Review: End-to-End Internet Packet Dynamics

The authors conduct a large scale study of Internet packet dynamics by
tracing 20000 TCP bulok transfers between 35 Internet sites. They chose
TCP as it represents real world traffic and TCP packet streams allow
fine-scale probing without unduly loading the network because it deals
with congestion. However, the latter reason also makes it more difficult
to separate effects of the transport protocol from that of the network.
Furthermore, TCP packets are sent over a wide-range of time scales,
which complicates the correlational and frequency-domain analysis, and
so the paper does not deal with frequency-domain analysis.
The paper first discusses network pathologies including:
- Out-of-order delivery: is strongly site-dependent and correlated with
route fluttering.
- Packet replication: is also site-specific and occurs both at the
receiver and sender.
- Packet corruption: required further study and the paper provided no
overall corruption rates.
The authors then describe a robust receiver-based algorithm for
estimating bottleneck bandwidth called Packet Bunch Modes (PBM), which
is based on estimating bandwidth for ranges of packet bunch sizes and
allowing multiple bottleneck values. They briefly describe the popular
packet-pair algorithms, based on queuing of two packets at the
bottleneck link, and discuss some of the limitations:
- Noise on the return path for sender-based packet-pair
- Out-of-order delivery: packets might not take the same path
- Limitations due to clock resolution: the receiver can only report a
particular minimum difference in time.
- Changes in bottleneck bandwidth due to routing changes
- Multi-channel bottleneck links: packets may be directed over different
channels.
The paper concludes that PBM is more effective than sender-bases packet
pair and comparable to receiver-based packet pair.
The authors investigate patterns of packet loss and find that loss
events are not well-modeled as independent and the distribution of the
duration of loss events exhibits infinite variance. Another interesting
finding is that the loss rate of acks is higher than data packets, and
this may be due to the fact that acks do not observe any congetsion
control.
Finally, the paper analyzes variations in packet transit delays
(focusing on network dynamics rather than transport protocol dynamics)
as indicators of congestion periods and find that congestion periods
span a wide range of time scales.
The authors conclude:
- TCP-based measurements provide a viable means for assessing end-to-end
packet dynamics as long as packet filter errors and TCP effects can be
removed.
- It is necessary to exercise caution in regarding any aspect of packet
dynamics as "typical" because there is a wide range of behaviours.
- Common assumptions such as in-order packet delivery, FIFO bottleneck
queuing, independent loss events, single congestion time scales and path
symmetries are all violated.
- TCP retransmission strategies are conservative when implemented
correctly.
- Sender-only measuremnt techniques are inferior to those that include
receiver-cooperation due to path asymmetries and reverse-path noise.
 
The paper conducts an interesting study. However, I feel that some
findings are presented without any reasonable deductions/reasonings
being given. Also, at some points the authors seem unable to failry
compare their two sets of data.
Received on Wed Nov 15 2006 - 13:49:27 EST