I work in computer networking, with a particular interest in the
modelling and measurement of tele-traffic, in particular the TCP/IP
packet data flowing over the Internet. In general terms the aim
is to understand in greater detail how the traffic sources and network
structure and protocols interact, with a view to making the network,
and end applications, more efficient. This has lead to work in a
number of seemingly different areas including statistical estimation
and clock synchronisation.
Software clocks in computers are based
on local hardware synchronising to more accurate remote clocks.
Currently the NTP system is used to synchronise hosts to
remote servers across the Internet. The stability of modern PC
hardware however actually supports higher accuracy and robustness that
NTP currently delivers. We are developing replacement for the NTP
clients and servers based on new principles, in particular the need to
distinguish between difference clocks
and absolute clocks
, and the
associated primacy of rate stability over absolute clock error.
The RAD difference clock, for example, can measure RTTs to under a microsecond, even
if connectively to the time server is lost for periods of over a week!
The SyncLab Project
has as its aim to provide a complete new system for network timing.
Currently client software is available for Linux and BSD Unix which can connect to existing NTP
servers. Download details, documentation and a number publications can be found
on the project page.
This project has been made possible in part by a grants from
the Australian Reseach Council, the Cisco University Research Program Fund at Silicon Valley
Community Foundation, two Google Research Grant Awards, and a partnership with Symmetricom Inc. (now Microsemi).
By Network Inference
we mean the application of sophisticated statistical techniques for the
translation of imperfect network measurement data
of the operation,
mechanisms, state, use, performance, and fairness of the network. For
example, the inference
techniques of Network Tomography use data probes like X-rays to look
`inside' the network body
to locate overloaded links. Such a capability is valuable across the
spectrum of network users: for
the Internet public to determine who is responsible for slow downloads,
for network operators to
troubleshoot their networks, and for regulators to police compliance to
Service Level Agreements. I am active in following three
directions within network inference.
Here test packets or `probes' are injected
into the network, collected at a set of receivers around the network edge, and inferences made on
the end-to-end path based on measured end-to-end delays and/or losses.
My interests in this area range from the underlying measurement
infrastructure, the `heuristic' design of effective probe streams and
their analysis, and the rigorous application of queueing theory to
active probing problems. My colleagues in this area include Attila Pásztor, François Baccelli
Sridhar Machiraju, and Jean Bolot
. The current focus in on the theoretical side, trying to build up a science of convex networks
, a property which will allow optimal probing strategies to be well defined and devised.
Whereas in active probing inference probes,
typically, follow a single end-to-end path which is modelled as a
sequence of queues, by Network Tomography we mean a class of inversion
problems (which may or may not involve probing) which is much more
ambitous in the spatial dimension (multiple sources and receivers over
the network) but treats nodes using simple black box models for loss or
delay. For example a link may be characterised simply by a single
number, a loss probability. My work in this area primarily
concerns multicast probes which flow from a single source to multiple
receivers, with copies being made at each branch point, tracing out a
measurement tree in the process. I have worked on loss, delay,
and topology inference in this context, with a major focus on
generalising beyond the classical simplifying assumptions of perfect
spatial and temporal independence. My colleagures here include Vijay Arya
, Nick Duffield
, François Baccelli
, and Rhys Bowden.
Route Tracking (advanced Traceroute)
One of the oldest probe based inference tools is traceroute
which makes use of features of the TCP/IP/ICMP protocol suite to trace
out the IP-level path between a source and destination. However,
because of load balancing, a high proportion of routes in the Internet
today have multiple branches, and failing to take this into account can
produce meaningless topology inferences. Paris Traceroute
is a generalised Traceroute tool which attempts to trace routes as they
really are, whether branched or not. I work with Paris Traceroute
researchers Renata Teixeira
, Timur Friedman
, Christophe Diot
, and Ítalo Cunha
applying statistical ideas to the problem of controlling the error in
what is effectively topology estimation, and in efficiently tracking
(branched) route changes over time.
In resource constrained environments
such as within core Internet routers, accurate measurement of traffic
features and statistics can be difficult. Two canonical approaches to
fast approximate measurement are: sampling
of the data, and sketching
which means the use of compact data structures which are fast to
update, but which store information imperfectly. My work in this
area has focussed on the measurement of the flow size distribution
(number of packets in a flow such as a TCP connection). This is an important metric for numerous applications including traffic modelling, management, and attack detection.
We evaluate data collection mechanisms in a Fisher Information
framework, comparing various sampling and sketching approaches in order
to determine which inherently captures the most information about the
distribution. We developed the Dual Sampling (DS) and the
Optimised Flow Sampling Sketch (OFSS) (see below for OFSS code) methods which are both
capable of being implemented at high speed. This work is with Paul Tune
Packet traffic has scale invariance
features, in particular long range dependence (LRD), which impacts on
network performance, performance analysis, accuracy of simulation, and
parameter estimation. My underlying interest has been in traffic
modelling, but in the analysis of real data, the need for more powerful
estimation tools naturally arises and I have also worked extensively in
this area. Much of the work here involves wavelets and is in
with Patrice Abry
from the Signal Analysis group of the
Ecole Normale Supérieure de Lyon
We confirmed that fractal traffic is real - not just an artifact of
poor estimation tools - and introduced wavelet analysis to the area.
Other colleagues include Patrick Flandrin
, Murad Taqqu
, Walter Willinger
, and Matthew Roughan
. Associated Matlab code is available for download at the links below.
The Flow Sampling Sketch, or FSS, is a skampling
method, that is a hybrid between sampling and sketching, which allows
the flow size distribution to be estimated with very low resource
requirements in both time and memory. The OFSS method is an optimally
tuned/calibrated FSS with a statistical performance which is within a constant factor of
Flow Sampling, which is known to be optimal.
The OFSS Matlab code
the critical calibration parameter of OFSS, pf*, to be calculated
for any given input load alpha. It also gives the associated
amount of information, and minimal variance, of any estimator using the
OFSS method. More details are given on the code page.