Improved Silent Data Error Detection Through Test Optimization Using Reinforcement Learning

Silent Data Errors (SDE) are a serious concern for at-scale computing in data centers because they may cause data loss or data corruption. It is therefore critical to identify defective devices that may experience SDE events and remove them before customer workloads are impacted. Because of the subtle nature of these defects, and the specific combination of events required for SDE to occur, detection is time consuming and costly. This paper reports the results of applying reinforcement learning methods to a family of open-source Eigen tests, and how this work improved the effectiveness of these tests at detecting SDE events.