Software failure avoidance using discrete control theory

• Authors: Stephane Lafortune;Yin Wang
• Venue: Software failure avoidance using discrete control theory
• Year: 2009

Software reliability is an increasingly pressing concern as the multicore revolution forces parallel programming upon the average programmer. Many existing approaches to software failure are ad hoc, based on best-practice heuristics. Often these approaches impose onerous burdens on developers, entail high runtime performance overheads, or offer no help for unmodified legacy code. We demonstrate that discrete control theory can be applied to software failure avoidance problems. Discrete control theory is a branch of control engineering that addresses the control of systems with discrete state spaces and event-driven dynamics. Typical modeling formalisms used in discrete control theory include automata and Petri nets, which are well suited for modeling software systems. In order to use discrete control theory for software failure avoidance problems, formal models of computer programs must first be constructed. Next, control logic must be synthesized from the model and given behavioral specifications. Finally, the control logic must be embedded into the execution engine or the program itself. At runtime, the provably correct control logic guarantees that the given failure-avoidance specifications are enforced. This thesis employs the above methodology in two different application domains: failure avoidance in information technology automation workflows and deadlock avoidance in multithreaded C programs. In the first application, we model workflows using finite-state automata and synthesize controllers for safety and nonblocking specifications expressed as regular languages using an automata-based discrete control technique, called Supervisory Control. The second application addresses the problem of deadlock avoidance in multithreaded C programs that use lock primitives. We exploit compiler technology to model programs as Petri nets and establish a correspondence between deadlock avoidance in the program and the absence of reachable empty siphons in its Petri net model. The technique of Supervision Based on Place Invariants is then used to synthesize the desired control logic, which is implemented using source-to-source translation. Empirical evidence confirms that the algorithmic techniques of Discrete Control Theory employed scale to programs of practical size in both application domains. Furthermore, comprehensive experiments in the deadlock avoidance problem demonstrate tolerable runtime overhead, no more than 18%, for a benchmark and several real-world C programs.