Network Management

1
Network Management

• Jennifer Rexford
• Princeton University
• http//www.cs.princeton.edu/jrex

2
My Research Interests Network Management

• The boundary between network elements and the
  systems that manage them
• Internet not built with management in mind
• Leading to expensive, convoluted solutions
• Clean-slate approach design for managability
• New architectures that are easier to manage
• Routing, packet forwarding, traffic management
• Ways my research would use GENI
• Running my own backbone, with my own routers

3
Today Inside a Single Network

• Data Plane
• Packet handling by routers
• Forwarding, filtering, queuing

Packet filters
4
Example 1 Network-Wide Control

• Getting rid of most of the control plane
• Real-time, network-wide management service
• Control plane merely provides measurement data
• Routing Control Platform (RCP)
• Has routing adjacencies with neighboring ISPs
• and directly configures the forwarding tables
• Based on network-wide policy objectives

RCP
RCP
BGP with ISPs
BGP with ISPs
5
Example 2 Secure, Robust Interdomain Routing

• Todays interdomain routing
• Single-path routing with convoluted decision
  rules
• No connection with behavior of the data plane
• Leading to vulnerabilities and poor performance
• Changing the routing protocol
• Anomaly detection to avoid suspicious routes
• Propagating multiple paths to customer
• Changing the data plane
• End-to-end monitoring of integrity performance
• Direct influence over the path-selection process

6
Example 3 Distributed Adaptive Traffic
Engineering

• Todays division of labor
• Users TCP congestion control
• Routers link-state routing protocols
• Operators tuning link weights on coarse
  timescale
• Top-down design from optimization theory
• Multiple paths between pairs of nodes
• Direct feedback from links about congestion
• Dynamic changes to sending rate along each path

7
Example 4 New Data-Plane Functionality

• Direct control over packet forwarding
• Management service can write forwarding tables
• Flexible splitting of traffic over multiple paths
• Packet measurement support on links
• Traffic statistics, such as congestion price
• UFO Underlay Friendly to Overlays
• Tunnel tables that subscribed services can modify
• Explicit feedback about network conditions
• Graceful handling of planned changes
• Advanced notice to subscribed overlay services
• Network-wide commit of new data-plane state

8
GENI Programmable Packet Processor
General purpose processors
C P U
C P U
C P U
C P U
N P
Switch Fabric
High-speed packet forwarding
F P G A
L C
L C
L C
L C
Links to other backbone nodes, edge sites, and
legacy Internet
9
How Id Use the Programmable Packet Processor

• Line cards
• Links, perhaps with bandwidth guarantees
• General-purpose processors
• Control/management plane functionality
• Overlay services subscribed to my network
• Network processors and FPGAs
• First, conventional IP- and MAC-based forwarding
• Then, my nifty customized data-plane design
• Support for connecting to legacy Internet
• BGP sessions, address translation, tunneling

10
Attracting Real Users

• How would I attract real users
• to my nerdy new routing architecture???
• Real users
• Connect from a GENI edge site over via a tunnel
• Treat my experiment as their service provider
• Distributed services, take me to your users!
• New services running on GENI (e.g., new CDN)
• Could run over my backbone architecture
• for better performance, reliability, security,
  flexibility, visibility, and many other X-ities
  ?

11
Conclusions

• Network management is hard
• Because the Internet wasnt designed for it
• Retrofitting management can only go so far
• Clean-slate design for manageability
• Revisiting the roles of operators, routers, users
• Novel management, control, and data planes
• Capitalizing on the GENI backbone
• Building an experimental service-provider network
• That runs my proposed refactoring of the roles