 12 percent (up to 64 percent and 52 percent) compared to the state-of-the-art soft-error reliability and performance aware compilation techniques, respectively. Moreover, we have shown that the soft-error reliability of a GPU is not related to the performance, but to the fine-grained timing behavior of an application.