Control versus data flow in parallel database machines

摘要

The execution of a query in a parallel database machine can be controlled in either a control flow way, or in a data flow way. In the former case a single system node controls the entire query execution. In the latter case the processes that execute the query, although possibly running on different nodes of the system, trigger each other. Lately, many database research projects focus on data flow control since it should enhance response times and throughput. The authors study control versus data flow with regard to controlling the execution of database queries. An analytical model is used to compare control and data flow in order to gain insights into the question which mechanism is better under which circumstances. Also, some systems using data flow techniques are described, and the authors investigate to which degree they are really data flow. The results show that for particular types of queries data flow is very attractive, since it reduces the number of control messages and balances these messages over the nodes.