Transactional Memory Scheduling Using Machine Learning Techniques

Current shared memory multi-core systems require powerful software and hardware techniques to support the performance parallel computation and consistency simultaneously. The use of transactional memory results in significant improvement of performance by avoiding thread synchronization and locks overhead. Also, transactions scheduling apparently influences the performance of transactional memory. In this paper, we study the fairness of transactions' scheduling using Lazy Snapshot Algorithm. The fairness of transactions' scheduling aims to balance between transactions types which are read-only and update transactions. Indeed, we support the fairness of the scheduling procedure by a machine learning technique. The machine learning techniques improve the fairness decisions according to transactions' history. The experiments in this paper show that the throughput of the Lazy Snapshot Algorithm is improved with a machine learning support. Indeed, our experiments show that the learning significantly affects the performance if the durations of update transactions are much longer than read-only ones. We also study several machine learning techniques to investigate the fairness decisions accuracy. In fact, K-Nearest Neighbor machine learning technique shows more accuracy and more suitability, for our problem, than Support Vector Machine Model and Hidden Markov Model.