Pitstop++: Enabling stable and fair deadlock freedom on fault-tolerant NoC

Abstract

It is essential to design a deadlock-free solution with high performance and low overhead, as a deadlock in NoC systems will result in severe performance degradation or even worse. Traditional deadlock recovery solutions ignore the promptness of deadlock detection and the efficiency of recovery, and they lack fairness and stability when the network is overloaded. In addition, they devote an excessive amount of resources to extremely rare protocol-level deadlocks, resulting in an increase in hardware overhead with little performance improvement. In order to address these issues, Pitstop++ designs a dual-deadlock detection mechanism and a buffer alternating mechanism, which fully considers the timeliness of deadlock detection and the efficiency of recovery. At the same time, Pitstop++ provides fast deadlock detection and efficient deadlock recovery mechanisms for the more common route-level deadlocks, as well as a conservative, low-overhead strategy for the extremely rare protocol-level deadlocks. Compared to Pitstop, Pitstop++ provides an average 51.0% increase in oversaturation throughput and a minimum 44.8% increase in network saturation throughput.