Mechanisms for Unbounded, Conflict-Robust Hardware Transactional Memory

Conventional lock implementations serialize access to critical sections guarded by the same lock, presenting programmers with a difficult tradeoff between granularity of synchronization and amount of parallelism realized. Recently, researchers have been investigating an emerging synchronization mechanism called transactional memory as an alternative to such conventional lock-based synchronization. Memory transactions have the semantics of executing in isolation from one another while in reality executing speculatively in parallel, aborting when necessary to maintain the appearance of isolation. This combination of coarse-grained isolation and optimistic parallelism has the potential to ease the tradeoff presented by lock-based programming. This dissertation studies the hardware implementation of transactional memory, making three main contributions. First, we propose the permissions-only cache, a mechanism that efficiently increases the size of transactions that can be handled in the local cache hierarchy to optimize performance. Second, we propose OneTM, an unbounded hardware transactional memory system that serializes transactions that escape the local cache hierarchy. Finally, we propose RetCon, a novel mechanism for detecting conflicts that reduces conflicts by allowing transactions to commit with different values than those with which they executed as long as dataflow and control-flow constraints are maintained.