NTPv5 use cases and requirements

NTP version 4 has seen active use for over a decade, and within this time period the protocol has not only been extended to support new requirements but also fallen victim to vulnerabilities that have made it used for distributed denial of service (DDoS) amplification attacks.

The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “NOT RECOMMENDED”, “MAY”, and “OPTIONAL” in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

There are several common scenarios for exxisting NTPv4 deployments; publicly accessible NTP services such as the NTP Pool are used to offer clock synchronisation for end users and embedded devices, ISP provided servers to synchronise devices such as customer-premesis equipment where reduced accuracy may be tollerable. Depending on the network and path these deployments may be affected by variable latency as well as throttling or blocking by providers. Data centres and cloud computing providers also have deployed and offer NTP services both for internal use and for customers, particularly where the network is unable to offer or does not require PTP . As these deployments are less likely to be constrained by network latency or power the potential for higher levels of accuracy and precision within the bounds of the protocol are possible.

At a high level, NTPv5 should be a protocol that is capable of operating in both local networks and also over public internet connections where packet loss, delay, and even filtering may occur. Timestamp resolution SHOULD either match or exceed NTPv4, and be extensible to represent any specified timescale. The protocol SHOULD NOT transmit time zone information and should focus on providing clock synchronisation as TZDIST already provides this ability.

Servers SHOULD have a new identifier that peers use as reference, this SHOULD NOT be a FQDN, an IP address, or identifier tied to a certificate. Servers SHOULD be able to migrate and change their identifiers as stratum topologies or network configuration changes occur. Clients SHOULD re-establish connections with servers at an interval to prevent attempting to maintain connectivity to dead host and give network operators the ability to move traffic away from IP addresses in a timely manner. This functionality should also compliment having a “Kiss of Death” or similar message from servers.

Algorithms describing functions such as clock filtering, selection and clustering SHOULD be omitted from the specification; the specification should instead only provide only what is necessary to describe protocol semantics and normative behaviours. The working group should consider creating a separate informational document to describe an algorithm to assist with implementation, and to consider adopting future documents which describe new algorithms as they are developed. Specifying client algorithms separately from the protocol allows will allow NTPv5 to meet the needs of applications with a variety of network properties and performance requirements. It also allows for innovation in implementations without sacrificing basic interoperability.

Support SHOULD be available for other timescales in addition to UTC - this should include, but not limited to the use of TAI or Modified Julian Date as defined in , however UTC SHALL be the default timescale and MUST be supported by all implementations. Consideration should be made to include listing the supported timescales either as part of specific IANA parameter registry, or as part of the extension registry.

The specification or the protocol SHOULD be explicit about when a leap second is being applied, and the protocol should allow for transmiting an upcoming leap second ahead of the day it is to be applicable. Nevertheless, due to network delays and the polling interval, applications with NTP clients will need to manage the leap second event at their local clock.

Server responses SHOULD include not only an indicator as to wether the server supports smearing, but also if the current time being transmitted is smeared. The protocol may also transmit the start/end or duration of the smearing ahead of time. It MUST be possible for clients to determine the unsmeared time of the timescale.

The support for compatibility with other protocols SHOULD NOT prevent addressing issues that have previous caused issues in deployments or cause ossification of the protocol. Protocol ossification MUST be addressed to prevent existing NTPv4 deployments which incorrectly respond to clients posing as NTPv5 from causing issues. Forward prevention of ossification (for a potential NTPv6 protocol in the future) SHOULD also be taken into consideration. The model for backward compatibility is servers that support mutliple versions NTP and send a response in the same version as the request. This does not preclude high stratum servers from acting as a client in one version of NTP and a server in another.

To provide the protocol MUST have the capability to be extended. The specification should specify that implementations MUST ignore unknown extensions. Unknown extensions received by a server from a lower stratum server SHALL not be added to response messages sent by the server receiving these extensions.

Considerations should be made about the future of the existing IANA registry for NTPv4 parameters. If NTPv5 becomes incompatible with these parameters a new registry SHOULD be created.

Encryption and authentication MUST be provided by the protocol specification as a default and MUST be resistent to downgrade attacks. The encryption used must have agility, allowing for the protocol to update as more secure cryptography becomes known and vulnerabilities are discovered. The specification MAY consider leaving room for middleboxes which may deliberately modify packets in flight for legitimate purposes. Thought must be given around how this will be incorporated into any applicable trust model. Downgrading attacks that could lead to an adversary disabling or removing encryption or authentication MUST NOT be possible in the design of the protocol. Detection and reporting of server malfeasence SHOULD remain out of scope of this specification as already provides this capability as a core functionality of the protocol. —back

The author would like to thank Doug Arnold for contributions to this document, and would like to acknowledge Daniel Franke, Watson Ladd, Miroslav Lichvar for their existing documents and ideas. The author would also like to thank Angelo Moriondo, Franz Karl Achard, and Malcom McLean for providing the author with motivation.