High performance ephemeral logging for database systems

We study ephemeral logging in the context of disk-resident database systems and disk-resident logs. We introduce new ephemeral logging algorithms that use system resources much more efficiently than the Extended Ephemeral Logging (XEL) algorithm. Through simulation, we show how our algorithms reduce main-memory usage and increase transaction throughput during normal operation of a transaction processing system.; XEL minimizes the size of a log by compressing it as new recovery data is written to it. XEL partitions the log into a series of logically circular generations and copies longer-lived log data from one generation to another in order to reclaim more quickly the space occupied by shorter-lived log data. As a result of copying, records in the log lose their original ordering; this leads to main-memory and log space overhead for obsolete recovery data.; We introduce the notion of Garbage Removal Dependencies (GRDs) to describe the effects of reordering log records within an XEL-managed log. We use GRDs to classify log records and to characterize main-memory and log space usage in XEL. We introduce a new ephemeral logging algorithm called Temporal Ephemeral Logging (TEL ) that eliminates explicit GRDs by using a non-partitioned log and by preserving the order of dependent log records during copying. As a result, TEL eliminates main-memory and log space overhead for obsolete recovery data.; We study the effects that ephemeral logging has on transaction throughput. We introduce an enhanced ephemeral logging algorithm based on TEL called Cache Ephemeral Logging (CEL). CEL eliminates log reads by relogging log records using data cached in main-memory. This makes all log bandwidth available for writing log data. We show how CEL can use this extra write bandwidth to improve transaction throughput or further reduce the size of the log.