``Efficient Distributed Disk Caching in Data Grid Management"

Song Jiang and Xiaodong Zhang

Proceedings of IEEE International COnference on Cluster Computing
(Cluster2003), Hong Kong, China, December 1-4, 2003.

Abstract

Effectively utilizing disk caches is critical for delivering and sharing
data in data-grids considering the large sizes of requested files and
excessively prolonged file transmission time. An essential component
in the disk cache management is its replacement policy that determines
which file(s) are least valuable and should be evicted to create space
for incoming files. Though a large number of replacement algorithms
for data objects of different sizes have been proposed recently in the
domain of Web-caching and disk caching in data grids, they inherit the
shortcomings of the LRU and LFU replacements in characterizing access
patterns.  In order to address this limit, we propose a technique to
measure relative file access locality strength -- how soon a file is to
be re-accessed before being evicted compared with other files.  When we
estimate the in-cache re-access probability, we take the disk space
consumed by accessed files as well as disk cache size into consideration.
Using a relative locality strength estimation, we are able to accurately
rank the value of each file for being cached, and select the file(s)
with least values for replacement.

Our simulation results show that our proposed policy is the most
effective one among existing policies in interpreting access patterns,
and consistently achieves performance improvement measured by hit ratios
and byte hit ratios.

Back to the HPCS Main Page at the Ohio State University.