On Scheduling in Distributed Transactional Memory: Techniques and Tradeoffs

Abstract

Data centers have been increasingly employed in distributed services to support a vast of amount of consumer requests. The requests range from web services to gaming for computation intensive applications. In order to process these requests, the data centers exploit in-memory data for high performance and ensure transactional properties for concurrent requests such as atomicity, consistency, and isolation. Traditionally lock-based synchronization has been used for the consistency of data, but is inherently error-prone. For example, coarse-grained locking, in which a large data structure is protected using a single lock is simple and easy to use, but permits little concurrency. In contrast, with fine-grained locking, in which each component of a data structure (e.g., a hash table bucket) is protected by a lock, programmers must acquire only necessary and sufficient locks to obtain maximum concurrency without compromising safety, and must avoid deadlocks when acquiring multiple locks. Both these situations are highly prone to programmer errors.