A Concurrent Skip List Balanced on Search

Abstract

We introduce a skip list, T-list, that updates the index on the search process by recording critical positions in the traverse of the index nodes. T-list uses a step counter to decide when and where to build a new index node and guarantees that the index node is generated in critical position unlike the probabilistic skip list, resulting in an efficient index structure. Meanwhile T-list does not enforce strong constraints to the overall structure unlike the deterministic skip list, thus eliminates a lot of maintaining work. Worst case in T-list can be efficiently repaired by a few requests that traverse the most part of list. Building a new index node in T-list only modifies the contents of two adjacent nodes, enabling the algorithm friendly to concurrent accessing. Experimental results show that compared to the skip list used in a popular application - LevelDB, T-list can construct a more efficient and stable index structure and the insertion and search performances are improved by 17.8% and 33.3% respectively. T-list also scales well with the threads number in the multi-core machine.