Emulating Border Flow
Policing using Re-PCN on Bulk DataBTB54/77, Adastral ParkMartlesham HeathIpswichIP5 3REUK+44 1473 645196bob.briscoe@bt.comhttp://bobbriscoe.net/
Transport
PCN Working GroupQuality of ServiceQoSCongestion ControlDifferentiated ServicesIntegrated ServicesAdmission Control PolicingFlow Rate PolicingInter-domainTrustTheft of ServiceSignallingProtocolCongestion NotificationScalabilityScaling per flow admission control to the Internet is a hard problem.
The approach of combining Diffserv and pre-congestion notification (PCN)
provides a service slightly better than Intserv controlled load that
scales to networks of any size without needing Diffserv's usual
overprovisioning, but only if domains trust each other to comply with
admission control and rate policing. This memo claims to solve this
trust problem without losing scalability. It provides a sufficient
emulation of per-flow policing at borders but with only passive bulk
metering rather than per-flow processing. Measurements are sufficient to
apply penalties against cheating neighbour networks.The IETF PCN working group is initially chartered to consider PCN
domains only under a single trust authority. However, after its initial
work is complete the charter says the working group may re-charter to
consider concatenated Diffserv domains, amongst other new work items.
The charter ends by stating "The details of these work items are outside
the scope of the initial phase; but the WG may consider their
requirements to design components that are sufficiently general to
support such extensions in the future."This memo is therefore contributed to describe how PCN could be
extended to inter-domain. We wanted to document the solution to reduce
the chances that something else eats up the codepoint space needed
before PCN re-charters to consider inter-domain. Losing the chance to
standardise this simple, scalable solution to the problem of
inter-domain flow admission control would be unfortunate
(understatement), given it took years to find, and even then it was very
difficult to find codepoint space for it.The scheme described here ()
requires the PCN ingress gateway to re-echo any PCN feedback it receives
back into the forward stream of IP packets (hence we call this scheme
re-PCN). Re-PCN works in a very similar way to the re-ECN proposal on
which it is based , the
only difference being that PCN might encode three states of congestion,
whereas ECN encodes two. This document is written to stand alone from
re-ECN, so that readers do not have to read .The authors seek comments from the Internet community on whether
combining PCN and re-ECN to create re-PCN in this way is a sufficient
solution to the problem of scaling microflow admission control to the
Internet as a whole. Here we emphasise that scaling is not just an issue
of numbers of flows, but also the number of security
entities—networks and users—who may all have conflicting
interests.This memo is posted as an Internet-Draft with the intent to
eventually be broken down in two documents; one for the standards track
and one for informational status. But until it becomes an item of IETF
working group business the whole proposal has been kept together to aid
understanding. Only the text of
of this document is intended to be normative (requiring
standardisation). The rest of the sections are merely informative,
describing how a system might be built from these protocols by the
operators of an internetwork. Note in particular that the policing and
monitoring functions proposed for the trust boundaries between operators
would not need standardisation by the IETF. They simply represent one
possible way that the proposed protocols could be used to extend the PCN
architecture to span multiple domains without
mutual trust between the operators.To realise the system described, this document also depends on other
documents chartered in the IETF Transport Area progressing along the
standards track: Pre-congestion notification (PCN) marking on interior nodes , chartered for
standardisation in the PCN w-g;The baseline encoding of pre-congestion notification in the IP
header , also
chartered for standardisation in the PCN w-g;Feedback of aggregate PCN measurements by suitably extending the
admission control signalling protocol (e.g. RSVP extension or NSIS extension ).The baseline encoding makes no new demands on codepoint space in the
IP header but provides just two PCN encoding states (not marked and
marked). The PCN architecture recognises that operators might want PCN
marking to trigger two functions (admission control and flow
termination) at different levels of pre-congestion, which seems to
require three encoding states. A scheme has been proposed that can do both functions
with just two encoding states, but simulations have shown it performs
poorly under certain conditions that might be typical. As it seems
likely that PCN might need three encoding states to be fully
operational, we want to be sure that three encoding states can be
extended to work inter-domain. Therefore, we have defined a three-state
extension encoding scheme in this document, then we have added the
re-PCN scheme to it. The three-state encoding we have chosen depends on
standardisation of yet another document in the IETF Transport Area:Propagation beyond the tunnel decapsulator of any changes in the
ECN field to ECT(0) or ECT(1) made within a tunnel (the ideal
decapsulation rules of );Full diffs of incremental changes between drafts are available at
URL: <http://www.cs.ucl.ac.uk/staff/B.Briscoe/pubs.html#repcn>Updated
references and other minor changes.Considerably updated the 'Status' note to explain the
relationship of this draft to other documents in the IETF
process (or not) and to chartered PCN w-g activity.Split out the dependencies into a separate note and added
dependencies on new PCN documents in progress.Made scalability motivation in the introduction clearer,
explaining why Diffserv over-provisioning doesn't scale unless
PCN is used.Clarified that the standards action in is to define the meanings of
the combination of fields in the IP header: the RE flag and
2-level congestion marking in the ECN field. And that it is not
characterised by a particular feedback style in the
transport.Switched round the two ECT codepoints to be compatible with
the new PCN baseline encoding and used less confusing naming for
re-PCN codepoints ().Generalised rules for encoding probes when bootstrapping or
re-starting aggregates & flows ().Downgraded drop sanction behaviour from MUST to conditional
SHOULD ().Added incremental deployment safety justification for choice
of which way round the RE flag works ().Added possible vulnerability to brief attacks and possible
solution to security considerations ().Updated references and terminology, particularly taking
account of recent new PCN w-g documents;Replaced suggested Ingress Gateway Algorithm for Blanking the
RE flag ()Clarifications throughout;Updated references.Changed filename to associate it with the new IETF PCN w-g,
rather than the TSVWG w-g.Introduction: Clarified that bulk policing only replaces
per-flow policing at interior inter-domain borders, while
per-flow policing is still needed at the access interface to the
internetwork. Also clarified that the aim is to neutralise any
gains from cheating using local bilateral contracts between
neighbouring networks, rather than merely identifying remote
cheaters.: Described the
traditional per-flow policing problem with inter-domain
reservations more precisely, particularly with respect to
direction of reservations and of traffic flows.Clarified status of onwards, in
particular that policers and monitors would not need
standardisation, but that the protocol in would require
standardisation. on competitive
routing: Added discussion of direct incentives for a receiver to
switch to a different provider even if the provider has a
termination monopoly.Clarified that "Designing in security from the start" merely
means allowing codepoint space in the PCN protocol encoding.
There is no need to actually implement inter-domain security
mechanisms for solutions confined to a single domain.Updated some references and added a ref to the Security
Considerations, as well as other minor corrections and
improvements.Added subsection on Border Accounting Mechanisms ()
on the re-ECN wire protocol clarified and re-organised to
separately discuss re-ECN for default ECN marking and for
pre-congestion marking (PCN).Router Forwarding Behaviour subsection added to re-organised
section on Protocol Operation (). Extensions section moved
within Protocol Operations.Emulating Border Policing () reorganised,
starting with a new Terminology subsection heading, and a
simplified overview section. Added a large new subsection on
Border Accounting Mechanisms within a new section bringing
together other subsections on Border Mechanisms generally (). Some text moved from old
subsections into these new ones.Added section on Incremental Deployment (), drawing together relevant points
about deployment made throughout.Sections on Design Rationale () and Security Considerations () expanded with some
new material, including new attacks and their defences.Suggested Border Metering Algorithms improved () for resilience to newly
identified attacks.The Internet community largely lost interest in the Intserv
architecture after it was clarified that it would be unlikely to scale
to the whole Internet . Although Intserv
mechanisms proved impractical, the bandwidth reservation service it
aimed to offer is still very much required.A recently proposed approach combines
Diffserv and pre-congestion notification (PCN) to provide a service
slightly better than Intserv controlled load . PCN does not require the considerable
over-provisioning that is normally required for admission control over
Diffserv to be robust against re-routes or
variation in the traffic matrix. It has been proved that Diffserv's
over-provisioning requirement grows linearly with the network diameter
in hops .A number of PCN domains can be concatenated into a larger PCN region
without any per-flow processing between them, but only if each domain
trusts the ingress network to have checked that upstream customers
aren't taking more bandwidth than they reserved, either accidentally or
deliberately. Unfortunately, networks can gain considerably by breaking
this trust. One way for a network to protect itself against others is to
handle flow signalling at its own border and police traffic against
reservations itself. However, this reintroduces the per-flow
unscalability at borders that Intserv over Diffserv suffers from.This memo describes a protocol called re-PCN that enables bulk border
measurements so that one network can protect its interests, even if
networks around it are deliberately trying to cheat. The approach
provides a sufficient emulation of flow rate policing at trust
boundaries but without per-flow processing. Per-flow rate policing for
each reservation is still expected to be used at the access edge of the
internetwork, but at the borders between networks bulk policing can be
used to emulate per-flow policing. The emulation is not perfect, but it
is sufficient to ensure that the punishment is at least proportionate to
the severity of the cheat. Re-PCN neither requires the unscalable
over-provisioning of Diffserv nor the per-flow processing at borders of
Intserv over Diffserv.It should therefore scale controlled load service to the whole
internetwork without the cost of Diffserv's linearly increasing
over-provisioning, or the cost of per-flow policing at each border. To
achieve such scaling, this memo combines two recent proposals, both of
which it briefly recaps: The pre-congestion notification (PCN) architecture describes how bulk pre-congestion notification
on routers within an edge-to-edge Diffserv region can emulate the
precision of per-flow admission control to provide controlled load
service without unscalable per-flow processing;Re-ECN: Adding Accountability to TCP/IP .We coin the term re-PCN for the combination of PCN and
re-ECN.The trick that addresses cheating at borders is to recognise that
border policing is mainly necessary because cheating upstream networks
will admit traffic when they shouldn't only as long as they don't
directly experience the downstream congestion their misbehaviour can
cause. The re-ECN protocol ensures a network can be made to experience
the congestion it causes in other networks. Re-ECN requires the sending
node to declare expected downstream congestion in all packets and it
makes it in its interest to declare this honestly. At the border between
upstream network 'A' and downstream network 'B' (say), both networks can
monitor packets crossing the border to measure how much congestion 'A'
is causing in 'B' and beyond. 'B' can then include a limit or penalty
based on this metric in its contract with 'A'. This is how 'A'
experiences the effect of congestion it causes in other networks. 'A' no
longer gains by admitting traffic when it shouldn't, which is why we can
say re-PCN emulates flow policing, even though it doesn't measure
flows.The aim is not to enable a network to identify some remote cheating party, which would
rarely be useful given the victim network would be unlikely to be able
to seek redress from a cheater in some remote part of the world with
whom no direct contractual relationship exists. Rather the aim is to
ensure that any gain from cheating will be cancelled out by penalties
applied to the cheating party by its local network. Further, the
solution ensures each of the chain of networks between the cheater and
the victim will lose out if it doesn't apply penalties to its neighbour.
Thus the solution builds on the local bilateral contractual
relationships that already exist between neighbouring networks.Rather than the end-to-end arrangement used when re-ECN was specified
for the TCP transport , this memo specifies re-ECN in
an edge-to-edge arrangement, making it applicable to deployment models
where admission control over Diffserv is based on pre-congestion
notification. Also, rather than using a TCP transport for regular
congestion feedback, this memo specifies re-ECN using RSVP as the
transport for feedback . RSVP is used to
be concrete, but a similar deployment model, but with a different
transport for signalling congestion feedback could be used (e.g.
Arumaithurai and
RMD both use NSIS).This memo aims to do two things: i) define how to apply the re-PCN
protocol to the admission control over Diffserv scenario; and ii)
explain why re-PCN sufficiently emulates border policing in that
scenario. Most of the memo is taken up with the second aim; explaining
why it works. Applying re-PCN to the scenario actually involves quite a
trivial modification to the ingress gateway. That modification can be
added to gateways later, so our immediate goal is to convince everyone
to have the foresight to define the PCN wire protocol encoding to
accommodate the extended codepoints defined in this document, whether
first deployments require border policing or not. Otherwise, when we
want to add policing, we will have built ourselves a legacy problem. In
other words, we aim to convince people to "Design in security from the
start."The body of this memo is structured as follows: describes the border policing
problem. We recap the traditional, unscalable view of how to solve
the problem, and we recap the admission control solution which has
the scalability we do not want to lose when we add border
policing; specifies the re-PCN
protocol solution in detail; explains how to
use the protocol to emulate border policing, and why it works; analyses the security of the
proposed solution; explains the sometimes subtle
rationale behind our design decisions; comments on the
overall robustness of the security assumptions and lists specific
security issues.It must be emphasised that we are not evangelical about removing
per-flow processing from borders. Network operators may choose to do
per-flow processing at their borders for their own reasons, such as to
support business models that require per-flow accounting. Our aim is to
show that per-flow processing at borders is no longer necessary in order to provide end-to-end QoS using
flow admission control. Indeed, we are absolutely opposed to
standardisation of technology that embeds particular business models
into the Internet. Our aim is merely to provide a new useful metric
(downstream congestion) at trust boundaries. Given the well-known
significance of congestion in economics, operators can then use this new
metric in their interconnection contracts if they choose. This will
enable competitive evolution of new business models (for examples
see ), even for sets of flows running
alongside another set across the same border but using the more
traditional model that depends on more costly per-flow processing at
each border.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 .If we claim to be able to emulate per-flow policing with bulk
policing at trust boundaries, we need to know exactly what we are
emulating. So, we will start from the traditional scenario with
per-flow policing at trust boundaries to explain why it has always
been considered necessary.To be able to take advantage of a reservation-based service such as
controlled load, a source-destination pair must reserve resources
using a signalling protocol such as RSVP . An RSVP signalling request refers to a flow of
packets by its flow ID tuple (filter spec ) (or its security parameter index (SPI) if port numbers are hidden by IPSec encryption).
Other signalling protocols use similar flow identifiers. But, it is
insufficient to merely authorise and admit a flow based on its
identifiers, for instance merely opening a pin-hole for packets with
identifiers that match an admitted flow ID. Because, once a flow is
admitted, it cannot necessarily be trusted to send packets within the
rate profile it requested.The packet rate must also be policed to keep the flow within the
requested flow spec . For instance,
without data rate policing, a source-destination pair could reserve
resources for an 8kbps audio flow but the source could transmit a
6Mbps video (theft of service). More subtly, the sender could generate
bursts that were outside the profile requested.In traditional architectures, per-flow packet rate-policing is
expensive and unscalable but, without it, a network is vulnerable to
such theft of service (whether malicious or accidental). Perhaps more
importantly, if flows are allowed to send more data than they were
permitted, the ability of admission control to give assurances to
other flows will break.Just as sources need not be trusted to keep within the requested
flow spec, whole networks might also try to cheat. We will now set up
a concrete scenario to illustrate such cheats. Imagine reservations
for unidirectional flows, through at least two networks, an edge
network and its downstream transit provider. Imagine the edge network
charges its retail customers per reservation but also has to pay its
transit provider a charge per reservation. Typically, both the charges
for buying from the transit and selling to the retail customer might
depend on the duration and rate of each reservation. The level of the
actual selling and buying prices are irrelevant to our discussion
(most likely the network will sell at a higher price than it buys, of
course).A cheating ingress network could systematically reduce the size of
its retail customers' reservation signalling requests (e.g. the
SENDER_TSPEC object in RSVP's PATH message) before forwarding them to
its transit provider and systematically reinstate the responses on the
way back (e.g. the FLOWSPEC object in RSVP's RESV message). It would
then receive an honest income from its upstream retail customer but
only pay for fraudulently smaller reservations downstream. A similar
but opposite trick (increasing the TSPEC and decreasing the FLOWSPEC)
could be perpetrated by the receiver's access network if the
reservation was paid for by the receiver.Equivalently, a cheating ingress network may feed the traffic from
a number of flows into an aggregate reservation over the transit that
is smaller than the total of all the flows. Because of these fraud
possibilities, in traditional QoS reservation architectures the
downstream network polices traffic at each border. The policer checks
that the actual sent data rate of each flow is within the signalled
reservation.Reservation signalling could be authenticated end to end, but this
wouldn't prevent the aggregation cheat just described. For this
reason, and to avoid the need for a global PKI, signalling integrity
is typically only protected on a hop-by-hop basis .A variant of the above cheat is where a router in an honest
downstream network denies admission to a new reservation, but a
cheating upstream network still admits the flow. For instance, the
networks may be using Diffserv internally, but Intserv admission
control at their borders . The cheat
would only work if they were using bulk Diffserv traffic policing at
their borders, perhaps to avoid the cost/complexity of Intserv border
policing. As far as the cheating upstream network is concerned, it
gets the revenue from the reservation, but it doesn't have to pay any
downstream wholesale charges and the congestion is in someone else's
network. The cheating network may calculate that most of the flows
affected by congestion in the downstream network aren't likely to be
its own. It may also calculate that the downstream router has been
configured to deny admission to new flows in order to protect
bandwidth assigned to other network services (e.g. enterprise VPNs).
So the cheating network can steal capacity from the downstream
operator's VPNs that are probably not actually congested.All the above cheats are framed in the context of RSVP's receiver
confirmed reservation model, but similar cheats are possible with
sender-initiated and other models.To summarise, in traditional reservation signalling architectures,
if a network cannot trust a neighbouring upstream network to
rate-police each reservation, it has to check for itself that the data
rate fits within each of the reservations it has admitted.We will now describe a generic internetworking scenario that we
will use to describe and to test our bulk policing proposal. It
consists of a number of networks and endpoints that do not fully trust
each other to behave. In we will tie
down exactly what we mean by partial trust, and we will consider the
various combinations where some networks do not trust each other and
others are colluding together.An ingress and egress gateway (Ingr G/W and Egr G/W in ) connect the interior Diffserv region
to the edge access networks where routers (not shown) use per-flow
reservation processing. Within the Diffserv region are three interior
domains, 'A', 'B' and 'C', as well as the inward facing interfaces of
the ingress and egress gateways. An ingress and egress border router
(BR) is shown interconnecting each interior domain with the next.
There will typically be other interior routers (not shown) within each
interior domain.In two paragraphs we now briefly recap how pre-congestion
notification is intended to be used to control flow admission to a
large Diffserv region. The first paragraph describes data plane
functions and the second describes signalling in the control plane. We
omit many details from including behaviour
during routing changes. For brevity here we assume other flows are
already in progress across a path through the Diffserv region before a
new one arrives, but how bootstrap works is described in . shows a single simplex
reserved flow from the sending (Sx) end host to the receiving (Rx) end
host. The ingress gateway polices incoming traffic and colours
conforming traffic within an admitted reservation to a combination of
Diffserv codepoint and ECN field that defines the traffic as
'PCN-enabled'. This redefines the meaning of the ECN field as a PCN
field, which is largely the same as ECN , but
with slightly different semantics defined in (or various extensions that
are currently experimental). The Diffserv region is called a
PCN-region because all the queues within it are PCN-enabled. This
means the per-hop behaviour they apply to PCN-enabled traffic consists
of both a scheduling behaviour and a new ECN marking behaviour that we
call `pre-congestion notification' . A PCN-enabled queue
typically re-uses the definition of expedited forwarding
(EF) for its scheduling behaviour. The
new congestion marking behaviour sets the PCN field of an increasing
proportion of PCN packets to the PCN-marked (PM) codepoint as their load approaches a
threshold rate that is lower than the line rate . This can be achieved with
an algorithm similar to a token-bucket called a virtual queue. The aim
is for a queue to start marking PCN traffic to trigger admission
control before the real queue builds up any congestion delay. The
level of a queue's pre-congestion marking is detected at the egress of
the Diffserv region and used by the signalling system to control
admission of further traffic that would otherwise overload that queue,
as follows.The end-to-end QoS signalling for a new reservation (to be concrete
we will use RSVP) takes one giant hop from ingress to egress gateway,
because interior routers within the Diffserv region are configured to
ignore RSVP. The egress gateway holds flow state because it takes part
in the end-to-end reservation. So it can classify all packets by flow
and it can identify all flows that have the same previous RSVP hop (an
ingress-egress-aggregate). For each ingress-egress-aggregate of flows
in progress, the egress gateway maintains a per-packet moving average
of the fraction of pre-congestion-marked traffic. Once an RSVP PATH
message for a new reservation has hopped across the Diffserv region
and reached the destination, an RSVP RESV message is returned. As the
RESV message passes, the egress gateway piggy-backs the relevant
pre-congestion level onto it . Again,
interior routers ignore the RSVP message, but the ingress gateway
strips off the pre-congestion level. If the pre-congestion level is
above a threshold, the ingress gateway denies admission to the new
reservation, otherwise it returns the original RESV signal back
towards the data sender.Once a reservation is admitted, its traffic will always receive low
delay service for the duration of the reservation. This is because
ingress gateways ensure that traffic not under a reservation cannot
pass into the PCN-region with a Diffserv codepoint that gives it
priority over the capacity used for PCN traffic.Even if some disaster re-routes traffic after it has been admitted,
if the PCN traffic through any PCN resource tips over a higher,
fail-safe threshold, pre-congestion notification can trigger flow
termination to very quickly bring every router within the whole
PCN-region back below its operating point. The same marking process
and ECN codepoint can be used for both admission control and flow
termination, by simply triggering them at different fractions of
marking . However
simulations have confirmed that this approach is not robust in all
circumstances that might typically be encountered, so approaches with
two thresholds and two congestion encodings are expected to be
required in production networks.The whole admission control system just described deliberately
confines per-flow processing to the access edges of the network, where
it will not limit the system's scalability. But ideally we want to
extend this approach to multiple networks, to take even more advantage
of its scaling potential. We would still need per-flow processing at
the access edges of each network, but not at the high speed interfaces
where they interconnect. Even though such an admission control system
would work technically, it would gain us no scaling advantage if each
network also wanted to police the rate of each admitted flow for
itself—border routers would still have to do complex packet
operations per-flow anyway, given they don't trust upstream networks
to do their policing for them.This memo describes how to emulate per-flow rate policing using
bulk mechanisms at border routers. Otherwise the full scalability
potential of pre-congestion notification would be limited by the need
for per-flow policing mechanisms at borders, which would make borders
the most cost-critical pinch-points. Instead we can achieve the long
sought-for vision of secure Internet-wide bandwidth reservations
without over-generous provisioning or per-flow processing. We still
use per-flow processing at the edge routers closest to the end-user,
but we need no per-flow processing at all in core or border routers—where scalability is most
critical.First we need to recap the way routers accumulate PCN congestion
marking along a path (it accumulates the same way as ECN). Each
PCN-capable queue into a link might mark some packets with a
PCN-marked (PM) codepoint, the marking probability increasing with the
length of the queue .
With a series of PCN-capable routers on a path, a stream of packets
accumulates the fraction of PCN markings that each queue adds. The
combined effect of the packet marking of all the queues along the path
signals congestion of the whole path to the receiver. So, for example,
if one queue early in a path is marking 1% of packets and another
later in a path is marking 2%, flows that pass through both queues
will experience approximately 3% marking over a sequence of
packets.(Note: Whenever the word 'congestion' is used in this document it
should be taken to mean congestion of the virtual resource assigned
for use by PCN-traffic. This avoids cumbersome repetition of the
strictly correct term 'pre-congestion'.)The packets crossing an inter-domain trust boundary within the
PCN-region will all have come from different ingress gateways and will
all be destined for different egress gateways. We will show that the
key to policing against theft of service is for a border router to be
able to directly measure the congestion that is about to be caused by
the packets it forwards into any of the downstream paths between
itself and the egress gateways that each packet is destined for. The
purpose of the re-PCN protocol is to make packets automatically carry
this information, which then merely needs to be counted locally at the
border.With the original PCN protocol, if a border router, e.g. that
between domains 'A' & 'B' ), counts PCN markings crossing
the border over a period, they represent the accumulated congestion
that has already been experienced by those packets (congestion
upstream of the border, u). The idea of re-PCN is to make the ingress
gateway continuously encode the path congestion it knows into a new
field in the IP header (in this case, `path' means the path from the
ingress to the egress gateway). This new field is not altered by queues along the path. Then at any
point on that path (e.g. between domains 'A' & 'B'), IP headers
can be monitored to measure both expected path congestion, p and
upstream congestion, u. Then congestion expected downstream of the
border, v, can be derived simply by subtracting upstream congestion
from expected path congestion. That is v ~= p - u.Importantly, it turns out that there is no need to monitor
downstream congestion on a per-flow, per-path or per-aggregate basis.
We will show that accounting for it in bulk by counting the volume of
all marked packet will be sufficient.In this section we define the names of the various codepoints of
the extended ECN field when used with pre-congestion notification,
deferring description of their semantics to the following sections.
But first we recap the re-ECN wire protocol proposed in .Re-ECN uses the two bit ECN field broadly as in
RFC3168 . It also uses a new re-ECN
extension (RE) flag. The actual position of the RE flag is different
between IPv4 & v6 headers so we will use an abstraction of the
IPv4 and v6 wire protocols by just calling it the RE flag. proposes using bit 48
(currently unused) in the IPv4 header for the RE flag, while for
IPv6 it proposes an congestion extension header.Unlike the ECN field, the RE flag is intended to be set by the
sender and remain unchanged along the path, although it can be read
by network elements that understand the re-ECN protocol. In the
scenario used in this memo, the ingress gateway is the 'sender' as
far as the scope of the PCN region is concerned, so it sets the RE
flag (as permitted for sender proxies in the specification of
re-ECN).Note that general-purpose routers do not have to read the RE
flag, only special policing elements at borders do. And no
general-purpose routers have to change the RE flag, although the
ingress and egress gateways do because in the edge-to-edge
deployment model we are using, they act as the endpoints of the PCN
region. Therefore the RE flag does not even have to be visible to
interior routers. So the RE flag has no implications on protocols
like MPLS. Congested label switching routers (LSRs) would have to be
able to notify their congestion with an ECN/PCN codepoint in the
MPLS shim , but like any interior IP
router, they can be oblivious to the RE flag, which need only be
read by border policing functions.Although the RE flag is a separate single bit field, it can be
read as an extension to the two-bit ECN field; the three
concatenated bits in what we will call the extended ECN field (EECN)
make eight codepoints available. When the RE flag setting is "don't
care", we use the RFC3168 names of the ECN codepoints, but proposes the following six
codepoint names for when there is a need to be more specific.ECN fieldRFC3168 codepointRE flagExtended ECN codepointRe-ECN meaning00Not-ECT0Not-RECTNot re-ECN-capable transport00Not-ECT1FNEFeedback not established10ECT(0)0---Legacy ECN use only 10ECT(0)1--CU--Currently unused
01ECT(1)0Re-EchoRe-echoed congestion and RECT01ECT(1)1RECTRe-ECN capable transport11CE0CE(0)Congestion experienced with Re-Echo11CE1CE(-1)Congestion experiencedAs permitted by the ECN specification and by the guidelines for specifying alternative
semantics for the ECN field , a proposal is
currently being advanced in the IETF to define different semantics
for how queues might mark the ECN field of certain packets. The idea
is to be able to notify congestion when the queue's load approaches
a logical limit, rather than the physical limit of the line. This
new marking is called pre-congestion notification and we will use the term
PCN-enabled queue for a queue that can apply pre-congestion
notification marking to the ECN fields of packets. recommends that a packet's Diffserv
codepoint should determine which type of ECN marking it receives. A
PCN-capable packet must meet two conditions; it must carry a DSCP
that has been associated with PCN marking and it must carry an ECN
field that turns on PCN marking.As an example, a packet carrying the VOICE-ADMIT DSCP would be
associated with expedited forwarding as
its scheduling behaviour and pre-congestion notification as its
congestion marking behaviour. PCN would only be turned on within a
PCN-region by an ECN codepoint other than Not-ECT (00). Then we
would describe packets with the VOICE-ADMIT DSCP and with ECN turned
on as PCN-capable packets. actually
proposes that two logical limits can be used for pre-congestion
notification, with the higher limit as a back-stop for dealing with
anomalous events. It envisages PCN will be used to admission control
inelastic real-time traffic, so marking at the lower limit will
trigger admission control, while at the higher limit it will trigger
flow termination.Because it needs two types of congestion marking, PCN needs four
states: Not PCN-capable (Not-PCN), PCN-capable but not PCN-marked
(NM), Admission Marked (AM) and Flow Termination Marked (TM). A
proposed encoding of the four required PCN states is shown on the
left of . Note that
these codepoints of the ECN field only take on the semantics of
pre-congestion notification if they are combined with a Diffserv
codepoint that the operator has configured to be associated with PCN
marking.This encoding only correctly traverses an IP in IP tunnel if the
ideal decapsulation rules in are followed when combining
the ECN fields of the outer and inner headers. If instead the
decapsulation rules in or are followed, any admission marking applied to
an outer header will be incorrectly removed on decapsulation at the
tunnel egress.The RFC3168 ECN field includes space for the experimental ECN
Nonce , which seems to require a fifth
state if it is also needed with re-PCN. But re-PCN supersedes any
need for the Nonce within the PCN-region. The ECN Nonce is an
elegant scheme, but it only allows a sending node (or its proxy) to
detect suppression of congestion marking in the feedback loop. Thus
the Nonce requires the sender (or in our case the PCN ingress) to be
trusted to respond correctly to congestion. But this is precisely
the main cheat we want to protect against (as well as many others).
Also, the ECN nonce only works once the receiver has placed packets
in the same order as they left the ingress, which cannot be done by
an edge node without adding unnecessary edge-edge packet ordering.
Nonetheless, if the ECN nonce were in use outside the PCN region
(end-to-end), the ingress would have to tunnel the arriving IP
header across the PCN region ().For the rest of this memo, to mean either Admission Marking or
Termination Marking we will call both "congestion marking" or "PCN
marking" unless we need to be specific. With the above encoding,
congestion marking can be read to mean any packet with the
right-most bit of the ECN field set.The re-ECN protocol can be used to control misbehaving sources
whether congestion is with respect to a logical threshold (PCN) or
the physical line rate (ECN). In either case the RE flag can be used
to create an extended ECN field. For PCN-capable packets, the 8
possible encodings of this 3-bit extended PCN (EPCN) field are
defined on the right of below. The purposes of
these different codepoints will be introduced in subsequent
sections.ECN fieldPCN codepointRE flagExtended PCN codepointRe-PCN meaning00Not-PCN0Not-PCNNot PCN-capable transport00Not-PCN1FNEFeedback not established10NM0Re-PCT-EchoRe-echoed congestion and Re-PCT10NM1Re-PCTRe-PCN capable transport01AM0AM(0)Admission Marking with Re-Echo01AM1AM(-1)Admission Marking 11TM0TM(0)Termination Marking with Re-Echo11TM1TM(-1)Termination MarkingNote that shows
re-PCN uses ECT(0) but shows re-ECN uses
ECT(1) for the unmarked state. The difference is
intended—although it makes it harder to remember the two
schemes, it makes them both safer during incremental deployment.The re-PCN protocol involves a simple addition to the action of
the gateway at the ingress edge of the PCN region (the
PCN-ingress-node). But first we will recap how PCN works without the
addition. For each active traffic aggregate across a PCN region
(ingress-egress-aggregate) the egress gateway measures the level of
PCN marking and feeds it back to the ingress piggy-backed as
'PCN-feedback-information' on any control signal passing between the
nodes (e.g. every flow set-up, refresh or tear-down). Therefore the
ingress gateway will always hold a fairly recent (typically at most
30sec) estimate of the ingress-egress-aggregate congestion level.
For instance, one aggregate might have been experiencing 3%
pre-congestion (that is, congestion marked octets whether Admission
Marked or Termination Marked).To comply with the re-PCN protocol, for all PCN packets in each
ingress-egress-aggregate the ingress gateway MUST clear the RE flag
to 0 for the same percentage of octets
as its current estimate of congestion on the aggregate (e.g. 3%) and
set it to 1 in the rest (97%). gives a simple pseudo-code
algorithm that the ingress gateway may use to do this.The RE flag is set and cleared this way round for incremental
deployment reasons (see ). To
avoid confusion we will use the term `blanking' (rather than
marking) when the RE flag is cleared to 0, so we will talk of the `RE blanking
fraction' as the fraction of octets with the RE flag cleared to
0.
illustrates our example. The horizontal axis represents the index of
each congestible resource (typically queues) along a path through
the Internet. The two superimposed plots show the fraction of each
extended PCN codepoint observed along this path, assuming there are
two congested routers somewhere within domains A and C. And below shows the
downstream pre-congestion measured at various border observation
points along the path. (later) shows the same
results of these subtractions, but in graphical form like the above
figure. The tabulated figures are actually reasonable approximations
derived from more precise formulae given in Appendix A of . The RE flag is not changed
by interior routers, so it can be seen that it acts as a reference
against which the congestion marking fraction can be compared along
the path.Border observation pointApproximate Downstream
pre-congestioningress -- A3% - 0% = 3%A -- B3% - 1% = 2%B -- C3% - 1% = 2%C -- egress3% - 3% = 0%Note that the ingress determines the RE blanking fraction for
each aggregate using the most recent feedback from the relevant
egress, arriving with each new reservation, or each refresh. These
updates arrive relatively infrequently compared to the speed with
which congestion changes. Although this feedback will always be out
of date, on average positive errors should cancel out negative over
a sufficiently long duration.In summary, the network adds pre-congestion marking in the
forward data path, the egress feeds its level back to the ingress in
RSVP (or similar signalling), then the ingress gateway re-echoes it
into the forward data path by blanking the RE flag. Then at any
border within the PCN-region, the pre-congestion marking that every
passing packet will be expected to experience downstream can be
measured to be the RE blanking fraction minus the congestion marking
fraction.When a new reservation PATH message arrives at the egress, if
there are currently no flows in progress from the same ingress,
there will be no state maintaining the current level of
pre-congestion marking for the aggregate. In the case of RSVP
reservation signalling, while the signal continues onward towards
the receiving host, the egress gateway can return an RSVP message to
the ingress with a flag asking the
ingress to send a specified number of data probes between them. The
more general possibilities for bootstrap behaviour are described in
the PCN architecture , including
using the reservation signal itself as a probe.However, with our new re-PCN scheme, the ingress does not know
what proportion of the data probes should have the RE flag blanked,
because it has no estimate yet of pre-congestion for the path across
the PCN-region.To be conservative, following the guidance for specifying other
re-ECN transports in ,
the ingress SHOULD set the FNE codepoint of the extended PCN header
in all probe packets (). As per the PCN deployment
model, the egress gateway measures the fraction of congestion-marked
probe octets and feeds back the resulting pre-congestion level to
the ingress, piggy-backed on the returning reservation response
(RESV) for the new flow. Probe packets are identifiable by the
egress because they carry the FNE codepoint.It may seem inadvisable to expect the FNE codepoint to be set on
probes, given legacy firewalls etc. might discard such packets
(because this flag had no previous legitimate use). However, in the
deployment scenarios envisaged, each domain in the PCN-region has to
be explicitly configured to support the admission controlled
service. So, before deploying the service, the operator MUST
reconfigure such a badly implemented middlebox to allow through
packets with the RE flag set.Note that we have said SHOULD rather than MUST for the FNE
setting behaviour of the ingress for probe packets. This entertains
the possibility of an ingress implementation having the benefit of
other knowledge of the path, which it re-uses for a newly starting
aggregate. For instance, it may hold cached information from a
recent use of the aggregate that is still sufficiently current to be
useful. If not all probe packets are set to FNE, the ingress will
have to ensure probe packets are identifiable by some other means,
perhaps by using the egress as the destination address.It might seem pedantic worrying about these few probe packets,
but this behaviour ensures the system is safe, even if the
proportion of probe packets becomes large.It might be expected that a new flow within an active aggregate
would need no special bootstrap behaviour. If there was an aggregate
already in progress between the gateways the new flow was about to
use, it would inherit the prevailing RE blanking fraction. And if
there were no active aggregate, the bootstrap behaviour for an
aggregate would be appropriate and sufficient for the new flow.However, for a number of reasons, at least the first packet of
each new flow SHOULD be set to the FNE codepoint, irrespective of
whether it is joining an active aggregate or not. If the first
packet is unlikely to be reliably delivered, a number of FNE packets
MAY be sent to increase the probability that at least one is
delivered to the egress gateway.If each flow does not start with an FNE packet, it will be seen
later that sanctions may be too strict at the interface before the
egress gateway. It will often be possible to apply sanctions at the
granularity of aggregates rather than flows, but in an
internetworked environment it cannot be guaranteed that aggregates
will be identifiable in remote networks. So setting FNE at the start
of each flow is a safe strategy. For instance, a remote network may
have equal cost multi-path (ECMP) routing enabled, causing different
flows between the same gateways to traverse different paths.After an idle period of more than 1 second, the ingress gateway
SHOULD set the EPCN field of the next packet it sends to FNE. This
allows the design of network policers to be deterministic (see ).However, if the ingress gateway can guarantee that the network(s)
that will carry the flow to its egress gateway all use a common
identifier for the aggregate (e.g. a single MPLS network without
ECMP routing), it MAY NOT set FNE when it adds a new flow to an
active aggregate. And an FNE packet need only be sent if a whole
aggregate has been idle for more than 1 second.Adding re-PCN works well with the regular PCN forwarding
behaviour of interior queues. However, below, two optional changes
are proposed when forwarding packets with a per-hop-behaviour that
requires pre-congestion notification:When a router cannot avoid
dropping PCN-capable packets, preferential dropping of packets
with different extended PCN codepoints SHOULD be implemented
between packets within a PHB that uses PCN marking. The drop
preference order to use is defined in . Note that to reduce
configuration complexity, Re-PCT-Echo and FNE MAY be given the
same drop preference, but if feasible, FNE SHOULD be dropped in
preference to Re-PCT-Echo.If this
proposal were advanced at the same time as PCN itself, we would
recommend that preferential drop based on extended PCN codepoint
SHOULD be added to router forwarding at the same time as PCN
marking. Preferential dropping can be difficult to implement,
but we RECOMMEND this security-related re-PCN improvement where
feasible as it is an effective defence against flooding
attacks.We propose that PCN-routers
SHOULD inspect the RE flag as well as the ECN field to decide
whether to drop or mark PCN DSCPs. They MUST choose drop if the
codepoint of this extended ECN field is Not-PCN. Otherwise they
SHOULD mark (unless, of course, buffer space is
exhausted).A PCN-capable router MUST
NOT ever congestion mark a packet carrying the Not-PCN codepoint
because the transport will only understand drop, not congestion
marking. But a PCN-capable router can mark rather than drop an
FNE packet, even though its ECN field when looked at in
isolation is '00' which appears to be a legacy Not-ECT packet.
Therefore, if a packet's RE flag is '1', even if its ECN field
is '00', a PCN-enabled router SHOULD use congestion marking.
This allows the `feedback not established' (FNE) codepoint to be
used for probe packets, in order to pick up PCN marking when
bootstrapping an aggregate.PCN marking
rather than dropping of FNE packets MUST only be deployed in
controlled environments, such as that in , where the presence of an egress node that
understands PCN marking is assured. Congestion events might
otherwise be ignored if the receiver only understands drop,
rather than PCN marking. This is because there is no guarantee
that PCN capability has been negotiated if feedback is not
established (FNE). Also, places the strong
condition that a router MUST apply drop rather than marking to
FNE packets unless it can guarantee that FNE packets are rate
limited either locally or upstream.PCN fieldRE flagExtended PCN codepointDrop PrefRe-PCN meaning100Re-PCT-Echo5/4Re-echoed congestion and Re-PCT001FNE4Feedback not established101Re-PCT3Re-PCN capable transport010AM(0)3Admission Marking with Re-Echo011AM(-1)3Admission Marking
110TM(0)2Termination Marking with Re-Echo111TM(-1)2Termination Marking 000Not-PCN1Not PCN-capable transportIf a different signalling system, such as NSIS, were used but it
provided admission control in a similar way using pre-congestion
notification (e.g. Arumaithurai or RMD ), we believe re-PCN could be used to
protect against misbehaving networks in the same way as proposed
above.The following sections are informative, not normative. The re-PCN
protocol described in above would
require standardisation, whereas operators acting in their own interests
would be expected to deploy policing and monitoring functions similar to
those proposed in the sections below without any further need for
standardisation by the IETF. Flexibility is expected in exactly how
policing and monitoring is done.In the rest of this memo, where the context makes it clear, we will
sometimes loosely use the term `congestion' rather than using the
stricter `downstream pre-congestion'. Also we will loosely talk of
positive or negative flows, meaning flows where the moving average of
the downstream pre-congestion metric is persistently positive or
negative. The notion of a negative metric arises because it is derived
by subtracting one metric from another. Of course actual downstream
congestion cannot be negative, only the metric can (whether due to
time lags or deliberate malice).Just as we will loosely talk of positive and negative flows, we
will also talk of positive or negative packets, meaning packets that
contribute positively or negatively to downstream pre-congestion.Therefore packets can be considered to have a `worth' of +1, 0 or
-1, which, when multiplied by their size, indicates their contribution
to downstream congestion. Packets will usually be initialised by the
PCN ingress with a worth of 0. Blanking the RE flag increments the
worth of a packet to +1. Congestion marking a packet decrements its
worth (whether admission marking or termination marking). Congestion
marking a previously blanked packet cancels out the positive worth
with the negative worth of the congestion marking (resulting in a
packet worth 0). The FNE codepoint is an exception. It has the same
positive worth as a packet with the Re-PCT-Echo codepoint. The table
below specifies unambiguously the worth of each extended PCN
codepoint. Note the order is different from the previous table to
emphasise how congestion marking processes decrement the worth (with
the exception of FNE).ECN fieldRE flagExtended PCN codepointWorthRe-PCN meaning000Not-PCNn/aNot PCN-capable transport100Re-PCT-Echo+1Re-echoed congestion and Re-PCT010AM(0)0Admission Marking with Re-Echo110TM(0)0Termination Marking with Re-Echo001FNE+1Feedback not established101Re-PCT0Re-PCN capable transport011AM(-1)-1Admission Marking 111TM(-1)-1Termination MarkingIt will be recalled that downstream congestion can be found by
subtracting upstream congestion from path congestion. displays the difference
between the two plots in to show downstream
pre-congestion across the same path through the Internet.To emulate border policing, the general idea is for each domain to
apply penalties to its upstream neighbour in proportion to the amount
of downstream pre-congestion that the upstream network sends across
the border. That is, the penalties should be in proportion to the
height of the plot. Downward arrows in the figure show the resulting
pressure for each domain to under-declare downstream pre-congestion in
traffic they pass to the next domain, because of the penalties.These penalties seem to encourage everyone to understate downstream
congestion in order to reduce the penalties they incur. But a
balancing pressure is introduced by the last domain (strictly by any
domain), which applies sanctions to flows if downstream congestion
goes negative before the egress gateway. The upward arrow at Domain
C's border with the egress gateway represents the incentive the
sanctions would create to prevent negative traffic. The same upward
pressure can be applied at any domain border (arrows not shown).Any flow that persistently goes negative by the time it leaves a
domain must not have been marked correctly in the first place. A
domain that discovers such a flow can adopt a range of strategies to
protect itself. Which strategy it uses will depend on policy, because
it cannot immediately assume malice—there may be an innocent
configuration error somewhere in the system.This memo does not propose to standardise any particular mechanism
to detect persistently negative flows, but does give examples.
Note that we have used the term flow, but there will be no need to
bury into the transport layer for port numbers; identifiers visible in
the network layer will be sufficient (IP address pair, DSCP, protocol
ID). The appendix also gives a mechanism to limit the required flow
state, preventing state exhaustion attacks.Of course, some domains may trust other domains to comply with
admission control without applying sanctions or penalties. In these
cases, the protocol should still be used but no penalties need be
applied. The re-PCN protocol ensures downstream pre-congestion marking
is passed on correctly whether or not penalties are applied to it, so
the system works just as well with a mixture of some domains trusting
each other and others not.Providers should be free to agree the contractual terms they wish
between themselves, so this memo does not propose to standardise how
these penalties would be applied. It is sufficient to standardise the
re-PCN protocol so the downstream pre-congestion metric is available
if providers choose to use it. However, the next section () gives some examples of how
these penalties might be implemented.The re-PCN protocol has been chosen to solve the policing problem
because it embeds a downstream pre-congestion metric in passing PCN
traffic that is difficult to lie about and can be measured in bulk.
The ability to emulate border policing depends on network operators
choosing to use this metric as one of the elements in their contracts
with each other.Already many inter-domain agreements involve a capacity and a usage
element. The usage element may be based on volume or various measures
of peak demand. We expect that those network operators who choose to
use pre-congestion notification for admission control would also be
willing to consider using this downstream pre-congestion metric as a
usage element in their interconnection contracts for admission
controlled (PCN) traffic.Congestion (or pre-congestion) has the dimension of [octet], being
the product of volume transferred [octet] and the congestion fraction
[dimensionless], which is the fraction of the offered load that the
network isn't able to serve (or would rather not serve in the case of
pre-congestion). Measuring downstream congestion gives a measure of
the volume transferred but modulated by congestion expected
downstream. So volume transferred during off-peak periods counts as
nearly nothing, while volume transferred at peak times or over
temporarily congested links counts very highly. The re-PCN protocol
allows one network to measure how much pre-congestion has been
`dumped' into it by another network. And then in turn how much of that
pre-congestion it dumped into the next downstream network. describes mechanisms for
calculating border penalties referring to for suggested metering algorithms for
downstream congestion at a border router. Conceptually, it could
hardly be simpler. It broadly involves accumulating the volume of
packets with the RE flag blanked and the volume of those with
congestion marking then subtracting the two.Once this downstream pre-congestion metric is available, operators
are free to choose how they incorporate it into their interconnection
contracts . Some may include a threshold
volume of pre-congestion as a quality measure in their service level
agreement, perhaps with a penalty clause if the upstream network
exceeds this threshold over, say, a month. Others may agree a set of
tiered monthly thresholds, with increasing penalties as each threshold
is exceeded. But, it would be just as easy, and more resistant to
gaming, to do away with discrete thresholds, and instead make the
penalty rise smoothly with the volume of pre-congestion by applying a
price to pre-congestion itself. Then the usage element of the
interconnection contract would directly relate to the volume of
pre-congestion caused by the upstream network.The direction of penalties and charges relative to the direction of
traffic flow is a constant source of confusion. Typically, where
capacity charges are concerned, lower tier customer networks pay
higher tier provider networks. So money flows from the edges to the
middle of the internetwork, towards greater connectivity, irrespective
of the flow of data. But we advise that penalties or charges for usage
should follow the same direction as the data flow—the direction
of control at the network layer. Otherwise a network lays itself open
to `denial of funds' attacks. So, where a tier 2 provider sends data
into a tier 3 customer network, we would expect the penalty clauses
for sending too much pre-congestion to be against the tier 2 network,
even though it is the provider.It may help to remember that data will be flowing in the other
direction too. So the provider network has as much opportunity to levy
usage penalties as its customer, and it can set the price or strength
of its own penalties higher if it chooses. Usage charges in both
directions tend to cancel each other out, which confirms that
usage-charging is less to do with revenue raising and more to do with
encouraging load control discipline in order to smooth peaks and
troughs, improving utilisation and quality.Further, when operators agree penalties in their interconnection
contracts for sending downstream congestion, they should make sure
that any level of negative marking only equates to zero penalty. In
other words, penalties are always paid in the same direction as the
data, and never against the data flow, even if downstream congestion
seems to be negative. This is consistent with the definition of
physical congestion; when a resource is underutilised, it is not
negatively congested. Its congestion is just zero. So, although short
periods of negative marking can be tolerated to correct temporary
over-declarations due to lags in the feedback system, persistent
downstream negative congestion can have no physical meaning and
therefore must signify a problem. The incentive for domains not to
tolerate persistently negative traffic depends on this principle that
negative penalties must never be paid for negative congestion.Also note that at the last egress of the PCN-region, domain C
should not agree to pay any penalties to the egress gateway for
pre-congestion passed to the egress gateway. Downstream pre-congestion
to the egress gateway should have reached zero here. If domain C were
to agree to pay for any remaining downstream pre-congestion, it would
give the egress gateway an incentive to over-declare pre-congestion
feedback and take the resulting profit from domain C.To focus the discussion, from now on, unless otherwise stated, we
will assume a downstream network charges its upstream neighbour in
proportion to the pre-congestion it sends (V_b in the notation of
). Effectively tiered thresholds
would be just more coarse-grained approximations of the fine-grained
case we choose to examine. If these neighbours had previously agreed
that the (fixed) price per octet of pre-congestion would be L, then
the bill at the end of the month would simply be the product L*V_b,
plus any fixed charges they may also have agreed.We are well aware that the IETF tries to avoid standardising
technology that depends on a particular business model. Indeed, this
principle is at the heart of all our own work. Our aim here is to make
a new metric available that we believe is superior to all existing
metrics. Then, our aim is to show that bulk border policing can at
least work with the one model we have just outlined. Of course,
operators are free to complement this pre-congestion-based usage
element of their charges with traditional capacity charging, and we
expect they will. But if operators don't want to use this business
model at all, they don't have to do bulk border policing. We also
assume that operators might experiment with the metric in other
models.Also note well that everything we discuss in this memo only
concerns interconnection within the PCN-region. ISPs are free to sell
or give away reservations however they want on the retail market. But
of course, interconnection charges will have a bearing on that.
Indeed, in the present scenario, the ingress gateway effectively sells
reservations on one side and buys congestion penalties on the other.
As congestion rises, one can imagine the gateway discovering that
congestion penalties have risen higher than the (probably fixed)
revenue it will earn from selling the next flow reservation. This
encourages the gateway to cut its losses by blocking new calls, which
is why we believe downstream congestion penalties can emulate per-flow
rate policing at borders, as the next section explains.The important feature of charging in proportion to congestion
volume is that the penalty aggregates and disaggregates correctly
along with packet flows. This is because the penalty rises linearly
with bit rate (unless congestion is absolutely zero) and linearly with
congestion, because it is the product of them both. So if the packets
crossing a border belong to a thousand flows, and one of those flows
doubles its rate, the ingress gateway forwarding that flow will have
to put twice as much congestion marking into the packets of that flow.
And this extra congestion marking will add proportionately to the
penalties levied at every border the flow crosses in proportion to the
amount of pre-congestion remaining on the path.Effectively, usage charges will continuously flow from ingress
gateways to the places generating pre-congestion marking, in
proportion to the pre-congestion marking introduced and to the data
rates from those gateways.As importantly, pre-congestion itself rises super-linearly with
utilisation of a particular resource. So if someone tries to push
another flow into a path that is already signalling enough
pre-congestion to warrant admission control, the penalty will be a lot
greater than it would have been to add the same flow to a less
congested path. This makes the incentive system fairly insensitive to
the actual level of pre-congestion for triggering admission control
that each ingress chooses. The deterrent against exceeding whatever
threshold is chosen rises very quickly with a small amount of
cheating.These are the properties that allow re-PCN to emulate per-flow
border policing of both rate and admission control. It is not a
perfect emulation of per-flow border policing, but we claim it is
sufficient to at least ensure the cost to others of a cheat is borne
by the cheater, because the penalties are at least proportionate to
the level of the cheat. If an edge network operator is selling
reservations at a large profit over the congestion cost, these
pre-congestion penalties will not be sufficient to ensure networks in
the middle get a share of those profits, but at least they can cover
their costs.We will now explain with an example. When a whole inter-network is
operating at normal (typically very low) congestion, the
pre-congestion marking from virtual queues will be a little higher
than if the real queues had been used—still low, but more
noticeable. But low congestion levels do not imply that usage charges must also be low. Usage charges will
depend on the price L as well.If the metric of the usage element of an interconnection agreement
was changed from pure volume to pre-congested volume, one would expect
the price of pre-congestion to be arranged so that the total usage
charge remained about the same. So, if an average pre-congestion
fraction turned out to be 1/1000, one would expect that the price L
(per octet) of pre-congestion would be about 1000 times the previously
used (per octet) price for volume. We should add that a switch to
pre-congestion is unlikely to exactly maintain the same overall level
of usage charges, but this argument will be approximately true,
because usage charge will rise to at least the level the market finds
necessary to push back against usage.From the above example it can be seen why a 1000x higher price will
make operators become acutely sensitive to the congestion they cause
in other networks, which is of course the desired effect; to encourage
networks to avoid the congestion they
allow their users to cause to others.If any network sends even one flow at higher rate, they will
immediately have to pay proportionately more usage charges. Because
there is no knowledge of reservations within the PCN-region, no
interior router can police whether the rate of each flow is greater
than each reservation. So the system doesn't truly emulate
rate-policing of each flow. But there is no incentive to pack a higher
rate into a reservation, because the charges are directly proportional
to rate, irrespective of the reservations.However, if virtual queues start to fill on any path, even though
real queues will still be able to provide low latency service,
pre-congestion marking will rise fairly quickly. It may eventually
reach the threshold where the ingress gateway would deny admission to
new flows. If the ingress gateway cheats and continues to admit new
flows, the affected virtual queues will rapidly fill, even though the
real queues will still be little worse than they were when admission
control should have been invoked. The ingress gateway will have to pay
the penalty for such an extremely high pre-congestion level, so the
pressure to invoke admission control should become unbearable.The above mechanisms protect against rational operators. In we discuss how networks can protect
themselves from accidental or deliberate misconfiguration in
neighbouring networks.As PCN traffic leaves the last network before the egress gateway
(domain 'C' in ) the RE
blanking fraction should match the congestion marking fraction, when
averaged over a sufficiently long duration (perhaps ~10s to allow a
few rounds of feedback through regular signalling of new and refreshed
reservations).To protect itself, domain 'C' should install a monitor at its
egress. It aims to detect flows of PCN packets that are persistently
negative. If flows are positive, domain 'C' need take no
action—this simply means an upstream network must be paying more
penalties than it needs to. gives a suggested algorithm
for the monitor, meeting the criteria below. It SHOULD introduce minimal false positives for honest
flows;It SHOULD quickly detect and sanction dishonest flows (minimal
false negatives);It MUST be invulnerable to state exhaustion attacks from
malicious sources. For instance, if the dropper uses flow-state,
it should not be possible for a source to send numerous packets,
each with a different flow ID, to force the dropper to exhaust its
memory capacity;If drop is used as a sanction, it SHOULD introduce sufficient
loss in goodput so that malicious sources cannot play off losses
in the egress dropper against higher allowed throughput.
Salvatori describes this attack,
which involves the source understating path congestion then
inserting forward error correction (FEC) packets to compensate
expected losses.Note that the monitor operates on flows but with careful design we
can avoid per-flow state. This is why we have been careful to ensure
that all flows MUST start with a packet marked with the FNE codepoint.
If a flow does not start with the FNE codepoint, a monitor is likely
to treat it unfavourably. This risk makes it worth setting the FNE
codepoint at the start of a flow, even though there is a cost to
setting FNE (positive `worth').Starting flows with an FNE packet also means that a monitor will be
resistant to state exhaustion attacks from other networks, as the
monitor can then be designed to never create state unless an FNE
packet arrives. And an FNE packet counts positive, so it will cost a
lot for a network to send many of them.Monitor algorithms will often maintain a moving average across
flows of the fraction of RE blanked packets. When maintaining an
average across flows, a monitor MUST ignore packets with the FNE
codepoint set. An ingress gateway sets the FNE codepoint when it does
not have the benefit of feedback from the egress. So counting packets
with FNE cleared would be likely to make the average unnecessarily
positive, providing headroom (or should we say footroom?) for
dishonest (negative) traffic.If the monitor detects a persistently negative flow, it could drop
sufficient negative and neutral packets to force the flow to not be
negative. This is the approach taken for the `egress dropper' in , but for the scenario in this
memo, where everyone would expect everyone else to keep to the
protocol, a management alarm SHOULD be raised on detecting
persistently negative traffic and any automatic sanctions taken SHOULD
be logged. Even if the chosen policy is to take no automatic action,
the cause can then be investigated manually.Then all ingresses cannot understate downstream pre-congestion
without their action being logged. So network operators can deal with
offending networks at the human level, out of band. As a last resort,
perhaps where the ingress gateway address seems to have been spoofed
in the signalling, packets can be dropped. Drops could be focused on
just sufficient packets in misbehaving flows to remove the negative
bias while doing minimal harm.A future version of this memo may define a control message that
could be used to notify an offending ingress gateway (possibly via the
egress gateway) that it is sending persistently negative flows.
However, we are aware that such messages could be used to test the
sensitivity of the detection system, so currently we prefer silent
sanctions.An extreme scenario would be where an ingress gateway (or set of
gateways) mounted a DoS attack against another network. If their
traffic caused sufficient congestion to lead to drop but they
understated path congestion to avoid penalties for causing high
congestion, the preferential drop recommendations in would at least ensure
that these flows would always be dropped before honest flows..One of the main design goals of re-PCN was for border security
mechanisms to be as simple as possible, otherwise they would become
the pinch-points that limit scalability of the whole internetwork.
As the title of this memo suggests, we want to avoid per-flow
processing at borders. We also want to keep to passive mechanisms
that can monitor traffic in parallel to forwarding, rather than
having to filter traffic inline—in series with forwarding. As
data rates continue to rise, we suspect that all-optical
interconnection between networks will soon be a requirement. So we
want to avoid any new need for buffering (even though border
filtering is current practice for other reasons, we don't want to
make it even less likely that we will ever get rid of it).So far, we have been able to keep the border mechanisms simple,
despite having had to harden them against some subtle attacks on the
re-PCN design. The mechanisms are still passive and avoid per-flow
processing, although we do use filtering as a fail-safe to
temporarily shield against extreme events in other networks, such as
accidental misconfigurations ().The basic accounting mechanism at each border interface simply
involves accumulating the volume of packets with positive worth
(Re-PCT-Echo and FNE), and subtracting the volume of those with
negative worth: AM(-1) and TM(-1). Even though this mechanism takes
no regard of flows, over an accounting period (say a month) this
subtraction will account for the downstream congestion caused by all
the flows traversing the interface, wherever they come from, and
wherever they go to. The two networks can agree to use this metric
however they wish to determine some congestion-related penalty
against the upstream network (see for examples). Although the
algorithm could hardly be simpler, it is spelled out using
pseudo-code in .Various attempts to subvert the re-ECN design have been made. In
all cases their root cause is persistently negative flows. But,
after describing these attacks we will show that we don't actually
have to get rid of all persistently negative flows in order to
thwart the attacks.In honest flows, downstream congestion is measured as positive
minus negative volume. So if all flows are honest (i.e. not
persistently negative), adding all positive volume and all negative
volume without regard to flows will give an aggregate measure of
downstream congestion. But such simple aggregation is only possible
if no flows are persistently negative. Unless persistently negative
flows are completely removed, they will reduce the aggregate measure
of congestion. The aggregate may still be positive overall, but not
as positive as it would have been had the negative flows been
removed.In we
discussed how to sanction traffic to remove, or at least to
identify, persistently negative flows. But, even if the sanction for
negative traffic is to discard it, unless it is discarded at the
exact point it goes negative, it will wrongly subtract from
aggregate downstream congestion, at least at any borders it crosses
after it has gone negative but before it is discarded.We rely on sanctions to deter dishonest understatement of
congestion. But even the ultimate sanction of discard can only be
effective if the sender is bothered about the data getting through
to its destination. A number of attacks have been identified where a
sender gains from sending dummy traffic or it can attack someone or
something using dummy traffic even though it isn't communicating any
information to anyone: A network can simply create its own dummy traffic to congest
another network, perhaps causing it to lose business at no cost
to the attacking network. This is a form of denial of service
perpetrated by one network on another. The preferential drop
measures in
provide crude protection against such attacks, but we are not
overly worried about more accurate prevention measures, because
it is already possible for networks to DoS other networks on the
general Internet, but they generally don't because of the grave
consequences of being found out. We are only concerned if re-PCN
increases the motivation for such an attack, as in the next
example.A network can just generate negative traffic and send it over
its border with a neighbour to reduce the overall penalties that
it should pay to that neighbour. It could even initialise the
TTL so it expired shortly after entering the neighbouring
network, reducing the chance of detection further downstream.
This attack need not be motivated by a desire to deny service
and indeed need not cause denial of service. A network's main
motivator would most likely be to reduce the penalties it pays
to a neighbour. But, the prospect of financial gain might tempt
the network into mounting a DoS attack on the other network as
well, given the gain would offset some of the risk of being
detected.Note that we have not included DoS by Internet hosts in the above
list of attacks, because we have restricted ourselves to a scenario
with edge-to-edge admission control across a PCN-region. In this
case, the edge ingress gateways insulate the PCN-region from DoS by
Internet hosts. Re-ECN resists more general DoS attacks, but this is
discussed in .The first step towards a solution to all these problems with
negative flows is to be able to estimate the contribution they make
to downstream congestion at a border and to correct the measure
accordingly. Although ideally we want to remove negative flows
themselves, perhaps surprisingly, the most effective first step is
to cancel out the polluting effect negative flows have on the
measure of downstream congestion at a border. It is more important
to get an unbiased estimate of their effect, than to try to remove
them all. A suggested algorithm to give an unbiased estimate of the
contribution from negative flows to the downstream congestion
measure is given in .Although making an accurate assessment of the contribution from
negative flows may not be easy, just the single step of neutralising
their polluting effect on congestion metrics removes all the gains
networks could otherwise make from mounting dummy traffic attacks on
each other. This puts all networks on the same side (only with
respect to negative flows of course), rather than being pitched
against each other. The network where a flow goes negative as well
as all the networks downstream lose out from not being reimbursed
for any congestion this flow causes. So they all have an interest in
getting rid of these negative flows. Networks forwarding a flow
before it goes negative aren't strictly on the same side, but they
are disinterested bystanders—they don't care that the flow
goes negative downstream, but at least they can't actively gain from
making it go negative. The problem becomes localised so that once a
flow goes negative, all the networks from where it happens and
beyond downstream each have a small problem, each can detect it has
a problem and each can get rid of the problem if it chooses to. But
negative flows can no longer be used for any new attacks.Once an unbiased estimate of the effect of negative flows can be
made, the problem reduces to detecting and preferably removing flows
that have gone negative as soon as possible. But importantly,
complete eradication of negative flows is no longer
critical—best endeavours will be sufficient.Note that the guiding principle behind all the above discussion
is that any gain from subverting the protocol should be precisely
neutralised, rather than punished. If a gain is punished to a
greater extent than is sufficient to neutralise it, it will most
likely open up a new vulnerability, where the amplifying effect of
the punishment mechanism can be turned on others.For instance, if possible, flows should be removed as soon as
they go negative, but we do NOT RECOMMEND any attempts to discard
such flows further upstream while they are still positive. Such
over-zealous push-back is unnecessary and potentially dangerous.
These flows have paid their `fare' up to the point they go negative,
so there is no harm in delivering them that far. If someone
downstream asks for a flow to be dropped as near to the source as
possible, because they say it is going to become negative later, an
upstream node cannot test the truth of this assertion. Rather than
have to authenticate such messages, re-PCN has been designed so that
flows can be dropped solely based on locally measurable evidence. A
message hinting that a flow should be watched closely to test for
negativity is fine. But not a message that claims that a positive
flow will go negative later, so it should be dropped.With the above penalty system, each domain seems to have a
perverse incentive to fake pre-congestion. For instance domain 'B'
profits from the difference between penalties it receives at its
ingress (its revenue) and those it pays at its egress (its cost). So
if 'B' overstates internal pre-congestion it seems to increase its
profit. However, we can assume that domain 'A' could bypass 'B',
routing through other domains to reach the egress. So the
competitive discipline of least-cost routing can ensure that any
domain tempted to fake pre-congestion for profit risks losing all its incoming traffic. The least congested
route would eventually be able to win this competitive game, only as
long as it didn't declare more fake pre-congestion than the next
most competitive route.The competitive effect of interdomain routing might be weaker
nearer to the egress. For instance, 'C' may be the only route 'B'
can take to reach the ultimate receiver. And if 'C' over-penalises
'B', the egress gateway and the ultimate receiver seem to have no
incentive to move their terminating attachment to another network,
because only 'B' and those upstream of 'B' suffer the higher
penalties. However, we must remember that we are only looking at the
money flows at the unidirectional network layer. There are likely to
be all sorts of higher level business models constructed over the
top of these low level 'sender-pays' penalties. For instance, we
might expect a session layer charging model where the session
originator pays for a pair of duplex flows, one as receiver and one
as sender. Traditionally this has been a common model for telephony
and we might expect it to be used, at least sometimes, for other
media such as video. Wherever such a model is used, the data
receiver will be directly affected if its sessions terminate through
a network like 'C' that fakes congestion to over-penalise 'B'. So
end-customers will experience a direct competitive pressure to
switch to cheaper networks, away from networks like 'C' that try to
over-penalise 'B'.This memo does not need to standardise any particular mechanism
for routing based on re-PCN. Goldenberg et al refers to various commercial products and
presents its own algorithms for moving traffic between multi-homed
routes based on usage charges. None of these systems require any
changes to standards protocols because the choice between the
available border gateway protocol (BGP) routes is based on a
combination of local knowledge of the charging regime and local
measurement of traffic levels. If, as we propose, charges or
penalties were based on the level of re-PCN measured locally in
passing traffic, a similar optimisation could be achieved without
requiring any changes to standard routing protocols.We must be clear that applying pre-congestion-based routing to
this admission control system remains an open research issue.
Traffic engineering based on congestion requires careful damping to
avoid oscillations, and should not be attempted without adult
supervision :) Mortier & Pratt
have analysed traffic engineering based on congestion. But without
the benefit of re-ECN or re-PCN, they had to add a path attribute to
BGP to advertise a route's downstream congestion (actually they
proposed that BGP should advertise the charge for congestion, which
we believe wrongly embeds an assumption into BGP that the only thing
to do with congestion is charge for it).The mechanisms described so far create incentives for rational
operators to behave. That is, one operator aims to make another
behave responsibly by applying penalties and expects a rational
response (i.e. one that trades off costs against benefits). It is
usually reasonable to assume that other network operators will
behave rationally (policy routing can avoid those that might not).
But this approach does not protect against the misconfigurations and
accidents of other operators.Therefore, we propose the following two mechanisms at a network's
borders to provide "defence in depth". Both are similar: A small sample of positive
packets should be picked randomly as they cross a border
interface. Then subsequent packets matching the same source and
destination address and DSCP should be monitored. If the
fraction of positive marking is well above a threshold (to be
determined by operational practice), a management alarm SHOULD
be raised, and the flow MAY be automatically subject to focused
drop.A small sample of
congestion marked packets should be picked randomly as they
cross a border interface. Then subsequent packets matching the
same source and destination address and DSCP should be
monitored. If the RE blanking fraction minus the congestion
marking fraction is persistently negative, a management alarm
SHOULD be raised, and the flow MAY be automatically subject to
focused drop.Both these mechanisms rely on the fact that highly positive (or
negative) flows will appear more quickly in the sample by selecting
randomly solely from positive (or negative) packets.Note that there is no assumption that users behave rationally. The system is
protected from the vagaries of irrational user behaviour by the
ingress gateways, which transform internal penalties into a
deterministic, admission control mechanism that prevents users from
misbehaving, by directly engineered means.The domains in are not expected
to be completely malicious towards each other. After all, we can assume
that they are all co-operating to provide an internetworking service to
the benefit of each of them and their customers. Otherwise their routing
polices would not interconnect them in the first place. However, we
assume that they are also competitors of each other. So a network may
try to contravene our proposed protocol if it would gain or make a
competitor lose, or both. But only if it can do so without being caught.
Therefore we do not have to consider every possible random attack one
network could launch on the traffic of another, given anyway one network
can always drop or corrupt packets that it forwards on behalf of
another.Therefore, we only consider new opportunities for gainful attack that our proposal introduces. But to
a certain extent we can also rely on the in depth defences we have
described ( ) intended to mitigate the
potential impact if one network accidentally misconfiguring the workings
of this protocol.The ingress and egress gateways are shown in the most generic
arrangement possible in , without
any surrounding network. This allows us to consider more specific cases
where these gateways and a neighbouring network are operated by the same
player. As well as cases where the same player operates neighbouring
networks, we will also consider cases where the two gateways collude as
one player and where the sender and receiver collude as one. Collusion
of other sets of domains is less likely, but we will consider such
cases. In the general case, we will assume none of the nine trust
domains across the figure fully trust any of the others.As we only propose to change routers within the PCN-region, we assume
the operators of networks outside the region will be doing per-flow
policing. That is, we assume the networks outside the PCN-region and the
gateways around its edges can protect themselves. So given we are
proposing to remove flow policing from some networks, our primary
concern must be to protect networks that don't do per-flow policing (the
potential `victims') from those that do (the `enemy'). The ingress and
egress gateways are the only way the outer enemy can get at the middle
victim, so we can consider the gateways as the representatives of the
enemy as far as domains 'A', 'B' and 'C' are concerned. We will call
this trust scenario `edges against middles'.Earlier in this memo, we outlined the classic border rate policing
problem (). It will now be useful to
reiterate the motivations that are the root cause of the problem. The
more reservations a gateway can allow, the more revenue it receives. The
middle networks want the edges to comply with the admission control
protocol when they become so congested that their service to others
might suffer. The middle networks also want to ensure the edges cannot
steal more service from them than they are entitled to.In the context of this `edges against middles' scenario, the re-PCN
protocol has two main effects: The more pre-congestion there is on a path across the PCN-region,
the higher the ingress gateway must declare downstream
pre-congestion.If the ingress gateway does not declare downstream pre-congestion
high enough on average, it will `hit the ground before the runway',
going negative and triggering sanctions, either directly against the
traffic or against the ingress gateway at a management levelAn executive summary of our security analysis can be stated in three
parts, distinguished by the type of collusion considered. Here there is
no collusion or collusion is limited to neighbours in the feedback
loop. In other words, two neighbouring networks can be assumed to
act as one. Or the egress gateway might collude with domain 'C'. Or
the ingress gateway might collude with domain 'A'. Or ingress and
egress gateways might collude with each other.In these cases where only neighbours in the
feedback loop collude, we concludes that all parties have a positive
incentive to declare downstream pre-congestion truthfully, and the
ingress gateway has a positive incentive to invoke admission control
when congestion rises above the admission threshold in any network
in the region (including its own). No party has an incentive to send
more traffic than declared in reservation signalling (even though
only the gateways read this signalling). In short, no party can gain
at the expense of another.In the case of
other forms of collusion between middle networks (e.g. between
domain 'A' and 'C') it would be possible for say 'A' & 'C' to
create a tunnel between themselves so that 'A' would gain at the
expense of 'B'. But 'C' would then lose the gain that 'A' had made.
Therefore the value to 'A' & 'C' of colluding to mount this
attack seems questionable. It is made more questionable, because the
attack can be statistically detected by 'B' using the second
`defence in depth' mechanism mentioned already. Note that 'C' can
defend itself from being attacked through a tunnel by treating the
tunnel end point as a direct link to a neighbouring network (e.g. as
if 'A' were a neighbour of 'C', via the tunnel), which falls back to
the safety of the neighbour-only scenario.Collusion between networks or
gateways within the PCN-region and networks or users outside the
region has not yet been fully analysed. The presence of full
per-flow policing at the ingress gateway seems to make this a less
likely source of a successful attack.{ToDo: Due to lack of time, the full write up of the security
analysis is deferred to the next version of this memo.}Finally, it is well known that the best person to analyse the
security of a system is not the designer. Therefore, our confident
claims must be hedged with doubt until others with perhaps a greater
incentive to break it have mounted a full analysis.We believe ECN has so far not been widely deployed because it
requires end system and widespread network deployment just to achieve a
marginal improvement in performance. The ability to offer a new service
(admission control) would be a much stronger driver for ECN
deployment.As stated in the introduction, the aim of this memo is to "Design in
security from the start" when admission control is based on
pre-congestion notification. The proposal has been designed so that
security can be added some time after first deployment, but only if the
PCN wire protocol encoding is defined with the foresight to accommodate
the extended set of codepoints defined in this document. Given admission
control based on pre-congestion notification requires few changes to
standards, it should be deployable fairly soon. However, re-PCN requires
a change to IP, which may take a little longer :)We expect that initial deployments of PCN-based admission control
will be confined to single networks, or to clubs of networks that trust
each other. The proposal in this memo will only become relevant once
networks with conflicting interests wish to interconnect their admission
controlled services, but without the scalability constraints of per-flow
border policing. It will not be possible to use re-PCN, even in a
controlled environment between consenting operators, unless it is
standardised into IP. Given the IPv4 header has limited space for
further changes, current IESG policy is not to
allow experimental use of codepoints in the IPv4 header, as whenever an
experiment isn't taken up, the space it used tends to be impossible to
reclaim. Therefore, for IPv4 at least, we will need to find a way to run
an experiment so that the header fields it uses can be reclaimed if the
experiment is not a success.If PCN-based admission control is deployed before re-PCN is
standardised into IP, wherever a network (or club of networks) connects
to another network (or club of networks) with conflicting interests,
they will place a gateway between the two regions that does per-flow
rate policing and admission control. If re-PCN is eventually
standardised into IP, it will be possible for these separate regions to
upgrade all their ingress gateways to support re-PCN before removing the
per-flow policing gateways between them. Given the edge-to-edge
deployment model of PCN-based admission control, it is reasonable to
expect incremental deployment of re-PCN will be feasible on a domain-by
domain basis, without needing to cater for partial deployment of re-PCN
in just some of the gateways around one PCN-domain.Nonetheless, if the upgrade of one ingress gateway is accidentally
overlooked, the RE flag has been defined the safe way round for the
default legacy behaviour (leaving RE cleared as 0). A legacy ingress will appear to be declaring a
high level of pre-congestion into the aggregate. The fail-safe border
mechanism in might trigger management
alarms (which would help in tracking down the need to upgrade the
ingress), but all packets would continue to be delivered safely, as
overstatement of downstream congestion requires no sanction.Only the ingress edge gateways around a PCN-region have to be
upgraded to add re-PCN support, not interior routers. It is also
necessary to add the mechanisms that monitor re-PCN to secure a network
against misbehaving gateways and networks. Specifically, these are the
border mechanisms () and the
mechanisms to sanction dishonest marking ().We also RECOMMEND adding improvements to forwarding on interior
routers (). But the
system works whether all, some or none are upgraded, so interior routers
may be upgraded in a piecemeal fashion at any time.The primary insight of this work is that downstream congestion is the
metric that would be most useful to control an internetwork, and
particularly to police how one network responds to the congestion it
causes in a remote network. This is the problem that has previously made
it so hard to provide scalable admission control.The case for using re-feedback (a generalisation of re-ECN) to police
congestion response and provide QoS is made in .
Essentially, the insight is that congestion is a factor that crosses
layers from the physical upwards. Therefore re-feedback polices
congestion as it crosses the physical interface between networks. This
is achieved by bringing information about congestion of resources later
on the path to the interface, rather than trying to deal with congestion
where it happens by examining the notoriously unreliable source address
in packets. Then congestion crossing the physical interface at a border
can be policed at the interface, rather than policing the congestion on
packets that claim to come from an address (which may be spoofed). Also,
re-feedback works in the network layer independently of other
layers—despite its name re-feedback does not actually require
feedback. It makes a source to act conservatively before it gets
feedback.On the subject of lack of feedback, the feedback not established
(FNE) codepoint is motivated by arguments for a state set-up bit in IP
to prevent state exhaustion attacks. This idea was first put forward
informally by David Clark and developed by Handley and Greenhalgh in
. The idea is that network layer datagrams
should signal explicitly when they require state to be created in the
network layer or the layer above (e.g. at flow start). Then a node can
refuse to create any state unless a datagram declares this intent. We
believe the proposed FNE codepoint serves the same purpose as the
proposed state set-up bit, but it has been overloaded with a more
specific purpose, using it on more packets than just the first in a
flow, but never less (i.e. it is idempotent). In effect the FNE
codepoint serves the purpose of a `soft-state set-up codepoint'.The re-feedback paper also makes the
case for converting the economic interpretation of congestion into hard
engineering mechanism, which is the basis of the approach used in this
memo. The admission control gateways around the PCN-region use hard
engineering, not incentives, to prevent end users from sending more
traffic than they have reserved. Incentive-based mechanisms are only
used between networks, because they are expected to respond to
incentives more rationally than end-users can be expected to. However,
even then, a network can use fail-safes to protect itself from
excessively unusual behaviour by neighbouring networks, whether due to
an accidental misconfiguration or malicious intent.The guiding principle behind the incentive-based approach used
between networks is that any gain from subverting the protocol should be
precisely neutralised, rather than punished. If a gain is punished to a
greater extent than is sufficient to neutralise it, it will most likely
open up a new vulnerability, where the amplifying effect of the
punishment mechanism can be turned on others.The re-feedback paper also makes the case against the use of
congestion charging to police congestion if it is based on classic
feedback (where only upstream congestion is visible to network
elements). It argues this would open up receiving networks to `denial of
funds' attacks and would require end users to accept dynamic pricing
(which few would).Re-PCN has been deliberately designed to simplify policing at the
borders between networks. These trust boundaries are the critical
pinch-points that will limit the scalability of the whole internetwork
unless the overall design minimises the complexity of security functions
at these borders. The border mechanisms described in this memo run
passively in parallel to data forwarding and they do not require
per-flow processing.This whole memo concerns the security of a scalable admission control
system. In particular the analysis section. Below some specific security
issues are mentioned that did not belong elsewhere or which comment on
the overall robustness of the security provided by the design.Firstly, we must repeat the statement of applicability in the
analysis: that we only consider new opportunities for gainful attack that our proposal introduces,
particularly if the attacker can avoid being identified. Despite only
involving a few bits, there is sufficient complexity in the whole system
that there are probably numerous possibilities for other attacks.
However, as far as we are aware, none reap any benefit to the attacker.
For instance, it would be possible for a downstream network to remove
the congestion markings introduced by an upstream network, but it would
only lose out on the penalties it could apply to a downstream
network.When one network forwards a neighbouring network's traffic it will
always be possible to cause damage by dropping or corrupting it.
Therefore we do not believe networks would set their routing policies to
interconnect in the first place if they didn't trust the other networks
not to arbitrarily damage their traffic.Having said this, we do want to highlight some of the weaker parts of
our argument. We have argued that networks will be dissuaded from faking
congestion marking by the possibility that upstream networks will
route round them. As we have said, these arguments are based on
fairly delicate assumptions and will remain fairly tenuous until
proved in practice, particularly close to the egress where less
competitive routing is likely.Given the congestion feedback system is piggy-backed on flow
signalling, which can be fairly infrequent, sanctions may not be
appropriate until a flow has been persistently negative for perhaps
20s. This may allow brief attacks to go unpunished. However,
vulnerability to brief attacks may be reduced if the egress triggers
asynchronous feedback when the congestion level on an aggregate has
risen sufficiently since the last feedback, rather than waiting for
the next opportunity to piggy-back on a signal.We should also point out that the approach in this memo was only
designed to be robust for admission control. We do not claim the
incentives will always be strong enough to force correct flow
termination behaviour. This is because a user will tend to perceive
much greater loss in value if a flow is terminated than if admission
is denied at the start. However, in general the incentives for
correct flow termination are similar to those for admission
control.Finally, it may seem that the 8 codepoints that have been made
available by extending the ECN field with the RE flag have been used
rather wastefully. In effect the RE flag has been used as an orthogonal
single bit in nearly all cases. The only exception being when the ECN
field is cleared to 00. The mapping of the
codepoints in an earlier version of this proposal used the codepoint
space more efficiently, but the scheme became vulnerable to a network
operator focusing its congestion marking to mark more positive than
neutral packets in order to reduce its penalties (see Appendix B of
).With the scheme as now proposed, once the RE flag is set or cleared
by the sender or its proxy, it should not be written by the network,
only read. So the gateways can detect if any network maliciously alters
the RE flag. IPSec AH integrity checking does not cover the IPv4 option
flags (they were considered mutable—even the one we propose using
for the RE flag that was `currently unused' when IPSec was defined). But
it would be sufficient for a pair of gateways to make random checks on
whether the RE flag was the same when it reached the egress gateway as
when it left the ingress. Indeed, if IPSec AH had covered the RE flag,
any network intending to alter sufficient RE flags to make a gain would
have focused its alterations on packets without authenticating headers
(AHs).Therefore, no cryptographic algorithms have been exploited in the
making of this proposal.This memo includes no request to IANA.This memo solves the classic problem of making flow admission control
scale to any size network. It builds on a technique, called PCN, which
involves the use of Diffserv in a domain and uses pre-congestion
notification feedback to control admission into each network path across
the domain .Without PCN, Diffserv requires over-provisioning that must grow
linearly with network diameter to cater for variation in the traffic
matrix. However, even with PCN, multiple network domains can only join
together into one larger PCN region if all domains trust each other to
comply with the protocols, invoking admission control and flow
termination when requested. Domains could join together and still police
flows at their borders by requiring reservation signalling to touch each
border and only use PCN internally to each domain. But the per-flow
processing at borders would still limit scalability.Instead, this memo proposes a technique called re-PCN which enables a
PCN region to extend across multiple domains, without unscalable
per-flow processing at borders, and still without the need for linear
growth in capacity over-provisioning as the hop-diameter of the Diffserv
region grows.We propose that the congestion feedback used for PCN-based admission
control should be re-echoed into the forward data path, by making a
trivial modification to the ingress gateway. We then explain how the
resulting downstream pre-congestion metric in packets can be monitored
in bulk at borders to sufficiently emulate flow rate policing.We claim the result of combining these two approaches is an admission
control system that scales to any size network and any number of interconnected networks, even if
they all act in their own interests.This proposal aims to convince its readers to "Design in Security
from the start," by ensuring the PCN wire protocol encoding can
accommodate the extended set of codepoints defined in this document,
even if per-flow policing is used at first rather than the bulk border
policing described here. This way, we will not build ourselves
tomorrow's legacy problem.Re-echoing congestion feedback is based on a principled technique
called Re-ECN ,
designed to add accountability for causing congestion to the
general-purpose IP datagram service. Re-ECN proposes to consume the last
completely unused bit in the basic IPv4 header or it uses extension
header in IPv6.All the following have given helpful comments either on re-PCN or on
relevant parts of re-ECN that re-PCN uses: Arnaud Jacquet, Alessandro
Salvatori, Steve Rudkin, David Songhurst, John Davey, Ian Self, Anthony
Sheppard, Carla Di Cairano-Gilfedder (BT), Mark Handley (who identified
the excess canceled packets attack), Stephen Hailes, Adam Greenhalgh
(UCL), Francois Le Faucheur, Anna Charny (Cisco), Jozef Babiarz, Kwok-Ho
Chan, Corey Alexander (Nortel), David Clark, Bill Lehr, Sharon Gillett,
Steve Bauer (MIT) (who publicised various dummy traffic attacks), Sally
Floyd (ICIR) and comments from participants in the CFP/CRN
Inter-Provider QoS, Broadband and DoS-Resistant Internet working
groups.Comments and questions are encouraged and very welcome. They can be
addressed to the IETF Congestion and Pre-Congestion Notification working
group's mailing list <pcn@ietf.org>, and/or to the author(s).The ingress gateway receives regular feedback
'PCN-feedback-information' reporting the fraction of congestion marked
octets for each aggregate arriving at the egress. So for each
aggregate it should blank the RE flag on this fraction of octets. A
suitable pseudo-code algorithm for the ingress gateway is as
follows:To meter the bulk amount of downstream pre-congestion in traffic
crossing an inter-domain border, an algorithm is needed that
accumulates the size of positive packets and subtracts the size of
negative packets. We maintain two counters: V_b: accumulated pre-congestion volumeB: total data volume (in case it is needed)A suitable pseudo-code algorithm for a border router is as
follows:At the end of an accounting period this counter V_b represents
the pre-congestion volume that penalties could be applied to, as
described in .For instance, accumulated volume of pre-congestion through a
border interface over a month might be V_b = 5TB (terabyte = 10^12
byte). This might have resulted from an average downstream
pre-congestion level of 0.001% on an accumulated total data volume
of B = 500PB (petabyte = 10^15 byte).The following process is suggested to complement the simple
algorithm above in order to protect against the various attacks from
persistently negative flows described in . As explained in
that section, the most important and first step is to estimate the
contribution of persistently negative flows to the bulk volume of
downstream pre-congestion and to inflate this bulk volume as if
these flows weren't there. The process below has been designed to
give an unbiased estimate, but it may be possible to define other
processes that achieve similar ends.While the above simple metering algorithm () is counting the bulk of traffic
over an accounting period, the meter should also select a subset of
the whole flow ID space that is small enough to be able to
realistically measure but large enough to give a realistic sample.
Many different samples of different subsets of the ID space should
be taken at different times during the accounting period, preferably
covering the whole ID space. During each sample, the meter should
count the volume of positive packets and subtract the volume of
negative, maintaining a separate account for each flow in the
sample. It should run a lot longer than the large majority of flows,
to avoid a bias from missing the starts and ends of flows, which
tend to be positive and negative respectively.Once the accounting period finishes, the meter should calculate
the total of the accounts V_{bI} for the subset of flows I in the
sample, and the total of the accounts V_{fI} excluding flows with a
negative account from the subset I. Then the weighted mean of all
these samples should be taken a_S = sum_{forall I} V_{fI} /
sum_{forall I} V_{bI}.If V_b is the result of the bulk accounting algorithm over the
accounting period ()
it can be inflated by this factor a_S to get a good unbiased
estimate of the volume of downstream congestion over the accounting
period a_S.V_b, without being polluted by the effect of persistently
negative flows.{ToDo: Write up algorithms similar to Appendix E of for the negative flow
monitor with flow management algorithm and the variant with bounded
flow state.}