Verifying Deadlock-Freedom of Communication Fabrics

Avoiding message dependent deadlocks in communication fabrics is critical for modern microarchitectures. If discovered late in the design cycle, deadlocks lead to missed project deadlines and suboptimal design decisions. One approach to avoid this problem is to get high level of confidence on an early microarchitectural model. However, formal proofs of liveness even on abstract models are hard due to large number of queues and distributed control. In this work we address liveness verification of communication fabrics described in the form of high-level microarchitectural models which use a small set of well-defined primitives. We prove that under certain realistic restrictions, deadlock freedom can be reduced to unsatisfiability of a system of Boolean equations. Using this approach, we have automatically verified liveness of several non-trivial models (derived from industrial microarchitectures), where state-of-the-art model checkers failed and pen and paper proofs were either tedious or unknown.