Autonomous analysis of interactive systems with self-propelled instrumentation
Alexander V. Mirgorodskiy, Barton P. Miller · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2005
Finding the causes of intermittent bugs and performance problems in modern systems is a challenging task. Conventional profilers focus on improving aggregate performance metrics in an application and disregard many problems that are highly visible to users but are deemed statistically insignificant. Finding intermittent bugs is also hard -- breakpoint debuggers change the timing of events, often masking the problem. To address these limitations, we propose a novel approach called self-propelled instrumentation -- using an autonomous agent to perform self-directed exploration of the system. We inject the agent into a running application, and the agent starts propagating through the code, carried by the application's flow of control. As it propagates, it inserts instrumentation dynamically to collect and analyze detailed execution information. The key feature of this approach lies in its ability to meet three requirements: high level of detail, low overhead, and autonomy (here, little reliance on human help). Existing techniques for tracing and profiling violate at least one of the requirements. As a proof of concept, we implemented a tool called spTracer that uses self-propelled instrumentation to obtain function-level traces from applications and the kernel. The tool enabled us to locate and fix three problems in unfamiliar code in the Linux environment.