Flux Research Group / School of Computing

Deterministic Systems Analysis

Anton Burtsev

PhD Dissertation, University of Utah. May 2013.



A modern software system is a composition of parts that are themselves highly complex: operating systems, middleware, libraries, servers, and so on. In principle, compositionality of interfaces means that we can understand any given module independently of the internal workings of other parts. In practice, however, abstractions are leaky, and with every generation, modern software systems grow in complexity. Traditional ways of understanding failures, explaining anomalous executions, and analyzing performance are reaching their limits in the face of emergent behavior, unrepeatability, cross-component execution, software aging, and adversarial changes to the system at run time.

Deterministic systems analysis has a potential to change the way we analyze and debug software systems. Recorded once, the execution of the system becomes an independent artifact, which can be analyzed offline. The availability of the complete system state, the guaranteed behavior of re-execution, and the absence of limitations on the run-time complexity of analysis collectively enable the deep, iterative, and automatic exploration of the dynamic properties of the system.

This work creates a foundation for making deterministic replay a ubiquitous system analysis tool. It defines design and engineering principles for building fast and practical replay machines capable of capturing complete execution of the entire operating system with an overhead of several percents, on a realistic workload, and with minimal installation costs. To enable an intuitive interface of constructing replay analysis tools, this work implements a powerful virtual machine introspection layer that enables an analysis algorithm to be programmed against the state of the recorded system through familiar terms of source-level variable and type names. To support performance analysis, the replay engine provides a faithful performance model of the original execution during replay.