Robust Adaptation to Available Parallelism in Transactional Memory Applications

Applications using transactional memory may exhibit fluctuating (dynamic) available parallelism, i.e. the maximum number of transactions that can be committed concurrently may change over time. Executing large numbers of transactions concurrently in phases with low available parallelism will waste processor resources in aborted transactions, while executing few transactions concurrently in phases with high available parallelism will degrade execution time by not fully exploiting the available parallelism. Three questions come to mind: (1) Are there such transactional applications? (2) How can such behaviour be exploited? and (3) How can available parallelism be measured or calculated efficiently? The contributions of this paper constitute the answers to these questions. This paper presents a system, called transactional concurrency tuning, that adapts the number of transactions executing concurrently in response to dynamic available parallelism, in order to improve processor resource usage and execution time performance. Four algorithms, called controller models, that vary in response strength were presented in previous work and shown to maintain execution time similar to the best case non-tuned execution time, but improve resource usage significantly in benchmarks that exhibit dynamic available parallelism. This paper presents an analysis of the four controller models’ response characteristics to changes in dynamic available parallelism, and identifies weaknesses that reduce their general applicability. These limitations lead to the design of a fifth controller model, called P-only transactional concurrency tuning (PoCC). Evaluation of PoCC shows it improves upon performance and response characteristics of the first four controller models, making it a robust controller model suitable for general use.