Analytical Modeling for Buffer Hit Rate Prediction

Today's database management systerns (DBMSs) require careful database configuration to achieve optimal performance, while manually tuning the database becomes more and more impractical. The facility of "no knobs operation" enables the DBMS ta autornatically reallocate its resources to maintain acceptable performance in the face of changing conditions. The performance of a DBMS is greatly influenced by the effective use of main memory, especially the buffer area. Buffer pool size tuning is therefore crucial to achieving good performance for a DBMS. It is however a time-consuming process to monitor buffer pool hit rate in order to tune the system. In this thesis research. we use an analytical modeling approach to predict the buffer pool hit rate to simplify the process of estimating hit rate. Since the buffer replacement algorithm determines the buffer hit rate, the model must represent the algorithm. A variation of GCLOCK algorithm is currently used by DBS. We develop a Mrtrkov Chain model of GCLOCK to estimate the hit rates of multiple buffer pools. We evaluate accuracy of the rnodel's estimates with experiments carried out on DB2 with the TPC benchmark.