A Logging Scheme for Reducing Update Workloads in Flash Storage

Abstract

By caching dirty pages in memory space of a buffering pool, a database system can reduce expensive physical I/O’s required in page updates. If any data page cached has constant updates on itself, it seems to stay long in the buffering pool without flushing-out. Although the existence of such aged dirty pages can reduce the amount of physical updates in storage, it is apt to prolong time taken for recovery procedure after system failure. To prevent such a delayed recovery time, database systems usually take an approach of flushing aged dirty pages in a background mode. Even though the approach may be beneficial in the case of HDD storage, this may not be the case for flash storage because of its high update costs. To solve this problem, we proposed a new logging scheme and a recovery algorithm running with it. Since aged dirty pages in our method are written into a dedicated log file, rather than into data area in storage, we can evade frequent updating of them. To reduce the amount of log data written for that purpose, our logging scheme uses a small size of snapshot log. Since the write of a snapshot log record can put the redo start point forwards, we can guarantee the fast recovery procedure, while reducing the number of page updates. Due to reduced update workloads, our method can improve the overall throughput of flash storage.