Datagram Convergence Layers for the Delay- and Disruption&nbhy;Tolerant Networking (DTN) Bundle Protocol and Licklider Transmission Protocol (LTP)

DTN communication protocols include the Bundle Protocol described in RFC 5050, which provides transmission of application data blocks ("bundles") through optional intermediate custody transfer, and the Licklider Transmission Protocol (LTP) -- LTP Motivation, LTP Specification, and LTP Security -- which can be used to transmit bundles reliably and efficiently over a point-to-point link. It is often desirable to test these protocols over Internet Protocol links. "Delay Tolerant Networking TCP Convergence Layer Protocol" defines a method for transporting bundles over TCP. This document specifies the preferred method for transmitting either bundles or LTP blocks across the Internet using datagrams in place of TCP. shows the general protocol layering described in the DTN documents. DTN Applications interact with the Bundle Protocol Layer, which in turn uses a Convergence Layer to prepare a bundle for transmission. The Convergence Layer will typically rely on a lower-level protocol to carry out the transmission.

+-----------------------------------------+ | | | DTN Application | | | +-----------------------------------------+ +-----------------------------------------+ | | | Bundle Protocol (BP) | | | +-----------------------------------------+ +-----------------------------------------+ | | | Convergence Layer Adapter (CL) | | | +-----------------------------------------+ +-----------------------------------------+ | | | Local Data-Link Layer (Transport) | | | +-----------------------------------------+ This document provides guidance for implementation of the two protocol stacks illustrated in . In (a), the Convergence Layer Adapter is UDP or DCCP for direct transport of bundles over the Internet. In (b), the Convergence Layer Adapter is LTP, which then uses UDP or DCCP as the local data-link layer.

+-------------+ +-------------+ | | | | | DTN App | | DTN App | | | | | +-------------+ +-------------+ +-------------+ +-------------+ | | | | | BP | | BP | | | | | +-------------+ +-------------+ +-------------+ +-------------+ | | | | | UDP/DCCP | | LTP | | | | | +-------------+ +-------------+ +-------------+ | | | UDP/DCCP | | | +-------------+ (a) (b)

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.

In order to utilize DTN protocols across the Internet, whether for testing purposes or as part of a larger network path, it is necessary to encapsulate them into a standard Internet Protocol so that they travel easily across the Internet. This is particularly true for LTP, which provides no endpoint addressing. This encapsulation choice needs to be made carefully in order to avoid redundancy, since DTN protocols may provide their own reliability mechanisms. Congestion control is vital to the continued functioning of the Internet, particularly for situations where data will be sent at arbitrarily fast data rates. The Bundle Protocol delegates provision of reliable delivery and, implicitly, congestion control to the convergence layer used (Section 7.2 of RFC 5050). In situations where TCP will work effectively in communications between pairs of DTN nodes, use of the TCP convergence layer will provide the required reliability and congestion control for transport of bundles and would be the default choice in the Internet. Alternatives such as encapsulating bundles directly in datagrams and using UDP or DCCP are not generally appropriate because they offer limited reliability and, in the case of UDP, no congestion control. LTP, on the other hand, offers its own form of reliability. Particularly for testing purposes, it makes no sense to run LTP over a protocol like TCP that offers reliability already. In addition, running LTP over TCP would reduce the flexibility available to users, since LTP offers more control over what data is delivered reliably and what data is delivered best effort, a feature that TCP lacks. As such, it would be better to run LTP over an unreliable protocol. One solution would be to use UDP. UDP provides no reliability, allowing LTP to manage that itself. However, UDP also does not provide congestion control. Because LTP is designed to run over fixed-rate radio links, it does provide rate control but not congestion control. Lack of congestion control in network connections is a major problem that can cause artificially high loss rates and/or serious fairness issues. Previous standards documents are unanimous in recommending congestion control for protocols to be used on the Internet, see "Congestion Control Principles", "Unicast UDP Usage Guidelines", and "Queue Management and Congestion Avoidance", among others. RFC 5405, in particular, calls congestion control "vital" for "applications that can operate at higher, potentially unbounded data rates". Therefore, any Bundle Protocol implementation permitting the use of UDP to transport LTP segments or bundles outside an isolated network for the transmission of any non-trivial amounts of data MUST implement congestion control consistent with RFC 5405. Alternatively, the Datagram Congestion Control Protocol (DCCP) was designed specifically to provide congestion control without reliability for those applications that traverse the Internet but do not desire to retransmit lost data. As such, it is RECOMMENDED that, if possible, DCCP be used to transport LTP segments across the Internet.

The Bundle Protocol allows bundles with sizes limited only by node resource constraints. In IPv4, the maximum size of a UDP datagram is nearly 64 KB. In IPv6, when using jumbograms, UDP datagrams can technically be up to 4 GB in size , although this option is rarely used. (Note: RFC 2147 was obsoleted by RFC 2675.) It is well understood that sending large IP datagrams that must be fragmented by the network has enormous efficiency penalties. The Bundle Protocol specification provides a bundle fragmentation concept that allows a large bundle to be divided into bundle fragments. If the Bundle Protocol is being encapsulated in DCCP or UDP, it therefore SHOULD create each fragment of sufficiently small size that it can then be encapsulated into a datagram that will not need to be fragmented at the IP layer. IP fragmentation can be avoided by using IP Path MTU Discovery , which depends on the deterministic delivery of ICMP Packet Too Big (PTB) messages from routers in the network. To bypass a condition referred to as a black hole , a newer specification is available in to determine the IP Path MTU without the use of PTB messages. This document does not attempt to recommend one fragmentation avoidance mechanism over another; the information in this section is included for the benefit of implementers.

Because DCCP implementations are not required to support IP fragmentation and are not allowed to enable it by default, a DCCP Convergence Layer (we will use "CL" from here on) MUST NOT accept data segments that cannot be sent as a single MTU-sized datagram.

When an LTP CL is using UDP for datagram delivery, it SHOULD NOT create segments that will result in UDP datagrams that will need to be fragmented, as discussed above.

In general, the use of the Bundle Protocol over a datagram CL is discouraged in IP networks. Bundles can be of (almost) arbitrary length, and the Bundle Protocol does not include an effective retransmission mechanism. Whenever possible, the Bundle Protocol SHOULD be operated over the TCP Convergence Layer or over LTP. If a datagram CL is used for transmission of bundles, every datagram MUST contain exactly one bundle or 4 octets of zero bits as a keep-alive. Bundles that are too large for the path MTU SHOULD be fragmented and reassembled at the Bundle Protocol layer to prevent IP fragmentation.

The DCCP CL for Bundle Protocol use SHOULD use the IANA-assigned port 4556/DCCP and service code 1685351985; the use of other port numbers and service codes is implementation specific.

The UDP CL for Bundle Protocol use SHOULD use the IANA-assigned port 4556/UDP; the use of other port numbers is implementation specific.

LTP is designed as a point-to-point protocol within DTN, and it provides intrinsic acknowledgement and retransmission facilities. LTP segments are transported over a "local data-link layer" (RFC 5325); we will use the term "transport" from here on. Transport of LTP using datagrams is an appropriate choice. When a datagram transport is used to send LTP segments, every datagram MUST contain exactly one LTP segment or 4 octets of zero bits as a keep-alive. LTP MUST perform segmentation in such a way as to ensure that every LTP segment fits into a single packet which will not require IP fragmentation as discussed above.

The DCCP transport for LTP SHOULD use the IANA-assigned port 1113/DCCP and service code 7107696; the use of other port numbers and service codes is implementation specific.

The UDP transport for LTP SHOULD use the IANA-assigned port 1113/UDP; the use of other port numbers is implementation specific.

It may be desirable for a UDP or DCCP CL or transport to send "keep-alive" packets during extended idle periods. This may be needed to refresh a contact table entry at the destination, or to maintain an address mapping in a NAT or a dynamic access rule in a firewall. Therefore, the CL or transport MAY send a datagram containing exactly 4 octets of zero bits. The CL or transport receiving such a packet MUST discard this packet. The receiving CL or transport may then perform local maintenance of its state tables; these maintenance functions are not covered in this document. Note that packets carrying bundles or segments will always contain more than 4 octets of information (either the bundle or the LTP header); keep-alive packets will therefore never be mistaken for actual data packets. If UDP or DCCP is being used for communication in both directions between a pair of bundle agents, transmission and processing of keep-alives in the two directions occurs independently. Keep-alive intervals SHOULD be configurable, SHOULD default to 15 seconds, and MUST NOT be configured shorter than 15 seconds.

Both the core Bundle Protocol specification and core LTP specification assume that they are transmitting over an erasure channel, i.e., a channel that either delivers packets correctly or not at all.

A DCCP transmitter MUST, therefore, ensure that the entire packet is checksummed by setting the Checksum Coverage to zero. Likewise, the DCCP receiver MUST ignore all packets with partial checksum coverage.

A UDP transmitter, therefore, MUST NOT disable UDP checksums, and the UDP receiver MUST NOT disable the checking of received UDP checksums. Even when UDP checksums are enabled, a small probability of UDP packet corruption remains. In some environments, it may be acceptable for LTP or the Bundle Protocol to occasionally receive corrupted input. In general, however, a UDP implementation SHOULD use optional security extensions available in the Bundle Protocol or LTP to protect against message corruption.

DCCP supports pluggable congestion control modules in order to optimize its behavior to particular environments. The two most common congestion control modules (CCIDs) are TCP-like Congestion Control (CCID2) and TCP-Friendly Rate Control (CCID3). TCP-like Congestion Control is designed to emulate TCP's congestion control as much as possible. It is recommended for applications that want to send data as quickly as possible, while TCP-Friendly Rate Control is aimed at applications that want to avoid sudden changes in sending rate. DTN use cases seem to fit more into the first case, so DCCP CL's and transports SHOULD use TCP-like Congestion Control (CCID2) by default.

Port number assignments 1113/UDP and 4556/UDP have been registered with IANA. The assignment for 1113/UDP referenced ; this entry has been changed to add the present document in addition to . The assignment of 4556/UDP had no reference; this entry has been changed to point to the present document. The service name for 4556/UDP has been changed from dtn-bundle-udp to dtn-bundle. Port number 1113/DCCP (ltp-deepspace) with Service Code 7107696 has been assigned for the transport of LTP. Port number 4556/DCCP (dtn-bundle) with Service Code 1685351985 has been assigned for the transport of bundles. The port number assignment for 4556/TCP is addressed in the document.

This memo describes the use of datagrams to transport DTN application data. Hosts may be in the position of having to accept and process packets from unknown sources; the DTN Endpoint ID can be discovered only after the bundle has been retrieved from the DCCP or UDP packet. Hosts SHOULD use authentication methods available in the DTN specifications to prevent malicious hosts from inserting unknown data into the application. Hosts need to listen for and process DCCP or UDP data on the known LTP or Bundle Protocol ports. A denial-of-service scenario exists where a malicious host sends datagrams at a high rate, forcing the receiving hosts to use their resources to process and attempt to authenticate this data. Whenever possible, hosts SHOULD use IP address filtering to limit the origin of packets to known hosts.