Many research results indicate that pages reference behavior in a database buffer is quite different from the behavior in a program environment, when the measurement unit is the page faults. However, the I/0 time (which includes the seek time and rotation time on the disk, and transfer time between the main memory and disk) is generally the dominant factor in the performance of a buffer manager. In this study, we use the disk-oriented I/0 time as the performance metric to evaluate five popular page replacement algorithms. These five algorithms include three popular algorithms: FIFO, RANDOM, and LRU in operating systems and two particular algorithms CLOCK and GCLOCK [5] designed for database management systems.
Under various multi-transaction enviroments, these five algorithms are first evaluated; CLOCK, in general, has the least disk-oriented I/0 time. To reduce disk-oriented I/0 time further, an intelligent scheduler is designed. The scheduler has the features of a deferred I/0 per reference and the SSTF disk scheduling policy. With the scheduler, the simulation results show that the average disk-oriented I/0 time can be reduced to around 2% for CLOCK and 25% for RANDOM, which is than the conventional non-scheduling method. By comparing the methods without/with the scheduler, CLOCK with the scheduler generally has the least disk-oriented I/0 time.