Review "The Impact and Implications of the Growh in Residential User-to-User Traffic"

There is a widely held belief that peer-to-peer applications are
responsible for a significant amount of traffic on ISP backbones.
However, there have not been many studies to prove this observation or
to quantify the extent. Additionally, with the introduction of
fiber-to-the-home, it is probable that peer-to-peer traffic will
dominate an even greater share of backbone traffic. This study
documents the measurements and analysis of the aggregated traffic from 7
Japanese ISPs over a period of 21 months covering 42% of Japanese
backbone traffic. It also investigates per-customer traffic in one of
the ISPs by comparing DSL and fiber users, heavy hitters and normal
users, and geographic traffic matrices.

The aggregate data for the high level analysis was collected by
aggregating interface counters for residential traffic at edge routers
belonging to 7 ISPs. The detailed per-customer analysis was performed
at one ISP using the Cisco Sampled Netflow tool which allows users to
collect statistics for a subset of ingress IP traffic without incurring
significant CPU utilization overhead.

The authors' analysis has resulted in the following conclusions:

-Peak hours for residential traffic have shifted from office hours to
evening hours with traffic constantly flowing, probably because of p2p apps

-Inbound and outbound traffic is almost equal, contradicting the common
belief that there would be significantly more traffic outbound from the
ISP's network

-Fiber users account for 86% of inbound traffic and 75% of outbound traffic

-TCP dynamic ports account for 83% of total traffic (above 1024)

-A user's inbound and outbound traffic is positively correlated. Fiber
has more users with greater inbound than outbound traffic volume,
probably because these users are compensating for the DSL users where
the opposite is true

-63% of total residential volume is user to user traffic and 90% of
total traffic is between Japanese users (cultural and language barrier)

-There is poor locality (only 2-3% of total volume is within the same
prefecture) and most users engage in communication with a wide range and
number of peers

-Small segments of users dictate the overall behaviour: 4% of heavy
hitters account for 75% inbound volume and 60% outbound volume.

Analysis shows the existence of diverse and widespread heavy hitters who
appear to be casual users rather than more dedicated users; it is
difficult to draw a line between heavy users and the rest. Punishing
heavy users is not a long term solution because the trend is toward
everyone becoming a heavy user. Therefore, we can no longer view heavy
hitters as exceptional extremes since there are too many of them and
they are statistically distributed over a wide traffic volume range.

Additionally, not all heavy users on fiber are heavy DSL users who moved
to fibre. Some regular users signed up for fiber intially and then
sought ways to make use of the available capacity. This last
observation implies that a new bandwidth-intensive killer app could
create a dramatic surge in bandwidth use. Similarly, traffic could
decrease if a new breed of locality-aware p2p application became
popular. Furthermore, current p2p applications are tuned to asymmetric
connections such as cable and DSL and they are too aggressive in
capturing all the available upload bandwidth.

Although the layout of the paper makes it easy to get lost in a litany
of tables and figures, I enjoyed reading this study because it sheds
light on the future of broadband in North America. I also enjoyed it
because the data was collected from a unique and meaningful vantage point.
Received on Mon Nov 20 2006 - 22:11:41 EST