Wednesday, 4.11.2015, 11:30
Archival data once written are rarely accessed by user, and need to be
reliably retained for long periods of time. The challenge of using
inexpensive NAND flash to archive cold data was posed recently for
saving data center costs. Solid state drives are faster, more
power-efficient and mechanically reliable than hard drives (HDs).
However, flash of high density is vulnerable to charge leakage over
time, and can only be cost-competitive to HD in archival systems if
longer retention periods (RPs) are achieved. Moreover, the size of
archival data grows exponentially each year, which makes finding the
data we need more difficult.
This talk describes two examples of our on-going research to address
the issues above. We first present the implementation of a coding
technique named rank modulation (RM). RM reads data using the relative
order of cell voltages, and is more resilient to retention errors. We
show that combining RM and memory scrubbing provides more than 100
years of retention period for 1x-nm triple-level cell NAND flash. We
then demonstrate an associative memory framework. The framework
utilizes the random-access capability of flash memory, and solves word
association puzzles with good precision and fast speed using
crowd-sourced data. We show the similarities between puzzle solving
and data retrieval, and discuss our plans on expanding the current
framework for more realistic data retrieval applications.
Yue is a postdoctoral fellow at California Institute of
Technology. His research focuses on algorithms and data
representations for emerging non-volatile memories. Yue worked as a
research intern at LSI in 2013. He received Ph. D. in computer science
from Texas A&M University in 2014.