Architectural support for commutativity in hardware speculation

Abstract

Hardware speculative execution schemes (e.g., hardware transactional memory (HTM)) enjoy low run-time overheads but suffer from limited concurrency because they detect conflicts at the level of reads and writes. By contrast, software speculation schemes can reduce conflicts by exploiting that many operations on shared data are semantically commutative: they produce semantically equivalent results when reordered. However, software techniques often incur unacceptable run-time overheads. To bridge this dichotomy, this thesis presents CommTM, an HTM that exploits semantic commutativity. CommTM extends the coherence protocol and conflict detection scheme to allow multiple cores to perform user-defined commutative operations to shared data concurrently and without conflicts. CommTM preserves transactional guarantees and can be applied to arbitrary HTMs. This thesis details CommTM's implementation and presents its evaluation. The evaluation uses a series of micro-benchmarks that covers commonly used operations and a suite of full transactional memory applications. We see that CommTM scales many operations that serialize on conventional HTMs, such as counter increments, priority updates, and top-K set insertions. As a result, at 128 cores on full applications, CommTM outperforms a conventional eager-lazy HTM by up to 3.4x and reduces or eliminates aborts.