Browse Theses

Opportunistic sharing of Internet-connected resources is a low cost method for obtaining access to unprecedented-scale collections of resources. An essential service in any resource-sharing environment is resource discovery: given a description of the resources desired, a resource discovery mechanism returns locations of resources that match the description.

Two resource-sharing environments are particularly well defined by applications, user communities, and deployments: Grid and peer-to-peer systems. Grids are sharing environments that rely on persistent, standards-based service infrastructures that allow well-established, mainly professional communities to share computers, storage space, sensors, software applications, and data across organizational boundaries. Peer-to-peer systems are Internet applications that harness resources from millions of autonomous participants. Thus, Grids provide infrastructure to support a variety of applications on resources shared by relatively small communities; at the scale of the peer-to-peer communities, remarkable sharing patterns are exhibited, such as free riding and intermittent resource participation.

The focus of this dissertation is on solution design for resource discovery in Grids of the scale and lack of reliability of today's peer-to-peer networks. This hybrid target environment requires fully decentralized solutions that scale with the number of users and resources and tolerate intermittent resource participation.

To explore the solution space, we propose a taxonomy for resource discovery solutions. This taxonomy proves to be a useful tool for discussing and comparing existing solutions.

Using this taxonomy, we delimit and explore a portion of the solution space. We build a scalable Grid emulator to evaluate mechanism performance in this subspace. Large-scale experiments reveal that the performance of mechanisms in this subspace is strongly dependent on sharing characteristics.

For inspiration, we turned to studying user behavior in various communities. We uncovered a significant usage pattern in file-sharing communities: users naturally form interest-based groups. This pattern can be exploited for system design in a variety of problems: we designed a file-location mechanism, FLASK, that exploits and benefits from this naturally emerging pattern. Trace-driven evaluations show FLASK leads to lower response latency, good scalability, support for intermittent participation; and satisfies requirements typical of scientific usage of data.