Recovery in Transaction Processing Systems is a kind of overhead for their performance even if ensures the reliability of databases. In this thesis, we analyze performance issues involved in typical database recovery algorithms in shared everything multiprocessor system running on general purpose operating system. We first adapted conventional logging and shadowing algorithms for shared everything multiprocessor system to utilize parallelism. We investigated the performance and characteristics of these algorithms by event driven simulator SLAMII.
The result shows that logging algorithms is outperformed in shared everything multiprocessor environment. To utilize intra transaction parallelism, a transaction may be splitted a number of processes, which can be processed concurrently. We observed that the optimal degree of intra transaction parallelism is 3.