Key points are not available for this paper at this time.
One of the demands of database system transaction management is to achieve a high degree of concurrency by taking into consideration the semantics of high-level operations. On the other hand, the implementation of such operations must pay attention to conflicts on the storage representation levels below. To meet these requirements in a layered architecture, we propose a multilevel transaction management utilizing layer-specific semantics. Based on the theoretical notion of multilevel serializability, a family of concurrency control strategies is developed. Suitable recovery protocols are investigated for aborting single transactions and for restarting the system after a crash. The choice of levels involved in a multilevel transaction strategy reveals an inherent trade-off between increased concurrency and growing recovery costs. A series of measurements has been performed in order to compare several strategies. Preliminary results indicate considerable performance gains of the multilevel transaction approach.
Gerhard Weikum (Fri,) studied this question.