Growth Trends in Wide-Area TCP Connections
Vern Paxson, LBL
Back to Index
The paper summarizes observed growth trends with data points taken over a 3-
year period of TCP traffic in and out of LBL. A few interesting rules of
thumb emerge from the analysis:
- Network use grows exponentially until the capacity is saturated. This
was true even in the bad old days of low-bandwidth networks and relatively
few users, and is more true than ever given the WWW. Armando's
corollary: Design your network for steady-state saturated operation.
- Accounting is difficult. The authors worked hard to eliminate "spurious"
traffic, such as anomalous scripts, heavy traffic between a particular pair
of nodes due to node-specific activity, TCP packets that claimed infeasible
data payload sizes, etc. Armando's corollary: It's hard to regulate
this kind of use as well, and there's nothing to stop me from flooding my
subnet with packets, the effects of which could propagate to faraway places
on the Internet.
- A small number of hosts/sites account for a large percentage of the
traffic. This is Amdahl's Law applied to network traffic and is not
surprising, but it's good to see it supported by numbers.
- The WWW is the killer app for the high-speed connected Internet, just as
email and news were the killer apps for their audience in their day.
Armando's corollary: New apps will arise that soak up bandwidth as
fast as badnwidth can be added. This is also true of software and CPU's.
Relevance to Multimedia
Distributed multimedia apps rely on long-haul packet-switched networks, often
for close-to-real-time data delivery (e.g. MBONE video). We can't hope to
design such services without a good understanding of the traffic patterns
these apps will encounter over a WAN.
5 out of 5: IMHO exposed many interesting dualities between network capacity
vs. traffic and system performance vs. offered load. The lessons to be drawn
from each are similar and should guide us in network design.
Armando Fox (firstname.lastname@example.org)