(The paper: http://conferences.sigcomm.org/hotnets/2014/papers/hotnets-XIII-final74.pdf)
Vissicchio outlines the main idea of the work: "we use lies to combine the advantages of OpenFlow and of traditional networking."
"Fibbing enables flexible routing à la SDN, today," he says, without "the hurdles of deploying a full complete OpenFlow network":
- fine-trained traffic engineering (optimize the available resources)
- provision backup paths (quickly and predictably react)
- deploy advanced services (steer traffic through middleboxes)
These needs are impossible to satisfy with IGPs today, he says. "We are not discovering a new problem. Operators know that IGPs are not flexible."
One option for operators is to use virtual circuits (MPLS), but this incurs control- and data-plane overhead. OpenFlow-based systems can realize these applications, but have deployment costs and "new challenges of a novel paradigm."
Vissicchio explains the team's technique, which is "to have the controller trick the routers into seeing a fake topology that is carefully constructed to achieve the desired Forwarding Information Base (FIB)." This achieves many of the benefits of an all-OpenFlow network, without needing OpenFlow.
"Link-state IGPs are actually good for something: to control router behavior," he says.
The technique is very expressive, Vissicchio says. "Any set of forwarding DAGs can be enforced by fibbing... We can do fine-grained steering to middleboxes on a per-destination basis."
And there are some other benefits: because "good old [routing] protocols running," the result is "easier troubleshooting" than more novel systems. The paper gives many more details about how the system works.
"We don't have the full flexibility of SDN, because we are not able to match all the fields of the packet," Vissicchio says. "But our argument is that you may not need this fine-grained control or fine-grained monitoring ability. A good monitoring infrastructure but not so fine-grained, so OpenFlow is not so necessary."
Q (Bruce Maggs): I'm reminded of a paper from some people at AT&T some years ago. There was a push to use MPLS, and they said, we can achieve these goals by just putting weights on the edges for IGP."
A: Our system has other benefits relative to MPLS.
Q (Dave Oran): This appears to depend on a misfeature of the Cisco implementation, which is that we don't check that both ends announce it. If you enforced that according to spec, this would fail.
A: The spec is ambiguous.
Q (Oran): Not when I wrote it. (laughter) What about source-destination routing vs. what you can do with fibbing? This would give you some of the notions of MPLS without creating virtual circuits.
A: May be possible, but requires more processing power. We do not expand processing power.
Post a Comment