2013 IEEE 33rd International Conference on Distributed Computing Systems (2007)
June 25, 2007 to June 27, 2007
Anwitaman Datta , SCE, NTU Singapore
Mike Franklin , EECS, UC Berkeley
Ion Stoica , EECS, UC Berkeley
We propose a new genre of overlay network for disseminating information from popular but resource constrained sources. We call this communication primitive as Latency Gradated Overlay, where information consumers selforganize themselves according to their individual resource constraints and the latency they are willing to tolerate in receiving the information from the source. Such a communication primitive finds immediate use in applications like RSS feeds aggregation. We propose heuristic algorithms to construct LagOver based on preferably some partial knowledge of the network at users (no knowledge slows the construction process) but no global coordination. The algorithms are evaluated based on simulations and show good characteristics including convergence, satisfying peers? latency and bandwidth constraints even in presence of moderately high membership dynamics. There are two points worth noting. First, optimizing jointly for latency and capacity (i.e., placing nodes that have free capacity close to the source) as long as latency constraint of other nodes are not violated performs better than optimizing for latency only. The joint optimization strategy has faster convergence of the LagOver network, and can deal with adversarial workloads that optimization of only latency can not deal with. Secondly, somewhat counter-intuitively, in order to do the aforementioned joint optimization, it is sufficient to find random nodes based on only the latency constraint, since even if the capacity of individual nodes is saturated it does not matter since the LagOver network can potentially be reconfigured.
Overlay Networks, Peer-to-Peer, Heterogeneity, Web 2.0/RSS, Information dissemination, Algorithms
Anwitaman Datta, Mike Franklin, Ion Stoica, "LagOver: Latency Gradated Overlays", 2013 IEEE 33rd International Conference on Distributed Computing Systems, vol. 00, no. , pp. 13, 2007, doi:10.1109/ICDCS.2007.116