(A) recovery method in main memory databases

Abstract

Main memory databases store all of the data in main memory, and thus provide higher transaction processing and faster response time than conventional disk-resident databases. However, since main memory is volatile, all or some parts of the database are periodically flushed to disks for reducing the amount of the redo process as well as for preserving data against system crashes. These disk I/O related works, checkpointing and logging, have a great influence on the performance of recovery in main memory databases. So far, recovery methods based on fuzzy checkpointing have been shown to be quite efficient in main memory databases due to its asynchronous backup feature; dirty pages are flushed to the backup databases in disks without considering locks and other transaction activities. By this feature, most works on fuzzy checkpointing in the past have considered only physical logging schemes. This is because the last consistent database state in fuzzy checkpointing is difficult to be reestablished without physical logging. The size of physical log records, in general, is very large, and hence the space overhead as well as the recovery time is significant. Furthermore, due to the huge amount of data in main memory and the high transaction processing rate, the number of dirty pages generated is often very high. In this thesis, we propose two recovery-related schemes: hybrid logging scheme and penultimate fuzzy checkpointing. The hybrid logging scheme allows logical logging to be accommodated under fuzzy checkpointing. So, it significantly reduces the amount of log data, which results in speedup of recovery processing. When applying penultimate concept to fuzzy checkpointing, a frequently updated page are not flushed at every checkpointing. Penultimate fuzzy checkpointing reduces the number of pages to be flushed during checkpointing and makes inter-checkpointing period shorter. As inter-checkpointing period is shorter, the smaller size of log records are generated ...