Looking Over the Fence at Networking

1
Looking Over the Fence at Networking

• Jennifer Rexford

2
Internet Success Leads to Ossification

• Intellectual ossification
• Pressure for backwards compatibility with
  Internet
• Risks stifling innovative intellectual thinking
• Infrastructure ossification
• Limits on the ability to influence deployment
• E.g., multicast, IPv6, QoS, and secure routing
• System ossification
• Shoe-horn solutions that increase system
  fragility
• E.g., NATs and firewalls

3
A Need to Invigorate Networking Research

• Measurement
• Understanding the Internet artifact
• Better built-in measurement for the future
• Modeling
• Performance models faithful to Internet realities
• X-ities like manageability, evolvability,
  security,
• Prototyping
• Importance of creating disruptive technology
• Emphasis on enabling new applications

4
Challenges of Measurement

• Extreme scale
• Large number of routers, links, ASes, packets,
• Difficulty of identifying flows
• End-to-end design
• Statelessness of the IP datagram
• Routing asymmetry
• Multipath routing
• Limitations on collection and sharing of data
• User privacy
• Confidentiality of business data

5
Measurement Research Line-Card Support

• Efficient measurement to place in line cards
• Online data collection at high speed
• Ideally useful for many kinds of analysis
• E.g., trajectory sampling
• Sample based on a hash of packet contents
• Sampled packets are sampled at each hop
• E.g., psamp activity at the IETF
• Parallel banks of filter, sample, and record
• E.g., deep packet inspection
• Algorithms for identifying patterns in packets
• Useful for detecting worms, viruses, etc.

6
Measurement Research Tomography

• Inference based on limited measurements
• Inverse problems that are often underconstrained
• E.g., AS relationships (e.g., Gao paper)
• Given collection of AS paths
• Infer business relationship between AS pairs
• E.g., traffic matrix
• Given link load statistics and routing
  configuration
• Infer offered load between ingress-egress pairs
• E.g., link performance statistics
• Given path-level measurements (e.g., loss, delay)
• Infer the performance of the individual links

7
Measurement Research Anomaly Detection

• Mining large, heterogeneous, distributed data
• To detect and diagnose anomalies, in real time
• Flash crowd, DDoS attack, worm, failure,
• Applying a variety of analysis techniques
• Statistics (e.g., Fourier, Wavelets, PCA)
• AI (e.g., Machine Learning)
• Algorithms (e.g., sketches, streaming algorithms)
• To a variety of kinds of data
• Per link packet or flow traces
• Per path delay, loss, or throughput
• Network-wide link matrix or traffic matrix

8
Measurement Research Privacy Confidentiality

• Preserving privacy and confidentiality
• Respect user privacy and business confidentiality
• While still producing useful analysis results
• E.g., anonymization of the data
• Anonymization of multi-dimensional data
• While still preserving associations across data
• E.g., privacy-preserving data analysis
• Distributed computation that hides information
• Computing a sum without revealing the parts

9
Measurement Research Protocol Design

• Protocol design
• Incorporating self-measurement, analysis, and
  diagnosis in future systems and protocols
• E.g., Early Congestion Notification
• Marking TCP packets that encounter congestion
• To trigger the sender to decrease sending rate
• E.g., BGP cause tags
• Tagging BGP update messages with root cause
• To reduce path exploration during convergence

10
Performance Models

• Traditional models
• Single queue
• Exponential distributions
• Open loop
• Steady state analysis
• Well-behaved parties
• Packet models
• Protocol analysis

• Advanced models
• Network of queues
• Heavy-tail distributions
• Closed loop
• Transients dynamics
• Selfish/malicious parties
• Multi-timescale models
• Protocol design

11
Modeling The X-ities (or Ilities)

• Beyond higher speed to consider X-ities
• Reliability
• Scalability
• Manageability
• Configurability
• Predictability
• Non-fragility
• Security
• Evolvability
• Challenging to model, or even to quantify

12
A Need for Interdisciplinary Work

• Statistical analysis
• Artificial intelligence
• Maximum likelihood estimation
• Streaming algorithms
• Cryptography
• Optimization
• Information theory
• Game theory and mechanism design

13
Discussion

• Where should the intelligence reside?
• Traditional Internet says the edge
• What about middleboxes (e.g., NAT)?
• Need to assemble applications from components
  located in different parts of the network?
• Better isolation and diagnosis of faults?
• Decentralized Internet makes this difficult
• Need to detection, diagnosis, and accountability
• Challenges the end-to-end argument

14
Discussion

• Data as a first-class object?
• Tradition Internet simple moves the bytes
• Naming, search, location, management in the net
• Modifyingg the data as it traverse the network
• Does the Internet have a control plane?
• Traditional Internet stress data transport
• What about network management and control?
• Today we place more emphasis on designing new
  protocols and mechanisms than controlling them

15
Discussion

• Abstractions on topology and performance
• Traditional Internet hides details from end hosts
• Network properties are, at best, inferred
• Guidelines for placement of middleboxes?
• Feedback info about topology and performance?
• Beyond cooperative congestion control
• Traditional Internet places congestion control in
  the end hosts, and trusts them to behave
• Is this trust misguided?
• New alternatives to congestion control?

16
Discussion

• Incorporating economic factors in design
• Traditional Internet ignores competitive forces
• Many constraints are economic, not technical
• Better to construct/align economic incentives
• Ways to deploy disruptive technology
• Traditional core is not open to disruptive tech
• Overlay network as a deployment strategy
• Other approaches? Virtualization? Middleboxes?
  Speaking the legacy protocols with new logic?
• Experimental facilities? A do over?

17
The Innovators Dilemma

• Leading companies often miss next big thing
• E.g., disk-drive industry and excavation
  equipment
• Problem
• Listening to customers leads to incremental
  improvement on the existing technology curve
• Disruptive technologies are often less effective
  for the existing customers, so tend to be ignored
• New companies exploit the new technology for a
  new market (e.g., desktops, laptops)
• Eventually, the new technology curve overtakes
  the old technology, usurping the old technology
• Will this happen with the Internet?