Reliable and Accurate Fault Detection with GPGPUs and LLVM

As the scale and complexity of cloud systems are increasing, system faults are becoming unavoidable. Therefore, they should be detected as reliably and accurately as possible. Black-box monitoring can reliably monitor a target system from a remote host, but it is often coarse-grained and cannot identify the root causes of system faults. In contrast, white-box monitoring can accurately obtain fault information inside a target system, but it is largely affected by system faults. This paper proposes GPUSentinel for more reliable white-box monitoring using general-purpose GPUs. GPUSentinel runs fault detectors in an isolated GPU, which is not easily affected by faults of a target system. For accurate detection, fault detectors in a GPU analyze main memory and directly monitor the state of the operating system. To easily develop such fault detectors, GPUSentinel provides a development environment with LLVM. We have implemented GPUSentinel and seven fault detectors and then confirmed that GPUSentinel could detect various system faults and identify the root causes.