Toward a better understanding of the properties and impact of memory hardware errors

Thesis (Ph. D.)--University of Rochester. Dept. of Electrical and Computer Engineering, 2012In modern computers, memory modules are a major and relatively expensive component. While the literature includes rich coverage of transient memory errors in specially controlled equipment, and moderate coverage of the impact of memory errors on high-level software, it lacks a systematic assessment of computer hardware memory errors from the ground up. This thesis is dedicated to this purpose. It provides first-hand memory error data from industrial production systems, to enhance understanding of the occurrence of the transient errors in production environment, and to gain insight into the mostly neglected intermittent/permanent hardware memory errors. We discovered that transient error rates have been significantly reduced over the years, and non-transient errors are a greater cause of computer failures than transient errors. The thesis also discusses the effectiveness of various protection mechanisms, based on realistic data. We found that Chipkill ECC is extremely effective in protecting against memory errors. The thesis further discusses software susceptibility to those memory errors, by injecting errors into the software systems to validate or invalidate numerous studies in the past. At last, we present a new tainting approach that enabled us to zoom into the error propagation process within the software. Besides the evaluation of memory error impact, the thesis makes equally important contributions to techniques for memory error analysis, powered by novel memory error tracking mechanisms. The technical contributions are two-fold. First, the thesis proposes an emulation-based memory error tracking method that can extend the coverage of tracking beyond conventional symbolic taint tracking by an order of magnitude. Second, it proposes a high-speed realtime tracking method to significantly reduce the cost of online error tracking with up to 4.5 times speed-up. These methods play a crucial part in our error injection studies