Overview
Researching the Internet---building simulation and emulation
scenarios, for example---requires an accurate mental model of how
the network really behaves;
and the best way to get such a model and keep it up to date is through
ongoing, large-scale, and representative Internet measurement.
Application-centric measurement techniques are often unsuitable for this
purpose: for example, long-lived, large-scale active measurements can
generate too much traffic overhead.
M&M is a set of passive measurement tools suitable for large scale studies of
Internet path characteristics The first
tool, multiQ, uses equally-spaced mode gaps in TCP flows' packet
interarrival time distributions to detect multiple bottleneck capacities and
their relative order. Unlike previous tools, multiQ can discover up to
three bottlenecks from the tcpdump trace of a single flow, and can work
with acknowledgment as well as data interarrivals. The second tool,
mystery, robustly measures loss events, packet losses, and RTT changes
from a trace.
multiQ uses packet interarrivals to investigate questions about the
capacities along a path. Its basic insight is that packet
interarrival times, shown as a distribution, demonstrate
equally-spaced mode gaps
caused by intervening cross traffic packets on bottleneck links in the path. To
automate capacity estimation multiQ analyzes the interarrival PDF at a
progression of resolutions corresponding to a known set of common link speeds as
demonstrated in the figure above.
mystery reports loss events, lost packets, and fine-grained semi-RTT
measurements throughout the length of each flow. mystery concentrates on
fine-grained, accurate
measurements of basic properties. These measurements easily combine with
each other, and with multiQ's results.
Both tools have been calibrated
using over 10,000 experiments on 400 heteregenous Internet paths with known
likely link capacities. We also conducted four large-scale
(375 million-packet)
measurement studies of 258 diverse NLANR traces taken over the past two
years. The M&M suite makes it easy to summarize important properties
from these traces, including the distribution of bottleneck link capacities
(which has increased markedly over the last two years), the levels of
statistical multiplexing on bottlenecks (there is a wide range on both
small- and large-capacity bottlenecks), and loss event rates for packets
with different minimum-capacity bottlenecks.
Papers
Resources
-
NLANR has traces of flows collected from over 21 different sites representing a cross section of Internet traffic.
- RON offers a testbed of around 25 sites to conduct experiments.
People
Related research
Capacity
Measurement
- Internet measurements can be divided into two classes: active and passive.
Nettimer is the main passive tool for measuring capacity. Our work builds on the
insight from Nettimer but achoeves higher accuracy and can discover multiple
bottleneck capacities. Unlike Nettimer which requires receiver side traces to
achieve good accuracy, multiQ can work with reasonable accuracy with either
sender or receiver side traces. [Local pdf]
[Website].
There are many active tools for measuring capacity. While some focus on finding
the capacities of all paths, Pathrate discovers minimum capacity link along a
path. [Local pdf][Website]
TCP
Measurements
- The T-RAT tool uses passive traces to classify TCP flows based onthe main
factors limiting their rates. tcpanaly autimatically analyzes TCP behavior from
traces and focuses on finding implementation anomalies.
M. I. T. Computer Science and Artificial Intelligence Laboratory
·
200 Technology Square · Cambridge, MA 02139
·
USA