Towards Applying Machine Learning to Adaptive Transactional Memory ∗

There is tremendous diversity among the published algorithms for implementing Transactional Memory (TM). Each of these algo-rithms appears to be well suited to certain workloads and architec-tures. However, for programs that operate in distinct phases, exhibit input-dependent behavior, or must run on many different classes of machine, the best algorithm cannot be selected before the program actually runs. We introduce a mechanism for dynamic profiling of a running transactional program, and show how the profile can be used with machine learning techniques to select a TM implemen-tation at run-time. Our preliminary results on the STAMP bench-mark suite show good performance, providing a baseline for future research into adaptivity mechanisms for TM. 1