This module supports the computation of the skew between two clocks:
the (relative) skew is the speed with which two clocks diverge. For
instance, if yesterday two clocks, at the same time, showed
respectively 10:00 and 10:05, while today when the former shows 10:00
the latter shows 10:04, we say that their relative skew is 1 minute/24
hours, roughly 7E-4.

The module contains one single subroutine, which accepts as input a
pair of timestamps, associated to a message from host A to
host B: the timestamps correspond to the time when the message was
sent, and to the time when message is received. Each timestamp
reflects the value of the local clock where the operation takes place:
the clock of host A for the send, the clock of B for the receive.

Please note that the module does _not_ contain any message exchange
facility, but only the mathematics needed to perform the skew
approximation, once timestamps are known.

The subroutine takes as argument:

- a reference to a hash where values related to the timing of the
network path from A to B;

- a 2-elems array (a data point in the sequel) containing the
timestamp of the receive event, and the differece between the send
timestamp and the receive timestamp for one message;

- a stack containing some data points, those that form the convex hull.

The usage is very simple, and is illustrated by the following example:

#!/usr/bin/perl -w
use strict;
use Time::Skew;

# Initialize data
my $hull=[];
my $result={};
while ( 1 ) {
# Exchange message and acquire a new data point
  my $datapoint = acquire();
# Call the convexhull subroutine
  Time::Skew::convexhull($result,$datapoint,$hull);
# After first message some results are still undefined
  ( defined $result->{skewjitter} ) || next;
# here you can use the results

  };
}
 
The data returned in the "result" hash is the following:

result->{skew}       the clock skew;
result->{skewjitter} the variance of the skew estimate, used to estimate
                     convergence;
result->{jitter}     difference between the current delay and the
                     previous delay;
result->{delay}      the communication delay, decremented by a constant 
                     (yet unknown) value, used to compute communication
                     jitter;
result->{elems}      the number of data points in the convex hull;
result->{select}     the index of the data point in the convex hull used to 
                     compute the skew;
result->{itimestamp} the timestamp, first element in the data point
                     just passed to the subroutine;
result->{delta}      the timestamp difference, second element in the data
                     point just passed to the subroutine;

The data returned in the "hull" stack is a series of data points,
selected from those passed to successive calls of the subroutine. The
number of data points in the "hull" stack usually does not exceed 20
units.

The algorithm is very fast: each call consists in scanning at most all
data points in the "hull" stack, performing simple arithmetic operations
for each element.

The algorithm must be fed with a sequence of data points before
returning significant results. The accuracy of the estimate keeps
growing while new data points are passed to the subroutine. A rough
rule of thumb to evaluate estimate accuracy is to observe the skew
jitter, and assume it corresponds to the skew estimate accuracy. Paths
with quite regular communication delay (small jitter) converge faster.

INSTALLATION
============

perl Makefile.PL
make
make install

TESTING
=======

Move into the t/ directory and call the testing program tester.pl. 

$ ./tester.pl > /tmp/out

The program takes data points from the "test_data" file, consisting of
10000 data points. Their processing takes a few seconds. Five lines

XXX HULL ALGORITHM ERROR (negative delay -7.105427357601e-15)

are expected on the stderr. As long as they are very small numbers
these warnings can be ignored. Collect the stdout in a file and compare
with expected stdout in file check.data.

The input data points are obtained simulating two clocks with a skew
of 1E-4. After 10000 data skew is estimated with an error of 4E-7,
while the skewjitter (optimistically) promises an error of 4E-9.