Combining HTM and RCU to Implement Highly Efficient Balanced Binary Search Trees

In this paper we combine Hardware Transactional Memory (HTM) with Read-Copy-Update (RCU) to implement highly scalable concurrent balanced Binary Search Trees (BSTs). The two key features of our approach are: a) read-only operations require no synchronization or restarts and b) tree modifications are first performed in private copies of subtrees, then HTM is used to validate their consistency, and upon successful validation, the copy is installed back in the shared tree by modifying only a single pointer. Our approach can be applied to any type of balanced BSTs (e.g., Red-Black, AVL, B-trees) and for the purpose of this paper we test it in an AVL tree. As our experimental evaluation reveals, the proposed AVL tree implementation achieves higher throughput and scalability compared to previous approaches for parallelizing balanced BSTs with HTM. More specifically on a machine comprising 22 physical cores (44 hardware threads) our tree outperforms other alternatives by 70% and 220%. As shown, this gain is attributed both to the much faster read-only operations and to the reduction in abort ratios thanks to the reduction of transactions’ write-sets.