Efficiency optimizations for implementations of deadlock immunity

Abstract

Deadlock immunity is a property by which programs, once afflicted by a deadlock, develop resistance against future occurrences of that deadlock. Our deadlock immunity system, called Dimmunix, provides transparent immunization against deadlocks involving mutex locks. In this paper, we focus on efficiently protecting systems against deadlocks regardless of the rate of synchronization operations performed. We describe five optimizations that reduce the runtime overhead imposed by Dimmunix on the host system: (1) offline deadlock detection and signature extraction, which avoids runtime tracking of lock-to-thread allocations; (2) selective program instrumentation, whereby only vulnerable synchronization statements are monitored; (3) inline matching of deadlock signatures, which avoids expensive call stack retrieval; (4) false positive reduction, which avoids unnecessary thread serialization; and (5) safe early resumption of threads, allowing suspended threads to resume their execution more quickly than in the original Dimmunix. Our optimizations enable Dimmunix to achieve a reduction of 2.8x-5.2x in the runtime overhead it introduces for real-world systems like Eclipse, Vuze, and MySQL JDBC. © 2012 Springer-Verlag.