A Clean Slate 4D Approach to Network Control and Management

• architecture 2

http://www.cs.cmu.edu/~4D/papers/greenberg-ccr05.pdf

the 4D architecture — an extreme design point

today’s data network is surprisingly fragile and difficult to manage. the root of these problems lies in the complexity of the control and management planes — the software and protocols coordinating network elements — the way the decision logic and the distributed-system issues are inexorably inter-winded.

a complete refactoring of the functionality and three key principles

network-wide objectives, network-wide views, direct control

Network-level objectives

Each network should be configured via specification of the requirements and goals for its performance.

e.g., a traffic-engineering objective could be stated as “keep all links below 70% utilization, even under single-link failures.” e.g., do not allow hosts in subnet A to access accounting servers in subnet B

Network-wide views

borrowed from the database community and means having assembled a coherent snapshot of the the state of each network component … view of topology, traffic, events …

2 contro and management today

the functionality split into three planes

the dataplane handles the individual packets

the control plane implements the distributed routing algorithms across the network elements

the management plane monitors the network and configures the dataplane mechanism and control plane protocols

4 research agenda

4.1.1 Algorithms Satisfying Network-Level Objectives

the decision plane implements logic that converts network-wide views and network-wide objectives into directives for the data plane

an ambitious goal is to create a language or notion for expressing network-level objectives

e.g., Traffic engineering: Given a network topology and traffic ma- trix, compute a forwarding graph — a forwarding-table entry for each destination prefix at each router — that minimizes an objective function, such as the maximum link utilization.

e.g., Reachability policies: Given a network topology, a traffic matrix, and a reachability matrix, compute a forwarding graph and packet filters that minimize an objective function, while satisfying the reachability constraints

ePlanned maintenance: Given a network topology, a traffic matrix, and a planned event to disable certain equipment, compute a sequence of changes to the forwarding graph to avoid using the routers and links undergoing maintenance.

4.1.2 Coordination Between Decision Elements

Independent decision elements (DEs) operate independently, without any explicit coordination. A network element resolves commands from different DEs based on static priorities and/or a timeout mechanism. Initial studies in the context of BGP routing suggest that this approach is viable ¹,

4.1.3 Introducing Hierarchy in the Decision Plane