Transaction management in nested transaction systems

Abstract

The notion of nested transactions has been introduced to overcome the deficiencies of single-level transactions. Nested transactions allow intra-transaction parallelism and failure independence. Thus, a nested transaction mechanism must provide proper synchronization and recovery for nested transactions. In this thesis, for complete transacion management in nested transaction systems, concurrency control schemes. deadlock detection schemes, and recovery shceme are proposed. For concurrency control scheme, a new lock transformation shceme for nested transction model which allows parent/child parallelism and commitment independence is porposed. The basic principle of lock transfrrmation scheme is to permit early release of locks, which prevents commit deadlocks incurred by permitting parent/child parallelism. In addition, two locking rules based on the lock transformation scheme, pessimistic locking rule(PLR) and optimistic locking rule(OLR), are proposed. PLR strictly follows the policy of traditional locking rules in nested transations, whereas OLR follows a new policy in favor of a higher concurrency. Then, the performance of two locking rules is studied using a simulation approach. According to the simulation results, OLR is superior to PLR with respect to the number of aborts, the number of deadlocks, and the degree of concurrency. Deadlock detection algorithms based on edge-chasing method for nested trnasactions using two proposed locking rules are proposed. Then, they are refined to reduce the overhead incurred by deadlock detection. The proposed deadlock detection algorithms fully make use of nested structrues of transactions in detecting deadlocks. As a result, in our algorithms, messages do not need to traverse transaction tree to detect deadlocks. This implies that our algorithms are able to detect deadlocks with constant message passing overhead regardless of the depth of nesting in a transaction tree. In addition, the principle of victim selection is...