| Network Working Group | G. Fairhurst, Ed. |
| Internet-Draft | University of Aberdeen |
| Intended status: Informational | B. Trammell, Ed. |
| Expires: August 31, 2015 | M. Kuehlewind, Ed. |
| ETH Zurich | |
| February 27, 2015 |
Services provided by IETF transport protocols and congestion control mechanisms
draft-ietf-taps-transports-03
This document describes services provided by existing IETF protocols and congestion control mechanisms. It is designed to help application and network stack programmers and to inform the work of the IETF TAPS Working Group.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at http://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on August 31, 2015.
Copyright (c) 2015 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
Most Internet applications make use of the Transport Services provided by TCP (a reliable, in-order stream protocol) or UDP (an unreliable datagram protocol). We use the term “Transport Service” to mean the end-to-end service provided to an application by the transport layer. That service can only be provided correctly if information about the intended usage is supplied from the application. The application may determine this information at design time, compile time, or run time, and may include guidance on whether a feature is required, a preference by the application, or something in between. Examples of features of Transport Services are reliable delivery, ordered delivery, content privacy to in-path devices, integrity protection, and minimal latency.
The IETF has defined a wide variety of transport protocols beyond TCP and UDP, including TCP, SCTP, DCCP, MP-TCP, and UDP-Lite. Transport services may be provided directly by these transport protocols, or layered on top of them using protocols such as WebSockets (which runs over TCP) or RTP (over TCP or UDP). Services built on top of UDP or UDP-Lite typically also need to specify additional mechanisms, including a congestion control mechanism (such as a windowed congestion control, TFRC or LEDBAT congestion control mechanism). This extends the set of available Transport Services beyond those provided to applications by TCP and UDP.
Transport protocols can also be differentiated by the features of the services they provide: for instance, SCTP offers a message-based service that does not suffer head-of-line blocking when used with multiple stream, because it can accept blocks of data out of order, UDP-Lite provides partial integrity protection, and LEDBAT can provide low-priority “scavenger” communication.
The following terms are defined throughout this document, and in subsequent documents produced by TAPS describing the composition and decomposition of transport services.
[NOTE: The terminology below was presented at the TAPS WG meeting in Honolulu. While the factoring of the terminology seems uncontroversial, there may be some entities which still require names (e.g. information about the interface between the transport and lower layers which could lead to the availability or unavailability of certain transport protocol features). Comments are welcome via the TAPS mailing list.]
This section provides a list of known IETF transport protocol and transport protocol frameworks.
[EDITOR’S NOTE: Contributions to the subsections below are welcome]
TCP is an IETF standards track transport protocol. [RFC0793] introduces TCP as follows: “The Transmission Control Protocol (TCP) is intended for use as a highly reliable host-to-host protocol between hosts in packet-switched computer communication networks, and in interconnected systems of such networks.” Since its introduction, TCP has become the default connection-oriented, stream-based transport protocol in the Internet. It is widely implemented by endpoints and widely used by common applications.
TCP is a connection-oriented protocol, providing a three way handshake to allow a client and server to set up a connection, and mechanisms for orderly completion and immediate teardown of a connection. TCP is defined by a family of RFCs [RFC4614].
TCP provides multiplexing to multiple sockets on each host using port numbers. An active TCP session is identified by its four-tuple of local and remote IP addresses and local port and remote port numbers. The destination port during connection setup has a different role as it is often used to indicate the requested service.
TCP partitions a continuous stream of bytes into segments, sized to fit in IP packets. ICMP-based PathMTU discovery [RFC1191][RFC1981] as well as Packetization Layer Path MTU Discovery (PMTUD) [RFC4821] are supported.
Each byte in the stream is identified by a sequence number. The sequence number is used to order segments on receipt, to identify segments in acknowledgments, and to detect unacknowledged segments for retransmission. This is the basis of TCP’s reliable, ordered delivery of data in a stream. TCP Selective Acknowledgment [RFC2018] extends this mechanism by making it possible to identify missing segments more precisely, reducing spurious retransmission.
Receiver flow control is provided by a sliding window: limiting the amount of unacknowledged data that can be outstanding at a given time. The window scale option [RFC7323] allows a receiver to use windows greater than 64KB.
All TCP senders provide Congestion Control: This uses a separate window, where each time congestion is detected, this congestion window is reduced. A receiver detects congestion using one of three mechanisms: A retransmission timer, detection of loss (interpreted as a congestion signal), or Explicit Congestion Notification (ECN) [RFC3168] to provide early signaling (see [I-D.ietf-aqm-ecn-benefits])
A TCP protocol instance can be extended [RFC4614] and tuned. Some features are sender-side only, requiring no negotiation with the receiver; some are receiver-side only, some are explicitly negotiated during connection setup.
By default, TCP segment partitioning uses Nagle’s algorithm [RFC0896] to buffer data at the sender into large segments, potentially incurring sender-side buffering delay; this algorithm can be disabled by the sender to transmit more immediately, e.g. to enable smoother interactive sessions.
[EDITOR’S NOTE: add URGENT and PUSH flag (note [RFC6093] says SHOULD NOT use due to the range of TCP implementations that process TCP urgent indications differently.) ]
A checksum provides an Integrity Check and is mandatory across the entire packet. The TCP checksum does not support partial corruption protection as in DCCP/UDP-Lite). This check protects from misdelivery of data corrupted data, but is relatively weak, and applications that require end to end integrity of data are recommended to include a stronger integrity check of their payload data.
A TCP service is unicast.
A User/TCP Interface is defined in [RFC0793] providing six user commands: Open, Send, Receive, Close, Status. This interface does not describe configuration of TCP options or parameters beside use of the PUSH and URGENT flags.
In API implementations derived from the BSD Sockets API, TCP sockets are created using the SOCK_STREAM socket type.
The features used by a protocol instance may be set and tuned via this API.
(more on the API goes here)
The transport protocol components provided by TCP are:
[EDITOR’S NOTE: discussion of how to map this to features and TAPS: what does the higher layer need to decide? what can the transport layer decide based on global settings? what must the transport layer decide based on network characteristics?]
[EDITOR’S NOTE: a few sentences describing Multipath TCP [RFC6824] go here. Note that this adds transport-layer multihoming to the components TCP provides]
SCTP is a message oriented standards track transport protocol and the base protocol is specified in [RFC4960]. It supports multi-homing to handle path failures. An SCTP association has multiple unidirectional streams in each direction and provides in-sequence delivery of user messages only within each stream. This allows to minimize head of line blocking. SCTP is extensible and the currently defined extensions include mechanisms for dynamic re-configurations of streams [RFC6525] and IP-addresses [RFC5061]. Furthermore, the extension specified in [RFC3758] introduces the concept of partial reliability for user messages.
SCTP was originally developed for transporting telephony signalling messages and is deployed in telephony signalling networks, especially in mobile telephony networks. Additionally, it is used in the WebRTC framework for data channels and is therefore deployed in all WEB-browsers supporting WebRTC.
[EDITOR’S NOTE: Michael Tuexen and Karen Nielsen signed up as contributors for these sections.]
SCTP is a connection oriented protocol using a four way handshake to establish an SCTP association and a three way message exchange to gracefully shut it down. It uses the same port number concept as DCCP, TCP, UDP, and UDP-Lite do and only supports unicast.
SCTP uses the 32-bit CRC32c for protecting SCTP packets against bit errors. This is stronger than the 16-bit checksums used by TCP or UDP. However, a partial checksum coverage as provided by DCCP or UDP-Lite is not supported.
SCTP has been designed with extensibility in mind. Each SCTP packet starts with a single common header containing the port numbers, a verification tag and the CRC32c checksum. This common header is followed by a sequence of chunks. Each chunk consists of a type field, flags, a length field and a value. [RFC4960] defines how a receiver processes chunks with an unknown chunk type. The support of extensions can be negotiated during the SCTP handshake.
SCTP provides a message-oriented service. Multiple small user messages can be bundled into a single SCTP packet to improve the efficiency. User messages which would result in IP packets larger than the MTU will be fragmented at the sender side and reassembled at the receiver side. There is no protocol limit on the user message size. [RFC4821] defines a method to perform packetization layer path MTU discovery with probe packets using the padding chunks defined the [RFC4820].
[RFC4960] specifies a TCP friendly congestion control to protect the network against overload. SCTP also uses a sliding window flow control to protect receivers against overflow.
Each SCTP association has between 1 and 65536 uni-directional streams in each direction. The number of streams can be different in each direction. Every user-message is sent on a particular stream. User messages can be sent ordered or un-ordered upon request by the upper layer. Only all ordered messages sent on the same stream are delivered at the receiver in the same order as sent by the sender. For user messages not requiring fragmentation, this minimises head of line blocking. The base protocol defined in [RFC4960] doesn’t allow interleaving of user-messages, which results in sending a large message on one stream can block the sending of user messages on other streams. [I-D.ietf-tsvwg-sctp-ndata] overcomes this limitation and also allows to specify a scheduler for the sender side streams selection. The stream re-configuration extension defined in [RFC6525] allows to reset streams during the lifetime of an association and to increase the number of streams, if the number of streams negotiated in the SCTP handshake is not sufficient.
According to [RFC4960], each user message sent is either delivered to the receiver or, in case of excessive retransmissions, the association is terminated in a non-graceful way, similar to the TCP behaviour. In addition to this reliable transfer, the partial reliability extension defined in [RFC3758] allows the sender to abandon user messages. The application can specify the policy for abandoning user messages. Examples for these policies include:
SCTP supports multi-homing. Each SCTP end-point uses a list of IP-addresses and a single port number. These addresses can be any mixture of IPv4 and IPv6 addresses. These addresses are negotiated during the handshake and the address re-configuration extension specified in [RFC5061] can be used to change these addresses during the livetime of an SCTP association. This allows for transport layer mobility. Multiple addresses are used for improved resilience. If a remote address becomes unreachable, the traffic is switched over to a reachable one, if one exists. Each SCTP end-point supervises continuously the reachability of all peer addresses using a heartbeat mechanism.
For securing user messages, the use of TLS over SCTP has been specified in [RFC3436]. However, this solution does not support all services provided by SCTP (for example un-ordered delivery or partial reliability), and therefore the use of DTLS over SCTP has been specified in [RFC6083] to overcome these limitations. When using DTLS over SCTP, the application can use almost all services provided by SCTP.
For legacy NAT traversal, [RFC6951] defines the UDP encapsulation of SCTP-packets. Alternatively, SCTP packets can be encapsulated in DTLS packets as specified in [I-D.ietf-tsvwg-sctp-dtls-encaps]. The latter encapsulation is used with in the WebRTC context.
Having a well defined API is also a feature provided by SCTP as described in the next subsection.
[RFC4960] defines an abstract API for the base protocol. An extension to the BSD Sockets API is defined in [RFC6458] and covers:
For the following SCTP protocol extensions the BSD Sockets API extension is defined in the document specifying the protocol extensions:
Future documents describing SCTP protocol extensions are expected to describe the corresponding BSD Sockets API extension in a Socket API Considerations section.
The SCTP socket API supports two kinds of sockets:
One-to-one style sockets are similar to TCP sockets, there is a 1:1 relationship between the sockets and the SCTP associations (except for listening sockets). One-to-many style SCTP sockets are similar to unconnected UDP sockets as there is a 1:n relationship between the sockets and the SCTP associations.
The SCTP stack can provide information to the applications about state changes of the individual paths and the association whenever they occur. These events are delivered similar to user messages but are specifically marked as notifications.
A couple of new functions have been introduced to support the use of multiple local and remote addresses. Additional SCTP-specific send and receive calls have been defined to allow dealing with the SCTP specific information without using ancillary data in the form of additional cmsgs, which are also defined. These functions provide support for detecting partial delivery of user messages and notifications.
The SCTP socket API allows a fine-grained control of the protocol behaviour through an extensive set of socket options.
The SCTP kernel implementations of FreeBSD, Linux and Solaris follow mostly the specified extension to the BSD Sockets API for the base protocol and the corresponding supported protocol extensions.
The transport protocol components provided by SCTP are:
[EDITOR’S NOTE: update this list.]
The User Datagram Protocol (UDP) [RFC0768] [RFC2460] is an IETF standards track transport protocol. It provides a uni-directional, datagram protocol which preserves message boundaries. It provides none of the following transport features: error correction, congestion control, or flow control. It can be used to send broadcast datagrams (IPv4) or multicast datagrams (IPv4 and IPv6), in addition to unicast (and anycast) datagrams. IETF guidance on the use of UDP is provided in[RFC5405]. UDP is widely implemented and widely used by common applications, especially DNS.
[EDITOR’S NOTE: Kevin Fall signed up as a contributor for this section.]
UDP is a connection-less protocol which maintains message boundaries, with no connection setup or feature negotiation. The protocol uses independent messages, ordinarily called datagrams. The lack of error control and flow control implies messages may be damaged, re-ordered, lost, or duplicated in transit. A receiving application unable to run sufficiently fast or frequently may miss messages. The lack of congestion handling implies UDP traffic may cause the loss of messages from other protocols (e.g., TCP) when sharing the same network paths. UDP traffic can also cause the loss of other UDP traffic in the same or other flows for the same reasons.
Messages with bit errors are ordinarily detected by an invalid end- to-end checksum and are discarded before being delivered to an application. There are some exceptions to this general rule, however. UDP-Lite (see [RFC3828], and below) provides the ability for portions of the message contents to be exempt from checksum coverage. It is also possible to create UDP datagrams with no checksum, and while this is generally discouraged [RFC1122] [RFC5405], certain special cases permit its use [RFC6935]. The checksum support considerations for omitting the checksum are defined in [RFC6936]. Note that due to the relatively weak form of checksum used by UDP, applications that require end to end integrity of data are recommended to include a stronger integrity check of their payload data.
On transmission, UDP encapsulates each datagram into an IP packet, which may in turn be fragmented by IP. Applications concerned with fragmentation or that have other requirements such as receiver flow control, congestion control, PathMTU discovery/PLPMTUD, support for ECN, etc need to be provided by protocols other than UDP [RFC5405].
[RFC0768] describes basic requirements for an API for UDP. Guidance on use of common APIs is provided in [RFC5405].
A UDP endpoint consists of a tuple of (IP address, port number). Demultiplexing using multiple abstract endpoints (sockets) on the same IP address are supported. The same socket may be used by a single server to interact with multiple clients (note: this behavior differs from TCP, which uses a pair of tuples to identify a connection). Multiple server instances (processes) binding the same socket can cooperate to service multiple clients– the socket implementation arranges to not duplicate the same received unicast message to multiple server processes.
Many operating systems also allow a UDP socket to be “connected”, i.e., to bind a UDP socket to a specific (remote) UDP endpoint. Unlike TCP’s connect primitive, for UDP, this is only a local operation that serves to simplify the local send/receive functions and to filter the traffic for the specified addresses and ports [RFC5405].
The transport protocol components provided by UDP are:
The Lightweight User Datagram Protocol (UDP-Lite) [RFC3828] is an IETF standards track transport protocol. UDP-Lite provides a bidirectional set of logical unicast or multicast message streams over a datagram protocol. IETF guidance on the use of UDP-Lite is provided in [RFC5405].
[EDITOR’S NOTE: Gorry Fairhurst signed up as a contributor for this section.]
UDP-Lite is a connection-less datagram protocol, with no connection setup or feature negotiation. The protocol use messages, rather than a byte-stream. Each stream of messages is independently managed, therefore retransmission does not hold back data sent using other logical streams.
It provides multiplexing to multiple sockets on each host using port numbers. An active UDP-Lite session is identified by its four-tuple of local and remote IP addresses and local port and remote port numbers.
UDP-Lite fragments packets into IP packets, constrained by the maximum size of IP packet.
UDP-Lite changes the semantics of the UDP “payload length” field to that of a “checksum coverage length” field. Otherwise, UDP-Lite is semantically identical to UDP. Applications using UDP-Lite therefore can not make assumptions regarding the correctness of the data received in the insensitive part of the UDP-Lite payload.
As for UDP, mechanisms for receiver flow control, congestion control, PMTU or PLPMTU discovery, support for ECN, etc need to be provided by upper layer protocols [RFC5405].
Examples of use include a class of applications that can derive benefit from having partially-damaged payloads delivered, rather than discarded. One use is to support error tolerate payload corruption when used over paths that include error-prone links, another application is when header integrity checks are required, but payload integrity is provided by some other mechanism (e.g. [RFC6936].
A UDP-Lite service may support IPv4 broadcast, multicast, anycast and unicast.
There is no current API specified in the RFC Series, but guidance on use of common APIs is provided in [RFC5405].
The interface of UDP-Lite differs from that of UDP by the addition of a single (socket) option that communicates a checksum coverage length value: at the sender, this specifies the intended checksum coverage, with the remaining unprotected part of the payload called the “error-insensitive part”. The checksum coverage may also be made visible to the application via the UDP-Lite MIB module [RFC5097].
The transport protocol components provided by UDP-Lite are:
Datagram Congestion Control Protocol (DCCP) [RFC4340] is an IETF standards track bidirectional transport protocol that provides unicast connections of congestion-controlled unreliable messages.
[EDITOR’S NOTE: Gorry Fairhurst signed up as a contributor for this section.]
The DCCP Problem Statement describes the goals that DCCP sought to address [RFC4336]. It is suitable for applications that transfer fairly large amounts of data and that can benefit from control over the trade off between timeliness and reliability [RFC4336].
It offers low overhead, and many characteristics common to UDP, but can avoid “Re-inventing the wheel” each time a new multimedia application emerges. Specifically it includes core functions (feature negotiation, path state management, RTT calculation, PMTUD, etc): This allows applications to use a compatible method defining how they send packets and where suitable to choose common algorithms to manage their functions. Examples of suitable applications include interactive applications, streaming media or on-line games [RFC4336].
DCCP is a connection-oriented datagram protocol, providing a three way handshake to allow a client and server to set up a connection, and mechanisms for orderly completion and immediate teardown of a connection. The protocol is defined by a family of RFCs.
It provides multiplexing to multiple sockets on each host using port numbers. An active DCCP session is identified by its four-tuple of local and remote IP addresses and local port and remote port numbers. At connection setup, DCCP also exchanges the the service code [RFC5595] mechanism to allow transport instantiations to indicate the service treatment that is expected from the network.
The protocol segments data into messages, typically sized to fit in IP packets, but which may be fragmented providing they are less than the A DCCP interface MAY allow applications to request fragmentation for packets larger than PMTU, but not larger than the maximum packet size allowed by the current congestion control mechanism (CCMPS) [RFC4340].
Each message is identified by a sequence number. The sequence number is used to identify segments in acknowledgments, to detect unacknowledged segments, to measure RTT, etc. The protocol may support ordered or unordered delivery of data, and does not itself provide retransmission. There is a Data Checksum option, which contains a strong CRC, lets endpoints detect application data corruption. It also supports reduced checksum coverage, a partial integrity mechanisms similar to UDP-lIte.
Receiver flow control is supported: limiting the amount of unacknowledged data that can be outstanding at a given time.
A DCCP protocol instance can be extended [RFC4340] and tuned. Some features are sender-side only, requiring no negotiation with the receiver; some are receiver-side only, some are explicitly negotiated during connection setup.
DCCP supports negotiation of the congestion control profile, to provide Plug and Play congestion control mechanisms. examples of specified profiles include [RFC4341] [RFC4342] [RFC5662]. All IETF-defined methods provide Congestion Control.
DCCP use a Connect packet to start a session, and permits half-connections that allow each client to choose features it wishes to support. Simultaneous open [RFC5596], as in TCP, can enable interoperability in the presence of middleboxes. The Connect packet includes a Service Code field [RFC5595] designed to allow middle boxes and endpoints to identify the characteristics required by a session. A lightweight UDP-based encapsulation (DCCP-UDP) has been defined [RFC6773] that permits DCCP to be used over paths where it is not natively supported. Support in NAPT/NATs is defined in [RFC4340] and [RFC5595].
Upper layer protocols specified on top of DCCP include: DTLS [RFC5595], RTP [RFC5672], ICE/SDP [RFC6773].
A DCCP service is unicast.
A common packet format has allowed tools to evolve that can read and interpret DCCP packets (e.g. Wireshark).
API characteristics include: - Datagram transmission. - Notification of the current maximum packet size. - Send and reception of zero-length payloads. - Set the Slow Receiver flow control at a receiver. - Detect a Slow receiver at the sender.
There is no current API specified in the RFC Series.
The transport protocol components provided by DCCP are:
RTP provides an end-to-end network transport service, suitable for applications transmitting real-time data, such as audio, video or data, over multicast or unicast network services, including TCP, UDP, UDP-Lite, DCCP.
[EDITOR’S NOTE: Varun Singh signed up as contributor for this section.]
pseudo transport
[NOTE: A few words on TLS [RFC5246] and DTLS [RFC6347] here, and how they get used by other protocols to meet security goals as an add-on interlayer above transport.]
[EDITOR’S NOTE: No identified contributor for this section yet.]
[EDITOR’S NOTE: No identified contributor for this section yet.]
[EDITOR’S NOTE: this section will drawn from the candidate features provided by protocol components in the previous section – please discuss on taps@ietf.org list]
[EDITOR’S NOTE: Dave Thaler has signed up as a contributor for this section. Michael Welzl also has a beginning of a matrix which could be useful here.]
[EDITOR’S NOTE: The below is a strawman proposal below by Gorry Fairhurst for initial discussion]
The table below summarises protocol mechanisms that have been standardised. It does not make an assessment on whether specific implementations are fully compliant to these specifications.
| Mechanism | UDP | UDP-L | DCCP | SCTP | TCP |
|---|---|---|---|---|---|
| Unicast | Yes | Yes | Yes | Yes | Yes |
| Mcast/IPv4Bcast | Yes(2) | Yes | No | No | No |
| Port Mux | Yes | Yes | Yes | Yes | Yes |
| Mode | Dgram | Dgram | Dgram | Dgram | Stream |
| Connected | No | No | Yes | Yes | Yes |
| Data bundling | No | No | No | Yes | Yes |
| Feature Nego | No | No | Yes | Yes | Yes |
| Options | No | No | Support | Support | Support |
| Data priority | * | * | * | Yes | No |
| Data bundling | No | No | No | Yes | Yes |
| Reliability | None | None | None | Select | Full |
| Ordered deliv | No | No | No | Stream | Yes |
| Corruption Tol. | No | Support | Support | No | No |
| Flow Control | No | No | Support | Yes | Yes |
| PMTU/PLPMTU | (1) | (1) | Yes | Yes | Yes |
| Cong Control | (1) | (1) | Yes | Yes | Yes |
| ECN Support | (1) | (1) | Yes | TBD | Yes |
| NAT support | Limited | Limited | Support | TBD | Support |
| Security | DTLS | DTLS | DTLS | DTLS | TLS, AO |
| UDP encaps | N/A | None | Yes | Yes | None |
| RTP support | Support | Support | Support | ? | Support |
Note (1): this feature requires support in an upper layer protocol.
Note (2): this feature requires support in an upper layer protocol when used with IPv6.
This document has no considerations for IANA.
This document surveys existing transport protocols and protocols providing transport-like services. Confidentiality, integrity, and authenticity are among the features provided by those services. This document does not specify any new components or mechanisms for providing these features. Each RFC listed in this document discusses the security considerations of the specification it contains.
[Editor’s Note: turn this into a real contributors section with addresses once we figure out how to trick the toolchain into doing so]
This work is partially supported by the European Commission under grant agreement FP7-ICT-318627 mPlane; support does not imply endorsement.
| [RFC0791] | Postel, J., "Internet Protocol", STD 5, RFC 791, September 1981. |