A crash recovery algorithm based on multiple logs that exploits parallelism

The paper describes an efficient algorithm for recovery based on multiple logs. The standard WAL algorithm is designed to work with a single log. Since each transaction is required to write a commit record to the log, it imposes limits on the maximum achievable throughput even if group commits are allowed. These limits are well below the needs of high transaction rate systems in airlines and in other areas. Another problem with a `single log´ algorithm is that the tail of the log is a `hot spot´ because of severe contention for it. It is shown that both these problems are solved by extending the write-ahead logging method into a multiple log algorithm which allows transactions to write to one of several logs, and also permits each log to perform checkpoints and I/O operations in parallel. A further advantage of multiple logs is that parallelism can be exploited at recovery time. An algorithm is outlined that speeds up recovery by scanning the multiple logs in parallel