Weighted adaptive concurrency control for software transactional memory

Abstract

Transactional memory programs may have dynamic available parallelism, which is defined as the number of transactions that can be committed concurrently. Prior work presented adaptive concurrency control, which adapts the number of active threads at runtime, and thus the number of concurrently executing transactions, based on available parallelism. Reducing threads when available parallelism is low, and vice versa, improved speedup and reduced wasted work (in aborted transactions). However, prior work did not consider the case where individual threads exhibit dynamic available parallelism. Deactivating threads with low available parallelism, and vice versa, may improve speedup and reduce wasted work further. This paper introduces weighted adaptive concurrency control to exploit the variance in available parallelism between threads. Four algorithms are designed, implemented, and evaluated. They improve speedup and reduce wasted work over prior non-weighted algorithms in applications whose threads exhibit such variance, while maintaining performance parity in applications whose threads do not.