Review of End-to-End Internet Packet Dynamics

The authors report and discuss their findings of a study of Internet packet
dynamics performed with 20,000 TCP bulk transfers of 100 KBytes between 35
sites. The transfers were logged at both sender and receiver in order to
gain additional insight. In order to obtain certain measurements, the
authors have devised several interesting methods for obtaining measures of
"bottleneck bandwidth", packet delay, etc.

Their findings are summarized below:

Pathologies

1) ACK reordering is fairly prevalent in the Internet with data packets
being reordered more often than ACKs, but it has little impact on
performance except for retransmit ACK timing. The most common cause is
route flutter (better paths coming up). The sender cannot soundly infer
whether the packets it sends are likely to be ordered.

2) Packet replication seldom occurs, but the authors do not provide a
satisfying explanation of the cause.

3) The rate of packet corruption is higher than previously believed and
TCP's 16-bit checksum may be inadequate. Authors do not make a general
conclusion about the corruption rate on the Internet.

Bottleneck Bandwidth

Previous attempts at estimating bottleneck bandwidth using packet pair
techniques had problems because of out of order delivery, clock resolutions,
changes in bottleneck bandwidth and multi-channel bottleneck links. The PBM
algorithm for estimating bottleneck bandwidth is described. Similar to the
packet pair algorithm, it is based on the simple observation that two
packets sent from a fast link (provided they travel the same route) will
have their spacing at the receiver equal to the amount of time required to
forward a packet through the bottleneck element. Unlike packet pair, it
forms estimates based on a range of packet sizes and allows for multiple
bottleneck values. They found that the most common bottleneck is a T1 line,
that bottlenecks change infrequently, and that in many cases the bottlenecks
for the send and receive paths are different.

Packet Loss

1) Authors claim that ACK losses give us a better picture of overall
Internet loss patterns and that data losses are better at expressing
available bandwidth because of congestion control algorithms. They find
that the measured loss rate of a connection is not a good long term
predictor of the future loss rate, even for only a few minutes in the
future, but zero loss is a good predictor. Flows show considerable
probability of zero loss and packet loss on the forward and reverse paths is
sometimes fully independent.

2) Packet losses occur in bursts.

3) Ensuring proper RTO calculations and the SACK option together eliminate
almost all avoidable redundant retransmissions.

Packet Delay

1) Timing compression is a rare event and it could be handled with outlier
filtering.

2) Delay occurs mostly on a time scale of 0.1 to 1 secs but frequently
exceeds this range.

I believe the main contributions of this paper are the proposed methods for
measuring packet dynamics. The data itself has only been collected from 35
sites, the authors cannot explain some of the variation between their data
sets and the reported results are not too unexpected. The authors urge the
re-examination of some of the fundamental assumptions about Internet
networks but, in reality, beliefs such as "the Internet is a FIFO network"
are only used as simplifying abstractions.
Received on Thu Nov 16 2006 - 08:25:09 EST