Review: End-To-End Internet Path Dynamics

The paper presents the finding of a large scale internet measurement study
focusing on study of end-to-end internet packet dynamics. In order to
cater for heterogeneity of internet paths and reducing logistics
difficulties, author a developed a framework in which a daemon is made to
run on different internet sites to facilitate measurement. Analyses
presented in this paper are based on 35 diverse internet sites with
measurements taken in two different years, 1995 and 1996. Measurements
include TCP bulk transfer conducted between these 35 sites. As TCP is
being used for measurement so in order to model the network behavior
accurately author also developed a program that separates the transport
behavior from network behavior.

Author begins his analysis by considering unexpected network behavior that
includes out-of-order packet delivery, packet replication and packet
corruption. Author argues that out-of-packet delivery is fairly common
phenomenon and this behavior is asymmetric with data packets reordered
more that acknowledge (ack) packets. The reordering phenomenon is site
dependent and occurs mainly because of frequent changes in network paths.
Based on the observing of reordering events author has suggested that the
70% of reordered deliveries could be identified if the receiver waits for
20msec thus improving fast retransmit mechanism. Secondly redundant
retransmission could easily be eliminated if SACK is used for
acknowledgement. The author then analyses the packet loss in internet but
has not provided any conclusive remarks about the reasons of occurrence of
packet loss. In the analysis of packet corruption author argues that the
packet corruption is not mainly due to noise and slow links as he has
shown that link with much high bandwidth also suffer from packet
corruption. Secondly author has highlighted that TCP’s 16-bit checksum is
not sufficient enough to detect these corruption as on average out of
65,536 packets one will be accepted. So the author suggests use of 32-bit
checksum to improve TCP’s ability to detect corrupted packets.

Author has also proposed the bottleneck bandwidth measurement technique.
Author has purposed to use receiver based packet pairing (RBPP) instead of
sender based packet pairing (SBPP) to cater for asymmetric properties of
network path. Though RBPP eliminates errors due to asymmetric but it has
some inherent inconsistencies due to packet pairing. This includes out of
order delivery, clock resolution limitation and multi-channel bottleneck
links. To overcome these shortcomings author has purposed to use ‘packet
bunch modes’ (PBM) in which k>2 packets are sent for measurements. In the
last sections author discusses the packet loss and packet delays. Authors
argue that the increase in packet loss from 1995 to 1996 nearly doubled
and this increase was not because of increase in window size. Also the
probability of losing a loaded packet is much higher than that of unloaded
packet. Based on the packet lost analysis author has also concluded that
standard TCP RTO estimation algorithm works well and deploying the SACK
further eliminates redundant transmissions.

Although the author claims that the current study can be categorized as a
large scale internet measurement but I still believe the links used in
this study are almost of same nature as the author is studying interaction
between internet sites which I believe are well provisioned. The results
could have been better if the author has considered internet users who are
behind dialup and DSL lines. I believe inclusion of these path would have
some effect on the overall study.
Received on Thu Nov 16 2006 - 09:08:40 EST