How NTP Operates

The ntpd utility operates by exchanging messages with one or more configured servers over a range of designated poll intervals. When started, whether for the first or subsequent times, the program requires several exchanges from the majority of these servers so the signal processing and mitigation algorithms can accumulate and groom the data and set the clock. In order to protect the network from bursts, the initial poll interval for each server is delayed an interval randomized over a few seconds. At the default initial poll interval of 64s, several minutes can elapse before the clock is set. This initial delay to set the clock can be safely and dramatically reduced using the iburst keyword with the server configuration command, as described in ntp.conf(5).

Most operating systems and hardware of today incorporate a time-of-year (TOY) chip to maintain the time during periods when the power is off. When the machine is booted, the chip is used to initialize the operating system time. After the machine has synchronized to an NTP server, the operating system corrects the chip from time to time. In the default case, if ntpd detects that the time on the host is more than 1000s from the server time, ntpd assumes something must be terribly wrong, and the only reliable action is for the operator to intervene and set the clock by hand. (Reasons for this include there is no TOY chip, or its battery is dead, or that the TOY chip is just of poor quality.) This causes ntpd to exit with a panic message to the system log. The -g option overrides this check, and the clock will be set to the server time regardless of the chip time (up to 68 years in the past or future — this is a limitation of the NTPv4 protocol). However, and to protect against broken hardware, such as when the CMOS battery fails or the clock counter becomes defective, once the clock has been set an error greater than 1000s will cause ntpd to exit anyway.

Under ordinary conditions, ntpd adjusts the clock in small steps so that the timescale is effectively continuous and without discontinuities. Under conditions of extreme network congestion, the roundtrip delay jitter can exceed three seconds and the synchronization distance, which is equal to one-half the roundtrip delay plus error budget terms, can become very large. The ntpd algorithms discard sample offsets exceeding 128 ms, unless the interval during which no sample offset is less than 128 ms exceeds 900s. The first sample after that, no matter what the offset, steps the clock to the indicated time. In practice, this reduces the false alarm rate where the clock is stepped in error to a vanishingly low incidence.

As the result of this behavior, once the clock has been set it very rarely strays more than 128 ms even under extreme cases of network path congestion and jitter. Sometimes, in particular, when ntpd is first started without a valid drift file on a system with a large intrinsic drift the error might grow to exceed 128 ms, which would cause the clock to be set backwards if the local clock time is more than 128 ms in the future relative to the server. In some applications, this behavior may be unacceptable. There are several solutions, however. If the -x option is included on the command line, the clock will never be stepped and only slew corrections will be used. But this choice comes at a cost that should be carefully explored before deciding to use the -x option. The maximum slew rate possible is limited to 500 parts-per-million (PPM) as a consequence of the correctness principles on which the NTP protocol and algorithm design are based. As a result, the local clock can take a long time to converge to an acceptable offset, about 2,000 s for each second the clock is outside the acceptable range. During this interval, the local clock will not be consistent with any other network clock and the system cannot be used for distributed applications that require correctly synchronized network time.

In spite of the above precautions, sometimes when large frequency errors are present the resulting time offsets stray outside the 128-ms range and an eventual step or slew time correction is required. If following such a correction the frequency error is so large that the first sample is outside the acceptable range, ntpd enters the same state as when the ntp.drift file is not present. The intent of this behavior is to quickly correct the frequency and restore operation to the normal tracking mode. In the most extreme cases, there may be occasional step/slew corrections and subsequent frequency corrections. It helps in these cases to use the burst keyword when configuring the server, but ONLY when you have permission to do so from the owner of the target host.

Finally, in the past, many startup scripts would run a separate utility to get the system clock close to correct before starting ntpd(8), but this was never more than a mediocre hack and is no longer needed. If you are following the best current practice <#starting> and you still need to set the system time before starting ntpd, please open a bug report and document what is going on, and then look at using ntpdig(1).

There is a way to start ntpd(8) that often addresses all of the problems mentioned above.

Starting NTP (Best Current Practice)

First, use the iburst option on your server and pool entries.

If you can also keep a good ntp.drift file then ntpd(8) will effectively "warm-start" and your system’s clock will be stable in under 11 seconds' time.

As soon as possible in the startup sequence, start ntpd(8) with at least the -g and perhaps the -N options. Then, start the rest of your "normal" processes. This will give ntpd(8) as much time as possible to get the system’s clock synchronized and stable.

Finally, if you have processes like dovecot or database servers that require monotonically-increasing time, run ntpwait(8) as late as possible in the boot sequence (perhaps with the -v flag) and after ntpwait(8) exits successfully it is as safe as it will ever be to start any processes that require stable time.

Frequency Discipline

The ntpd behavior at startup depends on whether the frequency file, usually ntp.drift, exists. This file contains the latest estimate of clock frequency error. When the ntpd is started and the file does not exist, the ntpd enters a special mode designed to quickly adapt to the particular system clock oscillator time and frequency error. This takes approximately 15 minutes, after which the time and frequency are set to nominal values and the ntpd enters normal mode, where the time and frequency are continuously tracked relative to the server. After one hour the frequency file is created and the current frequency offset written to it. When the ntpd is started and the file does exist, the ntpd frequency is initialized from the file and enters normal mode immediately. After that, the current frequency offset is written to the file at hourly intervals.

Operating Modes

ntpd normally operates continuously while monitoring for small changes in frequency and trimming the clock for the ultimate precision. However, it can operate in a one-time mode where the time is set from an external server and frequency is set from a previously recorded frequency file.

By default, ntpd runs in continuous mode where each of possibly several external servers is polled at intervals determined by an intricate state machine. The state machine measures the incidental roundtrip delay jitter and oscillator frequency wander and determines the best poll interval using a heuristic algorithm. Ordinarily, and in most operating environments, the state machine will start with 64s intervals and eventually increase in steps to 1024s. A small amount of random variation is introduced in order to avoid bunching at the servers. In addition, should a server become unreachable for some time, the poll interval is increased in steps to 1024s in order to reduce network overhead.

In some cases, it may not be practical for ntpd to run continuously. The -q option is provided to support running ntpd periodically from a cron(8) job. Setting this option will cause ntpd to exit just after setting the clock for the first time. The procedure for initially setting the clock is the same as in continuous mode; most applications will probably want to specify the iburst keyword with the server configuration command. With this keyword, a volley of messages are exchanged to groom the data and the clock is set in about 10 sec. If nothing is heard after a couple of minutes, the daemon times out and exits.

When kernel support is available to discipline the clock frequency, which is the case for stock Solaris, Linux, and FreeBSD, a useful feature is available to discipline the clock frequency. First, ntpd is run in continuous mode with selected servers in order to measure and record the intrinsic clock frequency offset in the frequency file. It may take some hours for the frequency and offset to settle down. Then the ntpd is stopped and run in one-time mode as required. At each startup, the frequency is read from the file and initializes the kernel frequency.

Poll Interval Control

This version of NTP includes an intricate state machine to reduce the network load while maintaining a quality of synchronization consistent with the observed jitter and wander. There are a number of ways to tailor the operation in order enhance accuracy by reducing the interval or to reduce network overhead by increasing it. However, the user is advised to carefully consider the consequences of changing the poll adjustment range from the default minimum of 64 s to the default maximum of 1,024 s. The default minimum can be changed with the tinker minpoll command to a value not less than 16 s. This value is used for all configured associations, unless overridden by the minpoll option on the configuration command. Note that most device drivers will not operate properly if the poll interval is less than 64 s and that the broadcast server and manycast client associations will also use the default unless overridden.

In some cases involving dial up or toll services, it may be useful to increase the minimum interval to a few tens of minutes and maximum interval to a day or so. Under normal operation conditions, once the clock discipline loop has stabilized the interval will be increased in steps from the minimum to the maximum. However, this assumes the intrinsic clock frequency error is small enough for the discipline loop correct it. The capture range of the loop is 500 PPM at an interval of 64s decreasing by a factor of two for each doubling of the interval. At a minimum of 1,024 s, for example, the capture range is only 31 PPM. If the intrinsic error is greater than this, the drift file ntp.drift will have to be specially tailored to reduce the residual error below this limit. Once this is done, the drift file is automatically updated once per hour and is available to initialize the frequency on subsequent daemon restarts.

The huff-n'-puff Filter

In scenarios where a considerable amount of data are to be downloaded or uploaded over telephone modems, timekeeping quality can be seriously degraded. This occurs because the differential delays on the two directions of transmission can be quite large. In many cases, the apparent time errors are so large as to exceed the step threshold and a step correction can occur during and after the data transfer is in progress.

The huff-n'-puff filter is designed to correct the apparent time offset in these cases. It depends on knowledge of the propagation delay when no other traffic is present. In common scenarios, this occurs during other than work hours. The filter maintains a shift register that remembers the minimum delay over the most recent interval measured usually in hours. Under conditions of severe delay, the filter corrects the apparent offset using the sign of the offset and the difference between the apparent delay and minimum delay. The name of the filter reflects the negative (huff) and positive (puff) correction, which depends on the sign of the offset.

The filter is activated by the tinker command and huffpuff keyword, as described in ntp.conf(5).