Distributed Deadlock

This theme explores formal modeling and control techniques for deadlock prevention in distributed systems using Petri nets, particularly Gadara nets and siphon/trap methods. Preventing deadlocks in multithreaded and multi-agent manufacturing systems requires understanding resource allocation dynamics and employing efficient supervisory control policies that guarantee system liveness and maximal permissiveness. Petri net models allow structural analysis to identify problematic siphons or resource allocations that lead to deadlocks, which can be constrained with monitors or supervisor places to enable live operation without deadlocks.

With the emergence of fast RDMA networks, distributed lock managers must balance scalability, latency, starvation prevention, and deadlock handling while avoiding centralized bottlenecks. This theme investigates decentralized lock management algorithms that achieve starvation-free, first-come-first-serve scheduling without explicit queues, exploiting RDMA atomic primitives (fetch-and-add) rather than compare-and-swap. It evaluates performance improvements in throughput, latency, and fairness essential to modern distributed transactional systems.

Effective distributed deadlock handling requires balancing detection frequency to minimize combined overhead of detection message complexity and resolution duration. This theme studies stochastic and analytical models identifying optimal deadlock detection intervals as functions of system parameters like deadlock formation rate and process count. It also explores algorithms for efficient detection and resolution of generalized deadlocks with minimal message, time complexity, and streamlined termination detection.