NetBSD/usr.sbin/xntp/html/driver27.html

427 lines
19 KiB
HTML

<!-- $NetBSD: driver27.html,v 1.1 1998/12/30 20:20:34 mcr Exp $ -->
<!DOCTYPE HTML PUBLIC "-//IETF//DTD HTML Strict//EN">
<html><head><title>
Arcron MSF Receiver
</title></head><body><h3>
Arcron MSF Receiver
</h3><hr>
<p><h4>Synopsis</h4>
<p>Address: 127.127.27.<var>u</var>
<br>Reference ID: MSF
<br>Driver ID: ARCRON_MSF
<br>Serial Port: <code>/dev/arc<var>u</var></code>; 300 baud, 8-bits, 2-stop,
no parity
<br>Features: <code>tty_clk</code>
<p><h4>Description</h4>
<p>This driver supports the Arcron MSF receiver, and would probably
also support the DCF77 variant of the same clock. The clock reports
its ID as ``<code>MSFa</code>'' to indicate MSF as a source and the use
of the ARCRON driver.
<p>This documentation describes version V1.1 (1997/06/23) of the source
and has been tested (amongst others) against xntpd3-5.90 on Solaris-1 (SunOS 4.1.3_U1 on
an SS1 serving as a router and firewall) and against xntpd3-5.90 on
Solaris-2.5 (on a SS1+ and TurboSPARC 170MHz). This code will probably
work, and show increased stability, reduced jitter and more efficiency
(fewer context switches) with the <code>tty_clk</code>
discipline/STREAMS module installed, but this has not been tested. For
a to-do list see the comments at the start of the code.
<p>This code has been significantly slimmed down since the V1.0 version,
roughly halving the memory footprint of its code and data.
<p>This driver is designed to allow the unit to run from batteries as
designed, for something approaching the 2.5 years expected in the usual
stand-alone mode, but no battery-life measurements have been taken.
<p>Much of this code is originally from the other refclock driver files
with thanks. The code was originally made to work with the clock
by <a href="mailto:derek@toybox.demon.co.uk">Derek Mulcahy</a>, with
modifications by <a href="mailto:d@hd.org">Damon Hart-Davis</a>.
Thanks also to <a href="mailto:lyndond@sentinet.co.uk">Lyndon David</a>
for some of the specifications of the clock.
<p>There is support for a Tcl/Tk monitor written by Derek Mulcahy
that examines the output stats; see the
<a href="http://www2.exnet.com/NTP/ARC/ARC.html">ARC Rugby MSF Receiver</a>
page for more details and the code.
<p>Look at the notes at the start of the code for further information;
some of the more important details follow.
<p>The driver interrogates the clock at each poll (ie every 64s by
default) for a timestamp. The clock responds at the start of the next
second (with the start bit of the first byte being on-time). The time
is in `local' format, including the daylight savings adjustment when it
is in effect. The driver code converts the time back to UTC.
<p>The clock claims to be accurate to within about 20ms of the
MSF-broadcast time, and given the low data transmission speed from
clock to host, and the fact that the clock is not in continuous sync
with MSF, it seems sensible to set the `precision' of this clock to -5
or -4, -4 being used in this code, which builds in a reported
dispersion of over 63ms (ie says ``This clock is not very good.'').
You can improve the reported precision to -4 (and thus reduce the base
dispersion to about 31ms) by setting the fudge <code>flag3</code> to
<code>1</code>.
<p>Even a busy and slow IP link can yield lower dispersions than this
from polls of primary time servers on the Internet, which reinforces
the idea that this clock should be used as a backup in case of problems
with such an IP link, or in the unfortunate event of failure of more
accurate sources such as GPS.
<p>By default this clock reports itself to be at stratum 2 rather than
the usual stratum 0 for a refclock, because it is not really suited to
be used as other than a backup source. The stratum reported can be
changed with the <code>fudge</code> directive to be whatever you like.
After careful monitoring of your clock, and appropriate choice of the
<code>time1</code> fudge factor to remove systematic errors in the
clock's reported time, you might fudge the clock to stratum 1 to allow
a stratum-2 secondary server to sync to it.
<p>The driver code arranges to resync the clock to MSF at intervals of
a little less than an hour (deliberately avoiding the same time each
hour to avoid any systematic problems with the signal or host). Whilst
resyncing, the driver supplements the normal polls for time from the
clock with polls for the reception signal quality reported by the
clock. If the signal quality is too low (0--2 out of a range of 0--5),
we chose not to trust the clock until the next resync (which we bring
forward by about half an hour). If we don't catch the resync, and so
don't know the signal quality, we do trust the clock (because this
would generally be when the signal is very good and a resync happens
quickly), but we still bring the next resync forward and reduce the
reported precision (and thus increase reported dispersion).
<p>If we force resyncs to MSF too often we will needlessly exhaust the
batteries the unit runs from. During clock resync this driver tries to
take enough time samples to avoid <code>xntpd</code> losing sync in
case this clock is the current peer. By default the clock would only
resync to MSF about once per day, which would almost certainly not be
acceptable for NTP purposes.
<p>The driver does not force an immediate resync of the clock to MSF
when it starts up to avoid excessive battery drain in case
<code>xntpd</code> is going to be repeatedly restarted for any reason,
and also to allow enough samples of the clock to be taken for
<code>xntpd</code> to sync immediately to this clock (and not remain
unsynchronised or to sync briefly to another configured peer, only to
hop back in a few poll times, causing unnecessary disturbance). This
behaviour should not cause problems because the driver will not accept
the timestamps from the clock if the status flag delivered with the
time code indicates that the last resync attempt was unsuccessful, so
the initial timestamps will be close to reality, even if with up to a
day's clock drift in the worst case (the clock by default resyncs to
MSF once per day).
<p>The clock has a peculiar RS232 arrangement where the transmit lines
are powered from the receive lines, presumably to minimise battery
drain. This arrangement has two consequences:
<ul>
<li>Your RS232 interface must drive both +ve and -ve
<li>You must (in theory) wait for an echo and a further 10ms between
characters
</ul>
This driver, running on standard Sun hardware, seems to work fine; note
the use of the <code>send_slow()</code> routine to queue up command
characters to be sent once every two seconds.
<p>Three commands are sent to the clock by this driver. Each command
consists of a single letter (of which only the bottom four bits are
significant), followed by a CR (ASCII 13). Each character sent to the
clock should be followed by a delay to allow the unit to echo the
character, and then by a further 10ms. Following the echo of the
command string, there may be a response (ie in the cae of the
<code>g</code> and <code>o</code> commands below), which in the case of
the <code>o</code> command may be delayed by up to 1 second so as the
start bit of the first byte of the response can arrive on time.
The commands and their responses are:
<dl>
<dt><code>g</code> CR
<dd>Request for signal quality. Answer only valid during (late part
of) resync to MSF signal. The response consists of two characters as
follows:
<ol>
<li><dl compact>
<dt>bit 7</dt> <dd> parity
<dt>bit 6</dt> <dd> always 0
<dt>bit 5</dt> <dd> always 1
<dt>bit 4</dt> <dd> always 1
<dt>bit 3</dt> <dd> always 0
<dt>bit 2</dt> <dd> always 0
<dt>bit 1</dt> <dd> always 1
<dt>bit 0</dt> <dd> = 0 if no reception attempt at the moment,
= 1 if reception attempt (ie resync) in progress
</dl>
<li><dl compact>
<dt>bit 7</dt> <dd> parity
<dt>bit 6</dt> <dd> always 0
<dt>bit 5</dt> <dd> always 1
<dt>bit 4</dt> <dd> always 1
<dt>bit 3</dt> <dd> always 0
<dt>bit 2--0</dt> <dd> reception signal quality in the range 0--5
(very poor to very good); if in the range 0--2 no
successful reception is to be expected. The reported
value drops to zero when not resyncing, ie when first
returned byte is not `3'.
</dl>
</ol>
<dt><code>h</code> CR
<dd>Request to resync to MSF. Can take up from about 30s to 360s.
Drains batteries so should not be used excessively. After this the
clock time and date should be correct and the phase within 20ms of time
as transmitted from Rugby MSF (remember to allow for propagation
time). By default the clock resyncs once per day shortly after 2am
(presumably to catch transitions to/from daylight saving time
quickly). With this driver code we resync at least once per hour to
minimise clock wander.
<dt><code>o</code> CR
<dd>Request timestamp. Start bit of first byte of response is on-time,
so may be delayed up to 1 second. Note that when the BST mode is in
effect the time is GMT/UTC +0100, ie an hour ahead of UTC to reflect
local time in the UK. The response data is as follows:
<ol>
<li>hours tens (hours range from 00 to 23)
<li>hours units
<li>minutes tens (minutes range from 00 to 59)
<li>minutes units
<li>seconds tens (seconds presumed to range from 00 to 60 to allow for leap second)
<li>seconds units
<li>day of week 1 (Monday) to 7 (Sunday)
<li>day of month tens (day ranges from 01 to 31)
<li>day of month units
<li>month tens (months range from 01 to 12)
<li>month units
<li>year tens (years range from 00 to 99)
<li>year units
<li>BST/UTC status <dl compact>
<dt>bit 7</dt> <dd> parity
<dt>bit 6</dt> <dd> always 0
<dt>bit 5</dt> <dd> always 1
<dt>bit 4</dt> <dd> always 1
<dt>bit 3</dt> <dd> always 0
<dt>bit 2</dt> <dd> = 1 if UTC is in effect (reverse of bit 1)
<dt>bit 1</dt> <dd> = 1 if BST is in effect (reverse of bit 2)
<dt>bit 0</dt> <dd> = 1 if BST/UTC change pending
</dl>
<li>clock status<dl compact>
<dt>bit 7</dt> <dd> parity
<dt>bit 6</dt> <dd> always 0
<dt>bit 5</dt> <dd> always 1
<dt>bit 4</dt> <dd> always 1
<dt>bit 3</dt> <dd> = 1 if low battery is detected
<dt>bit 2</dt> <dd> = 1 if last resync failed (though officially
undefined for the MSF clock)
<dt>bit 1</dt> <dd> = 1 if at least one reception attempt since
0230 for the MSF clock was successful
(0300 for the DCF77 clock)
<dt>bit 0</dt> <dd> = 1 if the clock has valid time---reset to zero
when clock is reset (eg at power-up), and
set to 1 after first successful resync attempt.
</dl>
</ol>
The driver only accepts time from the clock if the bottom three bits of
the status byte are <code>011</code>. The leap-year logic for
computing day-in-year is only valid until 2099, and the clock will
ignore stamps from the clock that claim BST is in effect in the first
hour of each year. If the UK parliament decides to move us to
+0100/+0200 time as opposed to the current +0000/+0100 time, it is not
clear what effect that will have on the time broadcast by MSF, and
therefore on this driver's usefulness.
</dl>
A typical <code>ntp.conf</code> configuration file for this driver
might be:
<pre>
# hostname(n) means we expect (n) to be the stratum at which hostname runs.
#------------------------------------------------------------------------------
# SYNCHRONISATION PARTNERS
# ========================
# Our betters...
server 127.127.27.0 # ARCRON MSF radio clock(1).
# Fudge stratum and other features as required.
# ADJUST time1 VALUE FOR YOUR HOST, CLOCK AND LOCATION!
fudge 127.127.27.0 stratum 1 time1 0.016 flag3 1 flag4 1
peer 11.22.33.9 # tick(1--2).
peer 11.22.33.4 # tock(3), boot/NFS server.
# This shouldn't get swept away unless left untouched for a long time.
driftfile /var/tmp/ntp.drift
#------------------------------------------------------------------------------
# RESTRICTIONS
# ============
# By default, don't trust and don't allow modifications. Ignore in fact.
restrict default ignore notrust nomodify
# Allow others in our subnet to check us out...
restrict 11.22.33.0 mask 255.255.255.0 nomodify notrust
# Trust our peers for time. Don't trust others in case they are insane.
restrict 127.127.27.0 nomodify
restrict 11.22.33.4 nomodify
restrict 11.22.33.9 nomodify
# Allow anything from the local host.
restrict 127.0.0.1
</pre>
There are a few <code>#define</code>s in the code that you might wish
to play with:
<dl>
<dt><code>ARCRON_KEEN</code></dt>
<dd>With this defined, the code is relatively trusting of the clock,
and assumes that you will have the clock as one of a few time sources,
so will bend over backwards to use the time from the clock when
available and avoid <code>xntpd</code> dropping sync from the clock
where possible. You may wish to undefine this, especially if you have
better sources of time or your reception is ropey. However, there are
many checks built in even with this flag defined.
<dt><code>ARCRON_OWN_FILTER</code></dt>
<dd>When defined, the code uses its own median-filter code rather than
that available in <code>ntp_refclock.c</code> since the latter seems to
have a minor bug, at least in version 3-5.90. If this bug goes away
this flag should be turned off to avoid duplication of code. (The bug,
if that's what it is, causes the last raw offset to be used rather than
the median offset.)
<p>Without this defined (and without
<code>ARCRON_MULTIPLE_SAMPLES</code> below) a typical set of offsets
reported and used to drive the clock-filter algorithm is (oldest
last):
<pre>
filtoffset= -4.32 -34.82 -0.78 0.89 2.76 4.58 -3.92 -2.17
</pre>
Look at that spike!
<p>With this defined a typical set of offsets is:
<pre>
filtoffset= -7.06 -7.06 -2.91 -2.91 -2.91 -1.27 -9.54 -6.70
</pre>
with the repeated values being some evidence of outlyers being discarded.
<dt><code>ARCRON_MULTIPLE_SAMPLES</code></dt>
<dd>When is defined, we regard each character in the returned timecode
as at a known delay from the start of the second, and use the smallest
(most negative) offset implied by any such character, ie with the
smallest kernel-induced display, and use that. This helps to reduce
jitter and spikes.
<dt><code>ARCRON_LEAPSECOND_KEEN</code></dt>
<dd>When is defined, we try to do a resync to MSF as soon as possible
in the first hour of the morning of the first day of the first and
seventh months, ie just after a leap-second insertion or deletion would
happen if it is going to. This should help compensate for the fact
that this clock does not continuously sample MSF, which compounds the
fact that MSF itself gives no warning of an impending leap-second
event. This code did not seem functional at the leap-second insertion
of 30th June 1997 so is by default disabled.
<dt><code>PRECISION</code></dt>
<dd>Currently set to <code>-4</code>, but you may wish to set it to
<code>-5</code> if you are more conservative, or to <code>-6</code> if
you have particularly good experience with the clock and you live on
the edge. Note that the <code>flag3</code> fudge value will improve
the reported dispersion one notch if clock signal quality is known
good. So maybe just leave this alone. B^)
<dt><code>NSAMPLES</code></dt>
<dd>Should be at least 3 to help smooth out sampling jitters. Can be
more, but if made too long can make <code>xntpd</code> overshoot on
clock corrections and can hold onto bad samples longer than you would
like. With this set to 4 and <code>NKEEP</code> set to 3 this allows
the occasional bad sample (in my experience less than 1 value in 10) to
be dropped. (Note that there seems to be some sort of `beat' effect in
the offset with a periodicity of about 7 samples as of this writing
(1997/05/11) still under investigation; a filter of approximately this
length should be able to almost completely suppress this effect.) Note
that if the fudge-factor <code>flag3</code> is set to 1, a larger
<code>NSAMPLES</code> is used.
</dl>
<p><h4>Monitor Data</h4>
<p>Each timecode is written to the <code>clockstats</code> file with a
signal quality value appended (`0'--`5' as reported by the clock, or
`6' for unknown).
<p>Each resync and result (plus gaining or losing MSF sync) is logged
to the system log at level <code>LOG_NOTICE</code>; note that each
resync drains the unit's batteries, so the syslog entry seems justified.
<p>Syslog entries are of the form:
<pre>
May 10 10:15:24 oolong xntpd[615]: ARCRON: unit 0: sending resync command
May 10 10:17:32 oolong xntpd[615]: ARCRON: sync finished, signal quality 5: OK, will use clock
May 10 11:13:01 oolong xntpd[615]: ARCRON: unit 0: sending resync command
May 10 11:14:06 oolong xntpd[615]: ARCRON: sync finished, signal quality -1: UNKNOWN, will use clock anyway
May 10 11:41:49 oolong xntpd[615]: ARCRON: unit 0: sending resync command
May 10 11:43:57 oolong xntpd[615]: ARCRON: sync finished, signal quality 5: OK, will use clock
May 10 12:39:26 oolong xntpd[615]: ARCRON: unit 0: sending resync command
May 10 12:41:34 oolong xntpd[615]: ARCRON: sync finished, signal quality 3: OK, will use clock
</pre>
<p><h4>Fudge Factors</h4>
<dl>
<dt><code>time1 <i>time</i></code>
<dd>Specifies the time offset calibration factor, in seconds and
fraction, with default 0.0. On a Sun SparcStation 1 running SunOS
4.1.3_U1, with the receiver in London, a value of 0.020 (20ms)
seems to be appropriate.
<p><dt><code>time2 <i>time</i></code>
<dd>Not currently used by this driver.
<p><dt><code>stratum <i>number</i></code>
<dd>Specifies the driver stratum, in decimal from 0 to 15, with default
0. It is suggested that the clock be fudged to stratum 1 so this
it is used a backup time source rather than a primary when more
accurate sources are available.
<p><dt><code>refid <i>string</i></code>
<dd>Specifies the driver reference identifier, an ASCII string from one
to four characters, with default <code>MSFa</code>.
<p><dt><code>flag1 0 | 1</code>
<dd>Not currently used by this driver.
<p><dt><code>flag2 0 | 1</code>
<dd>Not currently used by this driver.
<p><dt><code>flag3 0 | 1</code>
<dd>If set to 1, better precision is reported (and thus lower dispersion)
while clock's received signal quality is known to be good.
<p><dt><code>flag4 0 | 1</code>
<dd>If set to 1, a longer-than-normal (8-stage rather than 4-stage)
median filter is used, to provide some extra smoothing of clock
output and reduction in jitter, at the cost of extra clock
overshoot. Probably not advisable unless the server using this
clock has other sources it can use to help mitigate the overshoot.
</dl>
<p><h4>Additional Information</h4>
<p><a href="refclock.html"> Reference Clock Drivers</a>
<p><a href="http://www2.exnet.com/NTP/ARC/ARC.html">ARC Rugby MSF Receiver</a> page
<hr><address>Damon Hart-Davis (d@hd.org)</address></body></html>