| Internet-Draft | Thoughts on ICN QoS Architecture | February 2020 |
| Oran | Expires 31 August 2020 | [Page] |
This is a position paper. It documents the author's personal views on how Quality of Service (QoS) capabilities ought to be accommodated in ICN protocols like CCNx or NDN which employ flow-balanced Interest/Data exchanges and hop-by-hop forwarding state as their fundamental machinery. It argues that such protocols demand a substantially different approach to QoS from that taken in TCP/IP, and proposes specific design patterns to achieve both classification and differentiated QoS treatment on both a flow and aggregate basis. It also considers the effect of caches as a resource in addition to memory, CPU and link bandwidth that should be subject to explicitly unfair resource allocation. The proposed methods are intended to operate purely at the network layer, providing the primitives needed to achieve both transport and higher layer QoS objectives. It explicitly excludes any discussion of Quality of Experience (QoE) which can only be assessed and controlled at the application layer or above.¶
This document is not a product of the IRTF Information-Centric Networking Research Group (ICNRG) but has been through formal last call and has the support of the participants in the research group for publication as an individual submission.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 31 August 2020.¶
Copyright (c) 2020 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document.¶
The TCP/IP protocol suite used on today's Internet has over 30 years of accumulated research and engineering into the provision of Quality of Service machinery, employed with varying success in different environments. ICN protocols like Named Data Networking (NDN [NDN]) and Content-Centric Networking (CCNx [RFC8569],[RFC8609]) have an accumulated 10 years of research and very little deployment. We therefore have the opportunity to either recapitulate the approaches taken with TCP/IP (e.g. IntServ [RFC2998] and Diffserv [RFC2474]) or design a new architecture and associated mechanisms aligned with the properties of ICN protocols which differ substantially from those of TCP/IP. This position paper advocates the latter approach and comprises the author's personal views on how Quality of Service (QoS) capabilities ought to be accommodated in ICN protocols like CCNx or NDN. Specifically, these protocols differ in fundamental ways from TCP/IP. The important differences are summarized in the following table:¶
| TCP/IP | CCNx or NDN |
|---|---|
| Stateless forwarding | Stateful forwarding |
| Simple Packets | Object model with optional caching |
| Pure datagram model | Request-response model |
| Asymmetric Routing | Symmetric Routing |
| Independent flow directions | Flow balance |
| Flows grouped by IP prefix and port | Flows grouped by name prefix |
| End-to-end congestion control | Hop-by-hop congestion control |
This document proposes specific design patterns to achieve both flow classification and differentiated QoS treatment for ICN on both a flow and aggregate basis. It also considers the effect of caches as a resource in addition to memory, CPU and link bandwidth that should be subject to explicitly unfair resource allocation. The proposed methods are intended to operate purely at the network layer, providing the primitives needed to achieve both transport and higher layer QoS objectives. It does not propose detailed protocol machinery to achieve these goals; it leaves these to supplementary specifications, such as [I-D.moiseenko-icnrg-flowclass] and [I-D.anilj-icnrg-dnc-qos-icn]. It explicitly excludes any discussion of Quality of Experience (QoE) which can only be assessed and controlled at the application layer or above.¶
Much of this document is derived from presentations the author has given at ICNRG meetings over the last few years that are available through the IETF datatracker (see, for example [Oran2018QoSslides]).¶
QoS in ICN is an important topic with a huge design space. ICNRG has been discussing different specific protocol mechanisms as well as conceptual approaches. This document presents architectural considerations for QoS, leveraging ICN properties instead of merely applying IP-QoS mechanisms - without defining a specific architecture or specific protocols mechanisms yet. However, there is consensus in ICNRG that this document, clarifying the author's views, could inspire such work and should hence be published as a position paper.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 [RFC2119].¶
Much of this background material is tutorial and can be simply skipped by readers familiar with the long and checkered history of quality of service in packet networks. Other parts of it are polemical yet serve to illuminate the author's personal biases and technical views.¶
All networking systems provide some degree of "quality of service" in that they exhibit non-zero utility when offered traffic to carry. The term therefore is used to describe systems that control the allocation of various resources in order to achieve managed unfairness. Absent explicit mechanisms to decide what traffic to be unfair to, most systems try to achieve some form of "fairness" in the allocation of resources, optimizing the overall utility delivered to all demand under the constraint of available resources. From this it should be obvious that you cannot use QoS mechanisms to create or otherwise increase resource capacity! In fact, all known QoS schemes have non-zero overhead and hence may (albeit slightly) decrease the total resources available to carry user traffic.¶
Further, accumulated experience seems to indicate that QoS is helpful in a fairly narrow range of network conditions:¶
Failures can rapidly shift your state from the first above to the second, in which case either:¶
Nevertheless, though not universally deployed, QoS is advantageous at least for some applications and some network environments. Some examples include:¶
Another factor in the design and deployment of QoS is the scalability and scope over which the desired service can be achieved. Here there are two major considerations, one technical, the other economic/political:¶
While this document adopts the narrow definition of QoS as managed unfairness, much of the networking literature uses the term more colloquially as applying to any mechanism that improves overall performance. Readers assuming this broader context will find a large class of proven techniques to be ignored. This is intentional. Among these are seamless producer mobility schemes like MAPME [Auge2018], and network coding of ICN data as discussed in [I-D.irtf-nwcrg-nwc-ccn-reqs].¶
Finally, the relationship between QoS and either accounting or billing is murky. Some schemes can accurately account for resource consumption and ascertain to which user to allocate the usage. Others cannot. While the choice of mechanism may have important practical economic and political consequences for cost and workable business models, this document considers none of those things and discusses QoS only in the context of providing managed unfairness.¶
Some further background on congestion control for ICN is below.¶
Congestion control is necessary in any packet network that multiplexes traffic among multiple sources and destinations in order to:¶
Before moving on to QoS, it is useful to consider how congestion control works in NDN or CCNx. Unlike the IP protocol family, which relies exclusively on end-to-end congestion control (e.g. TCP[RFC0793], DCCP[RFC4340], SCTP[RFC4960], QUIC[I-D.ietf-quic-transport]), CCNx and NDN can employ hop-by-hop congestion control. There is per-Interest/Data state at every hop of the path and therefore outstanding Interests provide information that can be used to optimize resource allocation for data returning on the inverse path, such as bandwidth sharing, prioritization and overload control. In current designs, this allocation is often done using Interest counting. By accepting one Interest packet from a downstream node, implicitly this provides a guarantee (either hard or soft) that there is sufficient bandwidth on the inverse direction of the link to send back one Data packet. A number of congestion control schemes have been developed for ICN that operate in this fashion, for example [Wang2013], [Mahdian2016], [Song2018], [Carofiglio2012]. Other schemes, like [Schneider2016] neither count nor police Interests, but instead monitor queues using AQM (active queue management) to mark returning Data packets that have experienced congestion. This later class of schemes is similar to those used on IP in the sense that they depend on consumers adequately reducing their rate of Interest injection to avoid Data packet drops due to buffer overflow in forwarders. The former class of schemes is (arguably) more robust against mis-behavior by consumers.¶
Given the stochastic nature of round trip times, and the ubiquity of wireless links and encapsulation tunnels with variable bandwidth, a simple scheme that admits interests only based on a time-invariant estimate of the returning link bandwidth will perform poorly. However, two characteristics of NDN and CCNx-like protocols can help substantially to improve the accuracy and responsiveness of the bandwidth allocation:¶
A useful discussion of these properties and how they demonstrate the advantages of ICN approaches to congestion control can be found in [Carofiglio2016]¶
QoS is achieved through managed unfairness in the allocation of resources in network elements, particularly in the routers doing forwarding of ICN packets. So, a first order question is what resources need to be allocated, and how to ascertain which traffic gets what allocations. In the case of CCNx or NDN the important network element resources are:¶
| Resource | ICN Usage |
|---|---|
| Communication Link capacity | buffering for queued packets |
| Content Store capacity | to hold cached data |
| Forwarder memory | for the Pending Interest Table (PIT) |
| Compute capacity | for forwarding packets, including the cost of Forwarding Information Base (FIB) lookups. |
For these resources, any QoS scheme has to specify two things:¶
Two critical facts of life come into play when designing a QoS scheme. First, the number of equivalence classes that can be simultaneously tracked in a network element is bounded by both memory and processing capacity to do the necessary lookups. One can allow very fine-grained equivalence classes, but not be able to employ them globally because of scaling limits of core routers. That means it is wise to either restrict the range of equivalence classes, or allow them to be aggregated, trading off accuracy in policing traffic against ability to scale.¶
Second, the flexibility of expressible treatments can be tightly constrained by both protocol encoding and algorithmic limitations. The ability to encode the treatment requests in the protocol can be limited (as it is for IP - there are only 6 of the TOS bits available for Diffserv treatments), but as or more important is whether there are practical traffic policing, queuing, and pacing algorithms that can be combined to support a rich set of QoS treatments.¶
The two considerations above in combination can easily be substantially more expressive than what can be achieved in practice with the available number of queues on real network interfaces or the amount of per-packet computation needed to enqueue or dequeue a packet.¶
TCP/IP has fewer resource types to manage than ICN, and in some cases the allocation methods are simpler, as shown in the following table:¶
| Resource | IP Relevant | TCP/IP Usage |
|---|---|---|
| Communication Link capacity | YES | buffering for queued packets |
| Content Store capacity | NO | no content store in IP |
| Forwarder memory | MAYBE | not needed for output-buffered designs |
| Compute capacity | YES | for forwarding packets, but arguably much cheaper than ICN |
For these resources, IP has specified three fundamental things, as shown in the following table:¶
| What | How |
|---|---|
| Equivalence classes | subset+prefix match on IP 5-tuple {SA,DA,SP,DP,PT} |
| Diffserv treatments | (very) small number of globally-agreed traffic classes |
| Intserv treatments | per-flow parameterized Controlled Load and Guaranteed service classes |
Equivalence classes for IP can be pairwise, by matching against both source and destination address+port, pure group using only destination address+port, or source-specific multicast with source adress+port and destination multicast address+port.¶
With Intserv, the signaling protocol RSVP [RFC2205] carries two data structures, the FLOWSPEC and the TSPEC. The former fulfills the requirement to identify the equivalence class to which the QoS being signaled applies. The latter comprises the desired QoS treatment along with a description of the dynamic character of the traffic (e.g. average bandwidth and delay, peak bandwidth, etc.). Both of these encounter substantial scaling limits, which has meant that Intserv has historically been limited to confined topologies, and/or high-value usages, like traffic engineering.¶
With Diffserv, the protocol encoding (6 bits in the TOS field of the IP header) artificially limits the number of classes one can specify. These are documented in [RFC4594]. Nonetheless, when used with fine-grained equivalence classes, one still runs into limits on the number of queues required.¶
While one could adopt an approach to QoS mirroring the extensive experience with TCP/IP, this would, in the author's view, be a mistake. The implementation and deployment of QoS in IP networks has been spotty at best. There are of course economic and political reasons as well as technical reasons for these mixed results, but there are several architectural choices in ICN that make it a potentially much better protocol base to enhance with QoS machinery. This section discusses those differences and their consequences.¶
First and foremost, hierarchical names are a much richer basis for specifying equivalence classes than IP 5-tuples. The IP address (or prefix) can only separate traffic by topology to the granularity of hosts, and not express actual computational instances nor sets of data. Ports give some degree of per-instance demultiplexing, but this tends to be both coarse and ephemeral, while confounding the demultiplexing function with the assignment of QoS treatments to particular subsets of the data. Some degree of finer granularity is possible with IPv6 by exploiting the ability to use up to 64 bits of address for classifying traffic. In fact, the hICN project ([I-D.muscariello-intarea-hicn]), while adopting the request-response model of CCNx, uses IPv6 addresses as the available namespace, and IPv6 packets (plus "fake" TCP headers) as the wire format.¶
Nonetheless, the flexibility of tokenized, variable length, hierarchical names allows one to directly associate classes of traffic for QoS purposes with the structure of an application namespace. The classification can be as coarse or fine-grained as desired by the application. While not always the case, there is typically a straightforward association between how objects are named, and how they are grouped together for common treatment. Examples abound; a number can be conveniently found in [I-D.moiseenko-icnrg-flowclass].¶
In ICN, QoS is not pre-bound to network topology since names are non-topological, unlike unicast IP addresses. This allows QoS to be applied to multi-destination and multi-path environments in a straightforward manner, rather than requiring either multicast with coarse class-based scheduling or complex signaling like that in RSVP-TE [RFC3209] that is needed to make point-to-multipoint MPLS work.¶
Because of IP's stateless forwarding model, complicated by the ubiquity of asymmetric routes, any flow-based QoS requires state that is decoupled from the actual arrival of traffic and hence must be maintained, at least as soft-state, even during quiescent periods. Intserv, for example, requires flow signaling with state O(#flows). ICN, even worst case, requires state O(#active Interest/Data exchanges), since state can be instantiated on arrival of an Interest, and removed lazily once the data hase been returned.¶
Unlike Intserv, Difserv eschews signaling in favor of class-based configuration of resources and queues in network elements. However, Diffserv limits traffic treatments to a few bits taken from the ToS field of IP. No such wire encoding limitations exist for NDN or CCNx, as the protocol is completely TLV-based, and one (or even more than one) new field can be easily defined to carry QoS treatment information.¶
Therefore, there are greenfield possibilities for more powerful QoS treatment options in ICN. For example, IP has no way to express a QoS treatment like "try hard to deliver reliably, even at the expense of delay or bandwidth". Such a QoS treatment for ICN could invoke native ICN mechanisms, none of which are present in IP, such as:¶
Such mechanisms are typically described in NDN and CCNx as forwarding strategies. However, little or no guidance is given for what application actions or protocol machinery is used to decide which forwarding strategy to use for which Interests that arrive at a forwarder. See [BenAbraham2018] for an investigation of these issues. Associating forwarding strategies with the equivalence classes and QoS treatments directly can make them more accessible and useful to implement and deploy.¶
Stateless forwarding and asymmetric routing in IP limits available state/feedback to manage link resources. In contrast, NDN or CCNx forwarding allows all link resource allocation to occur as part of Interest forwarding, potentially simplifying things considerably. For example, with symmetric routing, producers have no control over the paths their data packets traverse, and hence any QoS treatments intended to influence routing paths from producer to consumer will have no effect.¶
One complication in the handling of ICN QoS treatments is not present in IP and hence worth mention. CCNx and NDN both perform Interest aggregation (See Section 2.3.2 of [RFC8569]). If an Interest arrives matching an existing PIT entry, but with a different QoS treatment from an Interest already forwarded, it can be tricky to decide whether to aggregate the interest or forward it, and how to keep track of the differing QoS treatments for the two Interests. Exploration of the details surrounding these situations is beyond the scope of this document; further discussion can be found for the general case of flow balance and congestion control in [I-D.oran-icnrg-flowbalance], and specifically for QoS treatments in [I-D.anilj-icnrg-dnc-qos-icn].¶
IP has three forwarding semantics, with different QoS needs (Unicast, Anycast, Multicast). ICN has the single forwarding semantic, so any QoS machinery can be uniformly applied across any request/response invocation, whether it employs dynamic destination routing, multi-destination parallel requests, or even localized flooding (e.g. directly on L2 multicast mechanisms). Additionally, the pull-based model of ICN avoids a number of thorny multicast QoS problems that IP has ([Wang2000], [RFC3170], [Tseng2003]).¶
The Multi-destination/multi-path forwarding model in ICN changes resource allocation needs in a fairly deep way. IP treats all endpoints as open-loop packet sources, whereas NDN and CCNx have strong asymmetry between producers and consumers as packet sources.¶
IP has no caching in routers, whereas ICN needs ways to allocate cache resources. Treatments to control caching operation are unlikely to look much like the treatments used to control link resources. NDN and CCNx already have useful cache control directives associated with Data messages. The CCNx controls include:¶
See [RFC8569] for the formal definitions.¶
ICN flow classifiers, such as those in [I-D.moiseenko-icnrg-flowclass] can be used to achieve soft or hard partitioning of cache resources in the content store of an ICN forwarder. For example, cached content for a given equivalence class can be considered fate shared in a cache whereby objects from the same equivalence class can be purged as a group rather than individually. This can recover cache space more quickly and at lower overhead than pure per-object replacement when a cache is under extreme pressure and in danger of thrashing. In addition, since the forwarder remembers the QoS treatment for each pending Interest in its PIT, the above cache controls can be augmented by policy to prefer retention of cached content for some equivalence classes as part of the cache replacement algorithm.¶
Based on the observations made in the earlier sections, this summary section captures the author's ideas for clear and actionable architectural principles for how to incorporate QoS machinery into ICN protocols like NDN and CCNx. Hopefully, they can guide further work and focus effort on portions of the giant design space for QoS that have the best tradeoffs in terms of flexibility, simplicity, and deployability.¶
Define equivalence classes using the name hierarchy rather than creating an independent traffic class definition. This directly associates the specification of equivalence classes of traffic with the structure of the application namespace. It can allow hierarchical decomposition of equivalence classes in a natural way because of the way hierarchical ICN names are constructed. Two practical mechanisms are presented in [I-D.moiseenko-icnrg-flowclass] with different tradeoffs between security and the ability to aggregate flows. Either prefix-based (EC3) or explicit name component based (ECNT) or both could be adopted as the part of the QoS architecture for defining equivalence classes.¶
Put consumers in control of Link and Forwarding resource allocation. Do all link buffering and forwarding (both memory and CPU) resource allocations based on Interest arrivals. This is attractive because it provides early congestion feedback to consumers, and allows scheduling the reverse link direction ahead of time for carrying the matching data. It makes enforcement of QoS treatments a single-ended (i.e. at the consumer) rather than a double-ended problem and can avoid wasting resources on fetching data that will wind up dropped when it arrives at a bottleneck link.¶
Allow producers to influence the allocation of cache resources. Producers want to affect caching decisions in order to:¶
For caching to be effective, individual Data objects in an equivalence class need to have similar treatment; otherwise well-known cache thrashing pathologies due to self-interference emerge. Producers have the most direct control over caching policies through the caching directives in Data messages. It therefore makes sense to put the producer, rather than the consumer or network operator in charge of specifying these equivalence classes.¶
See [I-D.moiseenko-icnrg-flowclass] for specific mechanisms to achieve this.¶
Allow consumers to influence the allocation of cache resources. Consumers want to affect caching decisions in order to:¶
Consumers can have indirect control over caching by specifying QoS treatments in their Interests. Consider the following potential QoS treatments by consumers that can drive caching policies:¶
Give network operators the ability to match customer SLAs to cache resource availability. Network operators, whether closely tied administratively to producer or consumer, or constituting an independent transit administration, provide the storage resources in the ICN forwarders. Therefore, they are the ultimate arbiters of how the cache resources are managed. In addition to any local policies they may enforce, the cache behavior from the QoS standpoint emerges from how the producer-specified equivalence classes map onto cache space availability, including whether cache entries are treated individually, or fate-shared. Forwarders also determine how the consumer-specified QoS treatments map to the precedence used for retaining Data objects in the cache.¶
Besides utilizing cache resources to meet the QoS goals of individual producers and consumers, network operators also want to manage their cache resources in order to:¶
Re-think how to specify traffic treatments - don't just copy Diffserv. Some of the Diffserv classes may form a good starting point, as their mapping onto queuing algorithms for managing link buffering are well understood. However, Diffserv alone does not allow one to express latency versus reliability tradeoffs or other useful QoS treatments. Nor does it permit "TSPEC"-style traffic descriptions as are allowed in a signaled QoS scheme. Here are some examples:¶
As an aside, loose latency control can be achieved by bounding Interest Lifetime as long as it is not also used as an application mechanism to provide subscriptions or establish path traces for producer mobility. See [Krol2018] for a discussion of the network versus application timescale issues in ICN protocols.¶
Basic QoS treatments such as those summarized above may not be adequate to cover the whole range of application utility functions and deployment environments we expect for ICN. While it is true that one does not necessarily need a separate signaling protocol like RSVP given the state carried in the ICN data plane by forwarders, there are some potentially important capabilities not provided by just simple QoS treatments applied to per- Interest/Data exchanges. INTserv's richer QoS capabilities may be of value, especially if they can be provided in ICN at lower complexity and protocol overhead than INTServ+RSVP.¶
There are three key capabilities missing from Diffserv-like QoS treatments, no matter how sophisticated they may be in describing the desired treatment for a given equivalence class of traffic. INTserv-like QoS provides all of these:¶
Given the limited applicability of these capabilities in today's Internet, the author does not take any position as to whether any of these INTserv-like capabilities are needed for ICN to be succesful. However, a few things seem important to consider. The following paragraphs speculate about the consequences to the CCNx or NDN protocol architectures of incorporating these features.¶
Superficially, it would be quite straightforward to accommodate INTserv-equivalent traffic descriptions in CCNx or NDN. One could define a new TLV for the Interest message to carry a TSPEC. A forwarder encountering this, together with a QoS treatment request (e.g. as proposed in Section 6.3) could associate the traffic specification with the corresponding equivalence class derived from the name in the Interest. This would allow the forwarder to create state that not only would apply to the returning Data for that Interest when being queued on the downstream interface, but be maintained as soft state across multiple Interest/Data exchanges to drive policing and shaping algorithms at per-flow granularity. The cost in Interest message overhead would be modest, however the complications associated with managing different traffic specifications in different Interests for the same equivalence class might be substantial. Of course, all the scalability considerations with maintaining per-flow state also come into play.¶
Similarly, it would be equally straightforward to have a way to express the degree of divergence capability that INTserv provides through its controlled load and guaranteed service definitions. This could either be packaged with the traffic specification or encoded separately.¶
In contrast to the above, performing admission control for ICN flows is likely to be just as heavy-weight as it turned out to be with IP using RSVP. The dynamic multi-path, multi-destination forwarding model of ICN makes performing admission control particularly tricky. Just to illustrate:¶
Despite the challenges above, it may be possible to craft an admission control scheme for ICN that achieves the desired QoS goals of applications without the invention and deployment of a complex separate admission control signaling protocol. There have been designs in earlier network architectures that were capable of performing admission control piggybacked on packet transmission.¶
(The earliest example the author is aware of is [Autonet]).¶
Such a scheme might have the following general shape (warning: serious hand waving follows!):¶
This document does not require any IANA actions.¶
There are a few ways in which QoS for ICN interacts with security and privacy issues. Since QoS addresses relationships among traffic rather than the inherent characteristics of traffic, it neither enhances nor degrades the security and privacy properties of the data being carried, as long as the machinery does not alter or otherwise compromise the basic security properties of the associated protocols. The QoS approaches advocated here for ICN can serve to amplify existing threats to network traffic however:¶