| RMCAT WG | M. Zanaty |
| Internet-Draft | Cisco |
| Intended status: Informational | V. Singh |
| Expires: September 03, 2014 | Aalto University |
| S. Nandakumar | |
| Cisco | |
| March 02, 2014 |
RTP Application Interaction with Congestion Control
draft-zanaty-rmcat-app-interaction-00
Interactive real-time media applications that use the Real-time Transport Protocol (RTP) over the User Datagram Protocol (UDP) must use congestion control techniques above the UDP layer since it provides none. This memo describes the interactions and interfaces necessary between the application components that relate to congestion control, including the RTP layer, the higher-level media codec control layer, and the lower-level transport interface and the congestion control layer.
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 September 03, 2014.
Copyright (c) 2014 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.
Interactive real-time media applications most commonly use RTP [RFC3550] over UDP [RFC0768]. Since UDP provides no form of congestion control, which is essential for any application deployed on the Internet, these RTP applications have historically implemented one of the following options at the application layer to address their congestion control requirements.
The RTP Media Congestion Avoidance Techniques (RMCAT) working group was chartered to standardize appropriate and effective congestion control for RTP applications. It is expected such applications will migrate from the above historical solutions to the RMCAT solution(s).
The RMCAT requirements [I-D.ietf-rmcat-cc-requirements] include low delay, reasonably high throughput, fast reaction to capacity changes including routing or interface changes, stability without over-reaction or oscillation, fair bandwidth sharing with other instances of itself and TCP flows, sharing information across multiple flows when possible [I-D.welzl-rmcat-coupled-cc], and performing as well or better in networks which support Active Queue Management (AQM), Explicit Congestion Notification (ECN), or Differentiated Services Code Points (DSCP).
In order to meet these requirements, interactions are necessary between the application's congestion controller, the RTP layer, media codecs, other components, and the OS's UDP stack. This memo discusses these interactions, presents a conceptual model of the required interfaces based on a simplified application decomposition, and proposes specific information exchange across these interfaces along with corresponding component behavior.
Note that RTP can also operate over other transports with integrated congestion control such as TCP [RFC5681] and DCCP [RFC4340], but that is beyond the scope of RMCAT and this memo.
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].
It is useful to decompose an RTP application into several components to facilitate understanding and discussion of where congestion control functions operate, and how they interface with the other components. The conceptual model in Figure 1 consists of the following components.
+----------------------+
| Config |
| | |
| Codec |
| | | |
| APP | RTP |
| | | |
| Congestion---|---Shared
| Control | State
+----------------------+
|
+----------------------+
| OS UDP |
+----------------------+
Figure 1
There are advantages and drawbacks to implementing congestion control in the application layer. It avoids OS dependencies and allows for rapid experimentation, evolution and optimization for each application. However, it also puts the burden on all applications, which raises the risks of improper or divergent implementations. One motivation of this memo is to mitigate such risks by giving proper guidance on how the application components relating to congestion control should interact.
Another drawback of congestion control in the application layer is that any decomposition, including the one presented in Figure 1, is purely conceptual and illustrative, since implementations have differing designs and decompositions. Conversely, this can be viewed as an advantage to distribute congestion control functions wherever expedient without rigid interfaces. For example, they may be distributed within the RTP stack itself, so the separate components in Figure 1 are combined into a single RTP+CC component as shown in Figure 2.
+----------------------+
| Config |
| | |
| Codec |
| APP | |
| RTP+CC------|---Shared
+----------------------+ State
|
+----------------------+
| OS UDP |
+----------------------+
Figure 2
The conceptual model in Figure 1 will be used throughout this memo to establish clearer boundaries between functions. But actual implementations may be closer to the looser model in [Singh12].
The primary interactions between the config and the codec that are relevant to congestion control are the multiplexing of media streams [I-D.ietf-mmusic-sdp-bundle-negotiation] and RTP/RTCP [RFC5761] on the same UDP port, and the negotiated RTCP attributes such as reduced size [RFC5506], codec control [RFC5104], transmission time [RFC5450], etc.
The config also establishes limits for the codec such as maximum bit rate and other codec-specific parameters. For example, a video codec config often sets limits on maximum resolution and frame rate as well as bit rate.
Packetization of codec frames into RTP packets can be an important interaction. Some network interfaces may benefit from small packet sizes well below the MTU, while others may benefit from large packets approaching the MTU. Equalizing packet sizes of a frame may also be beneficial in some cases, rather than a combination of large and small packets. For example, in some FEC schemes, the FEC bandwidth overhead depends on the largest source packet size. Equalizing the source packet sizes can yield lower overhead than a combination of large and small packets.
Allowed Rate (from CC to Codec): The max transmit rate allowed over the next time interval. The time interval may be specified or may use a default, for example, one second. The rate may be specified in bytes or packets or both. The rate must never exceed permanent limits established in session signaling such as the SDP bandwidth attribute [RFC4566] nor temporary limits in RTCP such as TMMBR [RFC5104] or REMB [I-D.alvestrand-rmcat-remb]. This is the most important interface among all components, and is always required in any RMCAT solution. In the simplest possible solution, it may be the only CC interface required.
Media Elasticity (from Codec to CC): Many live media encoders are highly elastic, often able to achieve any target bit rate within a wide range, by adapting the media quality. For example, a video encoder may support any bit rate within a range of a few tens or hundreds of kbps up to several Mbps, with rate changes registering as fast as the next video frame, although there may be limitations in the frequency of changes. Other encoders may be less elastic, supporting a narrower rate range, coarser granularity of rate steps, slower reaction to rate changes, etc. Other media, particularly some audio codecs, may be fully inelastic with a single fixed rate. CC can beneficially use codec elasticity, if provided, to plan Allowed Rate changes, especially when there are multiple flows sharing CC state and bandwidth.
Startup Ramp (from Codec to CC, and from CC to Codec): Startup is an important moment in a conversation. Rapid rate adaptation during startup is therefore important. The codec should minimize its startup media rate as much as possible without adversely impacting the user experience, and support a strategy for rapid rate ramp. The CC should allow the highest startup media rate as possible without adversely impacting network conditions, and also support rapid rate ramp until stabilizing on the available bandwidth. Startup can be viewed as a negotiation between the codec and the CC. The codec requests a startup rate and ramp, and the CC responds with the allowable parameters which may be lower/slower. The RMCAT requirements also include the possibility of bandwidth history to further accelerate or even eliminate startup ramp time. While this is highly desirable from an application viewpoint, it may be less acceptable to network operators, since it is in essence a gamble on current congestion state matching historical state, with the potential for significant congestion contribution if the gamble was wrong. Note that startup can often commence before user interaction or conversation to reduce the chance of clipped media.
Delay Tolerance (from Codec to CC): An ideal CC will always minimize delay and target zero. However, real solutions often need a real non-zero delay tolerance. The codec should provide an absolute delay tolerance, perhaps expressed as an impairment factor to mix with other metrics.
Loss Tolerance (from Codec to CC): An ideal CC will always minimize packet loss and target zero. However, real solutions often need a real non-zero loss tolerance. The codec should provide an absolute loss tolerance, perhaps expressed as an impairment factor to mix with other metrics. Note this is unrecoverable post-repair loss after retransmission or forward error correction.
Throughput Sensitivity (from Codec to CC): An ideal CC will always maximize throughput. However, real solutions often need a trade-off between throughput and other metrics such as delay or loss. The codec should provide throughput sensitivity, perhaps expressed as an impairment factor (for low throughputs) to mix with other metrics.
Rate Stability (from Codec to CC): The CC algorithm must strike a balance between rate stability and fast reaction to changes in available bandwidth. The codec should provide its preference for rate stability versus fast and frequent reaction to rate changes, perhaps expressed as an impairment factor (for high rate variance over short timescales) to mix with other metrics.
Forward Error Correction (FEC): Simple FEC schemes like XOR Parity codes [RFC5109] may not handle consecutive or burst loss well. More complex FEC schemes like Reed-Solomon [RFC6865] or Raptor [RFC6330] codes are more effective at handling bursty loss. The sensitivity to packet loss therefore depends on the media (source) encoding as well as the FEC (channel) encoding, and this sensitivity may differ for different loss patterns like random, periodic, or consecutive loss. Expressing this sensitivity to the congestion controller may help it choose the right balance between optimizing for throughput versus low loss.
Probing for Available Bandwidth: FEC can also be used to probe for additional available bandwidth, if the application desires a higher target rate than the current rate. FEC is preferable to synthetic probes since any contribution to congestion by the FEC probe will not impact the post-repair loss rate of the source media flow while synthetic probes may adversely affect the loss rate [Nagy14].
RTP Circuit Breakers: The intent behind RTP circuit breakers [I-D.ietf-avtcore-rtp-circuit-breakers] is to provide a kill switch of last resort, not true congestion control. The breakers should never trip when an effective congestion control is operating. This may impose some boundaries on RMCAT solutions to ensure the congestion control never approaches situations which may trigger the breakers.
RTCP Feedback: The primary method of communicating CC information is RTCP.
RTP Header Extensions: While RTCP is likely to be the primary carrier of CC feedback, the RMCAT requirements also include the possibility of using RTP header extensions in bidirectional flows for CC feedback. Transmission time [RFC5450], or possibly absolute time, also use header extensions, as would any per packet priority markings expected to survive across different networks and administrative domains.
Pacing / Shaping: Simple pacing / shaping strategies delay the transmission of packets to equalize inter-packet time intervals, assuming the bottleneck is most sensitive to packet rate. More complex pacing strategies may go beyond simple even distribution of transmission times. For example, Sprout [Winstein13] attempts to predict the optimal transmission time (and rate) with the highest probability of success for each packet based on channel statistics. Pacing may be always on, or adaptively enabled / disabled based on congestion state to minimize delay. Pacing may be performed within the CC for a single flow or across multiple flows. It may also be performed across all or selective traffic over the network interface if the OS supports interface-level traffic shaping.
Detection of Transport Capabilities: The CC can query the OS for useful transport capabilities such as ECN, DSCP, traffic shaping, etc. This may also aid upper layers in making better decisions such as whether or not to multiplex media streams. For example, if audio can be given differentiated network treatment from video when using separate ports.
ECN: If the OS and transport path support ECN, the CC can react faster than a loss-based CC and more reliably to congestion onset and abatement.
DSCP: If the OS and transport path support DSCP, the CC can map per-packet priority from RTP header extensions to DSCP (and layer 2 QoS if available) for better network handling under congestion.
AQM: If AQM is present in the bottleneck, and working effectively, there should be little or no excess delay observed when varying the transmission rate. The loss of such delay signals may hinder the performance of congestion control algorithms that are highly dependent on delay variation for adapting transmission rate. If the application has knowledge of the presence of AQM, through any means which are beyond the scope of this memo, it should communicate this to the CC. The CC may use this to alter its signal collection and rate adaptation strategies. The CC must not rely solely on this as a reliable indicator. It must continue to monitor statistics to validate this application hint, and react appropriately if the statistics suggest different network behavior.
Multiple Flows: Sharing state across multiple flows within the application can yield better CC decisions. Sharing state across even more flows beyond the application can yield even better CC decisions. The actual benefits and mechanisms of state sharing and coupled CC are described in [I-D.welzl-rmcat-coupled-cc].
Weighted Fairness: An important consideration in CC of multiple flows is their relative application-specified weights. Within an application, it is likely the different flows have different rate requirements, so equal bandwidth sharing may not be fair nor desirable, and weighted fairness may be required.
The RMCAT design team discussions contributed to this memo.
This memo includes no request to IANA.
Amplification attacks often use UDP traffic to launch denial of service attacks. Attackers may attempt to subvert congestion control protocols in UDP applications to launch amplification attacks by signaling more bandwidth than is actually available. For example, sending a victim a forged REMB or a few fast ACKs may result in the victim sending a high rate RTP stream. Attacks on conference servers could lead to further amplification if it distributes the streams to many others. One mitigation is to use SRTCP for congestion control messages where supported. Even if SRTCP is only authenticated not encrypted, SRTCP packets should always pass authentication checks before any message contents are interpreted. Non-secure RTCP should be avoided where possible.