Exploiting Asynchronous IO using the Asynchronous Iterator Model

Abstract

Asynchronous IO (AIO) allows a process to continue to do other work while an IO operation initiated ear- lier completes. AIO allows a large number of random IO operations to be issued at once, allowing the disk subsystem to order access to data on disk, reducing av- erage seek times considerably, as well as allowing much better utilization of disks in a multi-disk RAID envi- ronments where reads can be done in parallel across disks. In this paper we address the issue of how to extend a database query execution engine to exploit asyn- chronous IO. To best exploit AIO, we propose a new it- erator model called the Asynchronous Iterator Model, where a getnext() call on an operator can return a sta- tus LATER instead of blocking on an IO, permitting other actions to be initiated while an IO is pending. We show how to modify the implementation of Index Nested Loop (INL) join by issuing asynchronous re- quests to a batch of tuple ids. We have prototyped the asynchronous iterator model for INL joins on the Post- greSQL database system, and present performance re- sults that clearly indicate the benefits to be had from exploiting AIO.