NTP Over PTP

The Precision Time Protocol (PTP) was designed for highly accurate synchronization of clocks in local networks. It relies on hardware timestamping supported in all network devices involved in the synchronization (e.g. network interface controllers, switches, and routers) to eliminate the impact of software, processing and queueing delays on accuracy of offset and delay measurements. PTP was originally designed for multicast communication. Later was added support for unicast messaging, which is useful in larger networks with partial on-path PTP support (e.g. telecom profiles G.8265.1 and G.8275.2). The Network Time Protocol does not rely on hardware timestamping support, but implementations can use it if it is available to avoid the impact of software, processing and queueing delays, similarly to PTP. When comparing PTP with the timing modes of NTP, PTP is functionally closest to the NTP broadcast mode. An issue for NTP is hardware that can specifically timestamp only PTP packets. This limitation comes from a hardware design which can provide receive timestamps only at a limited rate instead of the maximum rate possible at the network link speed. To avoid missing receive timestamps when the interface is receiving other traffic at a high rate, a filter is implemented in the hardware to inspect each received packet and capture a timestamp only for packets that need it. The hardware filter can be usually configured for specific PTP transport (e.g. UDPv4, UDPv6, 802.3) and sometimes even the PTP message type (e.g. sync message or delay request) to further reduce the timestamping rate on the server or client side in the case of multicast messaging, but it typically cannot be configured to timestamp NTP messages sent to the UDP port 123. Another issue for NTP is missing hardware support in network switches and routers. With PTP the devices operate either as boundary clocks or transparent clocks. Boundary clocks are analogous to NTP clients that work also as servers for other clients. Transparent clocks are much simpler. They only measure the delay in forwarding of PTP packets and write this delay to the correction field of either the packet itself (one-step mode) or a later packet in the PTP exchange (two-step mode). Transparent clocks are specific to the PTP delay mechanism used in the network, either end-to-end (E2E) or peer-to-peer (P2P). This document specifies a new transport for NTP to enable hardware timestamping on NICs which can timestamp only PTP messages and also take advantage of one-step E2E PTP unicast transparent clocks. It adds a new extension field (TLV) for PTP to contain NTP messages and adds a new extension field for NTP to provide clients and peers with the correction of their NTP requests from transparent clocks. The NTP broadcast mode is not supported. NTP over PTP does not require any PTP clocks to be present in the network. It does not disrupt their operation if they are present. If the network uses one-step E2E transparent clocks, NTP clients and peers can reach the same or better accuracy as PTP clocks. Hosts in the network can operate as PTP clocks and NTP servers, clients, or peers using NTP over PTP at the same time. The client-server mode of NTP, even if using the PTP transport, has several advantages to PTP using multicast or unicast messaging: NTP is more secure. It can use existing security mechanisms specified for NTP like Network Time Security, not losing any of its features. Like the NTP broadcast mode, PTP is more difficult to secure against delay attacks. The PTP unicast mode allows an almost-infinite traffic amplification, which can be exploited for denial-of-service attacks and can only be limited by security mechanisms requiring client authentication. NTP is more resilient to failures. Each client can use multiple servers and detect failed sources in its source selection. In PTP a single hardware or software failure can disrupt the whole PTP domain. Multiple independent domains have to be used to handle any failure. NTP is better suited for synchronization in networks which do not have full on-path PTP support, or where timestamping errors do not have a symmetric distribution (e.g. due to sensitivity to network load). NTP does not assume network delay is constant and the rate of measurements in opposite directions is symmetric. It can filter the measurements more effectively and is not sensitive to asymmetrically distributed network delays and timestamping errors. PTP has to measure the offset and delay separately to enable multicast messaging, which is needed to reduce the transmit timestamping rate. With PTP unicast messaging and NTP client-server mode a limited transmit timestamping rate on server limits the number of clients. NTP needs fewer messages to get the same number of timestamps. It uses less network bandwidth than PTP using unicast messaging. NTP provides clients with an estimate of the maximum error of the clock (root distance).

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

A new TLV is defined for PTP to contain NTP messages in the client (3), server (4), and symmetric modes (1 and 2). Using other NTP modes in the TLV is not specified. Any transport specified for PTP that supports unicast messaging can be used for NTP over PTP, e.g. UDP on IPv4 and IPv6. The type value of the NTP TLV is TBD. The TLV contains the whole NTP message as would normally be the UDP payload, without any modifications. The TLV does not propagate through boundary clocks. If the UDP transport is used for PTP, the UDP source and destination port numbers MUST be the PTP event port (319). If the client implemented port randomization, requests and/or responses would not get a hardware receive timestamp due to the filter matching only the PTP port. The NTP TLV MUST be included in a PTP delay request message. The originTimestamp field and all fields of the header SHOULD be zero, except: messageType is 1 (delay request) versionPTP is 2 messageLength is the length of the PTP message including the NTP TLV domainNumber is 123 flagField has the unicastFlag (0x4) bit set sequenceId is increased by one with each transmitted PTP message An NTP client or peer using the PTP transport sends NTP requests contained in PTP delay requests as the NTP TLV. An NTP server or peer receiving NTP requests over the PTP transport MUST check for the domainNumber of 123 and the NTP TLV. Its responses to these requests MUST be contained in PTP delay requests as the NTP TLV. It MUST NOT respond with PTP delay responses, or any other PTP messages. If a PTP clock receives an NTP-over-PTP request, it will not recognize the domain number and ignore the message. If it responded to messages in the domain (e.g. due to misconfiguration), it would send a delay response (to port 320 if using the UDP transport), which would be ignored by the client. Any authenticator fields included in the NTP messages MUST be calculated only over the NTP message following the header of the NTP TLV. Receive and transmit timestamps contained in the NTP messages SHOULD NOT be adjusted for the beginning of the NTP data in the PTP message. They SHOULD still correspond to the ending of the reception and beginning of the transmission of the whole frame (e.g. start frame delimiter in an Ethernet frame).

One-step E2E PTP transparent clocks modify the correction field in the header of the PTP delay requests containing NTP messages. To be able to verify and apply the corrections to an NTP measurement, the client or peer needs to know the correction of both the request and response. The correction of the response is in the PTP header of the message itself. The correction of the request is provided by the server or other peer in a new NTP extension field included in the response. The format of the Network Correction Extension Field is shown in Figure .

The length of padding is the minimum required to make a valid extension field in the protocol version. For NTPv4 that is 16 octets if it is the last extension field in the message per RFC 7822. The Network Correction field in the extension field uses the 64-bit NTP timestamp format (resolution of about 1/4th of a nanosecond). The correction field in PTP header has a different format (64-bit nanoseconds + 16-bit fraction). The value of the NTP network correction is the sum of PTP corrections provided by transparent clocks and the time it takes to receive the packet (i.e. packet length including the frame check sequence divided by the link speed). The reason for not using the PTP correction alone is to avoid an asymmetric correction when the server and client, or peers, are connected to the network with different link speeds. The receive duration included in the NTP correction cancels out the transposition of PTP receive timestamp corresponding to the beginning of the reception to NTP receive timestamp corresponding to the end of the reception. The Figure shows the NTP timestamps, transmit/receive durations, and processing and queuing delays included in PTP corrections for an NTP exchange made over two PTP transparent clocks. The link speed is increasing on the network path from the client to the server. The propagation delays in cables are not shown.

When an NTP server which supports the PTP transport receives an NTP request containing the Network Correction Extension Field, it SHOULD respond with the extension field providing the network correction of the client's request. The server MUST ignore the value of the network correction in the request. An NTP client or peer which supports the PTP transport and is configured to use the network correction for the association SHOULD include the extension field in its NTP requests. In the case of a client, the correction value in the extension field SHOULD be always zero. When the client or peer has the network correction of both the request and response, it can correct the measured NTP peer delay and offset: delta_c = delta - (nc_rs + nc_rq - dur_rs - dur_rq) * (1 - freq_tc) theta_c = theta + (nc_rs - nc_rq) / 2 where delta is the NTP peer delay from RFC 5905 theta is the NTP offset from RFC 5905 nc_rq is the network correction of the request nc_rs is the network correction of the response dur_rq is the transmit duration of the request dur_rs is the receive duration of the response freq_tc is the maximum assumed frequency error of transparent clocks The corrected delay (delta_c) and offset (theta_c) MUST NOT be accepted for synchronization if any of delta_c, nc_rs, and nc_rq is negative. This requirement limits the error caused by faulty transparent clocks and man-in-the-middle attacks. Root delay (DELTA) MUST NOT be corrected to not make the maximum assumed error (root distance) dependent on accurate network corrections. The scaling by the freq_tc constant (e.g. 100 ppm) is needed to make room for errors in corrections made by transparent clocks running faster than true time and avoid samples with larger corrections from getting a shorter delay than samples with smaller corrections, which would negatively impact their filtering and weighting. The dur_rq and dur_rs values make the corrected peer delay correspond to a direct connection to the server. If they were not used, a perfectly corrected delay on a short network path would be too close to zero and frequently negative due to frequency offset between the client and server. Note that NTP peers and PTP clocks using the E2E delay mechanism are more sensitive to frequency offsets due to longer measurement intervals. If dur_rq is unknown, it MAY be assumed to be equal to dur_rs.

The author would like to thank Martin Langer for his useful comments.

IANA is requested to allocate the following field in the NTP Extension Field Types Registry: Field Type Meaning Reference [[TBD]] Network correction [[this memo]]

This section records the status of known implementations of the protocol defined by this specification at the time of posting of this Internet-Draft, and is based on a proposal described in RFC 7942. The description of implementations in this section is intended to assist the IETF in its decision processes in progressing drafts to RFCs. Please note that the listing of any individual implementation here does not imply endorsement by the IETF. Furthermore, no effort has been spent to verify the information presented here that was supplied by IETF contributors. This is not intended as, and must not be construed to be, a catalog of available implementations or their features. Readers are advised to note that other implementations may exist. According to RFC 7942, "this will allow reviewers and working groups to assign due consideration to documents that have the benefit of running code, which may serve as evidence of valuable experimentation and feedback that have made the implemented protocols more mature. It is up to the individual working groups to use this information as they see fit".

chrony (https://chrony-project.org) added experimental support for NTP over PTP in version 4.2. As the type of the NTP TLV, it uses 0x2023 from the experimental "do not propagate" range. It was tested on Linux with the following network controllers, which have hardware timestamping limited to PTP packets: Intel XL710 (i40e driver) - works Intel X540-AT2 (ixgbe driver) - works Intel 82576 (igb driver) - works Broadcom BCM5720 (tg3 driver) - works Broadcom BCM57810 (bnx2x driver) - does not timestamp unicast PTP packets Solarflare SFC9250 (sfc driver) - works The network correction was tested with the following switches which support operation as a one-step E2E PTP unicast transparent clock: FS.COM IES3110-8TF-R - works Juniper QFX5200-32C-32Q - works

The PTP transport prevents NTP clients from randomizing their source port. The corrections provided by PTP transparent clocks cannot be authenticated. Man-in-the-middle attackers can modify the correction field, but only corrections smaller than the measured delay are accepted by clients. The impact is comparable to the impact of delaying unmodified NTP messages.