Network Working Group E. Kohler
Request for Comments: 4340 UCLA
Category: Standards Track M. Handley
March 2006 Datagram Congestion Control Protocol (DCCP)
Status of This Memo
This document specifies an Internet standards track protocol for the
Internet community, and requests discussion and suggestions for
improvements. Please refer to the current edition of the "Internet
Official Protocol Standards" (STD 1) for the standardization state
and status of this protocol. Distribution of this memo is unlimited.
Copyright (C) The Internet Society (2006).
The Datagram Congestion Control Protocol (DCCP) is a transport
protocol that provides bidirectional unicast connections of
congestion-controlled unreliable datagrams. DCCP is suitable for
applications that transfer fairly large amounts of data and that can
benefit from control over the tradeoff between timeliness and
Table of Contents
1. Introduction ....................................................52. Design Rationale ................................................63. Conventions and Terminology .....................................73.1. Numbers and Fields .........................................73.2. Parts of a Connection ......................................83.3. Features ...................................................93.4. Round-Trip Times ...........................................93.5. Security Limitation ........................................93.6. Robustness Principle ......................................104. Overview .......................................................104.1. Packet Types ..............................................104.2. Packet Sequencing .........................................114.3. States ....................................................124.4. Congestion Control Mechanisms .............................14
The Datagram Congestion Control Protocol (DCCP) is a transport
protocol that implements bidirectional, unicast connections of
congestion-controlled, unreliable datagrams. Specifically, DCCP
provides the following:
o Unreliable flows of datagrams.
o Reliable handshakes for connection setup and teardown.
o Reliable negotiation of options, including negotiation of a
suitable congestion control mechanism.
o Mechanisms allowing servers to avoid holding state for
unacknowledged connection attempts and already-finished
o Congestion control incorporating Explicit Congestion Notification
(ECN) [RFC3168] and the ECN Nonce [RFC3540].
o Acknowledgement mechanisms communicating packet loss and ECN
information. Acks are transmitted as reliably as the relevant
congestion control mechanism requires, possibly completely
o Optional mechanisms that tell the sending application, with high
reliability, which data packets reached the receiver, and whether
those packets were ECN marked, corrupted, or dropped in the
o Path Maximum Transmission Unit (PMTU) discovery [RFC1191].
o A choice of modular congestion control mechanisms. Two mechanisms
are currently specified: TCP-like Congestion Control [RFC4341] and
TCP-Friendly Rate Control (TFRC) [RFC4342]. DCCP is easily
extensible to further forms of unicast congestion control.
DCCP is intended for applications such as streaming media that can
benefit from control over the tradeoffs between delay and reliable
in-order delivery. TCP is not well suited for these applications,
since reliable in-order delivery and congestion control can cause
arbitrarily long delays. UDP avoids long delays, but UDP
applications that implement congestion control must do so on their
own. DCCP provides built-in congestion control, including ECN
support, for unreliable datagram flows, avoiding the arbitrary delays
associated with TCP. It also implements reliable connection setup,
teardown, and feature negotiation.
2. Design Rationale
One DCCP design goal was to give most streaming UDP applications
little reason not to switch to DCCP, once it is deployed. To
facilitate this, DCCP was designed to have as little overhead as
possible, both in terms of the packet header size and in terms of the
state and CPU overhead required at end hosts. Only the minimal
necessary functionality was included in DCCP, leaving other
functionality, such as forward error correction (FEC), semi-
reliability, and multiple streams, to be layered on top of DCCP as
Different forms of conformant congestion control are appropriate for
different applications. For example, on-line games might want to
make quick use of any available bandwidth, while streaming media
might trade off this responsiveness for a steadier, less bursty rate.
(Sudden rate changes can cause unacceptable UI glitches such as
audible pauses or clicks in the playout stream.) DCCP thus allows
applications to choose from a set of congestion control mechanisms.
One alternative, TCP-like Congestion Control, halves the congestion
window in response to a packet drop or mark, as in TCP. Applications
using this congestion control mechanism will respond quickly to
changes in available bandwidth, but must tolerate the abrupt changes
in congestion window typical of TCP. A second alternative, TCP-
Friendly Rate Control (TFRC) [RFC3448], a form of equation-based
congestion control, minimizes abrupt changes in the sending rate
while maintaining longer-term fairness with TCP. Other alternatives
can be added as future congestion control mechanisms are
DCCP also lets unreliable traffic safely use ECN. A UDP kernel
Application Programming Interface (API) might not allow applications
to set UDP packets as ECN capable, since the API could not guarantee
that the application would properly detect or respond to congestion.
DCCP kernel APIs will have no such issues, since DCCP implements
congestion control itself.
We chose not to require the use of the Congestion Manager [RFC3124],
which allows multiple concurrent streams between the same sender and
receiver to share congestion control. The current Congestion Manager
can only be used by applications that have their own end-to-end
feedback about packet losses, but this is not the case for many of
the applications currently using UDP. In addition, the current
Congestion Manager does not easily support multiple congestion
control mechanisms or mechanisms where the state about past packet
drops or marks is maintained at the receiver rather than the sender.
DCCP should be able to make use of CM where desired by the
application, but we do not see any benefit in making the deployment
of DCCP contingent on the deployment of CM itself.
We intend for DCCP's protocol mechanisms, which are described in this
document, to suit any application desiring unicast congestion-
controlled streams of unreliable datagrams. However, the congestion
control mechanisms currently approved for use with DCCP, which are
described in separate Congestion Control ID Profiles [RFC4341,
RFC4342], may cause problems for some applications, including high-
bandwidth interactive video. These applications should be able to
use DCCP once suitable Congestion Control ID Profiles are
3. Conventions and Terminology
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 [RFC2119].
3.1. Numbers and Fields
All multi-byte numerical quantities in DCCP, such as port numbers,
Sequence Numbers, and arguments to options, are transmitted in
network byte order (most significant byte first).
We occasionally refer to the "left" and "right" sides of a bit field.
"Left" means towards the most significant bit, and "right" means
towards the least significant bit.
Random numbers in DCCP are used for their security properties and
SHOULD be chosen according to the guidelines in [RFC4086].
All operations on DCCP sequence numbers use circular arithmetic
modulo 2^48, as do comparisons such as "greater" and "greatest".
This form of arithmetic preserves the relationships between sequence
numbers as they roll over from 2^48 - 1 to 0. Implementation
strategies for DCCP sequence numbers will resemble those for other
circular arithmetic spaces, including TCP's sequence numbers [RFC793]
and DNS's serial numbers [RFC1982]. It may make sense to store DCCP
sequence numbers in the most significant 48 bits of 64-bit integers
and set the least significant 16 bits to zero, since this supports a
common technique that implements circular comparison A < B by testing
whether (A - B) < 0 using conventional two's-complement arithmetic.
Reserved bitfields in DCCP packet headers MUST be set to zero by
senders and MUST be ignored by receivers, unless otherwise specified.
This allows for future protocol extensions. In particular, DCCP
processors MUST NOT reset a DCCP connection simply because a Reserved
field has non-zero value [RFC3360].
3.2. Parts of a Connection
Each DCCP connection runs between two hosts, which we often name DCCP
A and DCCP B. Each connection is actively initiated by one of the
hosts, which we call the client; the other, initially passive host is
called the server. The term "DCCP endpoint" is used to refer to
either of the two hosts explicitly named by the connection (the
client and the server). The term "DCCP processor" refers more
generally to any host that might need to process a DCCP header; this
includes the endpoints and any middleboxes on the path, such as
firewalls and network address translators.
DCCP connections are bidirectional: data may pass from either
endpoint to the other. This means that data and acknowledgements may
flow in both directions simultaneously. Logically, however, a DCCP
connection consists of two separate unidirectional connections,
called half-connections. Each half-connection consists of the
application data sent by one endpoint and the corresponding
acknowledgements sent by the other endpoint. We can illustrate this
+--------+ A-to-B half-connection: +--------+
| | --> application data --> | |
| | <-- acknowledgements <-- | |
| DCCP A | | DCCP B |
| | B-to-A half-connection: | |
| | <-- application data <-- | |
+--------+ --> acknowledgements --> +--------+
Although they are logically distinct, in practice the half-
connections overlap; a DCCP-DataAck packet, for example, contains
application data relevant to one half-connection and acknowledgement
information relevant to the other.
In the context of a single half-connection, the terms "HC-Sender" and
"HC-Receiver" denote the endpoints sending application data and
acknowledgements, respectively. For example, DCCP A is the
HC-Sender and DCCP B is the HC-Receiver in the A-to-B
A DCCP feature is a connection attribute on whose value the two
endpoints agree. Many properties of a DCCP connection are controlled
by features, including the congestion control mechanisms in use on
the two half-connections. The endpoints achieve agreement through
the exchange of feature negotiation options in DCCP headers.
DCCP features are identified by a feature number and an endpoint.
The notation "F/X" represents the feature with feature number F
located at DCCP endpoint X. Each valid feature number thus
corresponds to two features, which are negotiated separately and need
not have the same value. The two endpoints know, and agree on, the
value of every valid feature. DCCP A is the "feature location" for
all features F/A, and the "feature remote" for all features F/B.
3.4. Round-Trip Times
DCCP round-trip time measurements are performed by congestion control
mechanisms; different mechanisms may measure round-trip time in
different ways, or not measure it at all. However, the main DCCP
protocol does use round-trip times occasionally, such as in the
initial values for certain timers. Each DCCP implementation thus
defines a default round-trip time for use when no estimate is
available. This parameter should default to not less than 0.2
seconds, a reasonably conservative round-trip time for Internet TCP
connections. Protocol behavior specified in terms of "round-trip
time" values actually refers to "a current round-trip time estimate
taken by some CCID, or, if no estimate is available, the default
round-trip time parameter".
The maximum segment lifetime, or MSL, is the maximum length of time a
packet can survive in the network. The DCCP MSL should equal that of
TCP, which is normally two minutes.
3.5. Security Limitation
DCCP provides no protection against attackers who can snoop on a
connection in progress, or who can guess valid sequence numbers in
other ways. Applications desiring stronger security should use IPsec
[RFC2401]; depending on the level of security required, application-
level cryptography may also suffice. These issues are discussed
further in Sections 7.5.5 and 18.
3.6. Robustness Principle
DCCP implementations will follow TCP's "general principle of
robustness": "be conservative in what you do, be liberal in what you
accept from others" [RFC793].
DCCP's high-level connection dynamics echo those of TCP. Connections
progress through three phases: initiation, including a three-way
handshake; data transfer; and termination. Data can flow both ways
over the connection. An acknowledgement framework lets senders
discover how much data has been lost and thus avoid unfairly
congesting the network. Of course, DCCP provides unreliable datagram
semantics, not TCP's reliable bytestream semantics. The application
must package its data into explicit frames and must retransmit its
own data as necessary. It may be useful to think of DCCP as TCP
minus bytestream semantics and reliability, or as UDP plus congestion
control, handshakes, and acknowledgements.
4.1. Packet Types
Ten packet types implement DCCP's protocol functions. For example,
every new connection attempt begins with a DCCP-Request packet sent
by the client. In this way a DCCP-Request packet resembles a TCP
SYN, but since DCCP-Request is a packet type there is no way to send
an unexpected flag combination, such as TCP's SYN+FIN+ACK+RST.
Eight packet types occur during the progress of a typical connection,
shown here. Note the three-way handshakes during initiation and
(2) Data transfer
DCCP-Data, DCCP-Ack, DCCP-DataAck -->
<-- DCCP-Data, DCCP-Ack, DCCP-DataAck
The two remaining packet types are used to resynchronize after bursts
Every DCCP packet starts with a fixed-size generic header.
Particular packet types include additional fixed-size header data;
for example, DCCP-Acks include an Acknowledgement Number. DCCP
options and any application data follow the fixed-size header.
The packet types are as follows:
Sent by the client to initiate a connection (the first part of the
three-way initiation handshake).
Sent by the server in response to a DCCP-Request (the second part
of the three-way initiation handshake).
Used to transmit application data.
Used to transmit pure acknowledgements.
Used to transmit application data with piggybacked acknowledgement
Sent by the server to request that the client close the
Used by the client or the server to close the connection; elicits
a DCCP-Reset in response.
Used to terminate the connection, either normally or abnormally.
Used to resynchronize sequence numbers after large bursts of loss.
4.2. Packet Sequencing
Each DCCP packet carries a sequence number so that losses can be
detected and reported. Unlike TCP sequence numbers, which are byte-
based, DCCP sequence numbers increment by one per packet. For
DCCP A DCCP B
DCCP-Data(seqno 1) -->
DCCP-Data(seqno 2) -->
<-- DCCP-Ack(seqno 10, ackno 2)
DCCP-DataAck(seqno 3, ackno 10) -->
<-- DCCP-Data(seqno 11)
Every DCCP packet increments the sequence number, whether or not it
contains application data. DCCP-Ack pure acknowledgements increment
the sequence number; for instance, DCCP B's second packet above uses
sequence number 11, since sequence number 10 was used for an
acknowledgement. This lets endpoints detect all packet loss,
including acknowledgement loss. It also means that endpoints can get
out of sync after long bursts of loss. The DCCP-Sync and DCCP-
SyncAck packet types are used to recover (Section 7.5).
Since DCCP provides unreliable semantics, there are no
retransmissions, and having a TCP-style cumulative acknowledgement
field doesn't make sense. DCCP's Acknowledgement Number field equals
the greatest sequence number received, rather than the smallest
sequence number not received. Separate options indicate any
intermediate sequence numbers that weren't received.
DCCP endpoints progress through different states during the course of
a connection, corresponding roughly to the three phases of
initiation, data transfer, and termination. The figure below shows
the typical progress through these states for a client and server.
(0) No connection
REQUEST DCCP-Request -->
<-- DCCP-Response RESPOND
PARTOPEN DCCP-Ack or DCCP-DataAck -->
(2) Data transfer
OPEN <-- DCCP-Data, Ack, DataAck --> OPEN
<-- DCCP-CloseReq CLOSEREQ
CLOSING DCCP-Close -->
<-- DCCP-Reset CLOSED
The nine possible states are as follows. They are listed in
increasing order, so that "state >= CLOSEREQ" means the same as
"state = CLOSEREQ or state = CLOSING or state = TIMEWAIT". Section 8
describes the states in more detail.
Represents nonexistent connections.
Represents server sockets in the passive listening state. LISTEN
and CLOSED are not associated with any particular DCCP connection.
A client socket enters this state, from CLOSED, after sending a
DCCP-Request packet to try to initiate a connection.
A server socket enters this state, from LISTEN, after receiving a
DCCP-Request from a client.
A client socket enters this state, from REQUEST, after receiving a
DCCP-Response from the server. This state represents the third
phase of the three-way handshake. The client may send application
data in this state, but it MUST include an Acknowledgement Number
on all of its packets.
The central data transfer portion of a DCCP connection. Client
and server sockets enter this state from PARTOPEN and RESPOND,
respectively. Sometimes we speak of SERVER-OPEN and CLIENT-OPEN
states, corresponding to the server's OPEN state and the client's
A server socket enters this state, from SERVER-OPEN, to order the
client to close the connection and to hold TIMEWAIT state.
Server and client sockets can both enter this state to close the
A server or client socket remains in this state for 2MSL (4
minutes) after the connection has been torn down, to prevent
mistakes due to the delivery of old packets. Only one of the
endpoints has to enter TIMEWAIT state (the other can enter CLOSED
state immediately), and a server can request its client to hold
TIMEWAIT state using the DCCP-CloseReq packet type.
4.4. Congestion Control Mechanisms
DCCP connections are congestion controlled, but unlike in TCP, DCCP
applications have a choice of congestion control mechanism. In fact,
the two half-connections can be governed by different mechanisms.
Mechanisms are denoted by one-byte congestion control identifiers, or
CCIDs. The endpoints negotiate their CCIDs during connection
initiation. Each CCID describes how the HC-Sender limits data packet
rates, how the HC-Receiver sends congestion feedback via
acknowledgements, and so forth. CCIDs 2 and 3 are currently defined;
CCIDs 0, 1, and 4-255 are reserved. Other CCIDs may be defined in
CCID 2 provides TCP-like Congestion Control, which is similar to that
of TCP. The sender maintains a congestion window and sends packets
until that window is full. Packets are acknowledged by the receiver.
Dropped packets and ECN [RFC3168] indicate congestion; the response
to congestion is to halve the congestion window. Acknowledgements in
CCID 2 contain the sequence numbers of all received packets within
some window, similar to a selective acknowledgement (SACK) [RFC2018].
CCID 3 provides TCP-Friendly Rate Control (TFRC), an equation-based
form of congestion control intended to respond to congestion more
smoothly than CCID 2. The sender maintains a transmit rate, which it
updates using the receiver's estimate of the packet loss and mark
rate. CCID 3 behaves somewhat differently than TCP in the short
term, but is designed to operate fairly with TCP over the long term.
Section 10 describes DCCP's CCIDs in more detail. The behaviors of
CCIDs 2 and 3 are fully defined in separate profile documents
4.5. Feature Negotiation Options
DCCP endpoints use Change and Confirm options to negotiate and agree
on feature values. Feature negotiation will almost always happen on
the connection initiation handshake, but it can begin at any time.
There are four feature negotiation options in all: Change L, Confirm
L, Change R, and Confirm R. The "L" options are sent by the feature
location and the "R" options are sent by the feature remote. A
Change R option says to the feature location, "change this feature
value as follows". The feature location responds with Confirm L,
meaning, "I've changed it". Some features allow Change R options to
contain multiple values sorted in preference order. For example:
Change R(CCID, 2) -->
<-- Confirm L(CCID, 2)
* agreement that CCID/Server = 2 *
Change R(CCID, 3 4) -->
<-- Confirm L(CCID, 4, 4 2)
* agreement that CCID/Server = 4 *
Both exchanges negotiate the CCID/Server feature's value, which is
the CCID in use on the server-to-client half-connection. In the
second exchange, the client requests that the server use either CCID
3 or CCID 4, with 3 preferred; the server chooses 4 and supplies its
preference list, "4 2".
The Change L and Confirm R options are used for feature negotiations
initiated by the feature location. In the following example, the
server requests that CCID/Server be set to 3 or 2, with 3 preferred,
and the client agrees.
<-- Change L(CCID, 3 2)
Confirm R(CCID, 3, 3 2) -->
* agreement that CCID/Server = 3 *
Section 6 describes the feature negotiation options further,
including the retransmission strategies that make negotiation
4.6. Differences from TCP
DCCP's differences from TCP apart from those discussed so far include
o Copious space for options (up to 1008 bytes or the PMTU).
o Different acknowledgement formats. The CCID for a connection
determines how much acknowledgement information needs to be
transmitted. For example, in CCID 2 (TCP-like), this is about one
ack per 2 packets, and each ack must declare exactly which packets
were received. In CCID 3 (TFRC), it is about one ack per round-
trip time, and acks must declare at minimum just the lengths of
recent loss intervals.
o Denial of Service (DoS) protection. Several mechanisms help limit
the amount of state that possibly-misbehaving clients can force
DCCP servers to maintain. An Init Cookie option analogous to
TCP's SYN Cookies [SYNCOOKIES] avoids SYN-flood-like attacks.
Only one connection endpoint has to hold TIMEWAIT state; the
DCCP-CloseReq packet, which may only be sent by the server, passes
that state to the client. Various rate limits let servers avoid
attacks that might force extensive computation or packet
o Distinguishing different kinds of loss. A Data Dropped option
(Section 11.7) lets an endpoint declare that a packet was dropped
because of corruption, because of receive buffer overflow, and so
on. This facilitates research into more appropriate rate-control
responses for these non-network-congestion losses (although
currently such losses will cause a congestion response).
o Acknowledgeability. In TCP, a packet may be acknowledged only
once the data is reliably queued for application delivery. This
does not make sense in DCCP, where an application might, for
example, request a drop-from-front receive buffer. A DCCP packet
may be acknowledged as soon as its header has been successfully
processed. Concretely, a packet becomes acknowledgeable at Step 8
of Section 8.5's packet processing pseudocode. Acknowledgeability
does not guarantee data delivery, however: the Data Dropped option
may later report that the packet's application data was discarded.
o No receive window. DCCP is a congestion control protocol, not a
flow control protocol.
o No simultaneous open. Every connection has one client and one
o No half-closed states. DCCP has no states corresponding to TCP's
FINWAIT and CLOSEWAIT, where one half-connection is explicitly
closed while the other is still active. The Data Dropped option's
Drop Code 1, Application Not Listening (Section 11.7), can achieve
a similar effect, however.
4.7. Example Connection
The progress of a typical DCCP connection is as follows. (This
description is informative, not normative.)
0. [CLOSED] [LISTEN]
1. DCCP-Request -->
2. <-- DCCP-Response
3. DCCP-Ack -->
4. DCCP-Data, DCCP-Ack, DCCP-DataAck -->
<-- DCCP-Data, DCCP-Ack, DCCP-DataAck
5. <-- DCCP-CloseReq
6. DCCP-Close -->
7. <-- DCCP-Reset
1. The client sends the server a DCCP-Request packet specifying the
client and server ports, the service being requested, and any
features being negotiated, including the CCID that the client
would like the server to use. The client may optionally piggyback
an application request on the DCCP-Request packet. The server may
ignore this application request.
2. The server sends the client a DCCP-Response packet indicating that
it is willing to communicate with the client. This response
indicates any features and options that the server agrees to,
begins other feature negotiations as desired, and optionally
includes Init Cookies that wrap up all this information and that
must be returned by the client for the connection to complete.
3. The client sends the server a DCCP-Ack packet that acknowledges
the DCCP-Response packet. This acknowledges the server's initial
sequence number and returns any Init Cookies in the DCCP-Response.
It may also continue feature negotiation. The client may
piggyback an application-level request on this ack, producing a
4. The server and client then exchange DCCP-Data packets, DCCP-Ack
packets acknowledging that data, and, optionally, DCCP-DataAck
packets containing data with piggybacked acknowledgements. If the
client has no data to send, then the server will send DCCP-Data
and DCCP-DataAck packets, while the client will send DCCP-Acks
exclusively. (However, the client may not send DCCP-Data packets
before receiving at least one non-DCCP-Response packet from the
5. The server sends a DCCP-CloseReq packet requesting a close.
6. The client sends a DCCP-Close packet acknowledging the close.
7. The server sends a DCCP-Reset packet with Reset Code 1, "Closed",
and clears its connection state. DCCP-Resets are part of normal
connection termination; see Section 5.6.
8. The client receives the DCCP-Reset packet and holds state for two
maximum segment lifetimes, or 2MSL, to allow any remaining packets
to clear the network.
An alternative connection closedown sequence is initiated by the
5b. The client sends a DCCP-Close packet closing the connection.
6b. The server sends a DCCP-Reset packet with Reset Code 1, "Closed",
and clears its connection state.
7b. The client receives the DCCP-Reset packet and holds state for
2MSL to allow any remaining packets to clear the network.