| In this dissertation we investigate the usage of a known data mining technique, frequent sequence mining, for correlating database accesses in storage systems. We look at I/O prefetching as an application of disk block correlations. We design, implement and compare dynamic and static block correlation mining techniques. For the static techniques, we study schemes with and without correlation rule retraining.; In our experimental evaluation, we use the MySQL database engine and two applications: DBT-2, a TPC-C-like benchmark and the RUBiS auctions benchmark. We perform the mining at the storage level and measure the respective application hit rates in the storage cache. Our results show that, by using block correlations, we can improve the storage cache hit rate by 3-21% for DBT-2 and 5-47% for RUBiS, compared to the baseline. Furthermore, we show that dynamic mining outperforms static mining in terms of higher hit rates and more accurate block correlation rules. |
