Path conscious caching of B~+ tree indexes in a shared disks cluster

摘要

A shared disks (SD) cluster is a representative architecture to couple multiple computing nodes for high performance transaction processing. Every node in the SD cluster shares a common database at the disk level. To avoid frequent disk accesses, each node caches both data pages and index pages in its local memory buffer. In general, index pages are accessed more often and thus cached at more nodes than their corresponding data pages. Furthermore, tree-based indexes such as B~+ trees require complicated operations, e.g., root-to-leaf traversal and structure modification operation (SMO) due to a page split or a page concatenation. This means that it is strongly required to devise a dedicated cache coherency scheme (CCS) for index pages that takes advantage of the semantics of their access. In this paper, we propose a new CCS for B~+ tree indexes in the SD cluster. The proposed scheme can reduce the message traffic between nodes and the number of tree re-traversals. An extensive performance study shows that the proposed scheme exhibits substantial performance improvement over the previous schemes, especially when there are large number of nodes or the update probability is high.