Network Research

1
Network Research

• An Operators PerspectiveTerry Gray
• University of Washington
• Associate Vice President, Technology
  Engineering, CC
• Affiliate Professor, CSE
• Microsoft EdgeNet
  01 June 2006

2
Theme

• "Every day's a winding road" -S. Crow
• "Every day's a research experiment" -T. Gray

3
Questions

• Seen any 404s go away by themselves lately?
• Is the trend getting better or worse?
• What is the max retry count in MS Windows?
• TcpMaxDataRetransmissions ?
• Could there be a connection?

4
Thesis

• The (open) Internet died in 2003 at the hands of
  slammer and blaster
• It's no longer about pervasive connectivitynow
  its all about (selective) isolation
• Tolerance for, and frequency of, glitches.. is
  increasing
• Current Internet model is busted e.g. silent
  failures, poor diagnostics

5
Buzz Words

• trust-mediated transparency
• convergence/virtualization

6
Miscellany

• Perimeter Protection Paradox
• Good ol' days of the network utility model
• Defense-in-Depth vs MTTD, etcClaim
• M-T-to-penetration, innovation, diagnosis d2
• How many arbitrary PEPs before the overall system
  becomes non-deterministic?

7
Trends

• Software-defined networks
• Personal lambdas
• Port 80/443 tunneling to get thru firewalls
• Encryption
• End-point firewalls (changes diagnostic picture)
• More policy boundaries less effective
• Regulation/compliance

8
Imagine

• Being in a NOC with an irate customer on the
  phone
• Customer "Is the network broken?"
• NOC "I give up, is it?

9
Hard problems

• DDOS
• Worsening MTTG
• Diagnosis complexity (MTTD)
• Provisioning complexity (vs. network utility)
• Mismatch between best and common deployment
  e.g. tcp retry count, duplex mismatch,
  stack-tuning
• S/N ratio of behavioral IDS monitors
• Path policy discovery
• Trust management
• Policy enforcement is silent looks like net
  failure

10
Hard problems (contd)

• Managing heterogenity bugs, conventions (jumbos)
• Organizational vs. geographic topologies
• Scaling personal lambdas
• Managing/diagnosing with e2e encryption
• Exception management
• Seduction of more protocols (e.g. MTR)
• Attribution vs. anonymity
• Multicast fault containment
• Optimal fault-zone sizing

11
Conclusions

• Plenty of work to do
• Help needed now!
• Current problems must inform future/clean-slate
  proposals