End-to-End Internet Packet Dynamics

Paper Review: End-to-End Internet Packet Dynamics

The paper presents a large-scale measurement study to characterize
end-to-end packet
dynamics in the Internet. There is inherent difficulty in taking such
measurements due
to the dynamic nature of Internet paths, network heterogeneity and the
difficulty of
collecting results from a large number of Internet paths. Their
measurements are based
on TCP in order to closely model real-world traffic. However, using
TCP also adds complexities such as distinguishing between the effects
of the protocol and the network. Secondly, TCP packets are sent with
irregular spacing further complicating analysis. These reasons led
the authors to steer away from the types of analyses affected by these
issues.

The analysis first characterizes pathological behaviour to ensure the
results are not skewed by these events. One such event is
out-of-order delivery which was shown to occur frequently. Also, it
was more prevalent in data packets compared to acknowledgements since
they are sent much closer together. The frequency of out-of-order
delivery varied greatly across the various connections. The behaviour
is highly unpredictable and is due to various causes including route
fluctuations. Out-of-order packets can cause unnecessary
retransmissions. The fast retransmit mechanism waits for 3
duplicate-acks. The authors try varying values and timeouts to see if
TCP can be more fine-tuned. However, they conclude that the value
cannot be safely dropped to less than 3.

Packet replication, a rare but possible event, also has the potential
for skewing the analysis. These can occur as a result of sender
retransmission or by link-level technologies with retransmissions.
Packet corruption is the final pathology examined in which an
incorrect copy of the data is delivered by the network. The study
suggests that 1 in 5000 data packets are corrupted in the Internet.
This result suggests that TCP?s 16-bit checksum is not adequate.

The paper presents its method for estimating bottleneck bandwidth, an
upper bound on how fast data can be transmitted. The
self-interference time-constant is defined as the time it takes to
forward a given packet through the bottleneck link. This is necessary
to ensure they were not measuring delays caused by the
self-interference of their own packets. In previous works,
measurements for available bandwidth were taken using the packet-pair
method, which was pointed out to have certain flaws. The paper
introduces a novel approach known as ?packet bunch modes? (PBM) for
determining bottleneck bandwidth. Using PBM, they found that
receiver-based packet pair schemes worked well whereas sender-based
schemes provided much less accurate results.

The paper then looks into packet loss. One interesting result was
that a zero-loss connection at a given point in time is a good
predictor of observing a zero-loss connection up to several hours in
the future. There were other forms of prediction that were found but
the predicting power fades over time. Packet loss was easily
categorized as occurring in bursts and their sizes span several orders
of magnitude.

The final measurement looks in to packet delay in regards to one-way
transit times. One interesting phenomenon was packet compression. If
ack packets are subject to packet compression, there is a strong
chance the sender-side sliding window will begin to send out data
faster than the receiver can handle. Thus packet compression is
harmful to TCP operation and has the potential to induce congestion in
the network.

This paper is interesting in that it doesn?t provide definitive,
concrete answers but does provide several key insights. The Internet
is too large and complex to have simple answers regarding end-to-end
behaviour. The paper could have benefited from more graphs. The
qualitative analysis was indeed insightful; however, there were
several occasions where numbers were simply reported. The data could
have been presented more clearly using graphs and tables. Also, the
graphs that were provided did not seem to complement the analysis
well. The paper focuses on the interesting topic of end-to-end
behaviour in the Internet. Although the paper is a step in the right
direction, there is a great deal of work which can be furthered in
this area. One basic insight which can be gained from this work is
that the Internet is highly unpredictable due to several factors and
because of this it is very difficult to model its behaviour.
Furthermore, common simplifying assumptions in Internet models were
found to be often violated.

-- Nadeem Abji
Received on Thu Nov 16 2006 - 00:12:40 EST