Network Management

1
Network Management

• has always been and always will be essential to
  the Internet

FCC Broadband Industry Practices Hearing WC
Docket No. 07-52 Stanford University April 17,
2008
Testimony of George Ou Former Network
Engineer www.LANArchitect.net
2
Internet meltdown in 1980s

• Lack of adequate congestion control in TCP
  allowed too many FTP users to overload Internet
  around 1986
• Van Jacobson created congestion control algorithm
  for TCP in 1987
• Congested routers randomly dropped packets to
  force every TCP end-point (client) to cut flow
  rate in half
• TCP clients then slowly increased flow rate with
  every successful transmission until next packet
  drop
• Caused all TCP streams to home in towards equal
  flow rate
• Fair bandwidth sharing, but only for applications
  of its time
• Jacobsons algorithm saved the Internet in 1987
  and remains dominant standard after 20 years
• Early example of managing network congestion

3
World Wide Wait in 1990s

• First generation of web browsers were not
  optimized for Internet
• World Wide Web turned in to theWorld Wide Wait
• Version 1.1 of HTTP revamped to efficiently use
  resources over 1.0

4
Todays crisis on the Internet

• Video-induced congestion collapse
• Efficient existing broadcast model migrating to
  bandwidth-intensive Video on Demand model over IP
• Full migration of video could require 100- to
  1000-fold increase in Internet capacity
• Exponentially more bandwidth required as video
  bit-rate and resolution increase to improve
  quality
• P2P is the dominant distribution model because
  most of its content isfree (read pirated)
• Video can fill any amount of bandwidth

5
More bandwidth doesnt help
The few throttling the many
According to the Japanese Government 1 of users
account for 47 of traffic 10 of users account
for 75 of traffic 90 of users get leftover 25
6
Exploiting Jacobsons algorithm
50/50 Fair
80/20 Unfair
92/8 Unfair
7
Persistence advantage in P2P apps
Corporate VPN telecommuter worker using G.722
codec _at_ 64 kbps payload and 33.8 kbps
packetization overhead Vonage or Lingo
SIP-based VoIP service with G.726 codec _at_ 32 kbps
payload and 18.8 kbps packetization
overhead I calculated that I sent 29976
kilobytes of mail over the last 56 days averaging
0.04956 kbps
8
Weighted TCP Per-user fairness
92/8 Unfair
50/50 Fair

• BT chief researcher Bob Briscoe proposes TCP fix
  before the IETF to neutralize multi-stream
  loophole
• Changing TCP takes many years, but its even
  harder to get over a billion devices to switch to
  new TCP client
• Newer network-based solutions being implemented

9
Present and future solutions

• Present solutions use protocol throttling
• P2P applications use disproportionately large
  amounts of bandwidth so theyre throttled to
  balance them out
• Use conventional router de-prioritization
  techniques on P2P
• Use TCP resets to occasionally stop P2P seeders
• Potentially affect an extremely rare
  low-bandwidth P2P user
• Can be fooled by protocol obfuscation techniques
• Future solutions are protocol-agnostic
• Weighted packet dropping at router and/or fair
  upstream scheduling on CMTS accomplishes per-user
  fairness
• Only targets bandwidth hogs and forces them to
  back off
• Cannot be fooled by protocol obfuscation

10
What is reasonable network management?
Protocol-agnostic per-user fairness
Fair
Future
Intelligent
Reasonable
ISP throttles BW hogging protocols
Metered Internet
Unreasonable
No network management
Unfair
Dumb
Past
11
Network management ensures harmonious coexistence

• P2P applications need volume, not priority
• Interactive applications (Web) and real-time
  applications (VoIP) want priority and not volume
• P2P, Interactive, and real-time applications each
  get what they want under a managed network
• Interactive and real-time apps have small/fixed
  volume so no matter how much theyre prioritized,
  they cannot slow down a P2P download.
• Unmanaged networks regardless of capacity will
  always be unfair and hostile to interactive and
  real-time applications