Tuesday, January 15, 2013
Week 2 Papers
Rethinking the design of the Internet: The end to end arguments vs. the brave new world by David D. Clark and Marjory S. Blumenthal
In this paper the authors discussed the possibility the end to end principle could be compromised.
In this paper which is more than 10 years old, the authors discussed on some scenarios that implementing an application-level function in the core of the network is necessary. Some of the arguments have been proved true now. For example, as the authors said, the Internet today is an untrustworthy world, manufacturers such as Cisco have selling devices such as Cisco Spam & Virus Blocker. CDN serve content to end-users with high availability and high performance. CDNs serve a large fraction of the Internet content today so that delivery of data is accelerated.
On the other hand, some other problems have already been solved. For example, today ISPs are becoming tunnels for data, and ISP service differentiation is almost unnoticeable.
This paper successfully foresaw a more sophisticated Internet in which application-level function is implemented in the core of the Internet.
Consensus Routing: The Internet as a Distributed System by J. John, E. Katz-Basset, A. Krishnamurthy, T. Anderson and A. Venkataramani
This 2008 paper proposed to improve the consistency of the network to reduce routing loops and blackholes.
In this paper the authors proposed running a distributed coordination algorithm to ensure that a route is adopted only after all dependent routers have agreed upon a globally consistent view of global state in order to reduce loops and blackholes. In the meantime, a small fraction of the packets are heuristically forwarded to find a route if a safe route is not found yet. This is called a transient mode the function.
One concern of this paper in my eyes is that if the network is not very reliable, the mechanism might have to update the consistency information very frequently, and the cost for sending and receiving these packets might be high.
This paper reminds me of OpenFlow, which is an open standard that enables researchers to run experimental protocols in the campus networks we use every day. We could leverage OpenFlow to try more optimized protocols.
Subscribe to:
Post Comments (Atom)
I would like to talk about the paper "Consensus Routing: The Internet as a Distributed System". This is a very great paper, and it is, concisely, about how to effectively improve packet delivery performance between routers and ASes. One thing I would like to point out is that, as Chenjin mentioned, in the case of frequent link failures, forming SFT might take a while and this brings two concerns; 1)Frequent link failures implies frequent epochs and it means not having unreliable paths more often than not. Although this implies system to go on transient mode, transient mode itself is only compared to BGP, it would be better to see how better or worse it is than other protocols mentioned in related work. 2)A node failing to respond to message exchange during any k-th phase, as said in the paper, will be removed from the network. However, this itself will cause another update in the network, so it may lead to successive updates.
ReplyDeleteAnother remark I would like to make is about security. Even though authors make a great job about answering security concerns, I am not convinced how system would react to a malicious router that advertises false link breakage and/or disrupts trigger exchanges by not completing link updates on purpose.
Lastly, I really did not understand how only 3000 (out of 25000) can participate the protocol while being loop-free.
In response to your question about how only 2500 ASes need to participate in the protocol for the network to be loop free; I believe the explanation was that only 3000 ASes are "transit" ASes. My understanding of that term was that those 3000 ASes were the only ASes that forwarded other ASes traffic. The other 22000 ASes only carried traffic which had either destinations or sources within its own network. Only ASes that forward other networks traffic are relevant when trying to avoid routing loops and black holes.
DeleteHowever, that still leaves the question of how they determined that only 3000 were 'transit' ASes. If the physical peering connections were how they determined the 22,000 non-transit ASes (ie. those 22,000 are leaves, in the graph sense, in the network and only peer with 1 other AS, and are therefore incapable of being a transit AS)then that number may be meaningful, but if 'non-transit' was merely the advertised policy of those ASes at the time, that seems like something very transient, and which could change quite rapidly over time. If it were the latter, seems like that would have a real impact on how simple adoption of there routing system would be.
Sorry, late I know, but you are right in your interpretation; wrt the churn in the set, the set cardinality changes slowly (even peering relationships change slowly) so I am not sure this is a concern; the part that changes more rapidly is obviously eyeballs ASes
DeleteRegarding the End-to-End paper, the authors were correct to worry about the potential dangers of application level labeling, in light of 'recent' events such as comcast filtering of bittorrent traffic. The paper tries to stay very neutral on the roles of legal enforcement (such as blocking of pornography, filtering of 'terrorist' traffic, tax enforcement, etc) which may be ok for a survey paper, but may also not set a good stage for discussion of the roles of various agents, especially given the active role of 'net neutrality' in today's political landscape.
ReplyDelete“end to end arguments”
ReplyDeleteWe live in a world where different parties who don’t trust each other but still want to communicate, this paper describes such a picture and give some solutions in augmentation design of core-network and application layers to meet that demand.
The design principle is moving of functions “up” and “out” from the lower core network. As to main issues when adding designs to the core network , the author center on “control” and “filtering”, data of stream control point and the content visibility and biding to the device such as encryption for private information. A method to resolve the controdict is noted “Labels tagged information” (like PICS standard) by different users to fulfill different goals. Design in applications levels centers more on “privacy” and “performance enhancement”, like content filtering or cache by the server. The paper discusses well from many aspects, but I think in the core network design issues, only considering of monitor of track and visibility is not enough.
Hey tried to post this at about 7pm but blogspot was down or something
ReplyDeleteI think one of the main points of the Consensus Routing paper was that BGP already implicitly operates in "consensus" and "transient" states and by making these explicit one can reduce the cost of network changes. With BGP, once the routing information in the system has converged, the system is implicitly in a "consensus" state and all packets will reach their intended destinations. However, when there is a change in the network, this information is slowly and often unpredictably propagated through the network. The routers, however, have no way of knowing that the system is in this "transient" state and thus can end up routing packets in loops or blackholes. By distinguishing between these two states, routers can guarantee correct behavior in the "consensus" state and make smart rerouting decisions in the "transient" state.
I am skeptical of their transient forwarding schemes, however, especially detour rerouting to a tier-1 AS. While this basically guarantees delivery, the added latency is fairly high and would also greatly increase load near these ASs if all rerouting was done through them. This scheme should only really be considered if the ASs can be demonstrated to handle this increase and the gains in delivery probability outweigh the cost of the long detour when compared to more local schemes.