RFC 5104
Codec Control Messages in the RTP Audio-Visual Profile with Feedback (AVPF)
3.5. Feedback Messages
This section describes the semantics of the different feedback messages and how they apply to the different use cases.3.5.1. Full Intra Request Command
A Full Intra Request (FIR) Command, when received by the designated media sender, requires that the media sender sends a Decoder Refresh Point (see section 2.2) at the earliest opportunity. The evaluation of such an opportunity includes the current encoder coding strategy and the current available network resources. FIR is also known as an "instantaneous decoder refresh request", "fast video update request" or "video fast update request". Using a decoder refresh point implies refraining from using any picture sent prior to that point as a reference for the encoding process of any subsequent picture sent in the stream. For predictive media types that are not video, the analogue applies. For example, if in MPEG-4 systems scene updates are used, the decoder refresh point consists of the full representation of the scene and is not delta-coded relative to previous updates.
Decoder refresh points, especially Intra or IDR pictures, are in general several times larger in size than predicted pictures. Thus, in scenarios in which the available bit rate is small, the use of a decoder refresh point implies a delay that is significantly longer than the typical picture duration. Usage in multicast is possible; however, aggregation of the commands is recommended. A receiver that receives a request closely after sending a decoder refresh point -- within 2 times the longest round trip time (RTT) known, plus any AVPF-induced RTCP packet sending delays -- should await a second request message to ensure that the media receiver has not been served by the previously delivered decoder refresh point. The reason for the specified delay is to avoid sending unnecessary decoder refresh points. A session participant may have sent its own request while another participant's request was in-flight to them. Suppressing those requests that may have been sent without knowledge about the other request avoids this issue. Using the FIR command to recover from errors is explicitly disallowed, and instead the PLI message defined in AVPF [RFC4585] should be used. The PLI message reports lost pictures and has been included in AVPF for precisely that purpose. Full Intra Request is applicable in use-cases 1 and 2.3.5.1.1. Reliability
The FIR message results in the delivery of a decoder refresh point, unless the message is lost. Decoder refresh points are easily identifiable from the bit stream. Therefore, there is no need for protocol-level notification, and a simple command repetition mechanism is sufficient for ensuring the level of reliability required. However, the potential use of repetition does require a mechanism to prevent the recipient from responding to messages already received and responded to. To ensure the best possible reliability, a sender of FIR may repeat the FIR until the desired content has been received. The repetition interval is determined by the RTCP timing rules applicable to the session. Upon reception of a complete decoder refresh point or the detection of an attempt to send a decoder refresh point (which got damaged due to a packet loss), the repetition of the FIR must stop. If another FIR is necessary, the request sequence number must be increased. A FIR sender shall not have more than one FIR (different request sequence number) outstanding at any time per media sender in the session.
The receiver of FIR (i.e., the media sender) behaves in complementary fashion to ensure delivery of a decoder refresh point. If it receives repetitions of the FIR more than 2*RTT after it has sent a decoder refresh point, it shall send a new decoder refresh point. Two round trip times allow time for the decoder refresh point to arrive back to the requestor and for the end of repetitions of FIR to reach and be detected by the media sender. An RTP mixer or RTP switching MCU that receive a FIR from a media receiver is responsible to ensure that a decoder refresh point is delivered to the requesting receiver. It may be necessary for the mixer/MCU to generate FIR commands. From a reliability perspective, the two legs (FIR-requesting endpoint to mixer/MCU, and mixer/MCU to decoder refresh point generating endpoint) are handled independently from each other.3.5.2. Temporal-Spatial Trade-off Request and Notification
The Temporal-Spatial Trade-off Request (TSTR) instructs the video encoder to change its trade-off between temporal and spatial resolution. Index values from 0 to 31 indicate monotonically a desire for higher frame rate. That is, a requester asking for an index of 0 prefers a high quality and is willing to accept a low frame rate, whereas a requester asking for 31 wishes a high frame rate, potentially at the cost of low spatial quality. In general, the encoder reaction time may be significantly longer than the typical picture duration. See use case 3 for an example. The encoder decides whether and to what extent the request results in a change of the trade-off. It returns a Temporal-Spatial Trade-off Notification (TSTN) message to indicate the trade-off that it will use henceforth. TSTR and TSTN have been introduced primarily because it is believed that control protocol mechanisms, e.g., a SIP re-invite, are too heavyweight and too slow to allow for a reasonable user experience. Consider, for example, a user interface where the remote user selects the temporal/spatial trade-off with a slider. An immediate feedback to any slider movement is required for a reasonable user experience. A SIP re-INVITE [RFC3261] would require at least two round-trips more (compared to the TSTR/TSTN mechanism) and may involve proxies and other complex mechanisms. Even in a well-designed system, it could take a second or so until the new trade-off is finally selected. Furthermore, the use of RTCP solves the multicast use case very efficiently. The use of TSTR and TSTN in multipoint scenarios is a non-trivial subject, and can be achieved in many implementation-specific ways.
Problems stem from the fact that TSTRs will typically arrive unsynchronized, and may request different trade-off values for the same stream and/or endpoint encoder. This memo does not specify a translator's, mixer's, or endpoint's reaction to the reception of a suggested trade-off as conveyed in the TSTR. We only require the receiver of a TSTR message to reply to it by sending a TSTN, carrying the new trade-off chosen by its own criteria (which may or may not be based on the trade-off conveyed by the TSTR). In other words, the trade-off sent in a TSTR is a non-binding recommendation, nothing more. Three TSTR/TSTN scenarios need to be distinguished, based on the topologies described in [RFC5117]. The scenarios are described in the following subsections.3.5.2.1. Point-to-Point
In this most trivial case (Topo-Point-to-Point), the media sender typically adjusts its temporal/spatial trade-off based on the requested value in TSTR, subject to its own capabilities. The TSTN message conveys back the new trade-off value (which may be identical to the old one if, for example, the sender is not capable of adjusting its trade-off).3.5.2.2. Point-to-Multipoint Using Multicast or Translators
RTCP Multicast is used either with media multicast according to Topo-Multicast, or following RFC 3550's translator model according to Topo-Translator. In these cases, unsynchronized TSTR messages from different receivers may be received, possibly with different requested trade-offs (because of different user preferences). This memo does not specify how the media sender tunes its trade-off. Possible strategies include selecting the mean or median of all trade-off requests received, giving priority to certain participants, or continuing to use the previously selected trade-off (e.g., when the sender is not capable of adjusting it). Again, all TSTR messages need to be acknowledged by TSTN, and the value conveyed back has to reflect the decision made.3.5.2.3. Point-to-Multipoint Using RTP Mixer
In this scenario (Topo-Mixer), the RTP mixer receives all TSTR messages, and has the opportunity to act on them based on its own criteria. In most cases, the mixer should form a "consensus" of potentially conflicting TSTR messages arriving from different participants, and initiate its own TSTR message(s) to the media sender(s). As in the previous scenario, the strategy for forming
this "consensus" is up to the implementation, and can, for example, encompass averaging the participants' request values, giving priority to certain participants, or using session default values. Even if a mixer or translator performs transcoding, it is very difficult to deliver media with the requested trade-off, unless the content the mixer or translator receives is already close to that trade-off. Thus, if the mixer changes its trade-off, it needs to request the media sender(s) to use the new value, by creating a TSTR of its own. Upon reaching a decision on the used trade-off, it includes that value in the acknowledgement to the downstream requestors. Only in cases where the original source has substantially higher quality (and bit rate) is it likely that transcoding alone can result in the requested trade-off.3.5.2.4. Reliability
A request and reception acknowledgement mechanism is specified. The Temporal-Spatial Trade-off Notification (TSTN) message informs the requester that its request has been received, and what trade-off is used henceforth. This acknowledgement mechanism is desirable for at least the following reasons: o A change in the trade-off cannot be directly identified from the media bit stream. o User feedback cannot be implemented without knowing the chosen trade-off value, according to the media sender's constraints. o Repetitive sending of messages requesting an unimplementable trade-off can be avoided.3.5.3. H.271 Video Back Channel Message
ITU-T Rec. H.271 defines syntax, semantics, and suggested encoder reaction to a Video Back Channel Message. The structure defined in this memo is used to transparently convey such a message from media receiver to media sender. In this memo, we refrain from an in-depth discussion of the available code points within H.271 and refer to the specification text [H.271] instead. However, we note that some H.271 messages bear similarities with native messages of AVPF and this memo. Furthermore, we note that some H.271 message are known to require caution in multicast environments -- or are plainly not usable in multicast or multipoint scenarios. Table 1 provides a brief, simplified overview of the messages currently defined in H.271, their roughly corresponding AVPF or Codec Control Messages (CCMs) (the latter as specified in this memo), and an indication of our current knowledge of their multicast safety.
H.271 msg type AVPF/CCM msg type multicast-safe -------------------------------------------------------------------- 0 (when used for reference picture selection) AVPF RPSI No (positive ACK of pictures) 1 picture loss AVPF PLI Yes 2 partial loss AVPF SLI Yes 3 one parameter CRC N/A Yes (no required sender action) 4 all parameter CRC N/A Yes (no required sender action) 5 refresh point CCM FIR Yes Table 1: H.271 messages and their AVPF/CCM equivalents Note: H.271 message type 0 is not a strict equivalent to AVPF's Reference Picture Selection Indication (RPSI); it is an indication of known-as-correct reference picture(s) at the decoder. It does not command an encoder to use a defined reference picture (the form of control information envisioned to be carried in RPSI). However, it is believed and intended that H.271 message type 0 will be used for the same purpose as AVPF's RPSI -- although other use forms are also possible. In response to the opaqueness of the H.271 messages, especially with respect to the multicast safety, the following guidelines MUST be followed when an implementation wishes to employ the H.271 video back channel message: 1. Implementations utilizing the H.271 feedback message MUST stay in compliance with congestion control principles, as outlined in section 5. 2. An implementation SHOULD utilize the IETF-native messages as defined in [RFC4585] and in this memo instead of similar messages defined in [H.271]. Our current understanding of similar messages is documented in Table 1 above. One good reason to divert from the SHOULD statement above would be if it is clearly understood that, for a given application and video compression standard, the aforementioned "similarity" is not given, in contrast to what the table indicates. 3. It has been observed that some of the H.271 code points currently in existence are not multicast-safe. Therefore, the sensible thing to do is not to use the H.271 feedback message type in multicast environments. It MAY be used only when all the issues mentioned later are fully understood by the implementer, and properly taken into account by all endpoints. In all other cases, the H.271 message type MUST NOT be used in conjunction with multicast.
4. It has been observed that even in centralized multipoint
environments, where the mixer should theoretically be able to
resolve issues as documented below, the implementation of such a
mixer and cooperative endpoints is a very difficult and tedious
task. Therefore, H.271 messages MUST NOT be used in centralized
multipoint scenarios, unless all the issues mentioned below are
fully understood by the implementer, and properly taken into
account by both mixer and endpoints.
Issues to be taken into account when considering the use of H.271 in
multipoint environments:
1. Different state on different receivers. In many environments, it
cannot be guaranteed that the decoder state of all media receivers
is identical at any given point in time. The most obvious reason
for such a possible misalignment of state is a loss that occurs on
the path to only one of many media receivers. However, there are
other not so obvious reasons, such as recent joins to the
multipoint conference (be it by joining the multicast group or
through additional mixer output). Different states can lead the
media receivers to issue potentially contradicting H.271 messages
(or one media receiver issuing an H.271 message that, when
observed by the media sender, is not helpful for the other media
receivers). A naive reaction of the media sender to these
contradicting messages can lead to unpredictable and annoying
results.
2. Combining messages from different media receivers in a media
sender is a non-trivial task. As reasons, we note that these
messages may be contradicting each other, and that their transport
is unreliable (there may well be other reasons). In case of many
H.271 messages (i.e., types 0, 2, 3, and 4), the algorithm for
combining must be aware both of the network/protocol environment
(i.e., with respect to congestion) and of the media codec
employed, as H.271 messages of a given type can have different
semantics for different media codecs.
3. The suppression of requests may need to go beyond the basic
mechanisms described in AVPF (which are driven exclusively by
timing and transport considerations on the protocol level). For
example, a receiver is often required to refrain from (or delay)
generating requests, based on information it receives from the
media stream. For instance, it makes no sense for a receiver to
issue a FIR when a transmission of an Intra/IDR picture is
ongoing.
4. When using the non-multicast-safe messages (e.g., H.271 type 0
positive ACK of received pictures/slices) in larger multicast
groups, the media receiver will likely be forced to delay or even
omit sending these messages. For the media sender, this looks
like data has not been properly received (although it was received
properly), and a naively implemented media sender reacts to these
perceived problems where it should not.
3.5.3.1. Reliability
H.271 Video Back Channel Messages do not require reliable
transmission, and confirmation of the reception of a message can be
derived from the forward video bit stream. Therefore, no specific
reception acknowledgement is specified.
With respect to re-sending rules, section 3.5.1.1 applies.
3.5.4. Temporary Maximum Media Stream Bit Rate Request and Notification
A receiver, translator, or mixer uses the Temporary Maximum Media
Stream Bit Rate Request (TMMBR, "timber") to request a sender to
limit the maximum bit rate for a media stream (see section 2.2) to,
or below, the provided value. The Temporary Maximum Media Stream Bit
Rate Notification (TMMBN) contains the media sender's current view of
the most limiting subset of the TMMBR-defined limits it has received,
to help the participants to suppress TMMBRs that would not further
restrict the media sender. The primary usage for the TMMBR/TMMBN
messages is in a scenario with an MCU or mixer (use case 6),
corresponding to Topo-Translator or Topo-Mixer, but also to Topo-
Point-to-Point.
Each temporary limitation on the media stream is expressed as a
tuple. The first component of the tuple is the maximum total media
bit rate (as defined in section 2.2) that the media receiver is
currently prepared to accept for this media stream. The second
component is the per-packet overhead that the media receiver has
observed for this media stream at its chosen reference protocol
layer.
As indicated in section 2.2, the overhead as observed by the sender
of the TMMBR (i.e., the media receiver) may differ from the overhead
observed at the receiver of the TMMBR (i.e., the media sender) due to
use of a different reference protocol layer at the other end or due
to the intervention of translators or mixers that affect the amount
of per packet overhead. For example, a gateway in between the two
that converts between IPv4 and IPv6 affects the per-packet overhead
by 20 bytes. Other mechanisms that change the overhead include
tunnels. The problem with varying overhead is also discussed in
[RFC3890]. As will be seen in the description of the algorithm for use of TMMBR, the difference in perceived overhead between the sending and receiving ends presents no difficulty because calculations are carried out in terms of variables that have the same value at the sender as at the receiver -- for example, packet rate and net media rate. Reporting both maximum total media bit rate and per-packet overhead allows different receivers to provide bit rate and overhead values for different protocol layers, for example, at the IP level, at the outer part of a tunnel protocol, or at the link layer. The protocol level a peer reports on depends on the level of integration the peer has, as it needs to be able to extract the information from that protocol level. For example, an application with no knowledge of the IP version it is running over cannot meaningfully determine the overhead of the IP header, and hence will not want to include IP overhead in the overhead or maximum total media bit rate calculation. It is expected that most peers will be able to report values at least for the IP layer. In certain implementations, it may be advantageous to also include information pertaining to the link layer, which in turn allows for a more precise overhead calculation and a better optimization of connectivity resources. The Temporary Maximum Media Stream Bit Rate messages are generic messages that can be applied to any RTP packet stream. This separates them from the other codec control messages defined in this specification, which apply only to specific media types or payload formats. The TMMBR functionality applies to the transport, and the requirements the transport places on the media encoding. The reasoning below assumes that the participants have negotiated a session maximum bit rate, using a signaling protocol. This value can be global, for example, in case of point-to-point, multicast, or translators. It may also be local between the participant and the peer or mixer. In either case, the bit rate negotiated in signaling is the one that the participant guarantees to be able to handle (depacketize and decode). In practice, the connectivity of the participant also influences the negotiated value -- it does not make much sense to negotiate a total media bit rate that one's network interface does not support. It is also beneficial to have negotiated a maximum packet rate for the session or sender. RFC 3890 provides an SDP [RFC4566] attribute that can be used for this purpose; however, that attribute is not usable in RTP sessions established using offer/answer [RFC3264]. Therefore, an optional maximum packet rate signaling parameter is specified in this memo.
An already established maximum total media bit rate may be changed at any time, subject to the timing rules governing the sending of feedback messages. The limit may change to any value between zero and the session maximum, as negotiated during session establishment signaling. However, even if a sender has received a TMMBR message allowing an increase in the bit rate, all increases must be governed by a congestion control mechanism. TMMBR indicates known limitations only, usually in the local environment, and does not provide any guarantees about the full path. Furthermore, any increases in TMMBR-established bit rate limits are to be executed only after a certain delay from the sending of the TMMBN message that notifies the world about the increase in limit. The delay is specified as at least twice the longest RTT as known by the media sender, plus the media sender's calculation of the required wait time for the sending of another TMMBR message for this session based on AVPF timing rules. This delay is introduced to allow other session participants to make known their bit rate limit requirements, which may be lower. If it is likely that the new value indicated by TMMBR will be valid for the remainder of the session, the TMMBR sender is expected to perform a renegotiation of the session upper limit using the session signaling protocol.3.5.4.1. Behavior for Media Receivers Using TMMBR
This section is an informal description of behaviour described more precisely in section 4.2. A media sender begins the session limited by the maximum media bit rate and maximum packet rate negotiated in session signaling, if any. Note that this value may be negotiated for another protocol layer than the one the participant uses in its TMMBR messages. Each media receiver selects a reference protocol layer, forms an estimate of the overhead it is observing (or estimating it if no packets has been seen yet) at that reference level, and determines the maximum total media bit rate it can accept, taking into account its own limitations and any transport path limitations of which it may be aware. In case the current limitations are more restricting than what was agreed on in the session signaling, the media receiver reports its initial estimate of these two quantities to the media sender using a TMMBR message. Overall message traffic is reduced by the possibility of including tuples for multiple media senders in the same TMMBR message. The media sender applies an algorithm such as that specified in section 3.5.4.2 to select which of the tuples it has received are most limiting (i.e., the bounding set as defined in section 2.2). It modifies its operation to stay within the feasible region (as defined
in section 2.2), and also sends out a TMMBN to the media receivers indicating the selected bounding set. That notification also indicates who was responsible for the tuples in the bounding set, i.e., the "owner"(s) of the limitation. A session participant that owns no tuple in the bounding set is called a "non-owner". If a media receiver does not own one of the tuples in the bounding set reported by the TMMBN, it applies the same algorithm as the media sender to determine if its current estimated (maximum total media bit rate, overhead) tuple would enter the bounding set if known to the media sender. If so, it issues a TMMBR reporting the tuple value to the sender. Otherwise, it takes no action for the moment. Periodically, its estimated tuple values may change or it may receive a new TMMBN. If so, it reapplies the algorithm to decide whether it needs to issue a TMMBR. If, alternatively, a media receiver owns one of the tuples in the reported bounding set, it takes no action until such time as its estimate of its own tuple values changes. At that time, it sends a TMMBR to the media sender to report the changed values. A media receiver may change status between owner and non-owner of a bounding tuple between one TMMBN message and the next. Thus, it must check the contents of each TMMBN to determine its subsequent actions. Implementations may use other algorithms of their choosing, as long as the bit rate limitations resulting from the exchange of TMMBR and TMMBN messages are at least as strict (at least as low, in the bit rate dimension) as the ones resulting from the use of the aforementioned algorithm. Obviously, in point-to-point cases, when there is only one media receiver, this receiver becomes "owner" once it receives the first TMMBN in response to its own TMMBR, and stays "owner" for the rest of the session. Therefore, when it is known that there will always be only a single media receiver, the above algorithm is not required. Media receivers that are aware they are the only ones in a session can send TMMBR messages with bit rate limits both higher and lower than the previously notified limit, at any time (subject to the AVPF [RFC4585] RTCP RR send timing rules). However, it may be difficult for a session participant to determine if it is the only receiver in the session. Because of this, any implementation of TMMBR is required to include the algorithm described in the next section or a stricter equivalent.
3.5.4.2. Algorithm for Establishing Current Limitations
This section introduces an example algorithm for the calculation of a session limit. Other algorithms can be employed, as long as the result of the calculation is at least as restrictive as the result that is obtained by this algorithm. First, it is important to consider the implications of using a tuple for limiting the media sender's behavior. The bit rate and the overhead value result in a two-dimensional solution space for the calculation of the bit rate of media streams. Fortunately, the two variables are linked. Specifically, the bit rate available for RTP payloads is equal to the TMMBR reported bit rate minus the packet rate used, multiplied by the TMMBR reported overhead converted to bits. As a result, when different bit rate/overhead combinations need to be considered, the packet rate determines the correct limitation. This is perhaps best explained by an example: Example: Receiver A: TMMBR_max total BR = 35 kbps, TMMBR_OH = 40 bytes Receiver B: TMMBR_max total BR = 40 kbps, TMMBR_OH = 60 bytes For a given packet rate (PR), the bit rate available for media payloads in RTP will be: Max_net media_BR_A = TMMBR_max total BR_A - PR * TMMBR_OH_A * 8 ... (1) Max_net media_BR_B = TMMBR_max total BR_B - PR * TMMBR_OH_B * 8 ... (2) For a PR = 20, these calculations will yield a Max_net media_BR_A = 28600 bps and Max_net media_BR_B = 30400 bps, which suggests that receiver A is the limiting one for this packet rate. However, at a certain PR there is a switchover point at which receiver B becomes the limiting one. The switchover point can be identified by setting Max_media_BR_A equal to Max_media_BR_B and breaking out PR: TMMBR_max total BR_A - TMMBR_max total BR_B PR = ------------------------------------------- ... (3) 8*(TMMBR_OH_A - TMMBR_OH_B) which, for the numbers above, yields 31.25 as the switchover point between the two limits. That is, for packet rates below 31.25 per second, receiver A is the limiting receiver, and for higher packet rates, receiver B is more limiting. The implications of this behavior have to be considered by implementations that are going to
control media encoding and its packetization. As exemplified above, multiple TMMBR limits may apply to the trade-off between net media bit rate and packet rate. Which limitation applies depends on the packet rate being considered. This also has implications for how the TMMBR mechanism needs to work. First, there is the possibility that multiple TMMBR tuples are providing limitations on the media sender. Secondly, there is a need for any session participant (media sender and receivers) to be able to determine if a given tuple will become a limitation upon the media sender, or if the set of already given limitations is stricter than the given values. In the absence of the ability to make this determination, the suppression of TMMBRs would not work. The basic idea of the algorithm is as follows. Each TMMBR tuple can be viewed as the equation of a straight line (cf. equations (1) and (2)) in a space where packet rate lies along the X-axis and net bit rate along the Y-axis. The lower envelope of the set of lines corresponding to the complete set of TMMBR tuples, together with the X and Y axes, defines a polygon. Points lying within this polygon are combinations of packet rate and bit rate that meet all of the TMMBR constraints. The highest feasible packet rate within this region is the minimum of the rate at which the bounding polygon meets the X-axis or the session maximum packet rate (SMAXPR, measured in packets per second) provided by signaling, if any. Typically, a media sender will prefer to operate at a lower rate than this theoretical maximum, so as to increase the rate at which actual media content reaches the receivers. The purpose of the algorithm is to distinguish the TMMBR tuples constituting the bounding set and thus delineate the feasible region, so that the media sender can select its preferred operating point within that region Figure 1 below shows a bounding polygon formed by TMMBR tuples A and B. A third tuple C lies outside the bounding polygon and is therefore irrelevant in determining feasible trade-offs between media rate and packet rate. The line labeled ss..s represents the limit on packet rate imposed by the session maximum packet rate (SMAXPR) obtained by signaling during session setup. In Figure 1, the limit determined by tuple B happens to be more restrictive than SMAXPR. The situation could easily be the reverse, meaning that the bounding polygon is terminated on the right by the vertical line representing the SMAXPR constraint.
Net ^ Media|a c b s Bit | a c b s Rate | a c b s | a cb s | a c s | a bc s | a b c s | ab c s | Feasible b c s | region ba s | b a s c | b s c | b s a | bs +------------------------------> Packet rate Figure 1 - Geometric Interpretation of TMMBR Tuples Note that the slopes of the lines making up the bounding polygon are increasingly negative as one moves in the direction of increasing packet rate. Note also that with slight rearrangement, equations (1) and (2) have the canonical form: y = mx + b where m is the slope and has value equal to the negative of the tuple overhead (in bits), and b is the y-intercept and has value equal to the tuple maximum total media bit rate. These observations lead to the conclusion that when processing the TMMBR tuples to select the initial bounding set, one should sort and process the tuples by order of increasing overhead. Once a particular tuple has been added to the bounding set, all tuples not already selected and having lower overhead can be eliminated, because the next side of the bounding polygon has to be steeper (i.e., the corresponding TMMBR must have higher overhead) than the latest added tuple. Line cc..c in Figure 1 illustrates another principle. This line is parallel to line aa..a, but has a higher Y-intercept. That is, the corresponding TMMBR tuple contains a higher maximum total media bit rate value. Since line cc..c is outside the bounding polygon, it
illustrates the conclusion that if two TMMBR tuples have the same overhead value, the one with higher maximum total media bit rate value cannot be part of the bounding set and can be set aside. Two further observations complete the algorithm. Obviously, moving from the left, the successive corners of the bounding polygon (i.e., the intersection points between successive pairs of sides) lie at successively higher packet rates. On the other hand, again moving from the left, each successive line making up the bounding set crosses the X-axis at a lower packet rate. The complete algorithm can now be specified. The algorithm works with two lists of TMMBR tuples, the candidate list X and the selected list Y, both ordered by increasing overhead value. The algorithm terminates when all members of X have been discarded or removed for processing. Membership of the selected list Y is probationary until the algorithm is complete. Each member of the selected list is associated with an intersection value, which is the packet rate at which the line corresponding to that TMMBR tuple intersects with the line corresponding to the previous TMMBR tuple in the selected list. Each member of the selected list is also associated with a maximum packet rate value, which is the lesser of the session maximum packet rate SMAXPR (if any) and the packet rate at which the line corresponding to that tuple crosses the X-axis. When the algorithm terminates, the selected list is equal to the bounding set as defined in section 2.2. Initial Algorithm This algorithm is used by the media sender when it has received one or more TMMBRs and before it has determined a bounding set for the first time. 1. Sort the TMMBR tuples by order of increasing overhead. This is the initial candidate list X. 2. When multiple tuples in the candidate list have the same overhead value, discard all but the one with the lowest maximum total media bit rate value. 3. Select and remove from the candidate list the TMMBR tuple with the lowest maximum total media bit rate value. If there is more than one tuple with that value, choose the one with the highest overhead value. This is the first member of the selected list Y. Set its intersection value equal to zero. Calculate its maximum
packet rate as the minimum of SMAXPR (if available) and the value
obtained from the following formula, which is the packet rate at
which the corresponding line crosses the X-axis.
Max PR = TMMBR max total BR / (8 * TMMBR OH) ... (4)
4. Discard from the candidate list all tuples with a lower overhead
value than the selected tuple.
5. Remove the first remaining tuple from the candidate list for
processing. Call this the current candidate.
6. Calculate the packet rate PR at the intersection of the line
generated by the current candidate with the line generated by the
last tuple in the selected list Y, using equation (3).
7. If the calculated value PR is equal to or lower than the
intersection value stored for the last tuple of the selected list,
discard the last tuple of the selected list and go back to step 6
(retaining the same current candidate).
Note that the choice of the initial member of the selected list Y
in step 3 guarantees that the selected list will never be emptied
by this process, meaning that the algorithm must eventually (if
not immediately) fall through to step 8.
8. (This step is reached when the calculated PR value of the current
candidate is greater than the intersection value of the current
last member of the selected list Y.) If the calculated value PR
of the current candidate is lower than the maximum packet rate
associated with the last tuple in the selected list, add the
current candidate tuple to the end of the selected list. Store PR
as its intersection value. Calculate its maximum packet rate as
the lesser of SMAXPR (if available) and the maximum packet rate
calculated using equation (4).
9. If any tuples remain in the candidate list, go back to step 5.
Incremental Algorithm
The previous algorithm covered the initial case, where no selected
list had previously been created. It also applied only to the media
sender. When a previously created selected list is available at
either the media sender or media receiver, two other cases can be
considered:
o when a TMMBR tuple not currently in the selected list is a
candidate for addition;
o when the values change in a TMMBR tuple currently in the
selected list.
At the media receiver, these cases correspond, respectively, to those
of the non-owner and owner of a tuple in the TMMBN-reported bounding
set.
In either case, the process of updating the selected list to take
account of the new/changed tuple can use the basic algorithm
described above, with the modification that the initial candidate set
consists only of the existing selected list and the new or changed
tuple. Some further optimization is possible (beyond starting with a
reduced candidate set) by taking advantage of the following
observations.
The first observation is that if the new/changed candidate becomes
part of the new selected list, the result may be to cause zero or
more other tuples to be dropped from the list. However, if more than
one other tuple is dropped, the dropped tuples will be consecutive.
This can be confirmed geometrically by visualizing a new line that
cuts off a series of segments from the previously existing bounding
polygon. The cut-off segments are connected one to the next, the
geometric equivalent of consecutive tuples in a list ordered by
overhead value. Beyond the dropped set in either direction all of
the tuples that were in the earlier selected list will be in the
updated one. The second observation is that, leaving aside the new
candidate, the order of tuples remaining in the updated selected list
is unchanged because their overhead values have not changed.
The consequence of these two observations is that, once the placement
of the new candidate and the extent of the dropped set of tuples (if
any) has been determined, the remaining tuples can be copied directly
from the candidate list into the selected list, preserving their
order. This conclusion suggests the following modified algorithm:
o Run steps 1-4 of the basic algorithm.
o If the new candidate has survived steps 2 and 4 and has become
the new first member of the selected list, run steps 5-9 on
subsequent candidates until another candidate is added to the
selected list. Then move all remaining candidates to the
selected list, preserving their order.
o If the new candidate has survived steps 2 and 4 and has not
become the new first member of the selected list, start by
moving all tuples in the candidate list with lower overhead
values than that of the new candidate to the selected list,
preserving their order. Run steps 5-9 for the new candidate,
with the modification that the intersection values and maximum
packet rates for the tuples on the selected list have to be
calculated on the fly because they were not previously stored.
Continue processing only until a subsequent tuple has been
added to the selected list, then move all remaining candidates
to the selected list, preserving their order.
Note that the new candidate could be added to the selected
list only to be dropped again when the next tuple is
processed. It can easily be seen that in this case the new
candidate does not displace any of the earlier tuples in the
selected list. The limitations of ASCII art make this
difficult to show in a figure. Line cc..c in Figure 1 would
be an example if it had a steeper slope (tuple C had a higher
overhead value), but still intersected line aa..a beyond where
line aa..a intersects line bb..b.
The algorithm just described is approximate, because it does not take
account of tuples outside the selected list. To see how such tuples
can become relevant, consider Figure 1 and suppose that the maximum
total media bit rate in tuple A increases to the point that line
aa..a moves outside line cc..c. Tuple A will remain in the bounding
set calculated by the media sender. However, once it issues a new
TMMBN, media receiver C will apply the algorithm and discover that
its tuple C should now enter the bounding set. It will issue a TMMBR
to the media sender, which will repeat its calculation and come to
the appropriate conclusion.
The rules of section 4.2 require that the media sender refrain from
raising its sending rate until media receivers have had a chance to
respond to the TMMBN. In the example just given, this delay ensures
that the relaxation of tuple A does not actually result in an attempt
to send media at a rate exceeding the capacity at C.
3.5.4.3. Use of TMMBR in a Mixer-Based Multipoint Operation
Assume a small mixer-based multiparty conference is ongoing, as
depicted in Topo-Mixer of [RFC5117]. All participants have
negotiated a common maximum bit rate that this session can use. The
conference operates over a number of unicast paths between the
participants and the mixer. The congestion situation on each of
these paths can be monitored by the participant in question and by
the mixer, utilizing, for example, RTCP receiver reports (RRs) or the
transport protocol, e.g., Datagram Congestion Control Protocol (DCCP)
[RFC4340]. However, any given participant has no knowledge of the
congestion situation of the connections to the other participants.
Worse, without mechanisms similar to the ones discussed in this
document, the mixer (which is aware of the congestion situation on
all connections it manages) has no standardized means to inform media
senders to slow down, short of forging its own receiver reports
(which is undesirable). In principle, a mixer confronted with such a
situation is obliged to thin or transcode streams intended for
connections that detected congestion.
In practice, unfortunately, media-aware streaming thinning is a very
difficult and cumbersome operation and adds undesirable delay. If
media-unaware, it leads very quickly to unacceptable reproduced media
quality. Hence, a means to slow down senders even in the absence of
congestion on their connections to the mixer is desirable.
To allow the mixer to throttle traffic on the individual links,
without performing transcoding, there is a need for a mechanism that
enables the mixer to ask a participant's media encoders to limit the
media stream bit rate they are currently generating. TMMBR provides
the required mechanism. When the mixer detects congestion between
itself and a given participant, it executes the following procedure:
1. It starts thinning the media traffic to the congested participant
to the supported bit rate.
2. It uses TMMBR to request the media sender(s) to reduce the total
media bit rate sent by them to the mixer, to a value that is in
compliance with congestion control principles for the slowest
link. Slow refers here to the available bandwidth / bit rate /
capacity and packet rate after congestion control.
3. As soon as the bit rate has been reduced by the sending part, the
mixer stops stream thinning implicitly, because there is no need
for it once the stream is in compliance with congestion control.
This use of stream thinning as an immediate reaction tool followed up
by a quick control mechanism appears to be a reasonable compromise
between media quality and the need to combat congestion.
3.5.4.4. Use of TMMBR in Point-to-Multipoint Using Multicast or
Translators
In these topologies, corresponding to Topo-Multicast or Topo-
Translator, RTCP RRs are transmitted globally. This allows all
participants to detect transmission problems such as congestion, on a
medium timescale. As all media senders are aware of the congestion
situation of all media receivers, the rationale for the use of TMMBR
in the previous section does not apply. However, even in this case
the congestion control response can be improved when the unicast
links are using congestion controlled transport protocols (such as TCP or DCCP). A peer may also report local limitations to the media sender.3.5.4.5. Use of TMMBR in Point-to-Point Operation
In use case 7, it is possible to use TMMBR to improve the performance when the known upper limit of the bit rate changes. In this use case, the signaling protocol has established an upper limit for the session and total media bit rates. However, at the time of transport link bit rate reduction, a receiver can avoid serious congestion by sending a TMMBR to the sending side. Thus, TMMBR is useful for putting restrictions on the application and thus placing the congestion control mechanism in the right ballpark. However, TMMBR is usually unable to provide the continuously quick feedback loop required for real congestion control. Nor do its semantics match those of congestion control given its different purpose. For these reasons, TMMBR SHALL NOT be used as a substitute for congestion control.3.5.4.6. Reliability
The reaction of a media sender to the reception of a TMMBR message is not immediately identifiable through inspection of the media stream. Therefore, a more explicit mechanism is needed to avoid unnecessary re-sending of TMMBR messages. Using a statistically based retransmission scheme would only provide statistical guarantees of the request being received. It would also not avoid the retransmission of already received messages. In addition, it would not allow for easy suppression of other participants' requests. For these reasons, a mechanism based on explicit notification is used. Upon the reception of a TMMBR, a media sender sends a TMMBN containing the current bounding set, and indicating which session participants own that limit. In multicast scenarios, that allows all other participants to suppress any request they may have, if their limitations are less strict than the current ones (i.e., define lines lying outside the feasible region as defined in section 2.2). Keeping and notifying only the bounding set of tuples allows for small message sizes and media sender states. A media sender only keeps state for the SSRCs of the current owners of the bounding set of tuples; all other requests and their sources are not saved. Once the bounding set has been established, new TMMBR messages should be generated only by owners of the bounding tuples and by other entities that determine (by applying the algorithm of section 3.5.4.2 or its equivalent) that their limitations should now be part of the bounding set.
(next page on part 3)
