|
Abstract : |
This paper evaluates the performance of topology-d, a service for applications that require knowledge of the underlying communication and computing infrastructure in order to deliver adequate performance. Topologyd estimates the state of the network and networked resources by periodically computing end-to-end latency and available bandwidth. Using its estimates, topology-d computes a fault tolerant, high bandwidth, low delay topology connecting participating sites. Topologies are periodically re-computed to take into account network and server load dynamics. Network-aware applications can then make use of topology-d's estimates and logical topologies to ensure they get adequate service from the underlying network and computing infrastructure. We deployed topology-d on 27 Internet sites throughout the world and collected latency, bandwidth and topology information for a period of two and a half months. The results of these Internet-wide experiments show that topology-d's estimates compare quite well with latency and bandwidth measurements from existing tools. We also observe that the logical topologies computed by topology-d are consistent with current latency and bandwidth estimates. The topologies are stable in the sense that the set of neighbors a host sees is small (usually 4 or 5) and stays relatively constant over time. On the other hand, we observe that topologies are responsive to changes in network and server load, as well as in group membership. 1, |
