Network Architecture Lab

Internet Performance Accountability

 
The current Internet provides no information on the fate of transmitted packets. As a result, when packets get lost or delayed, there is no clean way for the affected parties to localize the problem and fix it (if it is local), ask for compensation (if a service-level agreement has been violated), or simply learn from it (e.g., re-assess a peering agreement with an under-performing neighbor). Probing tools like traceroute can help localize network failures, however, they draw their conclusions based on the fate of probes, not actual traffic, which makes them susceptible to manipulation by transit networks.  Moreover, such probing tools often reveal the internal structure
and routing policies of ISPs, giving the latter an incentive to render their networks opaque to probing.
 
The goal of this project is a way to change this lack of accountability in the Internet: a clean, yet practical solution that tells network entities what they need to know (who is responsible for losing or delaying their packets), but not what they shouldn't (the internal structure and policies of other networks or ISPs). The key idea behind our work is that this information need not (and should not) be extricated by ad-hoc probing tools that treat the Internet as a black box and try to reverse-engineer its structure and failures. Rather, it should be provided by a cooperative, incentive-based framework, where networks provide verifiable information on their own performance and, in exchange, learn how their own traffic is being treated by their neighbors.

People

Katerina Argyraki, EPFL

Olga Irzak (now at University of Toronto)

Petros Maniatis, Intel Labs Berkeley

Ankit Singla (now at UIUC)

Ashish Subramanian

Papers etc

Verifiable Network-Performance Measurements [slides], in the ACM International Conference on emerging Network EXperiments and Technologies (CoNEXT), November 2010. Design and implementation of Network Confessional, a protocol that enables participating ISPs to expose verifiable information on their loss and delay performance without maintaining per-packet, per-flow, or per-path state. Here are the slides from a more detailed talk given at MIT in April 2011.

Loss and Delay Accountability for the Internet [slides], in the IEEE International Conference on Network Protocols (ICNP), October 2007. Design and implementation of AudIt, an explicit accountability interface for providing accurate loss/delay feedback on forwarded TCP traffic. Here are the slides from a more detailed talk given at the Max Planck SWS seminar in September 2007.

An Accountability Interface for the Internet, Olga and Ashish's poster for EPFL Research Day, July 2007. Outlines the idea of an explicit accountability interface, through which ISPs report on their forwarding performance.

Providing Packet Obituaries [slides], in the ACM Workshop on Hot Topics in Networking (HotNets), November 2004. Introduces the idea of transit networks providing feedback on the traffic they forward and describes a preliminary mechanism for informing traffic sources where their packets are getting lost.

Related work

A more theoretical perspective on accountability. Uses tools from cryptography to prove the minimum set of requirements for providing accountability under the broadest threat model.

In PeerReview, members of a distributed system hold each other accountable for Byzantine behavior by replaying secure logs of exchanged messages.

An economic analysis of ISP business shows that, without accountability, innovation and competition in the Internet are impossible.

Availability-oriented path selection relies on accurate information on the performance of transit networks to optimize route selection in multi-path routing.

SOME RESULTS

► We argued for shifting the focus of network performance tomography from identifying congested links to a new, more realistic goal: identifying the frequency with which links are congested (CoNEXT '11). 

► We relaxed one of the strongest assumptions made until now by network performance tomography, namely that the status of a network link is independent from the status of any other link (IMC '10).

► We developed RouteBricks, a parallel router architecture made up entirely of commodity PCs, which achieves multi-Gbps line rates by parallelizing functionality both across and within PCs (SOSP '09).