PlanetLab: an overlay testbed for broad-coverage services
ACM SIGCOMM Computer Communication Review
INFOCOM'96 Proceedings of the Fifteenth annual joint conference of the IEEE computer and communications societies conference on The conference on computer communications - Volume 2
High-bandwidth data dissemination for large-scale distributed systems
ACM Transactions on Computer Systems (TOCS)
Everlab: a production platform for research in network experimentation and computation
LISA'07 Proceedings of the 21st conference on Large Installation System Administration Conference
Compact samples for data dissemination
Journal of Computer and System Sciences
Journal of Network and Computer Applications
| Hi-index | 0.00 |
Peer-to-peer content distribution networks are currently being used widely, drawing upon a large fraction of the Internet bandwidth. Unfortunately, these applications are not designed to be network-friendly. They optimize download time by using all available bandwidth. As a result, long haul bottleneck links are becoming congested and the load on the network is not well balanced. In this paper, we introduce the Julia content distribution network. The innovation of Julia is in its reduction of the overall communication cost, which in turn improves network load balance and reduces the usage of long haul links. Compared with the state-of-the-art BitTorrent content distribution network, we find that while Julia achieves slightly slower average finishing times relative to BitTorrent, Julia nevertheless reduces the total communication cost in the network by approximately 33%. Furthermore, the Julia protocol achieves a better load balancing of the network resources, especially over trans-Atlantic links. We evaluated the Julia protocol using real WAN deployment and by extensive simulation. The WAN experimentation was carried over the PlanetLab wide area testbed using over 250 machines. Simulations were performed using the the GT-ITM topology generator with 1200 nodes. A surprisingly good match was exhibited between the two evaluation methods (itself an interesting result), an encouraging indication of the ability of our simulation to predict scaling behavior.