CABO: Concurrent Architectures are

Better Than One

Nick Feamster (Georgia Tech) Jennifer Rexford (Princeton) Lixin Gao (U. Massachusetts)

Virtualization as a Building Block for Networks

• Multiple protocols, networks, or architectures in parallel– Multiple logical routers on a single platform– Resource isolation in CPU, forwarding table,

bandwidth– Programmability for custom protocols and

mechanisms

Some Ongoing Projects• Path Splicing (Georgia Tech)

– Run multiple instances of routing protocol– Path bits indicate which instance to use at each hop– Exponential diversity gain, modest complexity

• VROOM: Virtual Routers on the Move (Princeton)

– Support overlay functions in routers– Efficient forwarding and scalable monitoring

• HORN: Hybrid Routing for Overlay Networks (UMass Amherst)– Different nodes see different detailed subgraph– Availability of link state with good scalability

Path Splicing

• Step 1: Run multiple instances of the routing protocol, each with slightly perturbed versions of the configuration

• Step 2: Allow traffic to switch between instances at any node in the protocol

ts

Compute multiple forwarding trees per destination.Allow packets to switch slices midstream.

Reliability Close to Best Possible

Average stretch is only 1.3

Sprint topology,degree-based perturbations

VROOM: Virtual Routers on the Move

• All logical configurations/states remain the same before/after the migration– IP addresses and routing protocol

configurations remain the same– Routing-protocol adjacencies stay up– No protocol (BGP/IGP) reconvergence

• Network topology stays intact– Adjacent routers won’t know router has moved

• Virtually no disruption to traffic– Traffic downtime can be eliminated

VROOM Result Slide?

A knows a partial topology (adjacent sub-graph) via link state protocol,

and learns routes from the border nodes of the sub-graph.

C-H

E-G-H

D-G-H

A-B-C-H

HORN:Hybrid rOuting for oveRlay Networks

• Link state routing– Not scalable – A virtual network could be as large as the Internet

• Distance vector routing– Convergence delay – Scales better

• Main idea: a tunable routing protocol– Trade scalability for better availability– Each slice runs a fixed protocol with tunable

parameter

Scalability vs. Convergence

VINI/Trellis: Deployment Platform

• Lightweight containers running on top of a single OS

• Each virtual host sees a set of virtual ethernet links

• Packet forwarding, traffic shaping remain outside of containers

Trellis: Speed and Scalability

Packet forwarding rates close to those of raw kernel

Can support a large number of virtual containers

CABO as a New Architecture

• Virtualization– Multiple logical routers on shared hardware– Resource isolation in CPU, FIBs, and bandwidth

• Programmability– General-purpose CPUs for control and manipulation– Network processors and FPGAs for fast forwarding

• Economic refactoring– Infrastructure provider: manage routers and links– Service provider: offer end-to-end services

Deciding Not to Decide• Flexibility has been key to the Internet’s success

– Many different applications and services

– Beyond anything the initial designers ever envisioned

• Today this flexibility is limited to the end systems– Not surprisingly, this is where we have seen innovation

• And, the “inside” is quite difficult to change– Witness the fate of IPv6, QoS, multicast, secure routing

• Even if we could start over…– Maybe the design problem is over-constrained

– Too many goals, some conflicting?

It’s Hard to be a Routing Protocol

• Many design goals– Global reachability

– Fast convergence

– Efficient use of resources

– Low protocol overhead

– Secure control plane

– Flexible routing policies

– <your wish list here>

• Perhaps we cannot satisfy all goals

Convergence vs. Scalability

Voice over IP Gateway

Remaining Traffic

Properties Fast convergence for a few prefixes

Scalability to 200K prefixes

Dissemination Flooding Hierarchical

Routing Protocol

Link state (OSPF or IS-IS)

Path vector (iBGP with route reflectors)