Thursday, October 30, 2008
Wednesday, October 29, 2008
Resilient Overlay Networks
This paper presents the concept of a Resilient Overlay Network (RON) that provides reliable packet delivery by reducing detection and recovery time following network failures or outages. The authors discuss the scalability advantages that come along with the hierarchical organization of the Internet and then describe some disadvantages in terms of reliable packet delivery. Inter-domain routing with BGP can damp routing updates to minimize oscillations; this distortion of link-state information inevitably leads to suboptimal fault tolerance. RON nodes are deployed in various domains and communicate with each other to provide optimal packet delivery, via either direct Internet paths or other RON nodes.
These application-layer overlay networks are a neat concept and somewhat reminiscent of middleboxes - an intermediary structure to bypass certain shortcomings of the Internet infrastructure. One important point that the authors make is that forwarding via at most one RON node provides sufficient reliability gains. Even so, I'm not sure a) how the modern Internet would react to a heavy influx (not just a small set as in this paper) of RON nodes b) whether inter-domain routing latency and distortion is still a big enough problem today to warrant the use of RONs. Interesting paper, must say.
These application-layer overlay networks are a neat concept and somewhat reminiscent of middleboxes - an intermediary structure to bypass certain shortcomings of the Internet infrastructure. One important point that the authors make is that forwarding via at most one RON node provides sufficient reliability gains. Even so, I'm not sure a) how the modern Internet would react to a heavy influx (not just a small set as in this paper) of RON nodes b) whether inter-domain routing latency and distortion is still a big enough problem today to warrant the use of RONs. Interesting paper, must say.
A First-Principles Approach To Understanding the Internet's Router-level Topology
This paper presents a slightly different perspective on modeling the topology of the Internet. The authors initially point out flaws in the power-laws approach - essentially, several different network models can give rise to similar degree distributions, thereby reducing the effectiveness of power-laws. Instead, a "first-principles" approach with an emphasis on practical network engineering considerations is adopted here. One must identify which factors are fundamental to topology construction rather than simply searching for data-driven power-law analysis. In doing so, router technology and economic considerations must be taken into account.
Using probabilistic methods to generate graphs that model network behavior is a flawed approach as random rewiring between the nodes can damage the performance and efficiency of these models. Interestingly, the authors bring up the point that they have used "toy models" rather than actual network data to illustrate the new modeling approach. Even though they refute this criticism to a degree, there will remain some lack of credibility due to this facet of the research. Unlike the power-laws paper, this one appears to spend more time discussing the flaws of other approaches and not enough on concrete modeling techniques backed by network-based evaluation. Valid points are brought up but the paper feels a bit incomplete.
Using probabilistic methods to generate graphs that model network behavior is a flawed approach as random rewiring between the nodes can damage the performance and efficiency of these models. Interestingly, the authors bring up the point that they have used "toy models" rather than actual network data to illustrate the new modeling approach. Even though they refute this criticism to a degree, there will remain some lack of credibility due to this facet of the research. Unlike the power-laws paper, this one appears to spend more time discussing the flaws of other approaches and not enough on concrete modeling techniques backed by network-based evaluation. Valid points are brought up but the paper feels a bit incomplete.
On Power-Law Relationships of the Internet Topology
This paper addresses the need for accurate models of the Internet topology, for simulation purposes as well as for analyzing and optimizing protocols. Topology can be studied at the intra-domain or inter-domain level using metrics such as node out-degree and inter-node distance. Power-laws have previously been used to describe network traffic but not network topology; probabilistic distance-based models have been very popular for this purpose. The authors aim to identify useful relationships between graph parameters, choosing parameters that can be captured with one number. For example, some chosen parameters are the rank exponent, the out-degree exponent and the hop-plot exponent.
One naturally wonders how such strict relationships emerge in a chaotic network of both cooperative and antagonistic forces such as the modern Internet. The authors anticipate and respond to this quandary with a discussion of the dynamic equilibrium reached in the disorder of the Internet. Several arguments are made in favor of further exploring such power-laws, such as the ubiquity of self-similarity and chaos. I really enjoyed this read and agree that this could be very important for network research going forward. The emergence of these power-laws is reminiscent of Gaussian and Poisson distributions describing various environments in communication and biology - we all know how these mathematical models have guided research in those disciplines.
One naturally wonders how such strict relationships emerge in a chaotic network of both cooperative and antagonistic forces such as the modern Internet. The authors anticipate and respond to this quandary with a discussion of the dynamic equilibrium reached in the disorder of the Internet. Several arguments are made in favor of further exploring such power-laws, such as the ubiquity of self-similarity and chaos. I really enjoyed this read and agree that this could be very important for network research going forward. The emergence of these power-laws is reminiscent of Gaussian and Poisson distributions describing various environments in communication and biology - we all know how these mathematical models have guided research in those disciplines.
Directed Diffusion: A Scalable and Robust Communication Paradigm for Sensor Networks
This paper describes the directed diffusion approach for coordinated sensing and data dissemination in wireless sensor networks. The design of this new paradigm is driven by robustness, scalability and energy efficiency requirements. Data-centric communication in this scheme involves a request for "interests" followed by data propagation and local aggregation. The authors stress that energy efficiency requirements dictate the need for short-range hop-by-hop communication and local computation. Directed diffusion has a few features reminiscent of the TAG service discussed in the previous paper, such as in-network data aggregation and caching.
The evaluation suggests that novel features such as reinforcement-based best path adaptation do indeed provide substantial energy savings, especially given the task-specific nature of modern sensor networks. One interesting point of discussion would be the potential feasibility of the local approach adopted here for a general IP-based network. I found this paper interesting but it is perhaps not as essential in the syllabus as the TAG paper.
The evaluation suggests that novel features such as reinforcement-based best path adaptation do indeed provide substantial energy savings, especially given the task-specific nature of modern sensor networks. One interesting point of discussion would be the potential feasibility of the local approach adopted here for a general IP-based network. I found this paper interesting but it is perhaps not as essential in the syllabus as the TAG paper.
Subscribe to:
Posts (Atom)