Dependability Analysis of Virtual Memory Systems

Abstract

Recent research has shown that even modern hard disks have complex failure modes that do not conform to "fail-stop" operation. Disks exhibit partial failures like block access errors and block corruption. Commodity operating systems are required to deal with such failures as commodity hard disks are known to be failure-prone. An important operating system component that is exposed to disk failures is the virtual memory system. In this paper, we examine the failure handling policies of different virtual memory systems for different classes of partial disk errors. We use type and context aware fault injection to explore as many of the internal code paths as possible. From experiments, we find that failure handling policies in current virtual memory systems are at best simplistic, and often inconsistent or even absent. Our fault injection technique also identifies bugs in the failure handling code in these systems. The study identifies possible reasons for poor failure handling, which can help in the design of a failure-aware virtual memory system