|
CMU-CS-05-177
Computer Science Department
School of Computer Science, Carnegie Mellon University
CMU-CS-05-177
Refactoring Network Control and Management:
A Case for the 4D Architecture
Albert Greenberg, Gisli Hjalmtysson, David A. Maltz, Andy Myers,
Jennifer Rexford, Geoffrey Xie, Hong Yan, Jibin Zhan, Hui Zhang
September 2005
CMU-CS-05-177.ps
CMU-CS-05-177.pdf
Keywords: Network control plane, centralized versus distributed
control, decision plane, dissemination plane, routing design, routing
protocols
We argue for the refactoring of the IP control plane to support
network-wide objectives and control. We put forward a design that
refactors functionality into a novel 4D architecture composed of four
separate planes: decision, dissemination, discovery and data. All
decision-making logic is moved out of routers along with current
management plane functions to create a logically centralized decision
plane, where network-level objectives and policies are specified and
enforced by direct configuration of states on individual network
elements. Pullingmuch of the control state and logic out of the
routers enables both simpler protocols, which do not have to embed
decision-making logic, and more powerful decision algorithms for
implementing sophisticated goals. Remaining on the routers is a
wafer-thin class of intrinsically distributed control functions.
These support the discovery plane, consisting of elementary functions
to discover topology and network state, and the dissemination plane,
consisting of elementary functions to distribute explicit instructions
to manipulate the data plane forwarding mechanisms.
This paper motivates the need for a new approach to network control
and management, describes the 4D architecture, and sketches the design
space and challenges posed by the architecture. As a first exploration
of the design space and its challengs, we have constructed a working
prototype that implements the 4D architecture. Through evaluation of
this prototype on 9 different network topologies derived from production
networks, we demonstrate that (i) the 4D architecture can achieve
subsecond reconvergence times upon single link or router failures and
can adequately deal with other failure scenarios including network
partition; (ii) the 4D architecture is able to implement a network
design intent such as a reachability matrix more robustly than
currently possible; and (iii) the 4D architecture does not introduce
excess overhead.
30 pages
|