Multiprocessor Real-Time Locking Protocols for Replicated Resources

A real-time multiprocessor synchronization problem is studied herein that has not be extensively studied before, namely, the management of replicated resources where tasks may require multiple replicas to execute. In prior work on replicated resources, k-exclusion locks have been used, but this restricts tasks to lock only one replica at a time. To motivate the need for unrestricted replica sharing, two use cases are discussed that reveal an interesting tradeoff: in one of the use cases, blocking is the dominant lock-related factor impacting schedulability, while in the other, lock/unlock overheads are. Motivated by these use cases, three replica-allocation protocols are presented. In the first two, the lock/unlock logic is very simple, yielding low overheads, but blocking is not optimal. In the third, blocking is optimal (ignoring constant factors), but additional lock/unlock overhead is incurred to properly order lock requests. Experiments are presented that examine the overhead/blocking tradeoff motivated by these protocols in some detail.