Concurrent Search Tree by Lazy Splaying

In many search tree (maps) applications the distribution of items accesses is non-uniform, with some popular items accessed more frequently than others. Traditional self-adjusting tree algorithms adapt to the access pattern, but are not suitable for a concurrent setting since they constantly move items to the tree’s root, turning the root into a sequential hot spot. Here we present lazy splaying, a new search tree algorithm that moves frequently accessed items close to the root without making the root a bottleneck. Lazy splaying is fast and highly scalable making at most one local adjustment to the tree on each access. It can be combined with other sequential or concurrent search tree algorithms. In the experimental evaluation we integrated lazy splaying into Bronson et. al.’s optimistic search tree implementation to get a concurrent highly scalable search tree, and show that it significantly improves performance on realistic non-uniform access patterns, while only slightly degrading performance for uniform accesses.