Error propagation analysis for GPGPU applications

点击翻译按钮获取中文摘要

Hardware errors are increasing due to shrinking feature sizes and manufacturing variations. Conventional techniques to mask hardware errors from the software such as Dual Modular Redundancy are too expensive in terms of power consumption for commodity systems. Therefore, there is a critical need for software assisted techniques for handling hardware faults. This project will explore the use of software-based techniques for providing error resilience to applications running on Graphic Processing Units (GPUs), also known as GPGPU applications. GPUs are finding significant traction in accelerating applications in the data center and high-performance computing (HPC) space, where energy efficiency is vitally important. We will focus on errors that have the maximum impact on reliability to achieve both energy-efficiency and reliability. This research project will benefit Canada in three ways. First, NVIDIA (our industry partner) has an R&D presence in Toronto through a recent acquisition (30 engineers), and the proposed research is important for their long-term business needs. Second, this project lays the foundations of protecting general-purpose software applications from hardware errors. As hardware faults are expected to increase in future processor generations, it becomes essential to protect future software systems from hardware errors. The results of this research can thus be used in a number of critical domains such as aerospace, defence and bio-medical engineering. Finally, the students (HQP) involved in this project will be trained in the development of error resilience techniques for GPGPU applications, which are becoming more and more popular.