A Fast Commit Protocol for Distributed Main Memory Database Systems

Although the distributed database systems has been studyed for a long time, there has been only few commercial systems available. The main reason for this is that the distributed commit processing costs too much which results in little or no performance gain compared with single node database system. In this paper, we note the difference in the update and logging policy between disk based database and main memory database in the distributed environmenty and presents a fast distributed commit protocol for the main memory database. In the proposed protocol, instead of sending and receiving two sets of messages one after the other as in two phase commit, only one set of messages are sent after the coordinator completes committing a distributed transaction. The main idea of this fast commit processing is to send all the redo-logs to the coordinator so that the coordinator alone can make the decision to commit or abort when the time comes. As a result, the frequency of the communication and the disk access related to the commit processing can be significantly reduced. Our simulation study shows that the proposed commit protocol achieves the high performance as we expect.