Speculative Versioning through Perceptron Predictors

A well-know method to avoid inconsistent state in Software Transactional Memory (STM) is a globally shared version clock whose values are used to tag memory locations. While this method does not require frequent validation of transactional data, it results in contentions over the global clock. Each time that a transaction commits it updates the global clock which results in costly coherence misses. The alternative approach is local clock which requires access to local variables instead of a global version clock. However, as we show in this paper, the optimum validation policy changes not only across applications but also within an application and through different phases of a program. To counter this challenge, we introduce Speculative Versioning (SV) which dynamically selects one of the two validation techniques based on probability of conflicts. SV is a speculative approach and relies on perceptron predictors to predict future conflicts. We have incorporated SV into TL2 and compared the performance of the new implementation with the original STM using Stamp v0.9.10 benchmark suite. Our results reveal that SV is effective and improves execution time of transactional applications up to 31%.