Concurrent access to B-trees

Abstract

The authors give a detailed view of Y. Sagiv's (1985) approach to controlling concurrent access to B/sup link/-trees. It is argued that Sagiv's compression process does not guarantee the 50% load factor as he claimed. This is due to the fact that his compression process treats nodes with a common father pairwise. Furthermore, his compression process sometimes moves information to the left. It is shown here how, as a result of this, readers and updaters may end up too far to the right. This means that processors sometimes have to break off their locate or restructure phase to start again or continue at the root, and that it is necessary to have a low key in each node. In the approach proposed here, the maintenance process does not treat nodes pairwise and, if it moves information, it does so to the right only. Therefore, the present approach does not exhibit the problems associated with Sagiv's approach. Furthermore, it is shown that it has some additional advantages, such as being able to guarantee any desired load factor, offering higher dynamics with respect to deletions and not needing locks in the index set.<>