Purge-Rehab: Eager Software Transactional Memory with High Performance Under Contention

摘要

Transactional memory is a programming model that attempts to make parallel programming easier. Transactional memory uses either eager (at encounter time) or lazy (at commit time) validation to check for conflicting accesses between concurrent transactions, and researchers have been divided over which approach is best. Despite this, there is little work in the literature evaluating their comparative performance. One exploration of this topic using microbenchmarks found that lazy outperforms eager and increases its advantage as contention rises. Furthermore, eager was prone to livelock when accesses were irregular, which resulted in starvation and low throughput. We revisit the comparative performance of eager and lazy under contention using a larger set of microbenchmarks, more complex benchmarks from STAMP, and against multiple state-of-the-art STMs: word-based RSTM, TinySTM and SwissTM. We validate earlier findings that eager suffers from livelock, especially when accesses are irregular. This inspired Purge-Rehab: a lightweight mechanism for controlling transaction restarts in eager validation to reduce livelock and thus improve throughput and reduce starvation. Compared to lazy validation, Purge-Rehab achieves higher performance in five benchmarks, similar in four, and is lower in only one, showing that eager validation can achieve high performance under contention. Purge-Rehab is implemented in word-based RSTM, but is applicable to any eager STM.