Congestion Exposure
(ConEx) Concepts and Abstract MechanismGoogle, Inc1600 Amphitheater ParkwayMountain View93117CaliforniaUSAmattmathis at google.comBTB54/77, Adastral ParkMartlesham HeathIpswichIP5 3REUK+44 1473 645196bob.briscoe@bt.comhttp://bobbriscoe.net/
Transport
Congestion Exposure (ConEx) Working GroupQuality of ServiceQoSCongestion ControlSignalingProtocolEncodingAuditPolicingThis document describes an abstract mechanism by which senders inform
the network about the congestion encountered by packets earlier in the
same flow. Today, network elements at any layer may signal congestion to the receiver by
dropping packets or by ECN markings, and the receiver passes this
information back to the sender in transport-layer feedback. The
mechanism described here enables the sender to also
relay this congestion information back into the network in-band at the
IP layer, such that the total amount of congestion from all elements on the path is revealed to all IP elements along the path, where it could, for example, be used to provide input to traffic management. This mechanism is called congestion exposure or ConEx. The companion document "ConEx Concepts and Use Cases" provides the entry-point to the set of ConEx documentation.This document describes an abstract mechanism by which, to a first approximation, senders inform the network about the congestion encountered by packets earlier in the same flow. It is not a complete protocol specification, because it is known that designing an encoding (e.g. packet formats, codepoint allocations, etc) is likely to entail compromises that preclude some uses of the protocol. The goal of this document is to provide a framework for developing and testing algorithms to evaluate the benefits of the ConEx protocol and to evaluate the consequences of the compromises in various different encoding designs. A companion document provides the entry point to the set of ConEx documentation. It outlines concepts that are pre-requisites to understanding why ConEx is useful, and it outlines various ways that ConEx might be used.
As typical end-to-end transport protocols continually seek out more network capacity,
network elements signal whenever congestion results, and the
transports are responsible for controlling this network congestion .
The more a transport tries to use capacity that others want to use, the more congestion signals will be attributable to that transport. Likewise, the more transport sessions sustained by a user and the longer the user sustains them, the more congestion signals will be attributable to that user. The goal of ConEx is to ensure that the resulting congestion signals are sufficiently visible and robust, because they are an ideal metric for networks to use as the basis of traffic management or other related functions.Networks indicate congestion by three possible signals: packet loss, ECN marking or queueing delay. ECN marking and some packet loss may be the outcome of Active Queue Management (AQM), which the network uses to warn senders to reduce their rates. Packet loss is also the natural consequence of complete exhaustion of a buffer or other network resource. Some experimental transport protocols and TCP variants infer impending congestion from increasing queuing delay. However, delay is too amorphous to use as a congestion metric. ConEx is only concerned with ECN markings and packet losses, because they are unambiguous signals of congestion. In both cases the congestion signals follow the route indicated in
. A congested
network device sends a signal in the data stream on the forward path to
the transport receiver, the receiver passes it back to the sender
through transport level feedback, and the sender makes some congestion
control adjustment.This document extends the capabilities of the Internet protocol suite with the addition of a new Congestion Exposure signal. To a first approximation this signal, also shown in ,
relays the congestion information from the transport sender back through the internetwork layer where it is visible to any interested internetwork layer devices along the forward path. This document frames the engineering problem of designing the ConEx signal. The requirements are described in and some example encoding are presented in . describes all of the protocol components.This new signal is expressly designed to support a variety of new policy mechanisms that might be used to instrument, monitor or manage traffic. The policy devices are not shown in but might be placed anywhere along the forward data path. They are described in Since the policy devices can affect how traffic is treated it is assumed that there is an intrinsic motivation for users, applications or operating systems to understate the congestion that they are causing. It is important to be able to audit ConEx signals, and to be able apply sufficient sanction to discourage cheating of congestion policies. The general approach to auditing is to count signals on the forward path to confirm that there are never fewer ConEx signals than congestion signals.
Many ConEx design constraints come from the need to assure that the audit function is sufficiently robust. The audit function is described in , however significant portions of this document (and prior research ) is motivated by issues relating to the audit function and making it robust.The congestion and ConEx signals shown in represent a series of discrete events: ECN marks or lost packets, carried by the forward data stream and fed back into the Internetwork layer. The policy and audit functions are most likely to act on the accumulated values of these signals, for which we use the term "volume". For example traffic volume is the total number of bytes delivered, optionally over a specified time interval and over some aggregate of traffic (e.g. all traffic from a site). While loss-volume is the total amount of bytes discarded from some aggregate over an interval. The term congestion-volume is defined precisely in . Note that volume per unit time is (average) rate.One of the design goals of the ConEx protocol is that none of the important policy mechanisms requires per flow state, and that policy mechanisms can be implemented for heavily aggregated traffic in the core of the Internet with complexity akin to accumulating marking volumes per logical link. Ideally it would also be possible to audit ConEx signals without per flow state, however this is not always possible. Since auditing can be done near the edges of the network where traffic is less aggregated, per flow state is more easily tolerated. Also, the flow-state required for audit creates itself as it detects new flows. Therefore a flow will not fail if it is re-routed away from the audit box currently holding its flow-state. Flow-state for auditing is discussed further in . In summary: i) flow state for auditing does not require route pinning; ii) auditing at the edges, with limited per flow state, enables policy in the core, without any per flow state.
There is a long standing argument over units of congestion: bytes vs packets (see and its references). This document does not take a strong position on this issue. However, we make the following observations: the most expensive links in the Internet, in terms of cost per bit, are all at lower data rates, where transmission times are large and packet sizes are important. In order for a policy to consider wire time, it needs to know the number of congested bytes. However, high speed networking equipment and the transport protocols themselves sometimes gauge resource consumption and congestion in terms of packets, which may prove to be problematic for application protocols that have irregular packet sizes, such as BGP, SPDY and some variable rate video encoding schemes. The units of congestion must be an explicitly stated property of any proposed encoding, and the consequences of that design decision must be evaluated along with other aspects of the design.To be successful the ConEx protocol must have the property that the relevant stakeholders each have the incentive to unilaterally start on each stage of partial deployment, which in turn creates incentives for further deployment. Furthermore, legacy systems that will never be upgraded do not become a barrier to deploying ConEx. Issues relating to partial deployment are described in .
Note that ConEx signals are not intended to be used for fine-grained congestion control. They are anticipated to be most useful at longer time scales, for example the total congestion caused by a user might serve as an input to higher level policy or accountability functions, designed to create incentives for improving user behavior, such as choosing to send large quantities of data at off-peak times, at lower data rates or with less aggressive protocols such as LEDBAT (see ).Ultimately ConEx signals have the potential to provide a mechanism to regulate global Internet congestion. From the earliest days of congestion control research there has been a concern that there is no mechanism to prevent transport designers from incrementally making protocols more aggressive without bound and spiraling to a "tragedy of the commons" Internet congestion collapse. The "TCP friendly" paradigm was created in part to forestall this failure. However, it no longer commands any authority because it has little to say about the Internet of today, which has moved beyond the scaling range of standard TCP. As a consequence, many transports and applications are opening arbitrarily large numbers of connections or using arbitrary levels of aggressiveness. ConEx represents a recognition that the IETF cannot regulate this space directly because it concerns the behaviour of users and applications, not individual transport protocols. Instead the IETF can give network operators the protocol tools to arbitrate the space themselves, with better bulk traffic management. This in turn should create incentives for users, and designers of application and of transport protocols to be more mindful about contributing to congesting. 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 RFC 2119 .ConEx signals in IP packet headers from the sender to the network:The transport (or at least this packet) is not ConEx-capable.The transport is ConEx-Capable. This
is the opposite of Not-ConEx.A packet sent by a ConEx Capable transport.
It carries at least one of the following signals:
The transport has experienced a loss.The transport has experienced an ECN mark.The transport is building up credit to allow for any future delay in
expected ConEx signals (see )The transport is ConEx-capable
but is signaling none of Re-Echo-Loss, Re-Echo-ECN or Credit.At least one of Re-Echo-Loss, Re-Echo-ECN or Credit.First time readers may wish to skim this section, since it is more understandable having read the entire document.Ideally, all the following requirements would be met by a Congestion
Exposure Signal:The ConEx Signal SHOULD be visible to internetwork layer devices
along the entire path from the transport sender to the transport
receiver. Equivalently, it SHOULD be present in the IPv4 or IPv6
header, and in the outermost IP header if using IP in IP tunneling.
The ConEx Signal SHOULD be immutable once set by the transport
sender. A corollary of these requirements is that the chosen ConEx
encoding SHOULD pass silently without modification through pre-existing
networking gear.The ConEx Signal SHOULD be useful under only partial deployment.
A minimal deployment SHOULD only require changes to transport
senders. Furthermore, partial deployment SHOULD create incentives
for additional deployment, both in terms of enabling ConEx on more
devices and adding richer features to existing devices. Nonetheless,
ConEx deployment need never be universal, and it is anticipated that
some hosts and some transports may never support the ConEx Protocol
and some networks may never use the ConEx Signals.The ConEx signal SHOULD be timely. There will be a minimum delay of one RTT, and often longer if the transport protocol sends infrequent feedback (consider RTCP for example). This delay complicates auditing, and SHOULD be minimized.The ConEx signal SHOULD be accurate and auditable. The general approach is to observe the volume of congestion signals and ConEx signals on the forward data path and verify that the ConEx signals do not under-represent the congestion signals (see ). The simplest mechanism to compensate for the round trip delay between the signals is for the sender to include a "credit" signal to cover the yet to be observed congestion that might occur during this delay. (see for details). Furthermore, the ConEx signals for packet loss and ECN marking SHOULD have distinct encodings because they are likely to require different auditing techniques.
It is already known that implementing ConEx signals is likely to entail some compromises, and therefore all the requirements above are expressed with the
keyword 'SHOULD' rather than 'MUST'. The only mandatory requirement is
that a concrete protocol description MUST give sound reasoning if it
chooses not to meet some requirement.
The role and constraints on the audit function are described in . There is no intention to standardise the audit function. However, it is necessary to lay down the following normative constraints on audit behaviour so that transport designers will know what to design against and implementers of audit devices will know what pitfalls to avoid:
Audit SHOULD introduce minimal false hits for
honest flows; Audit SHOULD quickly detect and sanction dishonest
flows, ideally on the first dishonest packet; Audit SHOULD NOT be designed around one particular rate
response, such as any particular TCP congestion control algorithm
or one particular resource sharing regime such as
TCP-friendliness . An important goal is to give ingress
networks the freedom to unilaterally allow different rate responses to congestion and different
resource sharing regimes , without having to coordinate with other
networks over details of individual flow behaviour; Audit SHOULD introduce sufficient sanction (e.g. loss in goodput) such that senders cannot gain from understating congestion; To the extent that the audit might be subject to false hits, the sanction SHOULD be proportionate to the degree to which congestion is understated. If audit over-punishes, attackers will find ways to harness it into amplifying attacks on others.
Ideally audit should, in the long-run, cause the user to get no better performance than they would get by being accurate. Audit SHOULD be able to counter state exhaustion
attacks. For instance, if the audit function uses flow-state, it should not be possible for
senders to exhaust its memory capacity by gratuitously sending numerous packets, each
with a different flow ID. Audit SHOULD NOT be vulnerable to `identity
whitewashing', where a transport can label a flow with a new ID more cheaply than
paying the cost of continuing to use its current ID ;An experimental ConEx specification SHOULD describe the following protocol details:The specific ConEx signal encodings with packet formats, bit fields and/or code points; An inventory of invalid combinations of flags or invalid codepoints
in the encoding. Whether security gateways should normalise, discard or
ignore such invalid encodings, and what values they should be considered
equivalent to by ConEx-aware elements;An inventory of any conflated signals or any other effects that are known to compromise signal integrity;A specification for signal units (bytes vs packets, etc), any approximations allowed and algorithms to do any implied conversions or accounting;If the units are bytes a definition of which headers are included in the size of the packet;How tunnels should propagate the ConEx encoding;Whether the encoding fields are mutable or not, to ensure that header authentication, checksum calculation, etc. process them correctly.A statement that the ConEx encoding is only applicable to unicast and
anycast, and that forwarding elements should silently ignore any ConEx
signalling on multicast packets (they should be forwarded unchanged)Definition of any extensibility;Backward and forward compatibility and potential migration strategies;Any (optional) modification to data-plane forwarding dependent on the encoding (e.g. preferential discard, interaction with Diffserv, ECN etc.);Any warning or error messages relevant to the encoding.
Note regarding item H on multicast: A multicast tree may involve
different levels of congestion on each leg. Any traffic management can only
monitor or control multicast congestion at or near each receiver. It would
make no sense for the sender to try to expose "whole path congestion" in
sent packets, because it cannot hope to describe all the differing
congestion levels on every leg of the tree.
A specification of any required changes to congestion feedback in particular transport protocols.A specification (or minimally a recommendation) for how a transport should estimate credits at the beginning of a new connection.A specification of whether any other protocol options should (or
must) be enabled along with an implementation of ConEx (e.g. at least
attempting to negotiate ECN and SACK capability);A specification of any configuration that a ConEx stack may require
(or preferably confirmation that it requires no configuration);A specification of the statistics that a protocol stack should log
for each type of marking on a per-flow or aggregate basis.An example of a strong audit algorithm suitable for detecting if a single flow is misstating congestion. This algorithm should present minimal false results, but need not have optimal scaling properties (e.g. may need per flow state).An example of an audit algorithm suitable for detecting misstated congestion in a large aggregate (e.g. no per-flow state).The possibility exists that these specifications over constrain the ConEx design, and can not be fully satisfied. An important part of the evaluation of any particular design will be a thorough inventory of all ways in which it might fail to satisfy these specifications.
Most protocol specifications start with a description of packet
formats and codepoints with their associated meanings. This document
does not: It is already known that choosing the encoding for ConEx
is likely to entail some engineering compromises that have the
potential to reduce the protocol's usefulness in some settings. For instance the experimental ConEx encoding chosen for IPv6 had to make compromises on tunnelling. Rather than making these engineering choices prematurely, this document side
steps the encoding problem by making it abstract. It describes several different representations of
ConEx Signals, none of which are specified to the level of specific bits or code points.The goal of this approach is to be as complete as possible for
discovering the potential usage and capabilities of the ConEx protocol,
so we have some hope of making optimal design decisions when choosing
the encoding.
Even if experiments reveal particular problems due to the encoding, then this document will still serve as a reference model.For tutorial purposes, it is helpful to describe a naïve encoding of the ConEx protocol for TCP and similar protocols: set a bit (not specified here) in the IP header on each retransmission and on each ECN signaled window reduction. Network devices along the forward path can see this bit and act on it. For example any device along the path might limit the rate of all traffic if the rate of marked (congested) packets exceeds a threshold. This simple encoding is sufficient to illustrate many of the benefits envisioned for ConEx. At first glance it looks like it might motivate people to deploy and use it. It is a one line code change that a small number of OS developers and content providers could unilaterally deploy across a significant fraction of all Internet traffic. However, this encoding does not support auditing so it would also motivate users and/or applications to misrepresent the congestion that they are causing . As a consequence the naïve encoding is not likely to be trusted and thus creates its own disincentives for deployment.Nonetheless, this Naïve encoding does present a clear mental model of how the ConEx protocol might function under various uses. It is useful for thought experiments where it can be stipulated that all participants are honest and it does illustrate some of the incentives that might be introduced by ConEx.In limited contexts it is possible to implement ConEx-like functions without any signals at all by measuring rest-of-path congestion directly from TCP headers. The algorithm is to keep at least one RTT of past TCP headers and matching each new header against the history to count duplicate data.This could implement many ConEx policies, without any explicit protocol. It is fairly easy to implement, at least at low rate (e.g. in a software based edge router). However, it would only be useful in cases where the network operator can see the TCP headers. This is currently (2012) the vast majority of traffic because UDP, IPSEC and VPN tunnels are used far less than SSL or TLS over TCP/IP, which do not hide TCP sequence numbers from network devices. However, anyone specifically intending to avoid the attention of a congestion policy device would only have to hide their TCP headers from the network operator (e.g. by using a VPN tunnel).The re-ECN specification presents an encoding
of ConEx in IPv4 and IPv6 that was tightly integrated with ECN encoding
in order to fit into the IPv4 header. ConEx and ECN are orthogonal signals in the sense that any individual packet may need to represent any one of the 4 possible combinations of signal values. Ideally their encoding should be entirely
independent. However, given the limited number of header bits and/or
code points, re-ECN chooses to partially share code points and to re-echo both losses and ECN with just one codepoint. The central theme of the re-ECN work is an audit
mechanism that provides sufficient disincentives against
misrepresenting congestion . It is analyzed
extensively in Briscoe's PhD dissertation .
For a tutorial background on re-ECN motivation and techniques, see [, ].Re-ECN is an example of one chosen set of compromises
attempting to meet the requirements of . The present document
takes a step back, aiming to state the ideal requirements in order to
allow the Internet community to assess whether different compromises might be
better.The problem with Re-ECN is that it requires that receivers be ECN enabled in addition to sender changes. Newer encodings overcome this problem by being able to represent loss and ECN based congestion separately.This encoding involves flag bits, each of which the sender can
set independently to indicate to the network one of the following
four signals:The transport is (or is not)
using ConEx with this packet (the protocol MUST be arranged so
that legacy transport senders implicitly send Not-ConEx)The transport has
(or has not) experienced a lossThe transport has
(or has not) experienced ECN-signaled congestionThe transport is (or is not)
building up congestion credit (see on the audit function)This encoding does not imply any exclusion property among the signals. Multiple types of congestion (ECN, loss) can be signalled on the same ACK. However, there will be many invalid combinations of
flags (e.g. Not-ConEx combined with any of the ConEx-marked flags), which
could be used to advantage against naive policy devices that only check each
flag separately.As long as the packets in a flow have uniform sizes, it does not matter whether the units of congestion are packets or bytes. However, if an application sends very irregular packet sizes, it may be necessary for the sender to mark multiple packets to avoid being in technical violation of the audit function.This encoding involves signaling one of the following five
codepoints:ENUM {Not-ConEx, ConEx-Not-Marked, Re-Echo-Loss, Re-Echo-ECN, Credit}Each named codepoint has the same meaning as in the encoding
using independent bits in the previous section.
The use of any one codepoint implies the negative of all the others.Inherently, the semantics of most of the enumerated codepoints
are mutually exclusive. 'Credit' is the only one that might need to
be used in combination with either Re-Echo-Loss or Re-Echo-ECN, but
even that requirement is questionable. It must not be forgotten that
the enumerated encoding loses the flexibility to signal these two
combinations, whereas the encoding with four independent bits is not
so limited. Alternatively two extra codepoints could be assigned to
these two combinations of semantics. The comment in the previous section about units also applies. The following comments apply generally to all the other encodings.
Congestion can be due to exhaustion of bit-carry capacity, or exhaustion of packet processing power. When a packet is discarded or marked to indicate congestion, there is no easy way to know whether the lost or marked packet signifies bit-congestion or packet-congestion. The above ConEx encodings that rely on marking packets suffer from the same ambiguity.
This problem is most acute when audit needs to check that one count of markings matches another. For example if there are ConEx markings on three large (1500B) packets, is that sufficient to match the loss of 5 small (60B) packets? If a packet-marking is defined to mean all the bytes in the packet are marked, then we have 4500B of Conex marked data against 300B of lost data, which is easily sufficient. If instead we are counting packets, then we have 3 ConEx packets against 5 lost packets, which is not sufficient. This problem will not arise when all the packets in a flow are the same size, but a choice needs to be made for flows in which packet sizes vary.
Therefore a ConEx encoding SHOULD explicity specify whether it assumes units of bytes or packets for both congestion indications and ConEx markings.
advises that congestion indications SHOULD be interpreted in units of bytes when responding to congestion, at least on today's Internet. In any TCP implementation this is simple to achieve for varying size packets, given TCP SACK tracks losses in bytes.
If an encoding is specified in units of bytes, the encoding SHOULD also specify which headers to include in the size of a packet.The components shown in as well as policy and audit are described in more detail.Congestion signals originate from network devices as they do today. A congested router, switch or other network device can discard or ECN mark packets when it is congested.The sending transport needs to be modified to send Congestion
Exposure Signals in response to congestion feedback signals (e.g. for the case of a TCP transport see ). We want to permit ConEx without ECN (e.g. if the receiver does not support ECN). However, we want to encourage a ConEx sender to at least attempt to negotiate ECN (a ConEx transport protocol spec may require this), because it is believed that ConEx without ECN is harder to audit, and thus potentially exposed to cheating. Since honest users have the potential to benefit from stronger mechanisms to manage traffic they have an incentive to deploy ConEx and ECN together. This incentive is not sufficient to prevent a dishonest user from constructing (or configuring) a sender that enables ConEx after choosing not to negotiate ECN, but is should be sufficient to prevent this from being the sustained default case for any significant pool of users.Permitting ConEx without ECN is necessary to facilitate bootstrapping other parts of ConEx deployment.Any receiving transport may already feedback sufficiently useful
signals to the sender so that it does not need to be altered.If the transport receiver does not support ECN, then it's native loss signaling mechanism (required for compliance with existing congestion control standards) will be sufficient for the Sender to generate ConEx signals.A traditional ECN implementation (RFC 3168 for TCP) signals congestion no more
than once per round trip. The sender may require more precise
feedback from the receiver otherwise it is at risk of appearing to be understating
its ConEx Signals.Ideally, ConEx should be added to a transport like TCP without
mandatory modifications to the receiver. But an optional modification
to the receiver could be recommended for precision (see ). This was the
approach taken when adding re-ECN to TCP .Policy devices are characterised by a need to be configured with a
policy related to the users or neighboring networks being served. In
contrast, auditing devices solely enforce compliance with the ConEx protocol and do not need
to be configured with any client-specific policy.Policy devices can typically be decomposed into two functions i)
monitoring the ConEx signal to compare it with a policy then ii)
acting in some way on the result. Various actions might be invoked
against 'out of contract' traffic, such as policing (see ), re-routing, or downgrading the
class of service.Alternatively a policy device might not act directly on the
traffic, but instead report to management systems that are designed
to control congestion indirectly. For instance the reports might
trigger capacity upgrades, penalty clauses in contracts, levy
charges based on congestion, or merely send
warnings to clients who are causing excessive congestion.Nonetheless, whatever action is invoked, the congestion
monitoring function will always be a necessary part of any policy
device.ConEx signals indicate the level of congestion along a whole path
from source to destination. In contrast, ECN signals
monitored in the middle of a network indicate the level of
congestion experienced so far on the path (of course, only in ECN-capable traffic.If a monitor in the middle of a network (e.g. at a network border)
measures both of these signals, it can subtract the level of ECN
(path so far) from the level of ConEx (whole path) to derive a
measure of the congestion that packets are likely to experience
between the monitoring point and their destination (rest-of-path
congestion).It will often be preferable for policy devices to monitor
rest-of-path congestion if they can, because it is a measure of the
downstream congestion that the policy device can directly influence
by controlling the traffic passing through it.A congestion policer can be implemented in a very similar way to
a bit-rate policer, but its effect can be focused solely on traffic of users
causing congestion downstream, which ConEx signals make visible.
Without ConEx signals, the only way to mitigate congestion is to
blindly limit traffic bit-rate, on the assumption that high bit-rate
is more likely to cause congestion.A congestion policer monitors all ConEx traffic entering a
network, or some identifiable subset. Using ConEx signals (and
preferably subtracting ECN signals to yield rest-of-path congestion), it measures the amount of
congestion that this traffic is contributing somewhere downstream.
If this persistently exceeds a policy-configured 'congestion-bit-rate' the
congestion policer can limit all the monitored ConEx traffic.A congestion policer can be implemented by a simple token bucket applied to an aggregate.
But unlike a bit-rate policer, it removes tokens only when it
forwards packets that are ConEx-Marked, effectively treating
Not-ConEx-Marked packets as invisible. Consequently, because tokens
give the right to send congested bits, the fill-rate of the token
bucket will represent the allowed congestion-bit-rate. This should
provide sufficient traffic management without having to additionally
constrain the straight bit-rate at all. See for details.Note that the policing action is to introduce a throttle (delay through traffic) immediately upstream of the congestion policer. This throttle could include a queue with its own AQM, which potentially increases the whole path congestion. In effect the congestion policer has moved the congestion earlier in the path, and focused it on one user to protect downstream resources by reducing the congestion in the rest of the path.The most critical aspect of ConEx is the capability to support robust
auditing. It can be assumed that there will be an intrinsic
motivation for users to understate the congestion that they are
causing. Without strong audit functions the ConEx signal is likely
to become inaccurate to the point of being useless. The most
important feature of an encoding design is likely to be the
robustness of the auditing it supports.
The general approach is to compare the volume of ConEx signals to
direct measures of actual congestion volume. The credit approach
described in can be used to guarantee that this is a
strict bound: if the actual congestion exceeds the ConEx signal, then
some congestion was understated and some sanction should be applied
to the traffic. Although sanctions are beyond the scope of this
document, an example sanction might be to throttle the traffic
immediately upstream of the auditor to prevent the user from getting
any advantage by understating congestion. Such a throttle would
likely include some combination of delaying or dropping traffic.
A ConEx auditor might use one of the following techniques:
Directly observe and compare the volume of ECN and ConEx marks.
Since the volume of ECN marks rises monotonically along a path, ECN auditing is most accurate when located near the transport receiver. For this reason ECN should be monitored downstream of the predominant bottleneck.
For non-encrypted standard TCP traffic on a single path, an auditor could measure losses by
detecting retransmissions, which appear as duplicate sequence numbers upstream of the loss and out of order data downstream of the loss. Since some reordering is present in the Internet, such a loss estimator would be most accurate near the sender.
For networks designed so that losses predominantly occur due to Active Queue Management
under the control of one IP-aware node on the path, the
auditor could be located at this bottleneck. It could simply
compare ConEx Signals with actual local packet discards (and ECN marks). This is a good
model for most consumer access networks where audit accuracy could
well be sufficient even if losses occasionally occur elsewhere
in the network.
Although the auditor at the predominant
bottleneck would not be able to count losses at other
nodes, transports would not know where losses were occurring
either. Therefore a transport would not know which losses it
could cheat and which ones it couldn't without getting caught.
For congestion signaled by loss, totally accurate auditing is not
believed to be possible in the general case, because it involves a
network node detecting the absence of some packets, when it cannot
necessarily identify retransmissions or missing packets. Furthermore the missing packet might simply be taking a different route.
It is for this reason that it is desirable to motivate the deploying of ECN, even though ECN is not strictly required for ConEx.
In addition, other audit techniques may be identified in the future. gives a comprehensive inventory of attacks against audit proposed by various people. It includes pseudocode for both deterministic and statistical audit functions designed to thwart these attacks and analyses the effectiveness of an implementation. Although this work is specific to the re-ECN protocol, most of the material is useful for designing and assessing audit of other specific ConEx encodings, against both ECN and loss.
The auditing function should be able to trigger sufficient sanction to discourage understating congestion . This
seems to require designing the sanction in concert with the
policy functions, even though they might be implemented in different
parts of the network. However, proves audit and policy functions can be independent as long as audit drops sufficient traffic to 'normalise' actual congestion signals to be no greater than ConEx signals. Note that in the future it might prove to be
desirable to provide advice on uniformly implementing sanctions,
because otherwise insufficient sanctions impairs the ability to implement
policy elsewhere in the network.Some of the audit algorithms require per flow state. This cost is expected to be tolerable, because these techniques are most apropos near the edges of the network, where traffic is generally much less aggregated, so the state need not overwhelm any one device.Holding flow-state seems to create a vulnerability to attacks that exhaust the auditor's memory by opening numerous new short flows. The audit function can protect itself from this attack by not allocating new flow-state unless a ConEx-marked packet arrives (e.g. credit at the start of a flow). Because policy devices rate limit ConEx-marked packets, this sets a natural limit to the rate at which a source can create flow-state in audit devices.Auditing can be distributed and redundant. One flow may be audited in multiple places, using multiple techniques. Some audit techniques do not require any per flow state and can be applied to aggregate traffic. These might be able to detect the presence of understated congestion at large scale and support recursively hunting for individual flows that are understating their congestion. Even at large scales, flows can be randomly selected for individual auditing.Sampling techniques can also be used to bound the total auditing memory footprint, although the implementer must be wary of "identifier white washing when caught" tactics where a source cheats until caught by sampling, then simply discards that flow ID and starts cheating with a new one.At the audit function, there will be an inherent delay of at least one round trip between a congestion signal and the subsequent ConEx signal it triggers, as shown in . However, the audit function cannot be expected to wait for a round trip to check that one signal balances the other, because that requires excessive state and the auditor can't easily determine the RTT of each flow. The simplest mechanism to compensate for the round trip delay between the signals is to have the sender include a "credit" signal to cover the yet to be observed congestion that might occur during this delay. The transport signals sufficient credit in advance to cover congestion expected during its feedback delay. Then, the audit function does not need to make allowance for round trip delays that it cannot quantify. This design choice correctly makes the transport responsible for both minimizing feedback delay and for the risk that packets in flight will cause congestion to others before the source can react.Making the source responsible for allowing for the round trip delay in ConEx signals is a design choice that needs to be consistently applied. Any such requirement SHOULD be specified in a particular ConEx encoding specification.For example, imagine the audit function keeps a running account of the balance between actual congestion signals (loss or ECN), which it counts as negative, and ConEx signals, which it counts as positive. Having made the transport responsible for round trip delays, it will be expected to have pre-loaded the audit function with some credit at the start. Therefore, if the balance ever goes negative, the audit function can immediately start punishing a flow, without any grace period.The ConEx abstract protocol described so far is intended to support
incremental deployment in every possible respect. For convenience, the
following list collects together all the features of ConEx that support
incremental deployment, and points to further information on each:The wire protocol encoding allows each packet
to indicate whether it is using ConEx or not (see on
).ConEx requires a modification to the source
in order to send ConEx packet markings (see ). Although ConEx support can be
indicated on a packet-by-packet basis, it is likely that all the
packets in a flow will either consistently support ConEx or
consistently not. It is also likely that, if the implementation of a
transport protocol supports ConEx, all the packets sent from that
host using that protocol will be ConEx marked. The implementations of some of the transport
protocols on a host might not support ConEx (e.g. the implementation
of DNS over UDP might not support ConEx, while perhaps RTP over UDP
and TCP will). Any non-upgraded transports and non-upgraded hosts
will simply continue to send regular Not-ConEx packets as
always.A network operator can create
incentives for senders to voluntarily reveal ConEx information.
Without ConEx information, a network operator tends to have to limit
the bit-rate or volume from a site more than is necessary, just in
case it might congest others. With ConEx information, the operator
can solely limit congestion-causing traffic, and otherwise allow
complete freedom. This greater freedom acts as an inducement for the
source to volunteer ConEx information. An operator may also monitor whether a
source transport has sent ConEx packets, and treat the same transport with
greater suspicion (e.g. a more stringent rate-limit) whenever it selectively
sends packets without ConEx support.
A ConEx source should be able to work
without a modified receiver. However, without sufficiently precise
congestion feedback from the receiver, the source may have to
conservatively send extra ConEx markings in order to avoid
understating congestion. The need for more precise receiver feedback
is not exclusive to ConEx, for instance Data Centre TCP (DCTCP ) uses precise feedback to good effect.
Nonetheless, if a receiver offers precise feedback, it will be best
if ConEx uses it (see ).Although it was stated above that ConEx
requires a modification to the source, ConEx signals could
theoretically be introduced by a proxy for the source, as long as it
can intercept feedback from the receiver. Similarly, more precise
feedback could thoretically be provided by a proxy for the receiver
rather than modifying the receiver itself.No modification to forwarding or queuing is needed for
ConEx.
However, once ConEx is deployed, it
is possible that a queue implementation could optionally take advantage of the
ConEx information in packets. For instance, it has been suggested
that a queue
would be more robust against flooding if it preferentially discarded
Not-ConEx packets then Not-Marked ConEx packets.A ConEx sender re-echoes congestion whether the
queues signaling congestion are ECN-enabled or not. Nonetheless,
auditing works best if most congestion is indicated by ECN rather
than loss (see ). Also,
monitoring rest-of-path congestion is not accurate if there are
congested non-ECN queues upstream of the monitoring point ().If a subset of traffic sources (or proxies)
use ConEx signals to reveal congestion in the internetwork layer, a
network operator can choose (or not) to use this information for
traffic management. As long as the end-to-end ConEx signals are
present, each network can unilaterally choose to use
them—independently of whether other networks do. ConEx marked packets may safely traverse a network that
ignores them. Networks MUST NOT change ConEx marked packets to Not-ConEx.
If necessary, endpoints SHOULD be able to detect if a network is
removing ConEx signals.An operator can
deploy policy devices () wherever traffic enters
its network, in order to monitor the downstream congestion that
incoming traffic contributes to, and control it if necessary. See
for further
discussion of deployment incentives for networks and scenarios where
some networks use ConEx-based policy devices and others don't.
An operator can deploy audit devices () unilaterally within its own network
to verify that traffic sources are not understating ConEx
information. From the viewpoint of one network operator (say N_a),
it only cares that the level of ConEx signaling is sufficient to
cover congestion in its own network. If traffic continues into a
congested downstream network (say N_b), it is of no concern to the
first network (N_a) if the end-to-end ConEx signaling is
insufficient to cover the congestion in N_b as well. This is N_b's
concern, and N_b can both detect such anomalous traffic and deal
with it using ConEx-based policy devices ().This memo includes no request to IANA.Note to RFC Editor: this section may be removed on publication as an
RFC.The only known risk associated with ConEx is that users and applications are very likely to be motivated to under-represent the congestion that they are causing. Significant portions of this document are about mechanisms to audit the ConEx signals and create sufficient sanction to inhibit such under-representation. In particular see .Security attacks and their defences are best discussed against a concrete protocol specification, not the abstract mechanism of this document. A concrete ConEx protocol will need to be accompanied by a document describing how the protocol and its audit mechanisms defend against likely attacks.
will be a useful source for such a document. It gives a comprehensive inventory of attacks against audit that have been proposed by various parties. It includes pseudocode for both deterministic and statistical audit functions designed to thwart these attacks and analyses the effectiveness of an implementation.However, is specific to the re-ECN protocol, which signalled ECN & loss together, whereas ConEx signals them separately. Therefore, although likely attacks will be similar, there will be more combinations of attacks to worry about, and defences and their analysis are likely to be a little different for ConEx.The main known attacks that a security document for a concrete ConEx protocol will need to address are listed below, and should be referred to for how re-ECN was designed to defend against similar attacks:
Attacks on the audit function (see Section 7.5 of ):
Designing the audit function so that a source cannot gain from starting a new flow once audit has detected cheating in a previous flow.
Avoiding audit discarding packets from all flows within an aggregate, which would allow one flow to pull down the average so that the audit function would discard packets from all flows, not just the offending flow.
An attacker understates ConEx markings in packets that spoof another flow, which fools the audit function into dropping the genuine user's packets.
Attacks by networks on other networks (see Section 8.2 of ):
Sending dummy traffic across a border with understated ConEx markings to bring down the average ConEx markings in the aggregate of border traffic. This attack can be combined with a TTL that expires before the packets reach an audit function.
Sending high volumes of valid packets that are both ConEx-Marked and ECN-Marked, which seems to represent congestion upstream, but it makes these packets immune to being further ECN-Marked downstream.
It is planned to document all known attacks and their defences (including all the above) in the RFC series. In the interim, and its references should be referred to for details and ways to address these attacks.
This document was improved by review comments from Toby
Moncaster, Nandita Dukkipati, Mirja Kuehlewind, Caitlin Bestler and John Leslie.Comments and questions are encouraged and very welcome. They can be
addressed to the IETF Congestion Exposure (ConEx) working group mailing
list <conex@ietf.org>, and/or to the authors.
Data Center TCP (DCTCP)
Re-feedback: Freedom with Accountability for Causing
Congestion in a Connectionless InternetworkBT & UCLPolicing Congestion Response in an Internetwork Using
Re-FeedbackBT & UCLBTBTEurécom & BTBTBTA Fairer, Faster Internet ProtocolBT & UCL
The Social Cost of Cheap Pseudonyms
Resource pricing and the evolution of congestion control
Cam UniCam UniPolicing Freedom to Use the Internet Resource PoolBTBT & UCLBT
Closed Loop Traffic Policing
Eurécom & BT