Document Type

Technical Report

Publication Date

2006-08-01

Filename

wucse-2006-43.pdf

DOI:

10.7936/K7XD101K

Technical Report Number

WUCS-2006-43

Abstract

The continuous growth in the Internet’s size, the amount of data traffic, and the complexity of processing this traffic gives rise to new challenges in building high-performance network devices. One of the most fundamental tasks performed by these devices is searching the network data for predefined keys. Address lookup, packet classification, and deep packet inspection are some of the operations which involve table lookups and searching. These operations are typically part of the packet forwarding mechanism, and can create a performance bottleneck. Therefore, fast and resource efficient algorithms are required. One of the most commonly used techniques for such searching operations is the Ternary Content Addressable Memory (TCAM). While TCAM can offer very fast search speeds, it is costly and consumes a large amount of power. Hence, designing cost-effective, power-efficient, and high-speed search techniques has received a great deal of attention in the research and industrial community. In this thesis, we propose a generic search technique based on Bloom filters. A Bloom filter is a randomized data structure used to represent a set of bit-strings compactly and support set membership queries. We demonstrate techniques to convert the search process into table lookups. The resulting table data structures are kept in the off-chip memory and their Bloom filter representations are kept in the on-chip memory. An item needs to be looked up in the off-chip table only when it is found in the on-chip Bloom filters. By filtering the off-chip memory accesses in this fashion, the search operations can be significantly accelerated. Our approach involves a unique combination of algorithmic and architectural techniques that outperform some of the current techniques in terms of cost-effectiveness, speed, and power-efficiency.

Comments

Permanent URL: http://dx.doi.org/10.7936/K7XD101K

Share

COinS