Review - End-to-End Internet Packet Dynamics

Review of End-to-End Internet Packet Dynamics
by Ivan Hernández

The paper presents a measurement and analysis of Internet packet
dynamics. The authors used 35 computer nodes in nine countries, which
provided a good heterogeneity to the measurements. The measurements
were two with a separation of one year. From each pair of nodes
perform TCP bulk transfers; next, the authors remove TCP behavior
from the measurements and use these data to do the analysis.

The authors describe the pathological behavior on a network. From the
measurements the authors show that out of order delivery is fairly
prevalent in the Internet; this effect is correlated with route
updates, where some packets follow an alternative and sometimes better
path. Because of this out of order delivery, the receiver does not
know whether a missing packet is delayed or it was dropped. The
authors suggest that by increasing the duplicate ACK threshold to
disambiguate a not-yet-received packet from an actual lost
packet. Another pathologies observable in the measurement are (1)
Packet replication, it is when the network delivers multiple copies of
the same packet, this problem is associated to malfunctioning layer
two devices; and (2) Packet corruption, in which the network delivers
to the receiver an imperfect copy of the original packet. One remark
of the authors is that TCP's 16 bit checksum is no longer adequate
because, according to their measurements, it does not detect data
corruption. The authors present an interesting discussion about how to
avoid redundant TCP retransmission. The proposed solution to prevent
most of the redundant retransmission is to have good retransmission
time-outs TCP implementations and deploy selective acknowledgment. The
authors make a detailed analysis of the collected data, in which they
use queueing theory to support their observations. The authors propose
a procedure called Packet Bunch Modes (PBM) to do bottleneck
estimation. This new procedure outperforms the previous packet-pair
estimators and performs well in scenarios with out of order delivery,
changes in bottleneck bandwidth, and multi-channel bottleneck links.

The authors claim that bottleneck bandwidth is asymmetric, the time
difference in two directions is more than 20% about 20% of the time;
therefore, sender-based bottleneck measurements yield to inaccurate
results. One interesting thing about the paper, is to see how actual
networks behave different from the models. For example, the out of
order delivery violates the network abstraction as a series of FIFO
queueing servers. Finally, the authors suggestions are based only on
the observations of the measurements; nevertheless, is not clear that
this observations will remain valid with the time, i.e., as new
network technologies and speeds are available, how much these
measurements and their results remain valid?
Received on Wed Nov 15 2006 - 23:26:21 EST