Common tools used for evaluating end-to-end network behavior, such as ping
and traceroute, are inadequate and limited. While the Internet was not
designed with measurement as a primary goal, daily maintenance and
administration requires a certain level of measurement capability.
Constructs inherent in the Transmission Control Protocol (TCP) can allow us
to measure packet loss across the network. More importantly, this method
provides a mechanism for determining directional packet loss rates.
In this paper Stefan Savage derives several clever tricks with the TCP
protocol to allow the determination of bidirectional packet loss. Here,
forward packet loss is distinguished from reverse packet loss to provide
finer grained detail in measurements. This is an important characteristic
as, for example, acknowledge segments are more loss tolerant than data
segments.
It is important to note that in order for this tool to be useful, connection
establishment must be possible. However, I would guess that most
administrative measurements involve machines that have become unreachable
over the network. (At least this is my experience.) That is, most time is
spent troubleshooting issues where connections are not getting through at
all. In this case, the method outlined in this paper would provided no
directional indication of the failure and would be no more useful than ping,
and less useful than traceroute.
Also, If current tools are truly inadequate, then why have they been
consistently in use since the advent of the Internet? It is true that ping
does not work against every host on the Internet, and a traceroute may not
reach the destination node most of the time. But, use of these tools has
remained; these inadequacies have not completely rendered the tools useless.
Even though I do not agree that classic tools have worm out their use, I did
enjoy this paper very much. Savage does a good job of addressing my concerns
with elegant and innovative solutions. Interesting examples of this include
the notion of "fast ACK parity" and the use of packet trimming in the
receiver window. It will be interesting to see if these solutions can be
used in other Internet property measurement.
Received on Sun Oct 30 2005 - 10:25:51 EST
This archive was generated by hypermail 2.2.0 : Sun Oct 30 2005 - 13:01:06 EST