Replace struct timeval with struct timespec as the main data type for
timestamps. This will allow the NTP code to work with timestamps in
nanosecond resolution.
Estimate asymmetry of network jitter on the path to the source as a
slope of offset against network delay in multiple linear regression. If
the asymmetry is significant and its sign doesn't change frequently, the
measured offsets (which are used later to estimate the offset and
frequency of the clock) are corrected to correspond to the minimum
network delay. This can significantly improve the accuracy and stability
of the estimated offset and frequency.
The minsamples and maxsamples directives now set the default value,
which can be overriden for individual sources in the server/peer/pool
and refclock directives.
Add new functions to change source's reference ID/address and reset the
instance. Use that instead of destroying and creating a new instance
when the NTP address is changed.
Reachability is no longer a requirement for selection. An unreachable
source can remain selected if its newest sample is not older than the
oldest sample from all reachable sources.
This is useful to prevent reselection when an accurate source uses a
very short polling interval (e.g. refclock) and is occasionally
unreachable for short periods of time.
If the newly accumulated sample is not newer than than the last one,
discard the source history and start from scratch. This can happen after
loading an invalid dump or when the system clock was stepped.
With iburst and very jittery sources the source skew can reach very high
values which makes the NTP test4 fail even with relatively small delays.
Limit the skew to 2000 ppm to avoid getting state where a source is unable
to accept more than first three iburst samples.
Dividing the weights by variance or unweighted variance seems to have a
significant negative impact on response with normally distributed
network delays.
Divide by the difference between the mean and minimum distance instead.
It should be stable as there is no loop and the response seems to be a
good compromise between the original minimum distance weighting which
works well with normally distributed delays and the variance weighting
which works well with exponentially distributed delays.
Require that the ratio of the increase in delay from the minimum one in
the stats data register to the standard deviation of the offsets in the
register is less than maxdelaydevratio or the difference between
measured offset and predicted offset is larger than the increase in
delay. In the allowed delay increase is included also skew and maximum
clock frequency error.
maxdelaydevratio is 10.0 by default.
