Exploring Social Networks in Computer Systems

Abstract:

Since the first appearance of computers half a century ago, computing
technology continues to evolve rapidly. Having the most powerful and fastest
computers is no longer the only goal for a successful computer system. A
common characteristic of successful computer systems in the modern era is
the ability to utilize vast amounts of information efficiently.

We introduce two important concepts from sociology into computer science:
human capital and social capital. In a society, people who possess high
human capital are individuals with more ability. They are more intelligent,
more attractive, more articulate, and more skilled. In the human capital
metaphor, these people are more likely to do better than others. Social
capital is the contextual complement to human capital. In the social capital
metaphor, people who do better are somehow better connected. The asset they
possess is the advantageous location in the society. Today's computer
systems exhibit a strong resemblance to human society. The linkage
relationships among nodes in a system affect the performance of the system
to a large extent. A well-connected system can outperform a collection of
powerful nodes.

In this dissertation, we present results of applying the social capital
metaphor to two kinds of computer systems: focused web crawling and
peer-to-peer systems. We have designed and implemented a novel focused
crawler that captures the topical linkage in the web. We have introduced the
notion of "topical link". Topical links connect pages with similar content.
By combining topical links with hyperlinks, the connectivity of the web
graph is greatly enhanced. In our study of peer-to-peer systems, we have
introduced the concept of "buddy net". Buddy net connects peers with similar
interests in the peer-to-peer systems. It is an interest-based overlay on
top of the physical overlay in peer-to-peer systems. With the help of buddy
net, the efficiency and robustness of a peer-to-peer system is greatly
improved.