Internet-Draft | RAW Architecture | September 2024 |
Thubert | Expires 13 March 2025 | [Page] |
Reliable and Available Wireless (RAW) provides for high reliability and availability for IP connectivity across any combination of wired and wireless network segments. The RAW Architecture extends the Deterministic Networking (DetNet) Architecture and other standard IETF concepts and mechanisms to adapt to the specific challenges of the wireless medium, in particular intermittently lossy connectivity. This document defines a network control loop that optimizes the use of constrained spectrum and energy while maintaining the expected connectivity properties, typically reliability and latency. The loop involves DetNet Operational Plane functions, with a new recovery Function and a new Point of Local Repair operation, that dynamically selects the DetNet path(s) for the future packets to route around local degradations and failures.¶
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 13 March 2025.¶
Copyright (c) 2024 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. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Deterministic Networking aims to provide bounded latency and eliminate congestion loss, even when co-existing with best-effort traffic. It is getting traction in various industries including professional A/V, manufacturing, online gaming, and smartgrid automation, with both cost savings and complexity benefits (e.g., vs. loads of point-to-point (P2P) cables).¶
Bringing determinism in a packet network means minimizing the statistical effects of multiplexing that result in probabilistic jitter and loss. This can be approached with a tight control of the physical resources to maintain the amount of traffic within a budgeted volume of data per unit of time that fits the physical capabilities of the underlying network, and the use of time-shared resources (bandwidth and buffers) per circuit, and/or by shaping and/or scheduling the packets at every hop.¶
This innovation was initially introduced on wired networks, with IEEE 802.1 Time-Sensitive networking (TSN) - for Ethernet LANs - and IETF DetNet. But the wired and the wireless media are fundamentally different at the physical level and in the possible abstractions that can be built for IPv6 [IPv6], more in [IPoWIRELESS]. Nevertheless, deterministic capabilities are required in a number of wireless use cases as well [RAW-USE-CASES]. With scheduled radios such as Time Slotted Channel Hopping (TSCH) and Orthogonal Frequency Division Multiple Access (OFDMA) [RAW-TECHNOS] being developed to provide determinism over wireless links at the lower layers, providing DetNet capabilities is now becoming possible.¶
Wireless networks operate on a shared medium where uncontrolled interference, including the self-induced multipath fading cause random transmission losses. Fixed and mobile obstacles and reflectors may block or alter the signal, causing transient and unpredictable variations of the throughput and packet delivery ratio (PDR) of a wireless link. This adds new dimensions to the statistical effects that affect the quality and reliability of the link.¶
Reliable and Available Wireless (RAW) takes up the challenge of providing highly available and reliable end-to-end performances in a network with scheduled wireless segments. To achieve this, RAW can leverage multiple links and parallel transmissions, providing enough diversity and redundancy to ensure the timely packet delivery while preserving energy and optimizing the use of the shared spectrum.¶
Distance Vector (DV) protocols can enable more than one feasible successors along non-equal-cost multipath forwarding graphs. This provides redundancy and allows to dynamically adapt the forwarding operation to the state of the links. But this protection is limited since only a subset of the nodes along the path have a feasible alternate successor.¶
RAW solves that problem by defining Protection Paths that can be fully non-congruent and can be activated dynamically upon failures. This requires additional control to take the routing decision early enough along the possible paths to route around the failure. RAW defines an end-to-end control loop that dynamically controls the activation and deactivation of the feasible Protection Paths.¶
In addition, RAW introduces the RAW API, which is an interface between the lower layer wireless technology and the DetNet layers. The RAW API is RAW technology [RAW-TECHNOS] dependent as it can vary what the different RAW technologies expose towards the DetNet layers. Furthermore, the different RAW technologies are equipped with different reliability features, e.g., short range broadcast, MUMIMO, PHY rate and other Modulation Coding Scheme (MCS) adaptation, (H)ARQ, constructive interference and overhearing. The RAW API enables interactions between the reliability functions provided by the wireless technology and the reliability functions provided by DetNet. That is, the RAW API makes cross-layer optimization possible for the reliability functions of different layers depending on the actual exposure provided via the RAW API by the given RAW technology.¶
This document presents the RAW problem and associated terminology in Section 3.2, presents a conceptual model for RAW in Section 4, and, based on that model, elaborates on an in-network optimization control loop in Section 5.2.¶
RAW reuses terminology defined for DetNet in the "Deterministic Networking Architecture" [RFC8655], e.g., PREOF for Packet Replication, Elimination and Ordering Functions. RAW inherits and augments the IETF art of Protection as seen in DetNet and Traffic Engineering.¶
RAW also reuses terminology defined for MPLS in [RFC4427] such as the term recovery as covering both Protection and Restoration, a number of recovery types. That document defines a number of concepts like recovery domain that are used in the RAW works, and creates the new term recovery graph. A recovery graph associates a topological graph with usage metadata that represent how the paths within the recovery graph are built.¶
RAW also reuses terminology defined for RSVP-TE in [RFC4090] such as the Point of Local Repair (PLR). The concept of backup path is generalized with protection path, which is the term mostly found in recent standards and used in this document.¶
RAW also reuses terminology defined for 6TiSCH in [6TiSCH-ARCHI] and equates the 6TiSCH concept of a Track with that of a recovery graph.¶
In an quantic analogy, a recovery graph is to a path what an atomic orbital is to a planetary orbit, in that the electron has a probability of presence within a known shape as opposed to a deterministic trajectory.¶
The concept of recovery graph is agnostic to the underlaying technology and applies but is not limited to any fully or partially wireless mesh. RAW specifies strict and loose recovery graphs depending on whether the path is fully controlled by RAW or traverses an opaque network where RAW cannot observe and control the individual hops.¶
RAW uses the following terminology and acronyms:¶
Automatic Repeat Request, a well-known mechanism, enabling an acknowledged transmission with retries to mitigate errors and loss. ARQ may be implemented at various layers in a network. ARQ is typically implemented at Layer-2, per hop and not end-to-end in wireless networks. ARQ improves delivery on lossy wireless. Additionally, ARQ retransmission may be further limited by a bounded time to meet end-to-end packet latency constraints. Additional details and considerations for ARQ are detailed in [RFC3366].¶
Forward Error Correction, adding redundant data to protect against a partial loss without retries.¶
Hybrid Automatic Repeat Request, combining FEC and ARQ.¶
Modulation and Coding Scheme. Controls the throughput of the Link to maintain reliable transmissions.¶
OAM stands for Operations, Administration, and Maintenance, and covers the processes, activities, tools, and standards involved with operating, administering, managing and maintaining any system. This document uses the terms Operations, Administration, and Maintenance, in conformance with the 'Guidelines for the Use of the "OAM" Acronym in the IETF' [RFC6291] and the system observed by the RAW OAM is the recovery graph.¶
Observe, Orient, Decide, Act. The OODA Loop is a conceptual cyclic model developed by USAF Colonel John Boyd, and that is applicable in multiple domains where agility can provide benefits against brute force.¶
In the context of RAW, a link flaps when the reliability of the wireless connectivity drops abruptly for a short period of time, typically of a subsecond to seconds duration.¶
Connection from end-devices to a data communication equipment. In the context of wireless, uplink refers to the connection between a station (STA) and a controller (AP) or a User Equipment (UE) to a Base Station (BS) such as a 3GPP 5G gNodeB (gNb).¶
The reverse direction from uplink.¶
Following the direction of the flow data path along a recovery graph.¶
Against the direction of the flow data path along a recovery graph.¶
Quoting section 1.1.3 of [INT-ARCHI]:¶
At a given moment, all the IP datagrams from a particular source host to a particular destination host typically traverse the same sequence of gateways. We use the term "path" for this sequence. Note that a path is unidirectional; it is not unusual to have different paths in the two directions between a given host pair.¶
Section 2 of [I-D.irtf-panrg-path-properties] points to a longer, more modern definition of path, which begins as follows:¶
A sequence of adjacent path elements over which a packet can be transmitted, starting and ending with a node. A path is unidirectional. Paths are time-dependent, i.e., the sequence of path elements over which packets are sent from one node to another may change. A path is defined between two nodes.¶
It follows that the general acceptance of a path is a linear sequence of links and nodes, as opposed to a multi-dimensional graph, defined by the experience of the packet that went from a node A to a node B. In the context of this document, a path is observed by following one copy or one fragment of a packet that conserves its uniqueness and integrity. For instance, if C replicates to E and F and D eliminates on the way from A to B, a packet from A to B can experience 2 paths, A->C->E->D->B and A->C->F->D->B. The terms lane is used to clarify when dealing with such path.¶
With DetNet and RAW, a packet may be duplicated, fragmented and network-coded, and the various byproducts may travel different paths that are not necessarily end-to-end between A and B; we refer to that complex experience as a DetNet path. As such, the DetNet path extends the above description of a path, but it still matches the experience of a packet that traverses the network.¶
With RAW, that experience is subject to change from a packet to the next, but all the possible experiences are all contained within a finite set. Therefore, we introduce below the term of a recovery graph that coalesces that set and covers the overall topology where the possible DetNet paths are all contained. As such, the recovery graph coalesces all the possible paths a flow may experience, each with its own statistical probability to be used.¶
A networking graph that can be followed to transport packets with equivalent treatment, associated with usage metadata; as opposed to the definition of a path above, a recovery graph represents not an actual but a potential, it is not necessarily a linear sequence like a simple path, and is not necessarily fully traversed (flooded) by all packets of a flow like a Detnet Path. Still, and as a simplification, the casual reader may consider that a recovery graph is very much like a DetNet path, aggregating multiple paths that may overlap, fork and rejoin, for instance to enable a protection service by the PREOF operations.¶
Refining further, a recovery graph is defined as the coalescence of the collection of all the feasible DetNet Paths that a packet which flow is assigned to the recovery graph may be forwarded along. A packet that is assigned to the recovery graph experiences one of the feasible DetNet Paths based on the current selection by the PLR at the time the packet traverses the network.¶
Refining even further, the feasible DetNet Paths within the recovery graph may or may not be computed in advance, but decided upon the detection of a change from a clean slate. Furthermore, the PLR decision may be distributed, which yields a large combination of possible and dependant decisions, with no node in the network capable of reporting which is the current DetNet Path within the recovery graph.¶
In DetNet [RFC8655] terms, a recovery graph has the following properties:¶
Figure 2 illustrates the generic concept of a recovery graph, between an Ingress Node and an Egress Node The recovery graph is composed of forward Lanes and forward or crossing Segments, see the definition for those terms in the next sections. A Protection Path contains at least 2 Lanes as a main path and a backup path.¶
Forward refers to progress towards the recovery graph Egress. Forward links are directional, and packets that are forwarded along the recovery graph can only be transmitted along the link direction. Crossing links are bidirectional, meaning that they can be used in both directions, though a given packet may use the link in one direction only. A Segment can be forward, in which case it is composed of forward links only, or crossing, in which case it is composed of crossing links only. A lane is always forward, meaning that it is composed of forward links and Segments.¶
An end-to-end forward lane between the Ingress and Egress Nodes of a recovery graph. A lane in a recovery graph is expressed as a strict sequence of DetNet Relay nodes or as a loose sequence of DetNet Relay nodes that are joined by recovery graph Segments.¶
A strict sequence of DetNet Transit nodes between 2 DetNet Relay nodes; a Segment of a recovery graph is composed topologically of two vertices of the recovery graph and one edge of the recovery graph between those vertices.¶
This document reuses the terminology in section 2 of [RFC8557] and section 4.1.2 of [RFC8655] for deterministic networking and deterministic networks.¶
A collection of consecutive IP packets defined by the upper layers and signaled by the same 5 or 6-tuple, see section 5.1 of [RFC8939]. Packets of the same flow must be placed on the same recovery graph to receive an equivalent treatment from Ingress to Egress within the recovery graph. Multiple flows may be transported along the same recovery graph. The DetNet Path that is selected for the flow may change over time under the control of the PLR.¶
A tuple identified by a stream_handle, and provided by a bridge, in accordance with IEEE 802.1CB. The tuple comprises at least destination MAC and VLAN ID. Continuous streams are characterized by bandwidth and max packet size; scheduled streams are characterized by a repeating pattern of timed transmissions.¶
See section 3.3 of [DetNet-DP]. The classical IP 5-tuple that identifies a flow comprises the source IP, destination IP, source port, destination port, and the upper layer protocol (ULP). DetNet uses a 6-tuple where the extra field is the DSCP field in the packet. The IPv6 flow label is not used for that purpose.¶
TSN stands for Time-Sensitive Networking and denotes the efforts at IEEE 802 for deterministic networking, originally for use on Ethernet. Wireless TSN (WTSN) denotes extensions of the TSN work on wireless media such as the selected RAW technologies [RAW-TECHNOS].¶
In the context of the RAW work, Reliability and Availability are defined as follows:¶
In the context of RAW, an SLA (service level agreement) is a contract between a provider (the network) and a client, the application flow, about measurable metrics such as latency boundaries, consecutive losses, and packet delivery ratio (PDR).¶
A service level objective (SLO) is one term in the SLA, for which specific network setting and operations are implemented. For instance, a dynamic tuning of the packet redundancy addresses an SLO of consecutive losses in a row by augmenting the chances of delivery of a packet that follows a loss.¶
A service level indicator (SLI) measures the compliance of an SLO to the terms of the contract. It can be for instance the statistics of individual losses and losses in a row as time series.).¶
Reliability is a measure of the probability that an item (e.g., system, network) will perform its intended function with no failure for a stated period of time (or number of demands or load) under stated environmental conditions. In other words, reliability is the probability that an item will be in an uptime state (i.e., fully operational or ready to perform) for a stated mission, e.g., to provide an SLA. See more in [NASA1].¶
Availability is the probability of an item’s (e.g., a network’s) mission readiness (e.g., to provide an SLA), an uptime state with the likelihood of a recoverable downtime state. Availability is expressed as (uptime)/(uptime+downtime). Note that it is availability that addresses downtime (incl. time for maintenance, repair, and replacement activities) and not reliability. See more in [NASA2].¶
See [RFC7799]. In the context of RAW, Active OAM is used to observe a particular recovery graph, DetNet Path, or Segment of a recovery graph regardless of whether it is used for traffic at that time.¶
An active OAM packet is considered in-band for the monitored recovery graph when it traverses the same set of links and interfaces and if the OAM packet receives the same QoS and service protection treatment as the packets of the data flows that are injected in the recovery graph.¶
Out-of-band OAM is an active OAM whose path is not topologically congruent to the recovery graph, or its test packets receive a QoS and/or service protection treatment that is different from that of the packets of the data flows that are injected in the recovery graph, or both.¶
An active OAM packet is a Limited OAM packet when it observes the RAW operation over a node, a segment, or a DetNet Path of the recovery graph, though not from Ingress to Egress. It is injected in the datapath and extracted from the datapath around the particular function or subnetwork (e.g., around a relay providing a Service sub-layer replication point) that is being tested.¶
An upstream OAM packet is an Out-of-Band OAM packet that traverses the recovery graph from egress to ingress on the reverse direction, to capture and report OAM measurements upstream. The collection may capture all information along the whole recovery graph, or it may only learn select data across all, or only a particular DetNet Path, or Segment of a recovery graph.¶
A residence time (RT) is defined as the time interval between the reception of a packet starts and the transmission of the packet begins. In the context of RAW, RT is useful for a transit node, not ingress or egress.¶
The RAW Operational Plane elements (OAM Supervisor and local CPF (lCPF)) may gather aggregated information from lower layers about e.g., link quality. This information may be obtained from inside the device using specialized API (e.g., L2 triggers) or via control protocols such as BFD [RFC5880] or DLEP [DLEP]. It may then be massaged and exported through oOAM messaging, and passed to the Controller Plane using the lCPF.¶
[DetNet-OAM] provides additional terminology related to OAM in the context of DetNet and by extension of RAW, whereas [RFC7799] defines the Active, Passive, and Hybrid OAM methods.¶
The reliability criteria of a critical system pervades through its elements, and if the system comprises a data network then the data network is also subject to the inherited reliability and availability criteria. It is only natural to consider the art of high availability engineering and apply it to wireless communications in the context of RAW.¶
There are three principles [pillars] of high availability engineering:¶
These principles are common to all high availability systems, not just ones with Internet technology at the center. Examples of both non-Internet and Internet are included.¶
Physical and logical components in a system happen to fail, either as the effect of wear and tear, when used beyond acceptable limits, or due to a software bug. It is necessary to decouple component failure from system failure to avoid the latter. This allows failed components to be restored while the rest of the system continues to function.¶
IP Routers leverage routing protocols to reroute to alternate routes in case of a failure. There is a rather open-ended issue over alternate routes -- for example, when links are cabled through the same conduit, they form a shared risk link group (SRLG), and share the same fate if the bundle is cut. The same effect can happen with virtual links that end up in a same physical transport through the games of encapsulation. In a same fashion, an interferer or an obstacle may affect multiple wireless transmissions at the same time, even between different sets of peers.¶
Intermediate network Nodes such as routers, switches and APs, wire bundles and the air medium itself can become single points of failure. For High Availability, it is thus required to use physically link- and Node-disjoint paths; in the wireless space, it is also required to use the highest possible degree of diversity (time, space, code, frequency, channel width) in the transmissions over the air to combat the additional causes of transmission loss.¶
From an economics standpoint, executing this principle properly generally increases capitalization expense because of the redundant equipment. In a constrained network where the waste of energy and bandwidth should be minimized, an excessive use of redundant links must be avoided; for RAW this means that the extra bandwidth must be used wisely and with parsimony.¶
Having a backup equipment has a limited value unless it can be reliably switched into use within the down-time parameters. IP Routers execute reliable crossover continuously because the routers use any alternate routes that are available [RFC0791]. This is due to the stateless nature of IP datagrams and the dissociation of the datagrams from the forwarding routes they take. The "IP Fast Reroute Framework" [FRR] analyzes mechanisms for fast failure detection and path repair for IP Fast-Reroute, and discusses the case of multiple failures and SRLG. Examples of FRR techniques include Remote Loop-Free Alternate [RLFA-FRR] and backup label-switched path (LSP) tunnels for the local repair of LSP tunnels using RSVP-TE [RFC4090].¶
Deterministic flows, on the contrary, are attached to specific paths where dedicated resources are reserved for each flow. Therefore each DetNet path must inherently provide sufficient redundancy to provide the guaranteed SLA at all times. The DetNet PREOF typically leverages 1+1 redundancy whereby a packet is sent twice, over non-congruent paths. This avoids the gap during the fast reroute operation, but doubles the traffic in the network.¶
In the case of RAW, the expectation is that multiple transient faults may happen in overlapping time windows, in which case the 1+1 redundancy with delayed reestablishment of the second path does not provide the required guarantees. The Data Plane must be configured with a sufficient degree of redundancy to select an alternate redundant path immediately upon a fault, without the need for a slow intervention from the Controller Plane.¶
The execution of the two above principles is likely to render a system where the user rarely sees a failure. But someone needs to in order to direct maintenance.¶
There are many reasons for system monitoring (FCAPS for fault, configuration, accounting, performance, security is a handy mental checklist) but fault monitoring is sufficient reason.¶
"An Architecture for Describing Simple Network Management Protocol (SNMP) Management Frameworks" [STD 62] describes how to use SNMP to observe and correct long-term faults.¶
"Overview and Principles of Internet Traffic Engineering" [TE] discusses the importance of measurement for network protection, and provides an abstract method for network survivability with the analysis of a traffic matrix as observed by SNMP, probing techniques, FTP, IGP link state advertisements, and more.¶
Those measurements are needed in the context of RAW to inform the controller and make the long-term reactive decision to rebuild a recovery graph based on statistical and aggregated information. RAW itself operates in the DetNet Network Plane at a faster time scale with live information on speed, state, etc... This live information can be obtained directly from the lower layer, e.g., using L2 triggers, read from a protocol such as the Dynamic Link Exchange Protocol (DLEP) [DLEP], or transported over multiple hops using OAM and reverse OAM, as illustrated in Figure 11.¶
The terms Reliability and Availability are defined for use in RAW in Section 2 and the reader is invited to read [NASA1] and [NASA2] for more details on the general definition of Reliability. Practically speaking a number of nines is often used to indicate the reliability of a data link, e.g., 5 nines indicate a Packet Delivery Ratio (PDR) of 99.999%.¶
This number is typical in a wired environment where the loss is due to a random event such as a solar particle that affects the transmission of a particular packet, but does not affect the previous or next packet, nor packets transmitted on other links. Note that the QoS requirements in RAW may include a bounded latency, and a packet that arrives too late is a fault and not considered as delivered.¶
For a periodic networking pattern such as an automation control loop, this number is proportional to the Mean Time Between Failures (MTBF). When a single fault can have dramatic consequences, the MTBF expresses the chances that the unwanted fault event occurs. In data networks, this is rarely the case. Packet loss cannot be fully avoided and the systems are built to resist to some loss, e.g., using redundancy with Retries (HARQ) or Packet Replication and Elimination (PRE), or, in a typical control loop, by linear interpolation from the previous measurements.¶
But the linear interpolation method cannot resist multiple consecutive losses, and a high MTBF is desired as a guarantee that this does not happen, IOW that the number of losses-in-a-row can be bounded. In that case, what is really desired is a Maximum Consecutive Loss (MCL). (See also section 5.9.5 in [DLEP].) If the number of losses in a row passes the MCL, the control loop has to abort and the system, e.g., the production line, may need to enter an emergency stop condition.¶
Engineers that build automated processes may use the network reliability expressed in nines or as an MTBF as a proxy to indicate an MCL, e.g., as described in section 7.4 of the "Deterministic Networking Use Cases" [RFC8578].¶
In contrast with wired networks, errors in transmission are the predominant source of packet loss in wireless networks.¶
The root cause for the loss may be of multiple origins, calling for the use of different forms of diversity:¶
A destructive interference by a reflection of the original signal.¶
A radio signal may be received directly (line-of-sight) and/or as a reflection on a physical structure (echo). The reflections take a longer path and are delayed by the extra distance divided by the speed of light in the medium. Depending on the frequency, the echo lands with a different phase which may add up to (constructive interference) or cancel the direct signal (destructive interference).¶
The affected frequencies depend on the relative position of the sender, the receiver, and all the reflecting objects in the environment. A given hop suffers from multipath fading for multiple packets in a row till a physical movement changes the reflection patterns.¶
Energy in the spectrum used for the transmission confuses the receiver.¶
The wireless medium itself is a Shared Risk Link Group (SRLG) for nearby users of the same spectrum, as an interference may affect multiple co-channel transmissions between different peers within the interference domain of the interferer, possibly even when they use different technologies.¶
The optimal transmission happens when the Fresnel Zone between the sender and the receiver is free of obstacles.¶
As long as a physical object (e.g., a metallic trolley between peers) that affects the transmission is not removed, the quality of the link is affected.¶
In an environment that is rich of metallic structures and mobile objects, a single radio link provides a fuzzy service, meaning that it cannot be trusted to transport the traffic reliably over a long period of time.¶
Transmission losses are typically not independent, and their nature and duration are unpredictable; as long as a physical object (e.g., a metallic trolley between peers) that affects the transmission is not removed, or as long as the interferer (e.g., a radar) keeps transmitting, a continuous stream of packets are affected.¶
The key technique to combat those unpredictable losses is diversity. Different forms of diversity are necessary to combat different causes of loss and the use of diversity must be maximized to optimize the PDR.¶
A single packet may be sent at different times (time diversity) over diverse paths (spatial diversity) that rely on diverse radio channels (frequency diversity) and diverse PHY technologies, e.g., narrowband vs. spread spectrum, or diverse codes. Using time diversity defeats short-term interferences; spatial diversity combats very local causes such as multipath fading; narrowband and spread spectrum are relatively innocuous to one another and can be used for diversity in the presence of the other.¶
While the generic "Deterministic Networking Problem Statement" [RFC8557] applies to both the wired and the wireless media, the methods to achieve RAW must extend those used to support time-sensitive networking over wires, as a RAW solution has to address less consistent transmissions, energy conservation and shared spectrum efficiency.¶
Operating at the Layer-3, RAW does not change the wireless technology at the lower layers. OTOH, it can further increase diversity in the spatial, time, code, and frequency domains by enabling multiple link-layer wired and wireless technologies in parallel or sequentially, for a higher resilience and a wider applicability. RAW can also provide homogeneous services to critical applications beyond the boundaries of a single subnetwork, e.g., using diverse radio access technologies to optimize the end-to-end application experience.¶
RAW extends the DetNet services by providing elements that are specialized for transporting IP flows over deterministic radio technologies such as listed in [RAW-TECHNOS]. Conceptually, RAW is agnostic to the radio layer underneath though the capability to schedule transmissions is assumed. How the PHY is programmed to do so, and whether the radio is single-hop or meshed, are unknown at the IP layer and not part of the RAW abstraction. Nevertheless, cross-layer optimizations may take place to ensure proper link awareness (think, link quality) and packet handling (think, scheduling).¶
The "Deterministic Networking Architecture" [RFC8655] is composed of three planes: the Application (User) Plane, the Controller Plane, and the Network Plane. The DetNet Network Plane is composed a Dataplane (packet forwarding) and an Operational Plane where OAM operations take place. In the Network Plane, the DetNet service sub-layer focuses on flow protection (e.g., using redundancy) and can be fully operated at Layer-3, while the DetNet forwarding sub-layer establishes the paths, associates the flows to the paths, and ensures the availability of the necessary resources, leverages Layer-2 functionalities for timely delivery to the next DetNet system.¶
The RAW Architecture extends the DetNet Network Plane, to accommodate one or multiple hops of homogeneous or heterogeneous wired and wireless technologies. RAW adds reactivity to the DetNet Forwarding sub-layer to compensate the dynamics for the radio links in terms of lossiness and bandwidth. This may apply for instance to mesh networks as illustrated in Figure 4, or diverse radio access networks as illustrated in Figure 10.¶
As opposed to wired links, the availability and performance of an individual wireless link cannot be trusted over the long term; it varies with transient service discontinuity, and any lane that includes wireless hops is bound to face short periods of high loss. On the other hand, being broadcast in nature, the wireless medium provides capabilities that are atypical on modern wired links and that the RAW Architecture can leverage opportunistically to improve the end-to-end reliability over a collection of paths.¶
Those capabilities include:¶
RAW and DetNet route application flows that require a special treatment along the paths that provide that treatment. This may be seen as a form of Path Aware Networking and may be subject to impediments documented in [RFC9049].¶
The establishment of a path is not in-scope for RAW. It may be, e.g., the product of a centralized Controller Plane Function (CPF) like a Path computation Element (PCE) [RFC4655], or may be computed in a distributed fashion ala Resource ReSerVation Protocol (RSVP) [RFC2205]. On the other hand, RAW leverages DetNet Network Plane enhancements to optimize the use of the paths and match the quality of the transmissions over time.¶
As opposed to wired networks, the action of installing a path over a set of wireless links may be very slow relative to the speed at which the radio conditions vary, and it makes sense in the wireless case to provide redundant forwarding solutions along a (see Section 2.3) and to leave it to the Network Plane to select which of those forwarding solutions are to be used for a given packet based on the current conditions.¶
RAW distinguishes the longer time scale at which routes are computed from the shorter time scale where forwarding decisions are made. For a limited time, RAW Network Plane operations happen at a time scale that sits between the routing and the forwarding time scales, on one DetNet flow, to select a DetNet path within the resources delineated by a recovery graph (see Section 2.3.2). The recovery graph is preestablished and installed by means outside of the scope of RAW; it may be strict or loose depending on whether each or just a subset of the hops are observed and controlled by RAW.¶
The RAW Architecture is based on an abstract OODA Loop (Observe, Orient, Decide, Act). The generic concept involves:¶
The overall OODA Loop optimizes the use of redundancy to achieve the required reliability and availability Service Level Agreement (SLA) while minimizing the use of constrained resources such as spectrum and battery.¶
RAW inherits the conceptual model described in section 4 of the DetNet Architecture [RFC8655] as illustrated in Figure 3, which also shows example reliability Functions in the different layers. RAW extends DetNet with Point of Local Repair (PLR, see Section 5.5), which is a point of local reaction to provide additional agility against transmission loss. The PLR can act, e.g., based on indications from the wireless layer or based on OAM.¶
The RAW API enables interactions between the reliability functions provided by the wireless technology and the reliability functions provided by DetNet. Thus, the RAW API enables cross-layer optimizations to improve reliability and availability.¶
A RAW Network Plane may be strict (as illustrated in Figure 6 or loose (as illustrated in Figure 7, depending on whether RAW observes and takes actions on all hops or not. For instance, the packets between two wireless entities may be relayed over a wired infrastructure, in which case RAW observes and controls the transmission over the wireless first and last hops, as well as end-to-end metrics such as latency, jitter, and delivery ratio. This operation is loose since the structure and properties of the wired infrastructure are ignored, and may be either controlled by other means such as DetNet/TSN, or neglected in the face of the wireless hops.¶
The RAW Nodes are DetNet relays that operate in the RAW Network Plane and are capable of additional diversity mechanisms and measurement functions related to the radio interface. RAW leverages a CPF that operates inside the RAW Nodes (typically the Ingress Edge Nodes) to dynamically adapt the path of the packets and optimizes the resource usage.¶
An RAW-enabled rCPF interacts with RAW Nodes over a Southbound API. It consumes data and information from the network and generates knowledge and wisdom to help steer the traffic optimally inside a recovery graph.¶
When a new flow is defined, the rCPF uses its current knowledge of the network to build a new recovery graph between an Ingress End System and an Egress End System for that flow; it indicates to the RAW Nodes where the PREOF and/or radio diversity and reliability operations may be actioned in the Network Plane.¶
A local CPF (lCPF) in the RAW node reports the Link-Layer metrics to the rCPF in a time-aggregated, e.g., statistical fashion. Example Link-Layer metrics include typical Link bandwidth (the medium speed depends dynamically on the PHY mode), number of flows (bandwidth that can be reserved for a flow depends on the number and size of flows sharing the spectrum) and average and mean squared deviation of availability and reliability figures such as Packet Delivery Ratio (PDR) over long periods of time.¶
Based on those metrics, the DetNet rCPF installs the recovery graph with enough redundant forwarding solutions to ensure that the Network Plane can reliably deliver the packets within a System Level Agreement (SLA) associated to the flows that it transports. The SLA defines end-to-end reliability and availability requirements, where reliability may be expressed as a successful delivery in order and within a bounded delay of at least one copy of a packet.¶
Depending on the use case and the SLA, the recovery graph may comprise non-RAW segments, either interleaved inside the recovery graph (e.g. over tunnels), or all the way to the Egress End Node (e.g., a server in the local wired domain). RAW observes the Lower-Layer Links between RAW nodes (typically, radio links) and the end-to-end Network Layer operation to decide at all times which of the diversity schemes is actioned by which RAW Nodes.¶
Once a recovery graph is established, per-segment and end-to-end reliability and availability statistics are periodically reported to the rCPF to assure that the SLA can be met or if not, then have the recovery graph recomputed.¶
RAW improves the reliability of transmissions and the availability of the communication resources, but does not provide scheduling and shaping, so RAW itself does not provide guarantees such as latency for the application payload. Rather, it should be seen as a dynamic optimization of the use of redundancy to maintain it within certain boundaries. For instance, ARQ is operated by the lower layers and RAW only abstracts the concept and hints the lower layers on the desired outcome, as opposed to performing the retries at Layer-3.¶
Guarantees such as bounded latency depend on the upper layers (Transport or Application) to provide the payload in volumes and at times that match the contract with the DetNet sub-layers and the layers below. Excess of incoming traffic at the DetNet Ingress causes either dropping, queueing, or reclassification of the packets, and entail loss, latency, or jitter, and moot the guarantees that are provided inside the DetNet Network.¶
When the traffic from upper layers matches the expectation of the lower layers, RAW still depends on the lower layers to provide the timing and physical resources guarantees that are needed to match the traffic SLA. When the availability of the physical resource varies, RAW acts on the distribution of the traffic to leverage alternates within a finite set of potential resources.¶
RAW leverages the DetNet Forwarding sub-layer and requires the support of OAM in DetNet Transit Nodes (see fig 3 of [RFC8655] for the dynamic acquisition of link capacity and state to maintain a strict RAW service, end-to-end, over a DetNet Network. RAW extends DetNet to improve the protection against link errors such as transient flapping that are far more common in wireless links. Nevertheless, the RAW methods are for the most part applicable to wired links as well, e.g., when energy savings are desirable and the available path diversity exceeds 1+1 linear redundancy.¶
RAW adds sub-layer functions that operate in the DetNet Operational Plane. The RAW functions such as lCPF and OAM typically run only in the DetNet Ingress Edge Node or End System, though it may also run in DetNet Relay Nodes when the RAW operations are distributed along the recovery graph. The RAW functions include the PLR, which decides the DetNet Path for the future packets of a flows along the DetNet Path, and the OAM Supervisor, which triggers, and learns from, OAM observations, and feeds the PLR for its next decision.¶
As illustrated in Figure 5, RAW extends the DetNet Stack (see fig 4 of [RFC8655] and Figure 3) with additional functionality at the DetNet Service sub-layer for the actuation of PREOF based on the PLR decision. Layer-3 in general and DetNet in particular operates on abstractions of the lower layers and through APIs to control those abstractions. For instance, DetNet already leverages lower layers for time-sensitive operations such as time synchronization and traffic shapers. As the performances of the radio layers are subject to rapid changes, RAW needs more dynamic gauges and knobs. To that effect, the RAW API enables interactions between the reliability functions provided by the wireless technology and the reliability functions provided by DetNet. That is, the RAW API provides a radio abstraction to the DetNet layer. The RAW API can be used to push reliability and timing hints like suggest X retries (min, max) within a time window, or send unicast (one next hop) or multicast (for overhearing). The other way +-+ around RAW needs hints about the radio conditions like L2 triggers (RSSI, LQI, ETX…) over all the wireless hops. This information is useful to both the lCPF and the PLR.¶
RAW uses various OAM functionalities at the different layers. For instance, the OAM function in the DetNet Service sub-layer (re)generates the OAM information as it is formed and propagated In-Band or Out-of-Band. The RAW functions may be present in Service sub-layer in DetNet Edge and Relay Nodes.¶
There are 2 main proposed models to deploy RAW and DetNet. In the first model (strict) illustrated in Figure 6, RAW operates over a continuous DetNet Service end-to-end between the Ingress and the Egress Edge Nodes or End Systems.¶
A minimal Forwarding sub-layer service is provided at all DetNet Nodes to ensure that the OAM information flows. Relay Nodes may or may not support RAW services, and the Edge nodes do support RAW. DetNet guarantees such as latency are provided end-to-end, and RAW supports the DetNet Service to optimize the use of resources.¶
In the second model (loose), illustrated in Figure 7, RAW operates over a partial DetNet Service where typically only the Ingress and the Egress End Systems support RAW. The DetNet Domain may extend beyond the Ingress node, or there may be a DetNet domain starting at an Ingress Edge Node at the first hop after the End System.¶
In the loose model, RAW cannot observe the hops in network, and the path beyond the first hop is opaque; RAW can still observe the end-to-end behavior and use Layer-3 measurements to decide whether to replicate a packet and select the first hop interface(s).¶
With DetNet, the Controller Plane Function handles the routing computation and maintenance. With RAW, the routing part of the CPF (rCPF) is segregated from the RAW Control Loop, so it may reside outside of the RAW network. To achieve RAW capabilities, the rCPF is extended to generate the information required by the lCPF, which acts as the orientation component in the loop. The rCPF may, e.g., propose DetNet Paths to be used as a reflex action in response to network events, or by provide aggregated history that the lCPF can use to make an oriented decision.¶
In a wireless mesh, the path to the DetNet CPF can be expensive and slow, possibly going across the whole mesh and back. Reaching to the CPF can also be slow in regards to the speed of events that affect the forwarding operation at the radio layer. Note that a distributed routing protocol may also take time and consume excessive wireless resources to reconverge to a new optimized state.¶
As a result, the DetNet CPF is not expected to be aware of and to react to very transient changes. The abstraction of a link at the routing level is expected to use statistical metrics that aggregate the behavior of a link over long periods of time, and represent its properties as shades of gray as opposed to numerical values such as a link quality indicator, or a boolean value for either up or down.¶
The interaction with the (remote) RAW rCPF is handled by the lCPF, which builds reports to the rCPF and digests the control information back, to be used inside a forwarding control loop for traffic steering.¶
In the case of wireless, the changes that affect the forwarding decision can happen frequently and often for short durations, e.g., a mobile object moves between a transmitter and a receiver, and cancels the line of sight transmission for a few seconds, or a radar measures the depth of a pool and interferes on a particular channel for a split second.¶
There is thus a desire to separate the long-term computation of the route and the short-term forwarding decision. In that model, the routing operation computes a recovery graph that enables multiple Non-Equal Cost Multi-Path (N-ECMP) forwarding solutions along so-called protection paths, and leaves it to the Network Plane to make the per-packet decision of which of these possibilities should be used.¶
In the context of Traffic Engineering (TE), an alternate path can be used upon the detection of a failure in the main path, e.g., using OAM in MPLS-TP or BFD over a collection of SD-WAN tunnels. RAW formalizes a forwarding time scale that is an order(s) of magnitude shorter than the controller plane routing time scale, and separates the protocols and metrics that are used at both scales. Routing can operate on long-term statistics such as delivery ratio over minutes to hours, but as a first approximation can ignore flapping. On the other hand, the RAW forwarding decision is made at the scale of the packet rate, and uses information that must be pertinent at the present time for the current transmission(s).¶
OODA (Observe, Orient, Decide, Act) is a generic formalism to represent the operational steps in a Control Loop. The RAW Architecture applies that generic model to continuously optimize the spectrum and energy used to forward packets within a recovery graph, instantiating the OODA steps as follows:¶
The overall OODA Loop optimizes the use of redundancy to achieve the required reliability and availability Service Level Agreement (SLA) while minimizing the use of constrained resources such as spectrum and battery.¶
RAW In-situ OAM operation in the Network Plane may observe either a full recovery graph or the DetNet Path that is being used at this time. As packets may be load balanced, replicated, eliminated, and / or fragmented for Network Coding (NC) forward error correction (FEC), the RAW In-situ operation needs to be able to signal which operation occurred to an individual packet.¶
Active RAW OAM may be needed to observe the unused segments and evaluate the desirability of a rerouting decision.¶
Finally, the RAW Service sub-layer Assurance may observe the individual PREOF operation of a relay node to ensure that it is conforming; this might require injecting an OAM packet at an upstream point inside the recovery graph and extracting that packet at another point downstream before it reaches the egress.¶
This observation feeds the RAW PLR that makes the decision on which path is used at which RAW Node, for one a small continuous series of packets.¶
In the case of an End-to-End Protection in a Wireless Mesh, the recovery graph is strict and congruent with the path so all links are observed.¶
Conversely, in the case of Radio Access Protection illustrated in Figure 10, the recovery graph is Loose and only the first hop is observed; the rest of the path is abstracted and considered infinitely reliable. The loss of a packet is attributed to the first hop Radio Access Network (RAN), even if a particular loss effectively happens farther down the path. In that case, RAW enables technology diversity (e.g., Wi-Fi and 5G) which in turn improves the diversity in spectrum usage.¶
The Links that are not observed by OAM are opaque to it, meaning that the OAM information is carried across and possibly echoed as data, but there is no information capture in intermediate nodes. In the example above, the Tunnel underlay is opaque and not controlled by RAW; still the RAW OAM measures the end-to-end latency and delivery ratio for packets sent via RAN 1, RAN 2 and RAN 3, and determines whether a packet should be sent over either or a collection of those access links.¶
RAW separates the long time scale at which a recovery graph is elaborated and installed, from the short time scale at which the forwarding decision is taken for one or a few packets (see in Section 5.1) that experience the same path until the network conditions evolve and another path is selected within the same recovery graph.¶
The recovery graph computation is out of scope, but RAW expects that the CPF that installs the recovery graph also provides related knowledge in the form of meta data about the links, segments and possible DetNet Paths. That meta data can be a pre-digested statistical model, and may include prediction of future flaps and packet loss, as well as recommended actions when that happens.¶
The meta data may include:¶
The recovery graph is installed with measurable objectives that are computed by the rCPF to achieve the RAW SLA. The objectives can be expressed as any of maximum number of packet lost in a row, bounded latency, maximal jitter, maximum number of interleaved out of order packets, average number of copies received at the elimination point, and maximal delay between the first and the last received copy of the same packet.¶
The RAW OODA Loop operates at the path selection time scale to provide agility vs. the brute force approach of flooding the whole recovery graph. The OODA Loop controls, within the redundant solutions that are proposed by the local CPF, which is used for each packet to provide a Reliable and Available service while minimizing the waste of constrained resources.¶
To that effect, RAW defines the Point of Local Repair (PLR), which performs rapid local adjustments of the forwarding tables within the diversity that the lCPF has in store for the recovery graph. The PLR enables to exploit the richer forwarding capabilities at a faster time scale over a portion of the recovery graph, in either a loose or a strict fashion.¶
The PLR operates on metrics that evolve faster, but that need to be advertised at a fast rate but only locally, within the recovery graph, and reacts on the metrics updates by changing the DetNet path in use for the affected flows.¶
The rapid changes in the forwarding decisions are made and contained within the scope of a recovery graph and the actions of the PLR are not signaled outside the recovery graph. This is as opposed to the rCPF that must observe the whole network and optimize all the recovery graphs globally, which can only be done at a slow pace and using long-term statistical metrics, as presented in Table 1.¶
rCPF | PLR (In Scope) | |
---|---|---|
Operation | Typically Centralized | Source-Routed or Distributed |
Communication | Slow, expensive | Fast, local |
Time Scale (order) | routing computation + round trip, milliseconds to seconds | lookup + FIB installation, micro to milliseconds |
Network Size | Large, many recovery graphs to optimize globally | Small, within one recovery graph |
Considered Metrics | Averaged, Statistical, Shade of grey | Instant values / boolean condition |
The PLR sits in the DetNet Forwarding sub-Layer of Edge and Relay Nodes. The PLR it operates on the packet flow, learning the recovery graph and path selection information from the packet, possibly making local decision and retagging the packet to indicate so. On the other hand, the PLR interacts with the lower layers (through triggers and DLEP) and with its peers (through iOAM and oOAM) to obtain up-to-date information about its links and the quality of the overall recovery graph, respectively, as illustrated in Figure 11.¶
The main action by the PLR is the swapping of the DetNet Path within the recovery graph for the future packets. The candidate DetNet Paths represent different energy and spectrum profiles, and provide protection against different failures.¶
The RAW API enriches the DetNet protection services (PREOF) with potential possibility to interact with lower layer one-hop reliability functions that are more typical to wireless than wires, including Automatic Repeat reQuest (ARQ), Forward Error Correction (FEC), Hybrid ARQ (HARQ) that includes both, and other techniques such as overhearing and constructive interferences. Because RAW may be leveraged on wired links, e.g., to save power, it is not expected that all lower layers support all those capabilities.¶
RAW provides hints to the lower layer services on the desired outcome, and the lower layer acts on those hinHts to provide the best approximation of that outcome, e.g., a level of reliability for one-hop transmission within a bounded budget of time and/or energy. Thus, the RAW API makes possible cross-layer optimization for reliability depending on the actual abstraction provided. That is, some reliability functions are controlled from Layer-3 using an abstract interface, while they are really operated at the lower layers.¶
The RAW Path Selection can be implemented in both centralized and distributed approaches. In the centralized approach, the PLR may obtain a set of pre-computed DetNet paths matching a set of expected failures, and apply the appropriate DetNet paths for the current state of the wireless links. In the distributed approach, the signaling in the packet may be more abstract than an explicit Path, and the PLR decision might be revised along the select DetNet Path based on a better knowledge of the rest of the way.¶
The dynamic DetNet Path selection in RAW avoids the waste of critical resources such as spectrum and energy while providing for the guaranteed SLA, e.g., by rerouting and/or adding redundancy only when a spike of loss is observed.¶
RAW uses all forms of diversity including radio technology and physical path to increase the reliability and availability in the face of unpredictable conditions. While this is not done specifically to defeat an attacker, the amount of diversity used in RAW makes an attack harder to achieve.¶
Radio networks typically encrypt at the MAC layer to protect the transmission. If the encryption is per pair of peers, then certain RAW operations like promiscuous overhearing become impossible.¶
A RAW policy may typically select the cheapest collection of links that matches the requested SLA, e.g., use free Wi-Fi vs. paid 3GPP access. By defeating the cheap connectivity (e.g., PHY-layer interference) the attacker can force an End System to use the paid access and increase the cost of the transmission for the user.¶
This document has no IANA actions.¶
The editor wishes to thank the document co-authors:¶
for their contributions to the text and ideas exposed in this document.¶
This architecture could never have been completed without the support and recommendations from the DetNet Chairs Janos Farkas and Lou Berger, and Dave Black, the DetNet Tech Advisor. Many thanks to all of you.¶
The authors wish to thank Balazs Varga, Dave Cavalcanti, Don Fedyk, Nicolas Montavont, and Fabrice Theoleyre for their in-depth reviews during the development of this document.¶