Cristel Pelsser's blog

The CFP for the ACM CoNEXT ‘10 Student Workshop is out !

http://conferences.sigcomm.org/co-next/2010/Workshops/StudentWorkshop/

I just updated my professional webpage with the computer network’s 2010 paper. I also added some links to earlier publications.

Lately, I’ve been looking at solutions to avoid correctness issues that may arise in iBGP. These issues are explained in Tim Griffin’s et al., Sigcomm 2002’s paper.

In Pelsser_DRS_IEICE_200903, I proposed a solution relying on an overlay of distributed servers that compute routes on behalf of the routers. Today, each router computes for every prefix its own routes, based on a limited visibility of the routes. In this proposal, we rely on a different distribution of the route computation. Each server has full visibility of the routes for a subset of the prefixes. A server computes for its subset of the prefixes, the route to be used by every router of the domain. The server performs a per-router computation. Full routing visibility at the node performing route computation ensures iBGP correctness.

A presentation describing this proposal can be found here.

In March, I presented Laurent Vanbever’s work called Customized BGP Route Selection (CRS) at the WIDE camp. This work is about using BGP MPLS/VPNs to advertise neighbor-specific BGP routes.

The presentation can be found here.

We took the opportunity of Tim Griffin, Olaf Maennel and Despoina Perouli’s visit at IIJ to organize a BGP workshop. We had very interesting discussions with local operators and researchers. The presentations can be found here http://www.attn.jp/maz/p/c/bgpworkshop200904/.

Cristel Pelsser, Steve Uhlig, Tomonori Takeda, Bruno Quoitin and Kohei Shiomoto, “Providing scalable NH-diverse iBGP route redistribution to achieve sub-second switch-over time“, to appear in Computer Networks, Elsevier.

Here is a reference to an all public position article from Olivier Bonaventure. The article is on the enforcement of Internet quota by Internet Service Providers. The article is in French. It highlights the dangers of quota on network neutrality. Dangers arise when these quotas are not applied indifferently to all services.

http://bit.ly/bvrTdJ

Panel from Mark Handley

The data plane is well structured with the different layers of the OSI model. Comparatively, in the control plane, there are a lot of protocols but no structure. These protocols each solve a very specific issue. They do a good job about it but now is the time to take a step back.

The control plane protocols tend to solve three categories of problems:

• congestion control
• multi-homing
• traffic engineering

Or problems resulting from the intersection of two of these categories.

In the presentation, Mark focusses on the intersection between congestion control and multi-homing. He proposes multi-path transport (multipath TCP) to solve this issue. For solving the congestion control and TE intersection problem, he proposes the re-ECN solution from BoB Briscoe.

1. Pierre Francois: Clarification to the treatment of BGP non-transitive extended communities (draft-decraene-idr-rfc4360-clarification). The implementations tested by the authors do not behave in the same way. One implementation removes the community upon reception on an eBGP session. The authors of the draft highlight that section 6 of RFC 4360 says that non-transitive extended communities SHOULD not be transmitted on eBGP session. Since it is a “SHOULD” and not a “MUST”, the authors suggest that when such a community is received on an eBGP session, it is not stripped from the BGP update. It is important to fix this behavior should BGP graceful shutdown be performed using a non-transitive extended community instead of the current allocated community. Using a non-transitive extended community will simplify the filters that currently need to be configured to confine graceful shutdown to the local ASs.
2. Pierre Francois: Analysis of the path selection modes proposed for add-path (draft-vvds-add-paths-analysis). In the “session type” mode (not described in the draft), the ASBRs advertise all paths to their RRs and on iBGP sessions of type “over”. RRs advertise the 2 best paths to their iBGP clients. This mode seems the most interesting to me, if the RRs can manage to hold all the routes.
3. Bruno Decraene: Requirements for BGP graceful shutdown (draft-ietf-grow-bgp-graceful-shutdown-requirements). It is important to provide a solution to reduce connectivity/packet losses upon the maintenance of a BGP session. Bruno asked that the draft be last called. After the presentation of the requirements, Pierre presented one solution that does not require modifications to BGP, only configurations of BGP filters. The presentation was followed by operators coming to the mic. They were afraid that the solution be more complex than what they do today. For example, one of them uses AS-path prepending. He does not care about having 0 packet loss. Thus, tunnels are not needed (too expensive). However, he would benefit from the proposal relying on low loc-pref setting. This will ensure that traffic be redirected away from the maintained resources. This is not always the case with AS-path prepending. This is because the AS-path length comes after the loc-pref in the BGP decision process.