NTP Version 4 Release Notes

Alice's Adventures in Wonderland, by Lewis Carroll, illustrations by Sir John Tenniel

NTP Version 4 Release Notes

This release of the NTP Version 4 (NTPv4) daemon for Unix incorporates new features and refinements to the NTP Version 3 (NTPv3) algorithms. However, it continues the tradition of retaining backwards compatibility with older versions. The NTPv4 version has been under development for quite a while and isn't finished yet. In fact, quite a number of NTPv4 features have already been implemented in the current NTPv3. The primary purpose of this release is to verify the remaining new code compiles and runs in the various architectures, operating systems and hardware complement that can't be verified here. Of particular interest are Windows NT, VMS and various reference clock drivers. As always, corrections and bugfixes are warmly received, especially in the form of context diffs.

This note summarizes the differences between this software release of NTPv4, called ntp-4.x.x, and the previous NTPv3 version, called xntp3-5.x.x

  1. Most of the extensive calculations are now done using 64-bit floating-point format, rather than 64-bit fixed-point format. The motivation for this is to reduce size, improve speed and avoid messy bounds checking. Workstations of today are much faster than when the original NTP version was designed in the early 1980s, and it is rare to find a processor architecture that does not support it. The fixed-point format is still used with raw timestamps, in order to retain the full precision of about 212 picoseconds. However, the algorithms which process raw timestamps all produce fixed-point differences before converting to double. The differences are ordinarily quite small so can be expressed without loss of accuracy in double format.
  2. The clock discipline algorithm has been redesigned to improve accuracy, reduce the impact of network jitter and allow an increase in poll intervals to well over one day with only moderate sacrifice in accuracy. The NTPv4 design allows servers to increase the poll intervals even when synchronized directly to the peer. In NTPv3 the poll interval in such cases was clamped to the minimum, usually 64 s. For those servers with hundreds of clients, the new design can dramatically reduce the network load.
  3. A burst-mode feature is available which results in good accuracy with intermittent connections typical of PPP and ISDN services. When enabled, at each poll interval the server sends eight messages over the next 30-s interval and processes them in a batch. Outlyers due to initial dial-up delays, etc., are avoided and the server synchronizes with its peer generally within 30 s.
  4. In addition to the NTPv3 authentication scheme, which uses private-key cryptography, a new NTPv4 autokey authentication scheme is available. Autokey uses public-key technology and avoids the need to distribute keys by separate means. The design is such that full accuracy is available without degradation due to processing demands of the public-key routines. It can be used in any of the NTP association modes, but is most useful in broadcast/multicast modes.
  5. NTPv4 includes two new association modes which in most applications can avoid per-host configuration altogether. Both of these are based on multicast technology. They provide for automatic discovery and configuration of servers and clients. In multicast mode, a server sends a message at fixed intervals using specified multicast addresses, while clients listen on these addresses. Upon receiving the message, a client exchanges several messages with the server in order to calibrate the multicast propagation delay between the client and server. In manycast mode, a client sends a message and expects one or more servers to reply. Using engineered algorithms, the client selects an appropriate subset of servers from the messages received and continues in ordinary client/server operation with them. The manycast scheme can provide somewhat better accuracy than the multicast scheme at the price of additional network overhead.
  6. The reference clock driver interface is smaller, more rational and moreaccurate. Support for pulse-per-second (PPS) signals has been extended to all drivers as an intrinsic function. Most of the drivers in NTPv3 have been converted to this interface, but some, including the PARSE subinterface, have yet to be overhauled. New drivers have been added for several GPS receivers now on the market. Drivers for the Canadian standard time and frequency station CHU and for audio IRIG signals have been updated and capabilites added to allow direct connection of these signals to the Sun audio port /dev/audio.
  7. In all except a very few cases, all timing intervals are randomized, so that the tendency for NTPv3 to bunch messages, especially with a large number of configured associations, is minimized.
  8. In NTPv3 a large number of weeds and useless code had grown over the years since the original NTPv1 code was implemented almost ten years ago. Using a powerful weedwacker, much of the shrubbery has been removed, with effect a substantial reduction in size of almost 40 percent.
  9. The entire distribution has been converted to gnu automake, which should greatly ease the task of porting to new and different programming environments, as well as reduce the incidence of bugs due to improper handling of idiosyncratic kernel functions.

Nasty Surprises

There are a few things different about this release that have changed since the latest NTP Version 3 release. Following are a few things to worry about:

  1. As required by Defense Trade Regulations (DTR), the cryptographic routines supporting the Data Encryption Standard (DES) and Message Digest 5 (MD5) have been removed from the export version of the distribtution. These routines are readily available in most countries from RSA Laboratories. Directions for their use are in the Building and Installing the Distribution page.
  2. As the result of the above, the ./authstuff directory, intended as a development and testing aid for porting cryptographic routines to exotic architectures, has been removed. Developers should note the NTP authentication routines use the interface defined in the rsaref2.0 package available from RSA laboratories.
  3. The enable and disable commands have a few changes in their arguments see the ntpd Configuration Options page for details.
  4. The scheme for enabling the ppsclock line discipline/streams module has changed. Formerly, it was enabled by setting fudge flag3 for the serial port connected to the PPS signal. Now, there is an explicit command pps used to designate the device name. See the Reference Clock Options page for details.
  5. While in fact not a new problem, some obscure option combinations require the server and peer commands to follow the others; in particular, the enable and pps commands should preceed these commands.

Caveats

This release has been compiled and tested on several systems, including SunOS 4.1.3, Solaris 2.5.1 and 2.6, Alpha 4.0, Ultrix 4.4, Linux, FreeBSD and HP-UX 10.02. It has not been compiled for Windows NT or VMS. We are relying on the NTP volunteer brigade to do that. Known problems are summarized below:

  1. To work properly in all cases, the enable and pps commands, if used, should appear before the server and fudge commands in the configuration file.
  2. The precision time kernel modifications now in stock Solaris 2.6 have bugs. The kernel discipline has been disabled by default. For testing, the kernel can be enabled using the enable kernel command either in the configuration file or via ntpdc.
  3. On Alpha 4.0 with reference clocks configured, debugging with the -d options doesn't work.
  4. The autokey function requires an NTP header extensions field, which is documented in an internet draft and implemented in this release. This field holds the public-key signature and certificate; however, the detailed format for these data have not yet been determined. It is expected this will happen in the near future and that implementation of the required algorithms will quickly follow using available cryptographic algorithms.
  5. The manycast function still needs some work. Ideally, the existing I/O routines would be enhanced to include the capability to determine the source address on every multicast packet sent, so that the autokey function could reliably construct the correct cryptosum. Meanwhile, the utility of manycast in conjunction with autokey is limited to clients with only a single network interface.
  6. The HTML documentation has been partially updated. However, most of the NTPv3 documentation continues to apply to NTPv4. Until the update happens, what you see is what you get. We are always happy to accept comments, corrections and bug reports. However, we are most thrilled upon receipt of patches to fix the dang bugs.

David L. Mills <mills@udel.edu>