Abstract

Providing complex (resource-intensive) network services is challenging because the resources they need and the resources that are available can vary significantly from request to request. To address this issue, we have proposed a flexible mechanism, called active frames, that provides a basis for selecting a set of available distributed computing resources, and then mapping tasks onto those resources. As a proof of concept, we have used active frames to build a remote visualization service, called Dv, that allows users to visualize the contents of scientific datasets stored at remote locations. In this paper, we evaluate the performance of active frames, in the context of Dv. In particular, we address the following two questions: (1) What performance penalty do we pay for the flexibility of the active frames mechanism ? (2) Can the throughput of a service based on active frames be predicted with reasonable accuracy from microbenchmarks ? The results of the evaluation suggest that the overhead imposed by active frames is reasonable (roughly 5%), and that simple models based on micro-benchmarks can conservatively predict measured throughput with reasonable accuracy (at most 20%).

Citations