Introduction

The difficulty with realizing innovations within operational networks arises from the tension between the need to improve and the maintenance of network reliability. This tension is evident in (at least) three ways:

Insufficiency of virtual, open routers:
Researchers generally agree that changes to the fundamental architecture of the Internet are necessary. Recently, virtualization and availability of router software API to the public have been touted as the tools required for these changes. By virtue of the fact that experimental functionalities are co-located with production-level processes within the same physical devices and that virtualization technology is imperfect, these tools are likely to impact operational traffic in real-life large-scale networks. Furthermore, virtualization is typically used in scenarios where the physical device has an abundance of resources and can afford to run multiple logical instances. Unfortunately, in the real-world, vendors are continuously straining to catch up with the increase in network load (traffic as well as routing table size), hence multiple virtual routers co-existing on a physical operational router is infeasible in the foreseeable future. We believe that an alternate solution, for both experimentation and incremental deployment of new functionality into operational environments, is necessary.

Improved router scalability:
The scaling of routers today is hardware-centric: it revolves around the addition of more, and improvements to, router modules such as linecards. Unfortunately, different types of modules are still highly interdependent, resulting in a largely monolithic router architecture that increases the time to innovate and test new functionalities and components to ensure reliability of the entire router. Specifically, rather than simply adding more independent devices to handle additional load (say routing table entries), entire routers are redesigned. This increases the cost of equipment, manpower required, and time to deploy. More importantly, the frequency of network upgrades involving routers with new designs increases, directly increasing downtimes needed to resolve bugs revealed only after deployment at scale. We think that a complementary solution focusing instead on address space partitioning to distribute load can be used to improve scalability further.

Enhanced network management:
The standardization of network management tools such as Netflow/IP-FIX and SNMP results in widely used and reliable tools. However, this comes at the expense of customization flexibility, and the dependency on vendors to support these tools can present problems at times: (1) service providers may prefer that tools developed to provide competitive advantages not be incorporated into generic routers; (2) existence of certain tools at every router of interest may not be guaranteed due to heterogeneity. These factors call for the separate development of some management tools that would play a role in improving network reliability.

These issues can be addressed by decoupling innovation and reliability: we would like to introduce new network functionality without interfering with operations that affect many aspects of everyday life. Motivated by the need to develop router architectures that allow the former without sacrificing the latter, this paper begins the exploration of an extreme end of this space: we require first and foremost that the existing network functionalities, services and performance be preserved, then we aim to ease the deployment of new ones. We believe that a key factor used to distinguish solutions in this space is the type of interface between components. Here, we use the common packet interface, thus ensuring that router changes are unnecessary.

Based on this packet interface, we propose a general architecture we call Trochilus, which enables pushing of router functionality to adjacent devices.

People
Cheng Tien Ee (AT&T Labs Research)
Lee Breslau (AT&T Labs Research)
Kadangode Ramakrishnan (AT&T Labs Research)

Papers, Slides:
Report '08: Extending Network-Layer Functionalities Today [pdf]