Summary: End to End Internet packet dynamics

The paper presents findings of a large scale experiments to study end to end
Internet packet dynamics by tracing TCP bulk transfers between 35 Internet
suites.

**

The analysis is not based on UDP or ICMP traffic rather it is based on TCP
traffic, close to real Internet traffic. Having TCP traffic for measurements
forces to distinguish between intertwined effects of transport protocol and
network traffic. The authors use a tool called tcpanaly, a tool that
understands the specifics of TCP implementations to achieve this. The
tcpdump measurements were done at Poisson intervals to avoid any bias.

Key observations:

1) Internet paths are correlated to route flutters, and sometimes are
subject to high incidence of reordering, but the effect is strong
site-dependent. The significant momentary increases in networking delays are
due to effects from both route changes and queuing.

2) Although packet replication happens, it is very rare. Packet
corruptions are not primarily due to noisy, slow links. The proportion of
Internet data packets corrupted is around 1 in 5000 but the authors fail to
provide definite conclusion about the overall Internet packet corruption
rates.

3) Internet path delays are asymmetric – meaning they have different
RTT delays for source to destination and destination to source.

4) One of the key contributions of the paper is the robust bottleneck
bandwidth computation algorithm. Prior work by Keshav and Jacobson is based
on the measurements done only at the sender and try to measure available
bandwidth instead of bottleneck bandwidth. Robust bottleneck measurement
algorithm overcomes these problems using "packet bunch modes" (PBM). It
achieves this by forming range of estimates and by allowing for multiple
bottleneck values. The idea is keep track of the size of the packets and
when packets are sent relative to one another and how they arrive relative
to another.

5) Loss rate increases are essentially due to higher loss rates during
the already loaded busy periods. Observing a zero loss connection at a given
point of time is quite a good predictor of observing zero loss connections
up to several hours in the future. So, losses occur in bursts. Thus caching
loss information could be beneficial.

6) Ack loss rates are higher than data loss rates because ack stream
does not adapt to network traffic whereas data stream does. Though both
ack/data loss rates seem fairly close to exponential distribution, the
authors fail to explain the link between adaptive sampling and exponential
distribution.

7) Packet losses occur in bursts and the distribution of packet loss
outage distribution has infinite variance. Loss bursts are greatly shaped by
drop tail queuing.

8) Internet delays variations occur primarily between 0.1-1 sec, but
extend frequently to much larger times.
Received on Thu Nov 16 2006 - 01:04:47 EST