A Case Study of Server Selection by Tina Tyan

Replication is a commonly used technique to improve availability and performance in a distributed system. Server selection takes advantage of the replicas to improve the end-toend performance seen by users of the system. CFS is a distributed, cooperative file system that inherently replicates each piece of data and spreads it to machines dispersed around the network. It provides mechanisms for both locating and retrieving data. This thesis considers the application of server selection to improving performance in CFS, in both the data location and data retrieval steps. Various server selection metrics and methods were tested in an Internet testbed of 10-15 hosts to evaluate the relative benefits and costs of each method. For the lookup step, we find that the triangle inequality holds with good enough correlation that past latency data stored on intermediary nodes can be used to select each server along the lookup path and reduce overall latency. For the data retrieval step, we find that selecting based on an initial ping probe significantly improves performance over random selection, and is only somewhat worse than issuing parallel requests to every replica and taking the first to respond. We also find that it may be possible to use past latency data for data retrieval. Thesis Supervisor: M. Frans Kaashoek Title: Professor of Computer Science and Engineering