Why the future of the Internet is not multimedia, and other high heresies

High heresy #1: the future of the inet is not multimedia

• not "just like" anything else
  • attempts to migrate traditional media to inet are failing. (radio is not a talking newspaper; TV is not radio with pictures.) It's a new medium.
  • packets don't carry "sensory analog" traffic well
  • impications for e-commerce, e-advertising?
• too expensive
  • long haul infrastructure too expensive; free unlimited bandwidth not coming (who remembers "power too cheap to meter"?)
  • fast access to local caches/content: maybe. (Long haul is up to 3 orders of magnitude more expensive to deploy) So @Home's model may be viable. Jury still out.
  • Corollary: (Content providers will have to accept that in order for their content to be seen, it will have to be cachable!)
  • "local" vs "long distnace" internet access???
• technology growth
  • second derivative is wa positive. "If you aren't scared, you don't understand." [QUOTE]
  • UUNET-2000: 50 terabits; today: 5 gbits; look for 1000x growth over 3 years.
• Voice: $250B "niche"; grows only with human population, driven by population demographics (like other "human centric" businesses). Little value in adding voice quality.

Fundamental shift in bandwidth consumption

• Formerly, growth driven by "mothers day"
• Now: hungry silicon cockroaches! (computing power that attracts communications bw)
• short,fast, bursty vs. long,slow,smooth channel use
• limiting factor for growth: basically, rate of melting sand!
• pecking order
  • fax machines, cellphones: blew out the North Amer Numbering Plan
  • PC+modem: blowing out switched capacity
  • inet backbones: eat long haul capacity
  • webphones and digital daytimers: eat wireless capacity? (huge fraction of cell phone calls are to set up other meetings, etc., not conversations!) "Network externality" effect will cause this usage to overwheml multimedia.
  • capacity is getting used up as fast as it's provided
• implications
  • web content will be examined by robots, only shown to humans after comparison shopping! Automatic aggregation services willdominate.
  • therefore existing advertising models are broken...
  • cockroaches will be empowered by owners to spend money.
• conclusion: follow the money
  • Comm networks 21st c. will be built for cockroaches
  • human/cockroach symbiosis

Heresy #2: multicast everywhere

• bad scaling; "The only problem is scale; all other problems inherit from that." (O'Dells Problem Statement) [QUOTE]
• no business model. (Uunet is spending >$300M next year for infrastructure! Similarly AOL...)
• parts may be salvageable for large scale use

Heresy #3: all you need is IP routers: guided tour of UUNET

• No traffic mgmt except on toy topologies
• classic IP dest-only fwding induces "super-aggregation" f IP traffic; can't disaggregate them. Result is unbalanced network; Controlling traffic "spread" is just not feasible -- only realizes small percentage of network capacity.
• capacity and bandwidth are different
  • hard to add "targeted" capacity
  • BW+BW=capacity; but capacity is "incoherent" (basically,IP can't convert it into ordered-delivery bandwidth)

• deep systemic redundancy
  • redundancy in the switches (implementation)
  • redundancy in routing maps (architecture)
  • little redundancy in IP routers
  • all hardware in "lights-out, hands-off" locations, costs at least $1K to get a human to go do something.
• cost of ownership (operating cost) vs. capital. Be careful which you optimize!
• 4 layer network, top to bottom:
  • collector networks: a lot of muxing to bring customer IP traffic to network core
  • IP routing: fully redundant mesh; no customer-specific state in core routers; protected at edges only (route filters). Optimized for fast IP forwarding, and that's all. Two full-meshes with spoke "leaves" interconnected in pairs.
  • transit switching: map large traffic flows onto traffic fabric in a way that's invariant to IP topology; "surgical precision" for adding transmission capacity. Full dynamic routing with constraints.
  • transmission facilities (TDM bandwidth, telephony-to-TTL); OC12c, OC48 "soon" (year-end 97)
  • Observation: traffic flow engineering is an economic optimization problem. Collapsing the above layers removes most of the knobs that allow this problem to be solved.
• "Metro architecture" (now being deployed):
  • OC-12 "metro rings" (july 97); above "superhubs" are situated around the ring (OC48: year end 97)
  • applies backbone architecture recursively
  • regional traffic engineering
  • segregation of data, control, routing "planes"; no reason to collapse these functions into one box
• "Massively Parallel backbone" (tm)
  • 500 Gbits trunk, 2-10 Gbits strands, 20-50 Tbits aggregate
  • building blocks: 16-port OC48c switches, DWDM (wave division muxing) puts multiple OC48c on a fiber, wavelength-agile lasers on line modules
  • take the OC48c switches, stack them up, stack another bunch of them up perpendicular, and you get something that looks like a 3D version of the Myrinet crossbar using OC48c. Whoa.

Intersting open problems... (from Q&A session)

• Simluation and modeling: traditional network modeling/simulation techniques break at this scale. What do you measure/model? What behaviors or variables are interesting? The "phenomenology" of this scale of packet network is just not well understood. (You can observe things like "large Net events": Olympics, Superbowl, new release of Netscape...)

• "Internet weather report": what does it measure?
• "Why (internet) caches are impossible"? (someone else's talk)
• Someone asked "What do you think AOL's problem is?" He declined to comment, saying "There but for the grace of God..."