| Transport Area Working Group | J. Holland |
| Internet-Draft | Akamai Technologies, Inc. |
| Intended status: Experimental | April 23, 2017 |
| Expires: October 25, 2017 |
Circuit Breaker Assisted Congestion Control (CBACC): Protocol Specification
draft-jholland-cb-assisted-cc-01
This document specifies Circuit Breaker Assisted Congestion Control (CBACC), which provides bandwidth information from senders to intermediate network nodes to enable good decisions for fast-trip Network Transport Circuit Breaker activity ([I-D.ietf-tsvwg-circuit-breaker]) when necessary for network health. CBACC is specifically designed to support protocols using IP multicast, particularly as a supplement to receiver-driven congestion control protocols to help affected networks rapidly detect and mitigate the impact of scenarios in which a network is oversubscribed to flows which are not responsive to congestion.
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 October 25, 2017.
Copyright (c) 2017 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.
This document specifies Circuit Breaker Assisted Congestion Control (CBACC).
CBACC is a congestion control building block designed for use with IP traffic that has a known maximum bandwidth, which does not reduce its sending rate in response to congestion. CBACC is specifically designed to supplement protocols using receiver-driven multicast congestion control systems that rely on well-behaved receivers to achieve congestion control in a very highly scalable system (up to millions of receivers) without a feedback path that reduces sending rates by senders. Examples of congestion control systems fitting this description include PLM, RLM, RLC, FLID-DL, SMCC, ESMCC, QIRLM, and WEBRC [RFC3738].
CBACC addresses a vulnerability to "overjoining", a condition in which receivers (particularly malicious receivers) subscribe to traffic which, from the sending side, is non-responsive to congestion. Overjoining attacks and the challenges they present are discussed in more detail in Appendix A.
A careful reading of the congestion control requirements of UDP Best Practices [I-D.ietf-tsvwg-rfc5405bis] suggests that a network that forwards multicast traffic is required to operate a circuit breaker to maintain network health under a persistent overjoining condition, at a cost of cutting off some or all multicast traffic across the network during high congestion.
CBACC provides a mechanism for networks to mitigate the impact of overjoining within a network by introducing a mechanism for communicating the bandwidth of non-responsive flows from the sender of the flow to the transit nodes forwarding the flow. The bandwidth information is sufficient to implement a fast-trip circuit breaker [I-D.ietf-tsvwg-circuit-breaker] within a single network node which can specifically block or police flows when receivers have overjoined the network's capacity.
In conjunction with receiver counts (e.g. via [RFC6807]) such nodes can also provide much improved network fairness for circuit breaking decisions during an overjoining condition.
In addition to streams using multicast receiver-driven congestion control, CBACC may also be suitable for use with other traffic, both unicast and multicast, that does not respond to congestion by reducing sending rates, including certain profiles of RTP [RFC3550] over either unicast or multicast, as well as several tunneling protocols (e.g. AMT [RFC7450] and GRE [RFC2784]) when they are known to carry traffic that would be suitable for CBACC. A complete specification for use of CBACC with unicast protocols and with tunneling protocols is out of scope for this document, though the security issues section does mention a few special considerations for potential unicast usage.
CBACC-compliant senders transmit Bandwidth Advertisements through the same transport path as the data traffic, so that circuit breakers can make informed decisions about how flows should be prioritized for circuit breaking. Additionally, CBACC-compliant circuit breakers transmit information to receivers about flows which have been or might soon be circuit-broken, to encourage CBACC-aware applications to use alternate methods to retrieve equivalent (though probably lower-quality and possibly less efficient) data when possible.
This document describes a building block as defined in [RFC3048]. This document describes a congestion control building block that conforms to [RFC2357]. This document follows the general guidelines provided in [RFC3269], in addition to the requirements on RFCs from [RFC5226] and [RFC3552].
| Term | Definition |
|---|---|
| circuit breaker | See [I-D.ietf-tsvwg-circuit-breaker] |
| controlled environment | See [I-D.ietf-tsvwg-rfc5405bis] Section 3.6 |
| general internet | See [I-D.ietf-tsvwg-rfc5405bis] Section 3.6 |
| flow | traffic for a single (source,destination) IP pair, including destinations that are group addresses |
| upstream | along a network topology path in the direction of a flow's sender |
| downstream | along a network topology path in the direction of a flow's receiver |
| ingress interface | the (single) upstream interface for a flow in a circuit breaker |
| egress interface | a downstream interface for a flow in a circuit breaker |
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].
CBACC is defined as an independent congestion control building block because it would be a useful supplement a wide variety of receiver-driven multicast congestion control schemes, such as [PLM] or other methods based on receiver-driven conformance to a measurement of available network bandwidth or congestion.
CBACC is also potentially valuable, even without other congestion control systems, in controlled environments where congestion control may not be required (e.g. for certain profiles of RTP [RFC3550]), since CBACC can provide protection for such a network against congestion due to sender or network mis-configuration.
CBACC provides a new form of communication between senders and network transit nodes to facilitate fast-trip circuit breakers as described in section 5.1 of [I-D.ietf-tsvwg-circuit-breaker] which are not available via previously existing methods. When used in conjunction with compatible circuit breakers, CBACC can greatly improve the safety of a network that accepts and delivers interdomain massively scalable multicast traffic to potentially untrusted receivers.
CBACC relies on the presence of CBACC-aware circuit breakers on a flow's transit path in order to provide congestion control in a network. In the absence of any CBACC-aware circuit breakers on a network path, CBACC constitutes a small extra overhead to a flow without providing any additional value.
CBACC provides a form of congestion control for massively scalable protocols using the IP multicast service. CBACC is best used in conjunction with another receiver-driven multicast congestion control, but it is also suitable for use even without another congestion control mechanism, or when presence of another congestion control mechanism is unproven, such as when accepting multicast joins from untrusted receivers.
CBACC senders send Bandwidth Advertisement packets to advertise the maximum sending bandwidth along the data path for a flow through a network.
CBACC bandwidth information is monitored by CBACC circuit breakers along the network path, which may block the forwarding of traffic for some flows in order to maintain network health. When a flow is blocked, a CBACC circuit breaker sets a bit in Bandwidth Advertisement packets before they're forwarded downstream that indicates to subscribed receivers of that flow that traffic has been blocked.
The protocol also defines a way to notify downstream receivers when a flow is in danger of being circuit broken in the near future. A CBACC-capable transport node SHOULD send this information when it is known, as described in section [TBD]. This gives applications an opportunity to gracefully shift to a lower-bandwidth version of the same content, when possible, providing an early warning system for avoiding congestion more smoothly.
A Bandwidth Advertisement packet constitutes an "ingress meter" as described in section 3.1 of [I-D.ietf-tsvwg-circuit-breaker]. The configured bandwidth caps of egress interfaces likewise constitute "egress meters". However, the diagram in the referenced document is simplified by running the ingress and egress on the same network node. At the CBACC-aware circuit breaker, the CBACC node has both pieces of information as soon as a Bandwidth Advertisement is received, and can trip the circuit breaker if the aggregate advertised CBACC bandwidth exceeds the actual bandwidth available on any egress interfaces.
Bandwidth advertisements can appear as either an IPv4 header option (as in Section 3.1 of [RFC0791]) or as an IPv6 extension header option (as in section 4.2 of [RFC2460]). They have the same layout:
0 1 2 3 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Type | Length |B|D|P| Res | Priority | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Bandwidth | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Figure 1
Bandwidth advertisements sent as IPv4 header options use option value [TBD], with the "copied" bit set and the option class "control", as specified in [RFC0791] section 3.1. Until and unless IANA assigns a value, this will be option number 158 as described in section 8 of [RFC4727] for experiments using IPv4 Option types. The length field is 8.
Bandwidth advertisements sent as IPv6 header options use option value [TBD], with the "action" bits set to "skip" and the "change" bit set to 1, as specified in [RFC2460] section 4.2. Until and unless IANA assigns a value, this will be option number 0x3e as described in section 8 of [RFC4727] for experiments using IPv6 Option Types. The length field is 6.
Using an IP header option has the benefit of exposing the bandwidth to all CBACC-compatible routers, in much the same way the IP Router Alert option would, but without being processed or causing undue load in non-CBACC routers.
The IP Header encapsulations DO work with IPSEC. As described in Appendix A of [RFC4302], the IP header fields are properly treated as mutable and zeroed for the IPSEC ICV calculations. CBACC circuit breakers MAY change bits in transit. The Bandwidth Advertisement header itself IS NOT protected by IPSEC security services, but protection of other parts of the packet remain unchanged.
As in several other protocols sending bandwidth values such as OSPF-TE [RFC3630], the bandwidth is expressed in bytes per second (not bits), in IEEE floating point format. For quick reference, this format is as follows:
0 1 2 3 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |S| Exponent | Fraction | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
Figure 2
S is the sign, Exponent is the exponent base 2 in "excess 127" notation, and Fraction is the mantissa - 1, with an implied binary point in front of it. Thus, the above represents the value:
(-1)**(S) * 2**(Exponent-127) * (1 + Fraction)
For more details, refer to [IEEE.754.1985].
Figure 3
Indicates that the flow has been circuit-broken.
Indicates that the flow is in danger of being circuit-broken.
Indicates that the flow should be policed instead of blocked. Flows marked for policing by the sender should have traffic proportionally dropped when bandwidth is needed, according to their priority. [TBD] Flesh this concept out, and decide whether it's actually viable. This was my attempt at addressing a suggestion from Bob Briscoe at IETF 97 in ICCRG at the mic, IIRC. It probably requires more state, such as total desired policable bandwidth, total current policed bandwidth, and current policing bandwidth per-flow, plus some definition of how to decide between cutting off some flows and policing others. This may not be worth the hassle, but there are some use cases such as FEC repair traffic which might actually be nicer this way. However, it might also be possible to get the same effect by assigning priority to those repair flows. Things like video enhancement layers of course are probably better done as a complete cutoff.
The sender MUST set all reserved bits to 0 when sending a CBACC control packet. Receivers and CBACC-capable transit nodes MUST accept any value in the reserved bits.
The sender MAY indicate relative priorities of different streams from the same sender with this field. This is an 8-bit unsigned integer, and higher values are kept preferentially over other traffic from the same sender with lower priority values, so all flows with a lower priority value are circuit-broken before any flows with a higher priority value. Among multiple flows from the same sender with the same priority, the highest bandwidth flows are circuit- broken first.
A CBACC circuit breaker holds the following state for each interface, for both the inbound and outbound directions on that interface:
The following state is kept for flows that are joined from at least one downstream interface and for which at least one CBACC Bandwidth Advertisement packet has been received:
Flows additionally have a per-interface state for egress interfaces:
The CBACC building block on a sender MUST have access to the maximum bandwidth that may be sent at any time in the following 3 seconds. A CBACC sender MUST send this value in a Bandwidth Advertisement packet once per second. The end result of the traffic sent on the wire for a particular flow MUST honor this maximum bandwidth commitment, such that bandwidth measurements taken over any sliding window one-second period MUST NOT exceed any of prior 3 maximum Bandwidth Advertisements (or any of them, if fewer than 3 have been sent).
A CBACC circuit breaker MUST order its monitored flows based on per-flow estimates of network fairness and preferentially circuit break less fair flows when bandwidth limits are exceeded. A normative method to determine network fairness for a flow is out of scope for this document, but CBACC circuit breaker implementations SHOULD provide a capability for network operators to configure administrative biases for specific sets of flows, and network operators SHOULD consider fairness concerns as expressed in [RFC2914] section 3.2 and other relevant documents describing best practices.
In particular, fairness metrics SHOULD favor multicast flows with many receivers over multicast flows with few receivers and flows with low bandwidth over flows with high bandwidth. When receiver counts are known (for example via the experimental PIM extension specified in [RFC6807]) a RECOMMENDED metric is (bandwidth/receiver count), though other metrics MAY be used where deemed appropriate by network operators following internet best practices, or when receiver counts can't be determined.
A CBACC sender MUST send Bandwidth Advertisements once per second. (Implementation-specific jitter in timer implementations not exceeding .1s is acceptable.)
If a circuit breaker receives more than 5 Bandwidth Advertisement packets for a flow in two seconds, the circuit breaker SHOULD set the flow to "pruned" and leave the upstream channel, and MUST drop Bandwidth Advertisement packets in excess of one per second.
Flows which are currently circuit-broken on an egress interface are set to "blocked". When a flow on an egress interface is in blocked state, Bandwidth Advertisement packets MUST be forwarded except as described in the preceding paragraph, the "Blocked" bit MUST be set to 1 before forwarding, and other traffic for that flow MUST NOT be forwarded along that interface.
When a flow is blocked or pruned, the circuit breaker MAY truncate the Bandwidth Advertisement packet, keeping only the headers of the packet containing the Bandwidth Advertisement before forwarding.
When a flow is pruned, the circuit-breaker MUST generate and forward a Bandwidth Advertisement packet once per second with the "Blocked" bit set when there are still downstream receivers connected.
In flows which are not circuit-broken but which would be circuit-broken if the bandwidth warning threshold were the bandwidth limit, the Danger bit MUST be set to 1 before forwarding. Both data and control packets are forwarded for flows in this situation. The "Danger" bit MAY be used by receivers to take early action to avoid getting circuit-broken by shifting to a lower-bandwidth representation, if available.
When a flow is in the "blocked" state on every egress interface, the circuit breaker MAY set the flow to "pruned" on the ingress interface and leave the channel upstream.
In addition to monitoring the advertised bandwidth, a CBACC circuit breaker or other assisting nodes in the network SHOULD monitor the observed bandwidth per flow, and SHOULD circuit break "overactive" flows, defined as those which exceed their CBACC maximum bandwidth commitment. A circuit breaker MAY perform constant monitoring on all flows, or MAY use load sharing techniques such as random selection or round robin to monitor only a certain subset of flows at a time.
When detecting overactive flows, circuit breakers MUST use techniques to avoid false positives due to transient upstream network conditions such as packet compression or occasional packet duplication. For example, using an average of bandwidth measurements over the prior 3 seconds would qualify, where a half-second window would not. (A full listing of reasonable false-positive avoidance techniques is out of scope for this document.)
[TBD: examples with network diagrams and bandwidths?] [TBD: some internal structure on this section. "wall of text" was some feedback]
The sender needs to know the bandwidth, including any upcoming changes, at least 3 seconds in advance. There is no requirement on how building blocks define this functionality except on the packets on the wire--the advance knowledge might, for example, be implemented by buffering and pacing on the sending machine. Specifics of the sending bandwidth implementations are out of scope for this document, as it's intended to provide requirements that will be applicable to a broad range of possible implementations, including RTP and WEBRC.
This draft requests IANA to allocate an IPv6 packet header option number with the "action" bits set to "skip" and the "change" bit set to 1, as specified in [RFC2460] section 4.2. [TO BE REMOVED: This registration should take place at the following location: http://www.iana.org/assignments/ipv6-parameters/ipv6-parameters.xhtml#extension-header.]
This draft also requests IANA to allocate an IPv4 packet header option number with the "copied" bit set and the option class "control", as specified in [RFC0791] section 3.1. [TO BE REMOVED: This registration should take place at the following location: http://www.iana.org/assignments/ip-parameters/ip-parameters.xhtml#ip-parameters-1.]
If those are deemed unacceptable, as an alternative with some compromises described in Section 5.2.1, this draft instead requests IANA to allocate a UDP destination port number. [TO BE REMOVED: This registration should take place at the following location: http://www.iana.org/assignments/service-names-port-numbers/service-names-port-numbers.xhtml.]
Forged Bandwidth Advertisement packets that get accepted by CBACC circuit breakers which dramatically over-report or under-report the correct bandwidth would present a potential DoS against a CBACC flow, by making the circuit breaker believe the flow exceeds the node's capacity when over-reporting, or by letting the node notice an apparent violation of the commitment to remain under the advertised bandwidth when under-reporting.
Similarly, it is possible to forge a CBACC Bandwidth Advertisement for a non-CBACC flow, which likewise may constitute a DoS against that flow.
For multicast, attacker would have to be on-path in order to deliver a forged packet to a CBACC circuit breaker, because the join's reverse path propagation will only reach the sender on a legitimate network path to its source address.
For unicast, it's a bigger problem, because ANY sender along path that doesn't have RPF check BCP 38 [RFC2827] permits attack on the flow via forged packet that substantially under-reports or over-reports bandwidth.
For AMT tunnels, when RPF checks along a path to the gateway are not present, nothing stops forged packets from being forwarded by the gateway. If these packets contain CBACC control packets, it's possible to inject a forged packet into the network downstream from the gateway, combining the unicast hole with the multicast hole. This is a vulnerability that should probably be addressed by a new AMT version with some defense against forgery of data.
For IPSEC, since the Bandwidth Advertisement IP header option is mutable, it's not protected by the IPSEC security services, so the Bandwidth Advertisement can be forged for consumption by the circuit breakers, even though the packet will be rejected by the end host with the security association. This could mount a DoS via the intermediate circuit-breakers by over-reporting or under-reporting flow bandwidth, when processing CBACC traffic through untrusted network paths.
The unicast vulnerabilities would be much mitigated by RPF checks as recommended by BCP 38 [RFC2827] at every hop, or otherwise maintained by the network. Absent such checks, cheap DoS vulnerabilities may be present from any permissive network locations.
CBACC control packets are sent as part of the data stream so that they traverse the same intermediate network nodes as the rest of the data, but they also carry control information that must be processed by certain nodes along that path.
This creates potential problems very similar to the problems with the Router Alert IP option discussed in Section 3 of [RFC6398], where a circuit-breaker might have a "fast path" for forwarding that can handle a much higher traffic volume than the "slow path" necessary to process CBACC control packets, which is potentially vulnerable to overloading.
If a CBACC-compatible circuit breaker receives a high rate of CBACC control packets, the circuit breaker MUST maintain network health for other flows. A circuit-breaker MAY drop all packets, including all CBACC control packets, for a flow in which more than 5 CBACC control packets were received in less than a second. (This number is intended to allow for moderate IP packet duplication and packet compression by upstream routers, while still being slow enough for handling of packets on the slow path.)
Since CBACC flows require state, it may be possible for a set of receivers and/or senders, possibly acting in concert, to generate many flows in an attempt to overflow the circuit breakers' state tables.
It is permissible for a network node to behave as a CBACC circuit breaker for some CBACC flows while treating other CBACC flows as non-CBACC, as part of a load balancing strategy for the network as a whole, or simply as defense against this concern when the number of monitored flows exceeds some threshold.
The same techniques described in section 3.1 of [RFC4609] can be used to help mitigate this attack, for much the same reasons. It is RECOMMENDED that network operators implement measures to mitigate such attacks.
Many thanks to Devin Anderson and Ben Kaduk for detailed reviews and many great suggestions. Thanks also to Cheng Jin, Scott Brown, Miroslav Kaduk, and Bob Briscoe for their thoughtful contributions.
[I-D.ietf-tsvwg-rfc5405bis] describes several remedies for unicast congestion control under UDP, even though UDP does not itself provide congestion control. In general, any network node under congestion could in theory collect evidence that a unicast flow's sending rate is not responding to congestion, and would then be justified in circuit-breaking it.
With multicast IP, the situation is different, especially in the presence of malicious receivers. A well-behaved sender using a receiver-controlled congestion scheme such as WEBRC does not reduce its send rate in response to congestion, instead relying on receivers to leave the appropriate multicast groups.
This leads to a situation where, when a network accepts inter-domain multicast traffic, as long as there are senders somewhere in the world with aggregate bandwidth that exceeds a network's capacity, receivers in that network can join the flows and overflow the network capacity. A receiver controlled by an attacker could do this at the IGMP/MLD level without running the application layer protocol that participates in the receiver-controlled congestion control.
A network might be able to detect and defend against the most naive version of such an attack by blocking end users that try to join too many flows at once. However, an attacker can achieve the same effect by joining a few high-bandwidth flows, if those exist anywhere, and an attacker that controls a few machines in a network can coordinate the receivers so they join disjoint sets of non-responsive sending flows.
This scenario will produce congestion in a middle node in the network that can't be easily detected at the edge where the IGMP/MLD join is accepted. Thus, an attacker with a small set of machines in a target network can always trip a circuit breaker if present, or can induce excessive congestion among the bandwidth allocated to multicast. This problem gets worse as more multicast flows become available.
This is a significant barrier to multicast adoption because there is no present defense which does not itself constitute a denial of service attack.
Although the same can apply to non-responsive unicast traffic, network operators can assume that non-responsive sending flows are in violation of congestion control best practices, and can therefore cut off such flows. However, non-responsive multicast senders are likely to be well-behaved participants in receiver-controlled congestion control schemes.
However, receiver controlled congestion control schemes also show the most promise for efficient massive scale content distribution via multicast, provided network health can be ensured. Therefore, mechanisms to mitigate overjoining attacks while still permitting receiver-controlled congestion control are necessary. [TBD: this whole section should be expanded and moved to a separate informational draft]
TBD: network diagram
Figure 4