[an error occurred while processing this directive]

Skip to Content

Computer Science Colloquia

Wednesday, April 20, 2011
Vidyabhushan Mohan-Ph.D. Qualifying Examination
Advisor: Sudhanva Gurumurthi, Advisor
Attending Faculty: Kevin Skadron; Jack Davidson; Marty Humphrey

Olsson Hall 236D, 10:30:00

Modeling NAND Flash Memory Reliability

NAND flash based Solid State Disks (SSDs) are fast becoming the choice of primary storage replacing the traditional Hard Disk Drive (HDD) based storage media. The power and performance benefits of SSDs over HDDs are especially attractive for use in data centers, whose workloads are I/O intensive. However, the limited lifetime of SSDs is often cited as an obstacle in adopting them for data centers. Recent studies (including our own at HotStorage 2010) have debunked these claims and have shown that at least write endurance, which is one aspect of flash reliability, is sufficiently high to withstand heavy I/O loads typical of data center workloads. In this project, I have extended my own work on write endurance and studied SSD reliability more broadly, by exploring the impact of Stress Induced Leakage Current(SILC) on the data retention property. Based on prior physics and devices research, I have developed a reliability model framework called FENCE, for studying endurance and data retention property of NAND Flash. FENCE is suitable for use in architecture and system evaluations. Using FENCE, a storage system simulator, and a set of workloads collected from production servers, I have performed a characterization study of how these physical phenomena affect both endurance and data retention property of NAND flash based SSDs. Results from this study indicate that Single-Level Cell (SLC) based SSDs do not have significant endurance or data retention issues while the data retention period of Multi-Level Cell(MLC) based SSDs reduces to about 2.5 years for write intensive workloads after 40,000 P/E cycles.