Fast TCP Doesn’t Matter to ISPs (What Am I Missing?)

1
Fast TCP Doesnt Matter to ISPs(What Am I
Missing?)

• Albert Greenberg
• Microsoft

2
Traffic and Cost

• Consumer Broadband gtgt Business
• Consumer broadband growth trends are stable
• Individual consumer broadband flows have little
  impact on ISP backbones
• Transmission rates per broadband subscriber are
  tiered and rate limited
• High speed access costs are high
• TCP is likely not the factor stopping a Disney
  from last minute distribution of content to
  theatres world wide

3
Fast TCP A Win for ISPs?

• What if Im wrong?
• Biggest deal file transfer size
• Downloading (stored content) gtgt Streaming on the
  Internet
• Fast TCP might increase consumer bandwidth demand
• Economic wins
• Bandwidth is the killer app, so this brings more
  revenue for ISPs
• Fast, robust download protocols have potential to
  help video-on-demand services in IPTV framework
• What would the impact be on network engineering?

4
Backbone

• Sources of router stress
• PPS high rates of little packets (e.g., VoIP),
  yet Fast TCP wont promote this
• Protocols stresses CPU, yet this is in the
  control plane not the data plane
• Congestion yet backbone physical pipes gtgt
  individual flow
• Caveat stat mux trouble could arise if using a
  set of parallel links for higher capacity
• Core link load is low under normal conditions
• Backbones dimensioned to handle rare failure
  conditions
• Tremendous stat mux gains on backbones 40Gbps
  links rolling in now
• A 1Gbps flow is still 1/40th of this

Backbone
Backbone
Backbone
Could be this Traffic sea level (after high
speed TCP)
Not this Traffic sea level (after high speed TCP)
Traffic sea level
5
Access

• Access is where the cost is, and where the
  potential for problems is
• Few 100 broadband subs per 1st aggregation router
• Few QoS mechanisms deployed between the first IP
  aggregation router and broadband subscriber
  router
• Higher speed access attracts subs who are
  early-adopters/power users, and who increase
  overall burstiness
• E.g., Kensuke Fukada, Kenjiro Cho, Hiroshi Esaki,
  The Impact Of Residential Broadband Traffic On
  Japanese ISP Backbones, ACM SIGCOMM Computer
  Communication Review, Vol. 35, Issue 1, 2005, pp.
  15-22.
• Yet TCP dynamics are constrained by the access
  links
• Access is rate limited, per sub
• Fast TCP still backs off
• QoS mechanisms in the last mile are available (to
  date generally not needed)
• Capacity planning practice, which is largely
  empirical, can be adapted to accommodate
  increased burstiness
• E.g., Matthew Roughan, Charles R. Kalmanek,
  Pragmatic Modeling of Broadband Access Traffic,
  Computer Communications, Volume/Number 26(8),
  May, 2003, pp.804-816.

6
Takeaways

• Impact on Internet health
• Dont see potential downside
• Impact on Internet economics
• See potential upside
• Impact on network engineering?
• Access gtgt Backbone
• Attendant increase in sluggishness in TCP
  feedback loop (more bytes shipped per RTT) risks
  potential for
• TCP synchronization and oscillation
• QoS impact (jitter, loss) on legacy TCP flows
• QoS impact of jumbo frames (if fast TCP becomes
  common, jumbos may as well)