Anatomy of a Scalable Software Transactional Memory

Existing software transactional memory (STM) implementations often exhibit poor scalability, usually because of nonscalable mechanisms for read sharing, transactional consistency, and privatization; some STMs also have nonscalable centralized commit mechanisms. We describe novel techniques to eliminate bottlenecks from all of these mechanisms, and present SkySTM, which employs these techniques. SkySTM is the first STM that supports privatization and scales on modern multicore multiprocessors with hundreds of hardware threads on multiple chips. A central theme in this work is avoiding frequent updates to centralized metadata, especially for multi-chip systems, in which the cost of accessing centralized metadata increases dramatically. A key mechanism we use to do so is a scalable nonzero indicator (SNZI), which was designed for this purpose. A secondary contribution of the paper is a new and simplified SNZI algorithm. Our scalable privatization mechanism imposes only about 4% overhead in low-contention experiments; when contention is higher, the overhead still reaches only 35% with over 250 threads. In contrast, prior approaches have been reported as imposing over 100% overhead in some cases, even with only 8 threads.