2. Overview of ICE
In a typical ICE deployment, we have two endpoints (known as AGENTS in RFC 3264 terminology) that want to communicate. They are able to communicate indirectly via some signaling protocol (such as SIP), by which they can perform an offer/answer exchange of SDP [RFC3264] messages. Note that ICE is not intended for NAT traversal for SIP, which is assumed to be provided via another mechanism [RFC5626]. At the beginning of the ICE process, the agents are ignorant of their own topologies. In particular, they might or might not be behind a NAT (or multiple tiers of NATs). ICE allows the agents to discover enough information about their topologies to potentially find one or more paths by which they can communicate.
Figure 1 shows a typical environment for ICE deployment. The two endpoints are labelled L and R (for left and right, which helps visualize call flows). Both L and R are behind their own respective NATs though they may not be aware of it. The type of NAT and its properties are also unknown. Agents L and R are capable of engaging in an offer/answer exchange by which they can exchange SDP messages, whose purpose is to set up a media session between L and R. Typically, this exchange will occur through a SIP server.
In addition to the agents, a SIP server and NATs, ICE is typically used in concert with STUN or TURN servers in the network. Each agent can have its own STUN or TURN server, or they can be the same.
+-------+
| SIP |
+-------+ | Srvr | +-------+
| STUN | | | | STUN |
| Srvr | +-------+ | Srvr |
| | / \ | |
+-------+ / \ +-------+
/ \
/ \
/ \
/ \
/ <- Signaling -> \
/ \
/ \
+--------+ +--------+
| NAT | | NAT |
+--------+ +--------+
/ \
/ \
/ \
+-------+ +-------+
| Agent | | Agent |
| L | | R |
| | | |
+-------+ +-------+
Figure 1: ICE Deployment Scenario
The basic idea behind ICE is as follows: each agent has a variety of candidate TRANSPORT ADDRESSES (combination of IP address and port for a particular transport protocol, which is always UDP in this specification)) it could use to communicate with the other agent. These might include:
- A transport address on a directly attached network interface
- A translated transport address on the public side of a NAT (a "server reflexive" address)
- A transport address allocated from a TURN server (a "relayed address").
Potentially, any of L's candidate transport addresses can be used to communicate with any of R's candidate transport addresses. In practice, however, many combinations will not work. For instance, if L and R are both behind NATs, their directly attached interface addresses are unlikely to be able to communicate directly (this is why ICE is needed, after all!). The purpose of ICE is to discover which pairs of addresses will work. The way that ICE does this is to systematically try all possible pairs (in a carefully sorted order) until it finds one or more that work.
2.1. Gathering Candidate Addresses
In order to execute ICE, an agent has to identify all of its address candidates. A CANDIDATE is a transport address -- a combination of IP address and port for a particular transport protocol (with only UDP specified here). This document defines three types of candidates, some derived from physical or logical network interfaces, others discoverable via STUN and TURN. Naturally, one viable candidate is a transport address obtained directly from a local interface. Such a candidate is called a HOST CANDIDATE. The local interface could be ethernet or WiFi, or it could be one that is obtained through a tunnel mechanism, such as a Virtual Private Network (VPN) or Mobile IP (MIP). In all cases, such a network interface appears to the agent as a local interface from which ports (and thus candidates) can be allocated.
If an agent is multihomed, it obtains a candidate from each IP address. Depending on the location of the PEER (the other agent in the session) on the IP network relative to the agent, the agent may be reachable by the peer through one or more of those IP addresses. Consider, for example, an agent that has a local IP address on a private net 10 network (I1), and a second connected to the public Internet (I2). A candidate from I1 will be directly reachable when communicating with a peer on the same private net 10 network, while a candidate from I2 will be directly reachable when communicating with a peer on the public Internet. Rather than trying to guess which IP address will work prior to sending an offer, the offering agent includes both candidates in its offer.
Next, the agent uses STUN or TURN to obtain additional candidates. These come in two flavors: translated addresses on the public side of a NAT (SERVER REFLEXIVE CANDIDATES) and addresses on TURN servers (RELAYED CANDIDATES). When TURN servers are utilized, both types of candidates are obtained from the TURN server. If only STUN servers are utilized, only server reflexive candidates are obtained from them. The relationship of these candidates to the host candidate is shown in Figure 2. In this figure, both types of candidates are discovered using TURN. In the figure, the notation X:x means IP address X and UDP port x.
To Internet
|
|
| /------------ Relayed
Y:y | / Address
+--------+
| |
| TURN |
| Server |
| |
+--------+
|
|
| /------------ Server
X1':x1'|/ Reflexive
+------------+ Address
| NAT |
+------------+
|
| /------------ Local
X:x |/ Address
+--------+
| |
| Agent |
| |
+--------+
Figure 2: Candidate Relationships
When the agent sends the TURN Allocate request from IP address and port X:x, the NAT (assuming there is one) will create a binding X1':x1', mapping this server reflexive candidate to the host candidate X:x. Outgoing packets sent from the host candidate will be translated by the NAT to the server reflexive candidate. Incoming packets sent to the server reflexive candidate will be translated by the NAT to the host candidate and forwarded to the agent. We call the host candidate associated with a given server reflexive candidate the BASE.
Note: "Base" refers to the address an agent sends from for a particular candidate. Thus, as a degenerate case host candidates also have a base, but it's the same as the host candidate.
When there are multiple NATs between the agent and the TURN server, the TURN request will create a binding on each NAT, but only the outermost server reflexive candidate (the one nearest the TURN server) will be discovered by the agent. If the agent is not behind a NAT, then the base candidate will be the same as the server reflexive candidate and the server reflexive candidate is redundant and will be eliminated.
The Allocate request then arrives at the TURN server. The TURN server allocates a port y from its local IP address Y, and generates an Allocate response, informing the agent of this relayed candidate. The TURN server also informs the agent of the server reflexive candidate, X1':x1' by copying the source transport address of the Allocate request into the Allocate response. The TURN server acts as a packet relay, forwarding traffic between L and R. In order to send traffic to L, R sends traffic to the TURN server at Y:y, and the TURN server forwards that to X1':x1', which passes through the NAT where it is mapped to X:x and delivered to L.
When only STUN servers are utilized, the agent sends a STUN Binding request [RFC5389] to its STUN server. The STUN server will inform the agent of the server reflexive candidate X1':x1' by copying the source transport address of the Binding request into the Binding response.
2.2. Connectivity Checks
Once L has gathered all of its candidates, it orders them in highest to lowest priority and sends them to R over the signaling channel. The candidates are carried in attributes in the SDP offer. When R receives the offer, it performs the same gathering process and responds with its own list of candidates. At the end of this process, each agent has a complete list of both its candidates and its peer's candidates. It pairs them up, resulting in CANDIDATE PAIRS. To see which pairs work, each agent schedules a series of CHECKS. Each check is a STUN request/response transaction that the client will perform on a particular candidate pair by sending a STUN request from the local candidate to the remote candidate.
The basic principle of the connectivity checks is simple:
- Sort the candidate pairs in priority order.
- Send checks on each candidate pair in priority order.
- Acknowledge checks received from the other agent.
With both agents performing a check on a candidate pair, the result is a 4-way handshake:
L R
- -
STUN request -> \ L's
<- STUN response / check
<- STUN request \ R's
STUN response -> / check
Figure 3: Basic Connectivity Check
It is important to note that the STUN requests are sent to and from the exact same IP addresses and ports that will be used for media (e.g., RTP and RTCP). Consequently, agents demultiplex STUN and RTP/RTCP using contents of the packets, rather than the port on which they are received. Fortunately, this demultiplexing is easy to do, especially for RTP and RTCP.
Because a STUN Binding request is used for the connectivity check, the STUN Binding response will contain the agent's translated transport address on the public side of any NATs between the agent and its peer. If this transport address is different from other candidates the agent already learned, it represents a new candidate, called a PEER REFLEXIVE CANDIDATE, which then gets tested by ICE just the same as any other candidate.
As an optimization, as soon as R gets L's check message, R schedules a connectivity check message to be sent to L on the same candidate pair. This accelerates the process of finding a valid candidate, and is called a TRIGGERED CHECK.
At the end of this handshake, both L and R know that they can send (and receive) messages end-to-end in both directions.
2.3. Sorting Candidates
Because the algorithm above searches all candidate pairs, if a working pair exists it will eventually find it no matter what order the candidates are tried in. In order to produce faster (and better) results, the candidates are sorted in a specified order. The resulting list of sorted candidate pairs is called the CHECK LIST. The algorithm is described in Section 4.1.2 but follows two general principles:
- Each agent gives its candidates a numeric priority, which is sent along with the candidate to the peer.
- The local and remote priorities are combined so that each agent has the same ordering for the candidate pairs.
The second property is important for getting ICE to work when there are NATs in front of L and R. Frequently, NATs will not allow packets in from a host until the agent behind the NAT has sent a packet towards that host. Consequently, ICE checks in each direction will not succeed until both sides have sent a check through their respective NATs.
The agent works through this check list by sending a STUN request for the next candidate pair on the list periodically. These are called ORDINARY CHECKS.
In general, the priority algorithm is designed so that candidates of similar type get similar priorities and so that more direct routes (that is, through fewer media relays and through fewer NATs) are preferred over indirect ones (ones with more media relays and more NATs). Within those guidelines, however, agents have a fair amount of discretion about how to tune their algorithms.
2.4. Frozen Candidates
The previous description only addresses the case where the agents wish to establish a media session with one COMPONENT (a piece of a media stream requiring a single transport address; a media stream may require multiple components, each of which has to work for the media stream as a whole to be work). Typically (e.g., with RTP and RTCP), the agents actually need to establish connectivity for more than one flow.
The network properties are likely to be very similar for each component (especially because RTP and RTCP are sent and received from the same IP address). It is usually possible to leverage information from one media component in order to determine the best candidates for another. ICE does this with a mechanism called "frozen candidates".
Each candidate is associated with a property called its FOUNDATION. Two candidates have the same foundation when they are "similar" -- of the same type and obtained from the same host candidate and STUN server using the same protocol. Otherwise, their foundation is different. A candidate pair has a foundation too, which is just the concatenation of the foundations of its two candidates. Initially, only the candidate pairs with unique foundations are tested. The other candidate pairs are marked "frozen". When the connectivity checks for a candidate pair succeed, the other candidate pairs with the same foundation are unfrozen. This avoids repeated checking of components that are superficially more attractive but in fact are likely to fail.
While we've described "frozen" here as a separate mechanism for expository purposes, in fact it is an integral part of ICE and the ICE prioritization algorithm automatically ensures that the right candidates are unfrozen and checked in the right order.
2.5. Security for Checks
Because ICE is used to discover which addresses can be used to send media between two agents, it is important to ensure that the process cannot be hijacked to send media to the wrong location. Each STUN connectivity check is covered by a message authentication code (MAC) computed using a key exchanged in the signaling channel. This MAC provides message integrity and data origin authentication, thus stopping an attacker from forging or modifying connectivity check messages. Furthermore, if the SIP [RFC3261] caller is using ICE, and their call forks, the ICE exchanges happen independently with each forked recipient. In such a case, the keys exchanged in the signaling help associate each ICE exchange with each forked recipient.
2.6. Concluding ICE
ICE checks are performed in a specific sequence, so that high-priority candidate pairs are checked first, followed by lower-priority ones. One way to conclude ICE is to declare victory as soon as a check for each component of each media stream completes successfully. Indeed, this is a reasonable algorithm, and details for it are provided below. However, it is possible that a packet loss will cause a higher-priority check to take longer to complete. In that case, allowing ICE to run a little longer might produce better results. More fundamentally, however, the prioritization defined by this specification may not yield "optimal" results. As an example, if the aim is to select low-latency media paths, usage of a relay is a hint that latencies may be higher, but it is nothing more than a hint. An actual round-trip time (RTT) measurement could be made, and it might demonstrate that a pair with lower priority is actually better than one with higher priority.
Consequently, ICE assigns one of the agents in the role of the CONTROLLING AGENT, and the other of the CONTROLLED AGENT. The controlling agent gets to nominate which candidate pairs will get used for media amongst the ones that are valid. It can do this in one of two ways -- using REGULAR NOMINATION or AGGRESSIVE NOMINATION.
With regular nomination, the controlling agent lets the checks continue until at least one valid candidate pair for each media stream is found. Then, it picks amongst those that are valid, and sends a second STUN request on its NOMINATED candidate pair, but this time with a flag set to tell the peer that this pair has been nominated for use. This is shown in Figure 4.
L R
- -
STUN request -> \ L's
<- STUN response / check
<- STUN request \ R's
STUN response -> / check
STUN request + flag -> \ L's
<- STUN response / check
Figure 4: Regular Nomination
Once the STUN transaction with the flag completes, both sides cancel any future checks for that media stream. ICE will now send media using this pair. The pair an ICE agent is using for media is called the SELECTED PAIR.
In aggressive nomination, the controlling agent puts the flag in every STUN request it sends. This way, once the first check succeeds, ICE processing is complete for that media stream and the controlling agent doesn't have to send a second STUN request. The selected pair will be the highest-priority valid pair whose check succeeded. Aggressive nomination is faster than regular nomination, but gives less flexibility. Aggressive nomination is shown in Figure 5.
L R
- -
STUN request + flag -> \ L's
<- STUN response / check
<- STUN request \ R's
STUN response -> / check
Figure 5: Aggressive Nomination
Once all of the media streams are completed, the controlling endpoint sends an updated offer if the candidates in the m and c lines for the media stream (called the DEFAULT CANDIDATES) don't match ICE's SELECTED CANDIDATES.
Once ICE is concluded, it can be restarted at any time for one or all of the media streams by either agent. This is done by sending an updated offer indicating a restart.
2.7. Lite Implementations
In order for ICE to be used in a call, both agents need to support it. However, certain agents will always be connected to the public Internet and have a public IP address at which it can receive packets from any correspondent. To make it easier for these devices to support ICE, ICE defines a special type of implementation called LITE (in contrast to the normal FULL implementation). A lite implementation doesn't gather candidates; it includes only host candidates for any media stream. Lite agents do not generate connectivity checks or run the state machines, though they need to be able to respond to connectivity checks. When a lite implementation connects with a full implementation, the full agent takes the role of the controlling agent, and the lite agent takes on the controlled role. When two lite implementations connect, no checks are sent.
For guidance on when a lite implementation is appropriate, see the discussion in Appendix A.
It is important to note that the lite implementation was added to this specification to provide a stepping stone to full implementation. Even for devices that are always connected to the public Internet, a full implementation is preferable if achievable.