Modeling the performance of algorithms on flash memory devices

Abstract

NAND flash memory is fast becoming popular as a component of large scale storage devices. For workloads requiring many random I/Os, flash devices can provide two orders of magnitude increased performance relative to magnetic disks. Flash memory has some unusual characteristics. In particular, general updates require a page write, while updates of 1 bits to 0 bits can be done in-place. In order to measure how well algorithms perform on such a device, we propose the "EWOM" model for analyzing algorithms on flash memory devices. We introduce flash-aware algorithms for counting, listmanagement, and B-trees, and analyze them using the EWOM model. This analysis shows that one can use the incremental 1-to-0 update properties of flash memory in interesting ways to reduce the required number of page-write operations.