Abstract:
The Linux 2.6 kernel supports asynchronous I/O as a result of propositions from the database industry. This is a positive evolution but is it a panacea? In the context of the Badger project, a collaboration between MySQL AB and University of Copenhagen, we evaluate how MySQL/InnoDB can best take advantage of Linux asynchronous I/O and how Linux can help MySQL/InnoDB best take advantage of the underlying I/O bandwidth. This is a crucial problem for the increasing number of MySQL servers deployed for very large database applications. In this paper, we first show that the conservative I/O submission policy used by InnoDB (as well as Oracle 9.2) leads to an under-utilization of the available I/O band-width. We then show that introducing prioritized asynchronous I/O in Linux will allow MySQL/InnoDB and the other Linux databases to fully utilize the available I/O bandwith using a more aggressive I/O submission policy.
