Adaptive Kalman Filter for Time
Synchronization over Packet-Switched Networks
Tese de Doutorado de Leandro Fabrício Auler
The
importance of Time synchronization for distributed applications is
growing. The
most currently used synchronization protocol is NTP. NTP and GPS satisfy the majority
of the world’s time synchronization needs, although, GPS have
limitations. GPS cannot function
inside
buildings, or in places where there are physical obstructions and high
amounts
of electromagnetic noise. In addition, GPS is a costly solution. This work
proposes a new approach to synchronize time among machines
interconnected by
computer networks, which minimizes the disturbance introduced by the
network with
no extra hardware cost. The methodology employs high-resolution clocks
and
stochastic methods to minimize the error introduced in time measurement
carried
out across the network. It also uses an adaptive approach to
continuously tune
the synchronization mechanism to the network traffic characteristics.
An
increase of time synchronization accuracy of two orders of magnitude,
in comparison
with traditional estimation methods was obtained as a result. In order to
synchronize two clocks over a computer network, it is required to
estimate some time parameters in the remote machine. The time
difference between two clocks at some time is known as the offset(t),
whose definition is given by (1), where t is a reference clock and t2
is a compared clock time. The Offset is sometimes referred as clock
error. The rate which the offset changes is called skew, and is shown
in equation (2). A skew equal to zero means that both clocks have the
same frequency. Temperature changes and electrical changes affect the
frequency of the clocks. The frequency deviation of a clock is
denominated drift, which can be defined as the second derivate of the
offset, and is shown in equation (3). For a crystal oscillator, the
frequency deviation is often very small and can be ignored. For ideal
purposes, the drift is considered zero through out this text. As a
direct consequence of this idealization, the skew is considered
constant. (1)
(2) (3) |
|
Main parameters used to compare two different clocks | Result of offset estimating
by
polling on the Internet on symetric paths and expected signat |
|
Example showing how the
packet
filter and the assistant clock acts on two machines |
Time
polling between two clocks on
different computers, a assistant clock and a reference clock |
NTP offset observed during 90h | Comparison between Kalman and NTP |
Artigos
publicados AULER,
Leandro Fabricio; D'AMORE, R. High
Resolution Time Estimation among Remote Machines Using Kalman Filtering.
In: WPerformance - V Workshop de Desempenho de Sistemas Computacionais
e de Comunicação 2006.
AULER, Leandro Fabricio; D'AMORE, R. Adaptive Kalman Filter for Time Synchronization over Packet-Switched Networks (An Heuristic Approach). In IEEE ICST COMSWARE 2007. |