Reducing hot-spot contention in shared-memory multiprocessor systems

In parallel systems it is possible for several processors to request concurrent access to a shared data structure such as a synchronization variable. Such an access pattern causes what is known as hotspot contention. In shared-memory multiprocessor systems that use a multistage interconnection network, hot-spot contention may result in "tree saturation" that degrades the system performance. It is important, therefore, to manage hot-spot contention properly. This paper reviews the existing strategies to reduce the effects of hot-spot contention. We first quantify the effects of hot-spot contention and identify the objectives of a hot-spot management strategy. We propose a taxonomy to categorize these strategies into one of avoidance-based, prevention-based, or detection-based methods. We then review and compare several representative strategies that have been proposed to reduce the effects of hot-spot contention. We conclude the paper by identifying several issues that need further research.