273 lines
13 KiB
Plaintext
273 lines
13 KiB
Plaintext
Leap Second Smearing with NTP
|
|
-----------------------------
|
|
|
|
By Martin Burnicki
|
|
with some edits by Harlan Stenn
|
|
|
|
The NTP software protocol and its reference implementation, ntpd, were
|
|
originally designed to distribute UTC time over a network as accurately as
|
|
possible.
|
|
|
|
Unfortunately, leap seconds are scheduled to be inserted into or deleted
|
|
from the UTC time scale in irregular intervals to keep the UTC time scale
|
|
synchronized with the Earth rotation. Deletions haven't happened, yet, but
|
|
insertions have happened over 30 times.
|
|
|
|
The problem is that POSIX requires 86400 seconds in a day, and there is no
|
|
prescribed way to handle leap seconds in POSIX.
|
|
|
|
Whenever a leap second is to be handled ntpd either:
|
|
|
|
- passes the leap second announcement down to the OS kernel (if the OS
|
|
supports this) and the kernel handles the leap second automatically, or
|
|
|
|
- applies the leap second correction itself.
|
|
|
|
NTP servers also pass a leap second warning flag down to their clients via
|
|
the normal NTP packet exchange, so clients also become aware of an
|
|
approaching leap second, and can handle the leap second appropriately.
|
|
|
|
|
|
The Problem on Unix-like Systems
|
|
--------------------------------
|
|
If a leap second is to be inserted then in most Unix-like systems the OS
|
|
kernel just steps the time back by 1 second at the beginning of the leap
|
|
second, so the last second of the UTC day is repeated and thus duplicate
|
|
timestamps can occur.
|
|
|
|
Unfortunately there are lots of applications which get confused it the
|
|
system time is stepped back, e.g. due to a leap second insertion. Thus,
|
|
many users have been looking for ways to avoid this, and tried to introduce
|
|
workarounds which may work properly, or not.
|
|
|
|
So even though these Unix kernels normally can handle leap seconds, the way
|
|
they do this is not optimal for applications.
|
|
|
|
One good way to handle the leap second is to use ntp_gettime() instead of
|
|
the usual calls, because ntp_gettime() includes a "clock state" variable
|
|
that will actually tell you if the time you are receiving is OK or not, and
|
|
if it is OK, if the current second is an in-progress leap second. But even
|
|
though this mechanism has been available for about 20 years' time, almost
|
|
nobody uses it.
|
|
|
|
|
|
NTP Client for Windows Contains a Workaround
|
|
--------------------------------------------
|
|
The Windows system time knows nothing about leap seconds, so for many years
|
|
the Windows port of ntpd provides a workaround where the system time is
|
|
slewed by the client to compensate the leap second.
|
|
|
|
Thus it is not required to use a smearing NTP server for Windows clients,
|
|
but of course the smearing server approach also works.
|
|
|
|
|
|
The Leap Smear Approach
|
|
-----------------------
|
|
Due to the reasons mentioned above some support for leap smearing has
|
|
recently been implemented in ntpd. This means that to insert a leap second
|
|
an NTP server adds a certain increasing "smear" offset to the real UTC time
|
|
sent to its clients, so that after some predefined interval the leap second
|
|
offset is compensated. The smear interval should be long enough,
|
|
e.g. several hours, so that NTP clients can easily follow the clock drift
|
|
caused by the smeared time.
|
|
|
|
During the period while the leap smear is being performed, ntpd will include
|
|
a specially-formatted 'refid' in time packets that contain "smeared" time.
|
|
This refid is of the form 254.x.y.z, where x.y.z are 24 encoded bits of the
|
|
smear value.
|
|
|
|
With this approach the time an NTP server sends to its clients still matches
|
|
UTC before the leap second, up to the beginning of the smear interval, and
|
|
again corresponds to UTC after the insertion of the leap second has
|
|
finished, at the end of the smear interval. By examining the first byte of
|
|
the refid, one can also determine if the server is offering smeared time or
|
|
not.
|
|
|
|
Of course, clients which receive the "smeared" time from an NTP server don't
|
|
have to (and even must not) care about the leap second anymore. Smearing is
|
|
just transparent to the clients, and the clients don't even notice there's a
|
|
leap second.
|
|
|
|
|
|
Pros and Cons of the Smearing Approach
|
|
--------------------------------------
|
|
The disadvantages of this approach are:
|
|
|
|
- During the smear interval the time provided by smearing NTP servers
|
|
differs significantly from UTC, and thus from the time provided by normal,
|
|
non-smearing NTP servers. The difference can be up to 1 second, depending
|
|
on the smear algorithm.
|
|
|
|
- Since smeared time differs from true UTC, and many applications require
|
|
correct legal time (UTC), there may be legal consequences to using smeared
|
|
time. Make sure you check to see if this requirement affects you.
|
|
|
|
However, for applications where it's only important that all computers have
|
|
the same time and a temporary offset of up to 1 s to UTC is acceptable, a
|
|
better approach may be to slew the time in a well defined way, over a
|
|
certain interval, which is what we call smearing the leap second.
|
|
|
|
|
|
The Motivation to Implement Leap Smearing
|
|
-----------------------------------------
|
|
Here is some historical background for ntpd, related to smearing/slewing
|
|
time.
|
|
|
|
Up to ntpd 4.2.4, if kernel support for leap seconds was either not
|
|
available or was not enabled, ntpd didn't care about the leap second at all.
|
|
So if ntpd was run with -x and thus kernel support wasn't used, ntpd saw a
|
|
sudden 1 s offset after the leap second and normally would have stepped the
|
|
time by -1 s a few minutes later. However, 'ntpd -x' does not step the time
|
|
but "slews" the 1-second correction, which takes 33 minutes and 20 seconds
|
|
to complete. This could be considered a bug, but certainly this was only an
|
|
accidental behavior.
|
|
|
|
However, as we learned in the discussion in http://bugs.ntp.org/2745, this
|
|
behavior was very much appreciated since indeed the time was never stepped
|
|
back, and even though the start of the slewing was somewhat undefined and
|
|
depended on the poll interval. The system time was off by 1 second for
|
|
several minutes before slewing even started.
|
|
|
|
In ntpd 4.2.6 some code was added which let ntpd step the time at UTC
|
|
midnight to insert a leap second, if kernel support was not used.
|
|
Unfortunately this also happened if ntpd was started with -x, so the folks
|
|
who expected that the time was never stepped when ntpd was run with -x found
|
|
this wasn't true anymore, and again from the discussion in NTP bug 2745 we
|
|
learn that there were even some folks who patched ntpd to get the 4.2.4
|
|
behavior back.
|
|
|
|
In 4.2.8 the leap second code was rewritten and some enhancements were
|
|
introduced, but the resulting code still showed the behavior of 4.2.6,
|
|
i.e. ntpd with -x would still step the time. This has only recently been
|
|
fixed in the current ntpd stable code, but this fix is only available with a
|
|
certain patch level of ntpd 4.2.8.
|
|
|
|
So a possible solution for users who were looking for a way to come over the
|
|
leap second without the time being stepped could have been to check the
|
|
version of ntpd installed on each of their systems. If it's still 4.2.4 be
|
|
sure to start the client ntpd with -x. If it's 4.2.6 or 4.2.8 it won't work
|
|
anyway except if you had a patched ntpd version instead of the original
|
|
version. So you'd need to upgrade to the current -stable code to be able to
|
|
run ntpd with -x and get the desired result, so you'd still have the
|
|
requirement to check/update/configure every single machine in your network
|
|
that runs ntpd.
|
|
|
|
Google's leap smear approach is a very efficient solution for this, for
|
|
sites that do not require correct timestamps for legal purposes. You just
|
|
have to take care that your NTP servers support leap smearing and configure
|
|
those few servers accordingly. If the smear interval is long enough so that
|
|
NTP clients can follow the smeared time it doesn't matter at all which
|
|
version of ntpd is installed on a client machine, it just works, and it even
|
|
works around kernel bugs due to the leap second.
|
|
|
|
Since all clients follow the same smeared time the time difference between
|
|
the clients during the smear interval is as small as possible, compared to
|
|
the -x approach. The current leap second code in ntpd determines the point
|
|
in system time when the leap second is to be inserted, and given a
|
|
particular smear interval it's easy to determine the start point of the
|
|
smearing, and the smearing is finished when the leap second ends, i.e. the
|
|
next UTC day begins.
|
|
|
|
The maximum error doesn't exceed what you'd get with the old smearing caused
|
|
by -x in ntpd 4.2.4, so if users could accept the old behavior they would
|
|
even accept the smearing at the server side.
|
|
|
|
In order to affect the local timekeeping as little as possible the leap
|
|
smear support currently implemented in ntpd does not affect the internal
|
|
system time at all. Only the timestamps and refid in outgoing reply packets
|
|
*to clients* are modified by the smear offset, so this makes sure the basic
|
|
functionality of ntpd is not accidentally broken. Also peer packets
|
|
exchanged with other NTP servers are based on the real UTC system time and
|
|
the normal refid, as usual.
|
|
|
|
The leap smear implementation is optionally available in ntp-4.2.8p3 and
|
|
later, and the changes can be tracked via http://bugs.ntp.org/2855.
|
|
|
|
|
|
Using NTP's Leap Second Smearing
|
|
--------------------------------
|
|
- Leap Second Smearing MUST NOT be used for public servers, e.g. servers
|
|
provided by metrology institutes, or servers participating in the NTP pool
|
|
project. There would be a high risk that NTP clients get the time from a
|
|
mixture of smearing and non-smearing NTP servers which could result in
|
|
undefined client behavior. Instead, leap second smearing should only be
|
|
configured on time servers providing dedicated clients with time, if all
|
|
those clients can accept smeared time.
|
|
|
|
- Leap Second Smearing is NOT configured by default. The only way to get
|
|
this behavior is to invoke the ./configure script from the NTP source code
|
|
package with the --enable-leap-smear parameter before the executables are
|
|
built.
|
|
|
|
- Even if ntpd has been compiled to enable leap smearing support, leap
|
|
smearing is only done if explicitly configured.
|
|
|
|
- The leap smear interval should be at least several hours' long, and up to
|
|
1 day (86400s). If the interval is too short then the applied smear offset
|
|
is applied too quickly for clients to follow. 86400s (1 day) is a good
|
|
choice.
|
|
|
|
- If several NTP servers are set up for leap smearing then the *same* smear
|
|
interval should be configured on each server.
|
|
|
|
- Smearing NTP servers DO NOT send a leap second warning flag to client time
|
|
requests. Since the leap second is applied gradually the clients don't even
|
|
notice there's a leap second being inserted, and thus there will be no log
|
|
message or similar related to the leap second be visible on the clients.
|
|
|
|
- Since clients don't (and must not) become aware of the leap second at all,
|
|
clients getting the time from a smearing NTP server MUST NOT be configured
|
|
to use a leap second file. If they had a leap second file they would apply
|
|
the leap second twice: the smeared one from the server, plus another one
|
|
inserted by themselves due to the leap second file. As a result, the
|
|
additional correction would soon be detected and corrected/adjusted.
|
|
|
|
- Clients MUST NOT be configured to poll both smearing and non-smearing NTP
|
|
servers at the same time. During the smear interval they would get
|
|
different times from different servers and wouldn't know which server(s) to
|
|
accept.
|
|
|
|
|
|
Setting Up A Smearing NTP Server
|
|
--------------------------------
|
|
If an NTP server should perform leap smearing then the leap smear interval
|
|
(in seconds) needs to be specified in the NTP configuration file ntp.conf,
|
|
e.g.:
|
|
|
|
leapsmearinterval 86400
|
|
|
|
Please keep in mind the leap smear interval should be between several and 24
|
|
hours' long. With shorter values clients may not be able to follow the
|
|
drift caused by the smeared time, and with longer values the discrepancy
|
|
between system time and UTC will cause more problems when reconciling
|
|
timestamp differences.
|
|
|
|
When ntpd starts and a smear interval has been specified then a log message
|
|
is generated, e.g.:
|
|
|
|
ntpd[31120]: config: leap smear interval 86400 s
|
|
|
|
While ntpd is running with a leap smear interval specified the command:
|
|
|
|
ntpq -c rv
|
|
|
|
reports the smear status, e.g.:
|
|
|
|
# ntpq -c rv
|
|
associd=0 status=4419 leap_add_sec, sync_uhf_radio, 1 event, leap_armed,
|
|
version="ntpd 4.2.8p3-RC1@1.3349-o Mon Jun 22 14:24:09 UTC 2015 (26)",
|
|
processor="i586", system="Linux/3.7.1", leap=01, stratum=1,
|
|
precision=-18, rootdelay=0.000, rootdisp=1.075, refid=MRS,
|
|
reftime=d93dab96.09666671 Tue, Jun 30 2015 23:58:14.036,
|
|
clock=d93dab9b.3386a8d5 Tue, Jun 30 2015 23:58:19.201, peer=2335,
|
|
tc=3, mintc=3, offset=-0.097015, frequency=44.627, sys_jitter=0.003815,
|
|
clk_jitter=0.451, clk_wander=0.035, tai=35, leapsec=201507010000,
|
|
expire=201512280000, leapsmearinterval=86400, leapsmearoffset=-932.087
|
|
|
|
In the example above 'leapsmearinterval' reports the configured leap smear
|
|
interval all the time, while the 'leapsmearoffset' value is 0 outside the
|
|
interval and increases from 0 to -1000 ms over the interval. So this can be
|
|
used to monitor if and how the time sent to clients is smeared. With a
|
|
leapsmearoffset of -.932087, the refid reported in smeared packets would be
|
|
254.196.88.176.
|