Algorithms and Methodology for Scalable Model Checking


Shaz Qadeer

Model checking algorithms for the verification of reactive systems proceed by a systematic and exhaustive exploration of the system state space. They do not scale to large designs because of the state explosion problem --the number of states grows exponentially with the number of components in the design. Consequently, the model checking problem is PSPACE-hard in the size of the design description. This dissertation proposes three novel techniques to combat the state explosion problem.

One of the most important advances in model checking in recent years has been the discovery of symbolic methods, which use a calculus of expressions, such as binary decision diagrams, to represent the state sets encountered during state space exploration. Symbolic model checking has proved to be effective for verifying hardware designs. Traditionally, symbolic checking of temporal logic specifications is performed by backward fixpoint reasoning with the operator Pre. Backward reasoning can be wasteful since unreachable states are explored. We suggest the use of forward fixpoint reasoning based on the operator Post. We show how all linear temporal logic specifications can be model checked symbolically by forward reasoning. In contrast to backward reasoning, forward reasoning performs computations only on the reachable states.

Heuristics that improve algorithms for application domains, such as symbolic methods for hardware designs, are useful but not enough to make model checking feasible on industrial designs. Currently, exhaustive state exploration is possible only on designs with about 50-100 boolean state variables. Assume-guarantee verification attempts to combat the state explosion problem by using the principle of "divide and conquer," where the components of the implementation are analyzed one at a time. Typically, an implementation component refines its specification only when its inputs are suitably constrained by other components in the implementation. The assume-guarantee principle states that instead of constraining the inputs by implementation components, it is sound to constrain them by the corresponding specification components, which can be significantly smaller. We extend the assume-guarantee proof rule to deal with the case where the specification operates at a coarser time scale than the implementation. Using our model checker Mocha, which implements this methodology, we verify VGI, a parallel DSP processor chip with 64 compute processors each containing approximately 800 state variables and 30K gates.

Our third contribution is a systematic model checking methodology for verifying the abstract shared-memory interface of sequential consistency on multiprocessor systems with three parameters --number of processors, number of memory locations, and number of data values. Sequential consistency requires that some interleaving of the local temporal orders of read/write events at different processors be a trace of serial memory. Therefore, it suffices to construct a non-interfering serializer that watches and reorders read/write events so that a trace of serial memory is obtained. While in general such a serializer must be unbounded even for fixed values of the parameters --checking sequential consistency is undecidable!-- we show that the paradigmatic class of snoopy cache coherence protocols has finite-state serializers. In order to reduce the arbitrary-parameter problem to the fixed-parameter problem, we develop a novel framework for induction over the number of processors and use the notion of a serializer to reduce the problem of verifying sequential consistency to that of checking language inclusion between finite state machines.

Ph.D. thesis, University of California at Berkeley, October 1999, 150 pages.


Download PostScript / PDF document.