LosaTM: A Hardware Transactional Memory Integrated with a Low-Overhead Scenario-Awareness Conflict Manager

Chao Fu,Li Wan,Jun Han
DOI: https://doi.org/10.1109/tpds.2022.3206777
IF: 5.3
2022-01-01
IEEE Transactions on Parallel and Distributed Systems
Abstract:The vigorous development of high compute-intensive applications has led to the demand for maximizing the concurrency of multicore processors. The best-effort hardware transactional memory(HTM) is an important technology adopted by vendors to improve the potential concurrency of multicore processors, but the HTM implementations on commercial products have some drawbacks for its simplicity and need some further optimizations to enable more exploitation of concurrency. In this article, we propose and evaluate a novel design of HTM, called LosaTM, which can provide a scenario-awareness conflict management strategy. By leveraging the proposed feature of multiple-grained coherency maintenance in the coherence protocol, LosaTM resolves most false conflicts at a half-cache-line granularity. Furthermore, we design a winner/aborter vector conflict management algorithm to improve the efficiency of LosaTM in handling friendly-fire and unfairness competition that we have newly defined. In order to coordinate these integrated conflict management strategies, a scheduling strategy is also proposed to adaptively select the appropriate management according to the specific conflict scenario. We use gem5 to simulate LosaTM in detail on an 8-core tiled CMP system, and the simulation result shows that it only causes 0.7% of the L1 cache size hardware overhead while achieving a 38% average execution time reduction on the native STAMP. The speedup also demonstrates that LosaTM outperforms the state-of-the-art designs in previous works.
What problem does this paper attempt to address?