InSight: A Framework for Application Diagnosis using Virtual Machine Record and Replay

Non-deterministic execution poses several challenges toward diagnosis—debugging, profiling and execution state mining, of software systems (user-level applications and operating systems). While several techniques using modified libraries, library wrappers, binary instrumentation and memory shadowing techniques exist, we aim to exploit the record and replay technique enabled by virtualization to provide a generalized diagnosis framework. Our solution, is motivated by the requirements of reproducibility of execution, no application source code modification and no or minimal binary instrumentation overheads—all of which are seldom provided by existing techniques. InSight , our diagnosis framework, consists of two stages—the first which records execution state of applications and the virtual machine and the second that replays and analyses the recorded execution. We implement InSight on the Linux kernel-based virtual machine (KVM) platform and as contributions implement an efficient record and replay substrate for KVM and a diagnosis framework using this substrate. This paper describes the design and implementation of these components and develops a set of diagnosis tools—potential deadlock detection, lock usage profiling and function profiling. We also present experiments to demonstrate correctness, the low overheads of InSight and the related diagnosis outcomes.