With the end of Moore's Law, many-core platforms have become widely used. These platforms organize their CPUs in a cache-like Non-Uniform Memory Access (NUMA) hierarchy. CPUs are divided in groups sharing a region of cache/memory. Subsequent levels further merge the CPUs into larger groups sharing higher latency memory.

To support scalable concurrent programs, locking primitives for such NUMA systems have been widely studied. The state-of-art solutions however don't consider all aspects of these systems, such as architectural and hierarchical differences between distinct platforms. Furthermore, the correctness of the solutions doesn't consider Weak Memory Models (WMM) present in some platforms.

CLoF, a Compositional Locking Framework, creates NUMA-aware locks optimized for target platforms. It first composes NUMA-oblivious locks into hundreds of NUMA-aware lock combinations and then it selects the best performing lock according to a user-tunable policy.

This talk shows an usage of CLoF for one x86 and one Armv8 platform. CLoF achieves on both platforms twice the throughput of CNA lock - a NUMA-aware lock on the process of merging into Linux. Measuring several different contention levels, CLoF outperforms the most efficient lock found, HMCS, by around 10% on both platforms across all scenarios.

The full paper has been presented on SOSP'21.
Rafael Lourenco de Lima Chehab, Antonio Paolillo, Diogo Behrens, Ming Fu, Hermann Härtig, and Haibo Chen. 2021. CLoF: A Compositional Lock Framework for Multi-level NUMA Systems. In Proceedings of the ACM SIGOPS 28th Symposium on Operating Systems Principles (SOSP '21). Association for Computing Machinery, New York, NY, USA, 851–865. DOI:https://doi.org/10.1145/3477132.3483557.