DetNet N. Finn
Internet-Draft Huawei Technologies Co. Ltd
Intended status: Informational J-Y. Le Boudec
Expires: September 12, 2019 E. Mohammadpour
J. Zhang
Huawei Technologies Co. Ltd
B. Varga
J. Farkas
March 11, 2019

DetNet Bounded Latency


This document presents a parameterized timing model for Deterministic Networking (DetNet), so that existing and future standards can achieve the DetNet quality of service features of bounded latency and zero congestion loss. It defines requirements for resource reservation protocols or servers. It calls out queuing mechanisms, defined in other documents, that can provide the DetNet quality of service.

Status of This Memo

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

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 September 12, 2019.

Copyright Notice

Copyright (c) 2019 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 ( in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.

Table of Contents

1. Introduction

The ability for IETF Deterministic Networking (DetNet) or IEEE 802.1 Time-Sensitive Networking (TSN, [IEEE8021TSN]) to provide the DetNet services of bounded latency and zero congestion loss depends upon A) configuring and allocating network resources for the exclusive use of DetNet/TSN flows; B) identifying, in the data plane, the resources to be utilized by any given packet, and C) the detailed behavior of those resources, especially transmission queue selection, so that latency bounds can be reliably assured. Thus, DetNet is an example of an IntServ Guaranteed Quality of Service [RFC2212]

As explained in [I-D.ietf-detnet-architecture], DetNet flows are characterized by 1) a maximum bandwidth, guaranteed either by the transmitter or by strict input metering; and 2) a requirement for a guaranteed worst-case end-to-end latency. That latency guarantee, in turn, provides the opportunity for the network to supply enough buffer space to guarantee zero congestion loss.

To be of use to the applications identified in [I-D.ietf-detnet-use-cases], it must be possible to calculate, before the transmission of a DetNet flow commences, both the worst-case end-to-end network latency, and the amount of buffer space required at each hop to ensure against congestion loss.

This document references specific queuing mechanisms, defined in other documents, that can be used to control packet transmission at each output port and achieve the DetNet qualities of service. This document presents a timing model for sources, destinations, and the DetNet transit nodes that relay packets that is applicable to all of those referenced queuing mechanisms. The parameters specified in this model:

Using the model presented in this document, it should be possible for an implementor, user, or standards development organization to select a particular set of queuing mechanisms for each device in a DetNet network, and to select a resource reservation algorithm for that network, so that those elements can work together to provide the DetNet service.

This document does not specify any resource reservation protocol or server. It does not describe all of the requirements for that protocol or server. It does describe requirements for such resource reservation methods, and for queuing mechanisms that, if met, will enable them to work together.

2. Terminology and Definitions

This document uses the terms defined in [I-D.ietf-detnet-architecture].

3. DetNet bounded latency model

3.1. Flow creation

There are two models for flow creation, static (Section 3.1.1) and dynamic (Section 3.1.2). Most of the mathematical analysis provided in this document is applicable to either flow creation model; any dependencies on the choice of flow creation model are pointed out in the text.

3.1.1. Static flow creation

The static problem:

Given a network and a set of DetNet flows, compute an end-to-end latency bound (if computable) for each flow, and compute the resources, particularly buffer space, required in each DetNet transit node to achieve zero congestion loss.

In this model, all of the DetNet flows are known before the calculation commences. This problem is of interest to relatively static networks, or static parts of larger networks. It gives the best possible worst-case behavior. The calculations can be extended to provide global optimizations, such as altering the path of one DetNet flow in order to make resources available to another DetNet flow with tighter constraints.

This calculation may be more difficult to perform than that of the dynamic model (Section 3.1.2), because the flows passing through one port on a DetNet transit node affect each others' latency. The effects can even be circular, from Flow A to B to C and back to A. On the other hand, the static calculation can often accommodate queuing methods, such as transmission selection by strict priority, that are unsuitable for the dynamic calculation.

The static flow creation model is not limited only to static networks; the entire calculation for all flows can be repeated each time a new DetNet flow is created or deleted. If some already-established flow would be pushed beyond its latency requirements by the new flow, then either the new flow is refused, or some other suitable action taken.

3.1.2. Dynamic flow creation

The dynamic problem:

Given a network whose maximum capacity for DetNet flows is bounded by a set of static configuration parameters applied to the DetNet transit nodes, and given just one DetNet flow, compute the worst-case end-to-end latency that can be experienced by that flow, no matter what other DetNet flows (within the network's configured parameters) might be created or deleted in the future. Also, compute the resources, particularly buffer space, required in each DetNet transit node to achieve zero congestion loss.

This model is dynamic, in the sense that flows can be added or deleted at any time, with a minimum of computation effort, and without affecting the guarantees already given to other flows.

The choice of queuing methods is critical to the applicability of the dynamic model. Some queuing methods (e.g. CQF, Section 7.1) make it easy to configure bounds on the network's capacity, and to make independent calculations for each flow. Other queuing methods (e.g., transmission selection by strict priority), make this calculation impossible, because the worst case for one flow cannot be computed without complete knowledge of all other flows. Other queuing methods (e.g. the credit-based shaper defined in [IEEE8021Q] section can be used for dynamic flow creation, but yield poorer latency and buffer space guarantees than when that same queuing method is used for static flow creation (Section 3.1.1).

The dynamic flow creation model assumes the use of the following paradigm for provisioning DetNet flows:

  1. Perform any configuration required by the DetNet transit nodes in the network for the classes of service to be offered, including one or more classes of DetNet service. This configuration is done beforehand, and not tied to any particular flow.
  2. Characterize the new DetNet flow in IntServ terms (Section 8).
  3. Establish the path that the DetNet flow will take through the network from the source to the destination(s). This can be a point-to-point or a point-to-multipoint path.
  4. Select one of the DetNet classes of service for the DetNet flow.
  5. Compute the worst-case end-to-end latency for the DetNet flow. In the process, determine whether sufficient resources are available for that flow to guarantee the required latency and to provide zero congestion loss.
  6. Assuming that the resources are available, commit those resources to the flow. This may or may not require adjusting the parameters that control the queuing mechanisms at each hop along the flow's path.

This paradigm can be static and/or dynamic, and can be implemented using peer-to-peer protocols or using a central server model. In some situations, backtracking and recursing through this list may be necessary.

Issues such as un-provisioning a DetNet flow in favor of another when resources are scarce are not considered, but are left to the static flow creation model (Section 3.1.1). How the path to be taken by a DetNet flow is chosen is not considered in this document.

3.2. Relay node model

A model for the operation of a DetNet transit node is required, in order to define the latency and buffer calculations. In Figure 1 we see a breakdown of the per-hop latency experienced by a packet passing through a DetNet transit node, in terms that are suitable for computing both hop-by-hop latency and per-hop buffer requirements.

      DetNet transit node A            DetNet transit node B
   +-------------------------+       +------------------------+
   |              Queuing    |       |              Queuing   |
   |   Regulator subsystem   |       |   Regulator subsystem  |
   |   +-+-+-+-+ +-+-+-+-+   |       |   +-+-+-+-+ +-+-+-+-+  |
-->+   | | | | | | | | | +   +------>+   | | | | | | | | | +  +--->
   |   +-+-+-+-+ +-+-+-+-+   |       |   +-+-+-+-+ +-+-+-+-+  |
   |                         |       |                        |
   +-------------------------+       +------------------------+
2,3  4      5        6      1    2,3   4      5        6     1   2,3
                1: Output delay       4: Processing delay
                2: Link delay         5: Regulation delay
                3: Preemption delay   6: Queuing delay.

Figure 1: Timing model for DetNet or TSN

In Figure 1, we see two DetNet transit nodes (typically, bridges or routers), with a wired link between them. In this model, the only queues we deal with explicitly are attached to the output port; other queues are modeled as variations in the other delay times. (E.g., an input queue could be modeled as either a variation in the link delay [2] or the processing delay [4].) There are six delays that a packet can experience from hop to hop.

1. Output delay

The time taken from the selection of a packet for output from a queue to the transmission of the first bit of the packet on the physical link. If the queue is directly attached to the physical port, output delay can be a constant. But, in many implementations, the queuing mechanism in a forwarding ASIC is separated from a multi-port MAC/PHY, in a second ASIC, by a multiplexed connection. This causes variations in the output delay that are hard for the forwarding node to predict or control.
2. Link delay

The time taken from the transmission of the first bit of the packet to the reception of the last bit, assuming that the transmission is not suspended by a preemption event. This delay has two components, the first-bit-out to first-bit-in delay and the first-bit-in to last-bit-in delay that varies with packet size. The former is typically measured by the Precision Time Protocol and is constant (see [I-D.ietf-detnet-architecture]). However, a virtual "link" could exhibit a variable link delay.
3. Preemption delay

If the packet is interrupted in order to transmit another packet or packets, (e.g. [IEEE8023] clause 99 frame preemption) an arbitrary delay can result.
4. Processing delay

This delay covers the time from the reception of the last bit of the packet to the time the packet is enqueued in the regulator (Queuing subsystem, if there is no regulation). This delay can be variable, and depends on the details of the operation of the forwarding node.
5. Regulator delay

This is the time spent from the insertion of the last bit of a packet into a regulation queue until the time the packet is declared eligible according to its regulation constraints. We assume that this time can be calculated based on the details of regulation policy. If there is no regulation, this time is zero.
6. Queuing subsystem delay

This is the time spent for a packet from being declared eligible until being selected for output on the next link. We assume that this time is calculable based on the details of the queuing mechanism. If there is no regulation, this time is from the insertion of the packet into a queue until it is selected for output on the next link.

Not shown in Figure 1 are the other output queues that we presume are also attached to that same output port as the queue shown, and against which this shown queue competes for transmission opportunities.

The initial and final measurement point in this analysis (that is, the definition of a "hop") is the point at which a packet is selected for output. In general, any queue selection method that is suitable for use in a DetNet network includes a detailed specification as to exactly when packets are selected for transmission. Any variations in any of the delay times 1-4 result in a need for additional buffers in the queue. If all delays 1-4 are constant, then any variation in the time at which packets are inserted into a queue depends entirely on the timing of packet selection in the previous node. If the delays 1-4 are not constant, then additional buffers are required in the queue to absorb these variations. Thus:

4. Computing End-to-end Latency Bounds

4.1. Non-queuing delay bound

End-to-end latency bounds can be computed using the delay model in Section 3.2. Here it is important to be aware that for several queuing mechanisms, the worst-case end-to-end delay is less than the sum of the per-hop worst-case delays. An end-to-end latency bound for one DetNet flow can be computed as

end_to_end_latency_bound = non_queuing_latency + queuing_latency

The two terms in the above formula are computed as follows. First, at the h-th hop along the path of this DetNet flow, obtain an upper bound per-hop_non_queuing_latency[h] on the sum of delays 1,2,3,4 of Figure 1. These upper-bounds are expected to depend on the specific technology of the DetNet transit node at the h-th hop but not on the T-SPEC of this DetNet flow. Then set non_queuing_latency = the sum of per-hop_non_queuing_latency[h] over all hops h.

4.2. Queuing delay bound

Second, compute queuing_latency as an upper bound to the sum of the queuing delays along the path. The value of queuing_latency depends on the T-SPEC of this flow and possibly of other flows in the network, as well as the specifics of the queuing mechanisms deployed along the path of this flow.

For several queuing mechanisms, queuing_latency is less than the sum of upper bounds on the queuing delays (5,6) at every hop. This occurs with (1) per-flow queuing, and (2) per-class queuing with regulators, as explained in Section 4.2.1, Section 4.2.2, and Section 6.

For other queuing mechanisms the only available value of queuing_latency is the sum of the per-hop queuing delay bounds. In such cases, the computation of per-hop queuing delay bounds must account for the fact that the T-SPEC of a DetNet flow is no longer satisfied at the ingress of a hop, since burstiness increases as one flow traverses one DetNet transit node.

4.2.1. Per-flow queuing mechanisms

With such mechanisms, each flow uses a separate queue inside every node. The service for each queue is abstracted with a guaranteed rate and a delay. For every flow the per-node delay bound as well as end-to-end delay bound can be computed from the traffic specification of this flow at its source and from the values of rates and latencies at all nodes along its path. Details of calculation for IntServ are described in Section 6.5.

4.2.2. Per-class queuing mechanisms

With such mechanisms, the flows that have the same class share the same queue. A practical example is the queuing mechanism in Time Sensitive Networking. One key issue in this context is how to deal with the burstiness cascade: individual flows that share a resource dedicated to a class may see their burstiness increase, which may in turn cause increased burstiness to other flows downstream of this resource. Computing latency upper bounds for such cases is difficult, and in some conditions impossible [charny2000delay][bennett2002delay]. Also, when bounds are obtained, they depend on the complete configuration, and must be recomputed when one flow is added.

A solution to deal with this issue is to reshape the flows at every hop. This can be done with per-flow regulators (e.g. leaky bucket shapers), but this requires per-flow queuing and defeats the purpose of per-class queuing. An alternative is the interleaved regulator, which reshapes individual flows without per-flow queuing ([Specht2016UBS], [IEEE8021Qcr]). With an interleaved regulator, the packet at the head of the queue is regulated based on its (flow) regulation constraints; it is released at the earliest time at which this is possible without violating the constraint. One key feature of per-flow or interleaved regulator is that, it does not increase worst-case latency bounds [le_boudec_theory_2018]. Specifically, when an interleaved regulator is appended to a FIFO subsystem, it does not increase the worst-case delay of the latter.

Figure 2 shows an example of a network with 5 nodes, per-class queuing mechanism and interleaved regulators as in Figure 1. An end-to-end delay bound for flow f, traversing nodes 1 to 5, is calculated as follows:

end_to_end_latency_bound_of_flow_f = C12 + C23 + C34 + S4

In the above formula, Cij is a bound on the aggregate response time of queuing subsystem in node i and interleaved regulator of node j, and S4 is a bound on the response time of the queuing subsystem in node 4 for flow f. In fact, using the delay definitions in Section 3.2, Cij is a bound on sum of the delays 1,2,3,6 of node i and 4,5 of node j. Similarly, S4 is a bound on sum of the delays 1,2,3,6 of node 4. A practical example of queuing model and delay calculation is presented Section 6.4.

+---+   +---+   +---+   +---+   +---+
| 1 |---| 2 |---| 3 |---| 4 |---| 5 |
+---+   +---+   +---+   +---+   +---+

Figure 2: End-to-end latency computation example

REMARK: The end-to-end delay bound calculation provided here gives a much better upper bound in comparison with end-to-end delay bound computation by adding the delay bounds of each node in the path of a flow [TSNwithATS].

4.3. Ingress considerations

A sender can be a DetNet node which uses exactly the same queuing methods as its adjacent DetNet transit node, so that the latency and buffer calculations at the first hop are indistinguishable from those at a later hop within the DetNet domain. On the other hand, the sender may be DetNet unaware, in which case some conditioning of the flow may be necessary at the ingress DetNet transit node.

This ingress conditioning typically consists of a FIFO with an output regulator that is compatible with the queuing employed by the DetNet transit node on its output port(s). For some queuing methods, simply requires added extra buffer space in the queuing subsystem. Ingress conditioning requirements for different queuing methods are mentioned in the sections, below, describing those queuing methods.

4.4. Interspersed non-DetNet transit nodes

It is sometimes desirable to build a network that has both DetNet aware transit nodes and DetNet non-aware transit nodes, and for a DetNet flow to traverse an island of non-DetNet transit nodes, while still allowing the network to offer latency and congestion loss guarantees. This is possible under certain conditions.

In general, when passing through a non-DetNet island, the island causes delay variation in excess of what would be caused by DetNet nodes. That is, the DetNet flow is "lumpier" after traversing the non-DetNet island. DetNet guarantees for latency and buffer requirements can still be calculated and met if and only if the following are true:

  1. The latency variation across the non-DetNet island must be bounded and calculable.
  2. An ingress conditioning function (Section 4.3) may be required at the re-entry to the DetNet-aware domain. This will, at least, require some extra buffering to accommodate the additional delay variation, and thus further increases the worst-case latency.

The ingress conditioning is exactly the same problem as that of a sender at the edge of the DetNet domain. The requirement for bounds on the latency variation across the non-DetNet island is typically the most difficult to achieve. Without such a bound, it is obvious that DetNet cannot deliver its guarantees, so a non-DetNet island that cannot offer bounded latency variation cannot be used to carry a DetNet flow.

5. Achieving zero congestion loss

When the input rate to an output queue exceeds the output rate for a sufficient length of time, the queue must overflow. This is congestion loss, and this is what deterministic networking seeks to avoid.

5.1. A General Formula

To avoid congestion losses, an upper bound on the backlog present in the regulator and queuing subsystem of Figure 1 must be computed during resource reservation. This bound depends on the set of flows that use these queues, the details of the specific queuing mechanism and an upper bound on the processing delay (4). The queue must contain the packet in transmission plus all other packets that are waiting to be selected for output.

A conservative backlog bound, that applies to all systems, can be derived as follows.

The backlog bound is counted in data units (bytes, or words of multiple bytes) that are relevant for buffer allocation. For every class we need one buffer space for the packet in transmission, plus space for the packets that are waiting to be selected for output. Excluding transmission and preemption times, the packets are waiting in the queue since reception of the last bit, for a duration equal to the processing delay (4) plus the queuing delays (5,6).


Then a bound on the backlog of traffic of all classes in the queue at this output port is

backlog_bound = ( nb_classes + nb_input_ports ) * max_packet_length + total_in_rate* max_delay45

6. Queuing model

6.1. Queuing data model

Sophisticated queuing mechanisms are available in Layer 3 (L3, see, e.g., [RFC7806] for an overview). In general, we assume that "Layer 3" queues, shapers, meters, etc., are precisely the "regulators" shown in Figure 1. The "queuing subsystems" in this figure are not the province solely of bridges; they are an essential part of any DetNet transit node. As illustrated by numerous implementation examples, some of the "Layer 3" mechanisms described in documents such as [RFC7806] are often integrated, in an implementation, with the "Layer 2" mechanisms also implemented in the same node. An integrated model is needed in order to successfully predict the interactions among the different queuing mechanisms needed in a network carrying both DetNet flows and non-DetNet flows.

Figure 3 shows the general model for the flow of packets through the queues of a DetNet transit node. Packets are assigned to a class of service. The classes of service are mapped to some number of regulator queues. Only DetNet/TSN packets pass through regulators. Queues compete for the selection of packets to be passed to queues in the queuing subsystem. Packets again are selected for output from the queuing subsystem.

|                    Class of Service Assignment                    |
   |      |          |         |           |     |       |       |
+--V-+ +--V-+     +--V--+   +--V--+     +--V--+  |       |       |
|Flow| |Flow|     |Flow |   |Flow |     |Flow |  |       |       |
|  0 | |  1 | ... |  i  |   | i+1 | ... |  n  |  |       |       |
| reg| | reg|     | reg |   | reg |     | reg |  |       |       |
+--+-+ +--+-+     +--+--+   +--+--+     +--+--+  |       |       |
   |      |          |         |           |     |       |       |
+--V------V----------V--+   +--V-----------V--+  |       |       |
|  Trans.  selection    |   | Trans. select.  |  |       |       |
+----------+------------+   +-----+-----------+  |       |       |
           |                      |              |       |       |
        +--V--+                +--V--+        +--V--+ +--V--+ +--V--+
        | out |                | out |        | out | | out | | out |
        |queue|                |queue|        |queue| |queue| |queue|
        |  1  |                |  2  |        |  3  | |  4  | |  5  |
        +--+--+                +--+--+        +--+--+ +--+--+ +--+--+
           |                      |              |       |       |
|                      Transmission selection                       |
           |                      |              |       |       |
           V                      V              V       V       V
     DetNet/TSN queue       DetNet/TSN queue    non-DetNet/TSN queues

Figure 3: IEEE 802.1Q Queuing Model: Data flow

Some relevant mechanisms are hidden in this figure, and are performed in the queue boxes:

Ideally, neither of these actions are performed on DetNet packets. Full queues for DetNet packets should occur only when a flow is misbehaving, and the DetNet QoS does not include "yellow" service for packets in excess of committed rate.

The Class of Service Assignment function can be quite complex, even in a bridge [IEEE8021Q], since the introduction of per-stream filtering and policing ([IEEE8021Q] clause In addition to the Layer 2 priority expressed in the 802.1Q VLAN tag, a DetNet transit node can utilize any of the following information to assign a packet to a particular class of service (queue):

The "Transmission selection" function decides which queue is to transfer its oldest packet to the output port when a transmission opportunity arises.

6.2. Preemption

In [IEEE8021Q] and [IEEE8023], the transmission of a frame can be interrupted by one or more "express" frames, and then the interrupted frame can continue transmission. This frame preemption is modeled as consisting of two MAC/PHY stacks, one for packets that can be interrupted, and one for packets that can interrupt the interruptible packets. The Class of Service (queue) determines which packets are which. Only one layer of preemption is supported -- a transmitter cannot have more than one interrupted frame in progress. DetNet flows typically pass through the interrupting MAC. Best-effort queues pass through the interruptible MAC, and can thus be preempted.

6.3. Time-scheduled queuing

In [IEEE8021Q], the notion of time-scheduling queue gates is described in section Below every output queue (the lower row of queues in Figure 3) is a gate that permits or denies the queue to present data for transmission selection. The gates are controlled by a rotating schedule that can be locked to a clock that is synchronized with other DetNet transit nodes. The DetNet class of service can be supplied by queuing mechanisms based on time, rather than the regulator model in Figure 3. These queuing mechanisms are discussed in Section 7, below.

6.4. Time-Sensitive Networking with Asynchronous Traffic Shaping

Consider a network with a set of nodes (DetNet transit nodes and hosts) along with a set of flows between hosts. Hosts are sources or destinations of flows. There are four types of flows, namely, control-data traffic (CDT), class A, class B, and best effort (BE) in decreasing order of priority. Flows of classes A and B are together referred to AVB flows. It is assumed a subset of TSN functions as described next.

It is also assumed that contention occurs only at the output port of a TSN node. Each node output port performs per-class scheduling with eight classes: one for CDT, one for class A traffic, one for class B traffic, and five for BE traffic denoted as BE0-BE4 (according to TSN standard). In addition, each node output port also performs per-flow regulation for AVB flows using an interleaved regulator (IR), called Asynchronous Traffic Shaper (ATS) in TSN. Thus, at each output port of a node, there is one interleaved regulator per-input port and per-class. The detailed picture of scheduling and regulation architecture at a node output port is given by Figure 4. The packets received at a node input port for a given class are enqueued in the respective interleaved regulator at the output port. Then, the packets from all the flows, including CDT and BE flows, are enqueued in a class based FIFO system (CBFS) [TSNwithATS].

      +--+   +--+ +--+   +--+
      |  |   |  | |  |   |  |
      |IR|   |IR| |IR|   |IR|
      |  |   |  | |  |   |  |
      +-++XXX++-+ +-++XXX++-+
        |     |     |     |
        |     |     |     |
+---+ +-v-XXX-v-+ +-v-XXX-v-+ +-----+ +-----+ +-----+ +-----+ +-----+
|   | |         | |         | |Class| |Class| |Class| |Class| |Class|
|CDT| | Class A | | Class B | | BE4 | | BE3 | | BE2 | | BE1 | | BE0 |
|   | |         | |         | |     | |     | |     | |     | |     |
+-+-+ +----+----+ +----+----+ +--+--+ +--+--+ +--+--+ +--+--+ +--+--+
  |        |           |         |       |       |       |       |
  |      +-v-+       +-v-+       |       |       |       |       |
  |      |CBS|       |CBS|       |       |       |       |       |
  |      +-+-+       +-+-+       |       |       |       |       |
  |        |           |         |       |       |       |       |
|                     Strict Priority selection                     |

Figure 4: Architecture of a TSN node output port with interleaved regulators (IRs)

The CBFS includes two Credit-Based Shaper (CBS) subsystems, one for each class A and B. The CBS serves a packet from a class according to the available credit for that class. The credit for each class A or B increases based on the idle slope, and decreases based on the send slope, both of which are parameters of the CBS. The CDT and BE0-BE4 flows in the CBFS are served by separate FIFO subsystems. Then, packets from all flows are served by a transmission selection subsystem that serves packets from each class based on its priority. All subsystems are non-preemptive. Guarantees for AVB traffic can be provided only if CDT traffic is bounded; it is assumed that the CDT traffic has leaky bucket arrival curve with two parameters r_h as rate and b_h as bucket size, i.e., the amount of bits entering a node within a time interval t is bounded by r_h t + b_h.

Additionally, it is assumed that the AVB flows are also regulated at their source according to leaky bucket arrival curve. At the source hosts, the traffic satisfies its regulation constraint, i.e. the delay due to interleaved regulator at hosts is ignored.

At each DetNet transit node implementing an interleaved regulator, packets of multiple flows are processed in one FIFO queue; the packet at the head of the queue is regulated based on its leaky bucket parameters; it is released at the earliest time at which this is possible without violating the constraint. The regulation parameters for a flow (leaky bucket rate and bucket size) are the same at its source and at all DetNet transit nodes along its path. A delay bound of CBFS for an AVB flow f of class A or B can be computed if the following condition holds:

sum of leaky bucket rates of all flows of this class at this node <= R, where R is given below for every class.

If the condition holds, the delay bound is:

d_f = T + (b_t-L_min_f)/R - L_min_f/c

where L_min_f is the minimum packet length of flow f; c is the output link transmission rate; b_t is the sum of the b term (bucket size) for all the flows having the same class as flow f at this node. Parameters R and T are calculated as follows for class A and class B, separately:

If f is of class A:

R = I_A (c-r_h)/ c
T = L_nA + b_h + r_h L_n/c)/(c-r_h)

where L_nA is the maximum packet length of class B and BE packets; L_n is the maximum packet length of classes A,B, and BE.

If f is of class B:

R = I_B (c-r_h)/ c
T = (L_BE + L_A + L_nA I_A/(c_h-I_A) + b_h + r_h L_n/c)/(c-r_h)

where L_A is the maximum packet length of class A; L_BE is the maximum packet length of class BE.

Then, an end-to-end delay bound is calculated by the formula Section 4.2.2, where for Cij:

Cij = max(d_f')

where f' is any flow that shares the same CBFS class with flow f at node i and the same interleaved regulator as flow f at node j.

More information of delay analysis in such a DetNet transit node is described in [TSNwithATS].

6.4.1. Flow Admission

The delay calculation requires some information about each node. For each node, it is required to know the idle slope of CBS for each class A and B (I_A and I_B), as well as the transmission rate of the output link (c). Besides, it is necessary to have the information on each class, i.e. maximum packet length of classes A, B, and BE. Moreover, the leaky bucket parameters of CDT (r_h,b_h) should be known. To admit a flow/flows, their delay requirements should be guaranteed not to be violated. As described in Section 3.1, the two problems static and dynamic are addressed separately. In either of the problems, the rate and delay should be guaranteed. Thus,

The static admission control:

The leaky bucket parameters of all flows are known, therefore, for each flow a delay bound can be calculated. The computed delay bound for every flow should not be more than its delay requirement. Moreover, the sum of the rate of each flow (r_f) should not be more than the rate allocated to each class (R). If these two conditions hold, the configuration is declared admissible.
The dynamic admission control:

For dynamic admission control, we allocate to every node and class A or B, static value for rate (R) and maximum burstiness (b_t). In addition, for every node and every class A and B, two counters are maintained:
R_acc is equal to the sum of the leaky-bucket rates of all flows of this class already admitted at this node; At all times, we must have:
R_acc <=R, (Eq. 1)
b_acc is equal to the sum of the bucket sizes of all flows of this class already admitted at this node; At all times, we must have:
b_acc <=b_t. (Eq. 2)

A new flow is admitted at this node, if Eqs. (1) and (2) continue to be satisfied after adding its leaky bucket rate and bucket size to R_acc and b_acc. A flow is admitted in the network, if it is admitted at all nodes along its path. When this happens, all variables R_acc and b_acc along its path must be incremented to reflect the addition of the flow. Similarly, when a flow leaves the network, all variables R_acc and b_acc along its path must be decremented to reflect the removal of the flow.

The choice of the static values of R and b_t at all nodes and classes must be done in a prior configuration phase; R controls the bandwidth allocated to this class at this node, b_t affects the delay bound and the buffer requirement. R must satisfy the constraints given in Annex L.1 of [IEEE8021Q].

6.5. IntServ

Integrated service (IntServ) is an architecture that specifies the elements to guarantee quality of service (QoS) on networks. To satisfied guaranteed service, a flow must conform to a traffic specification (T-spec), and reservation is made along a path, only if routers are able to guarantee the required bandwidth and buffer.

Consider the traffic model which conforms to token bucket regulator (r, b), with

The traffic specification can be described as an arrival curve:

alpha(t) = b + rt

This token bucket regulator requires that, during any time window t, the number of bit for the flow is limited by alpha(t) = b + rt.

If resource reservation on a path is applied, IntServ model of a router can be described as a rate-latency service curve beta(t).

beta(t) = max(0, R(t-T))

It describes that bits might have to wait up to T before being served with a rate greater or equal to R.

It should be noted that, the guaranteed service rate R is a share of link's bandwidth. The choice of R is related to the specification of flows which will transmit on this node. For example, in strict priority policy, considering a flow with priority j, its share of bandwidth may be R=c-sum(r_i), i<j, where c is the link bandwidth, r_i is the token bucket rate for the flows with priority higher than j. The choice of T is also related to the specification of all the flows traversing this node. For example, in a generalized processor sharing (GPS) node, T = L / R + L_max/c, where L is the maximum packet size for the flow, L_max is the maximum packet size in the node across all flows. Other choice of R and T are also supported, according to the specific scheduling of the node and flows traversing this node.

As mentioned previously in this section, delay bound and backlog bound can be easily obtained by comparing arrival curve and service curve. Backlog bound, or buffer bound, is the maximum vertical derivation between curves alpha(t) and beta(t), which is v=b+rT. Delay bound is the maximum horizontal derivation between curves alpha(t) and beta(t), which is h = T+b/R. Graphical illustration of the IntServ model is shown in Figure 5.

    + bit              .        *
    |                 .     *
    |                .  *
    |               *
    |           *  .
    |       *     .
    |   *   |    .        ..  Service curve
    *-----h-|---.         **  Arrival curve
    |       v  .           h  Delay_bound
    |       | .            v  Backlog_bound
    |       |.
    +-------.--------------------+ time

Figure 5: Computation of backlog bound and delay bound. Note that arrival and service curves are not necessary to be linear.

The output bound, or the next-hop arrival curve, is alpha_out(t) = b + rT + rt, where burstiness of the flow is increased by rT, compared with the arrival curve.

We can calculate the end-to-end delay bound for a path including N nodes, among which the i-th node offers service curve beta_i(t),

beta_i(t) = max(0, R_i(t-T_i)), i=1,...,N

By concatenating these IntServ nodes, an end-to-end service curve can be computed as

beta_e2e (t) = max(0, R_e2e(t-T_e2e) )


R_e2e = min(R_1,..., R_N)
T_e2e = T_1 + ... + T_N

Similarly, delay bound, backlog bound and output bound can be computed by using the original arrival curve alpha(t) and concatenated service curve beta_e2e(t).

7. Time-based DetNet QoS

7.1. Cyclic Queuing and Forwarding

Annex T of [IEEE8021Q] describes Cyclic Queuing and Forwarding (CQF), which provides bounded latency and zero congestion loss using the time-scheduled gates of [IEEE8021Q] section For a given DetNet class of service, a set of two or three buffers is provided at the output queue layer of Figure 3. A cycle time Tc is configured for each class c, and all of the buffer sets in a class swap buffers simultaneously throughout the DetNet domain at that cycle rate, all in phase.

0 time -->  0.7     1   (units of Tc)   2                   3
                        DetNet transit node A out port 1
|      a      <-DT->|        b          |          c        |       d
 \_____              \_____
       \_____              \_____  queue-to-queue delay = 1.3 Tc
             \_____              \_____
                   \_____              \_____  DetNet transit node B
                         \_                  \_ queue assignment, in
       |                   |            |<-DT->|  port 2 to out 3  |
      0.3  time-->        1.3          2.0    2.3                 3.3
      window to transfer
         to buffer c  --->  VVVVVVVVVVVV
       if dead time not                         window to transfer
          excessive         VVVVVVVVVVVVVVVVVVV <--- to buffer d
                        DetNet transit node B out port 3
|         a         |         b         |         c         |       d
0    time-->        1                   2                   3

Figure 6: CQF timing diagram and dead time computation

Figure 6 shows two DetNet transit nodes A and B, including three timelines for:

  1. The output queues on port 1 in node A.
  2. The input gate function ([IEEE8021Q], that assigns packets received on port 1 of transit node B to output queues on port 2 of transit node B.
  3. The output queues on port 2 of node B.

In this figure, the output ports on the two nodes are synchronized, and a new buffer starts transmitting at each tick, shown as 0, 1, 2, ... The output times shown for timelines 1 and 3 are the times at which packets are selected for output, which is the start point of the output time (1) of Figure 1. The queue assignments times on timeline 3 take place at the beginning of the queuing delay (6) of Figure 1. Time-based CQF, as described here, does not require any regulator queues. In the shown in the figure, the total time for delays 1 through 6 of Figure 1 is 1.3Tc. Of course, any value is possible.

7.1.1. CQF timing sequence

In general, as shown in Figure 6, the windows for buffer assignment do not align perfectly with the windows for buffer transmission. The input gates (the center timeline in Figure 6) must switch from using one buffer to using another buffer in sync with the (delayed) received data, at times offset by the dead time from the output buffer switching (the bottom timeline in Figure 6).

If the dead time DT in Figure 6 is not excessive, then it is feasible to subtract the dead time from the cycle time Tc, and use the remainder as the input window. In the example in Figure 6, packets from node A buffer a can be transferred from the input port to node B's buffer "c" during the window shown by the upper row "VVVV...". Input must cease by time = 2.0, because that is when transit node B starts transmitting the contents of buffer c. In this case, only two output buffers are in use, one filling and one outputting.

If the dead time is too large (e.g., if the delays placed the middle timeline's switching points at n+0.9, instead of n+0.3), three buffers are used by node B. This case is shown by the lower row "VVVV..." in Figure 6. In this case, node B places the data received from node A buffer a into node B buffer d between the times 1.3 and 2.3 in Figure 6. Buffer b starts outputting at time = 2.0, while buffer d is filling. Thus, three buffers are in use, one filling, one waiting, and one emptying.

7.1.2. Dead time computation

The time for switching input packet buffer assignments is equal to the minimum possible offset from transmission selection in node A to buffer assignment in node B, which is the sum of the minimum values for all of the delays 1 through 5 in Figure 1 (the queue-to-queue delay). All packets must be received and assigned to an output buffer before the next switching point, which means that all must be transmitted in time for them to arrive at buffer assignment even if worst-case (longest delay) is encountered for the queue-to-queue delay. Thus, the minimum dead time for the 3-buffer case is the sum of the worst-case variation in the queue-to-queue delay, plus the worst-case difference between the two transit nodes' buffer switching clocks.

For the 2-buffer case, we must add the offset (shown as "DT" in Figure 6) from the end of node B's output switch to the end of node B's input switch.

7.1.3. Tc computation

Given the dead time DT, there remains a transmit window of (Tc - DT - Int). The DT was explained in (Section 7.1.2). "Int" is the worst-case interference with the start of transmission, when the output buffers switch, caused by lower-priority traffic. This is equal to one worst-case transmission time, which means that the size of the packets in all lower-priority queues must be bounded. If Ethernet preemption ([IEEE8023] clause 99) is employed for lower-priority queues, then this worst-case interference is reduced to the size of the largest unfragmentable Ethernet frame.

The bandwidth requirement of any given DetNet flow has to be translated to CQF terms, in order to determine whether that flow can be accommodated at each port. A flow has to be characterized as using a maximum number of bit times on the wire per cycle time Tc. For Ethernet, for example ([IEEE8023]), this includes the preamble (8 bytes), destination MAC address through CRC (minimum 64 bytes) and the inter-packet gap (12 bytes). The total bit times per cycle Tc required by all of the DetNet flows passing through a given port cannot exceed the available transmit window (Tc - DT - Int).

7.1.4. CQF latency calculation

The per-hop latency is trivially determined by the wire delay plus the queuing delay. Since the wire delay is either absorbed into the queueing delay (dead time is small and two buffers are used) or padded out to a whole cycle time Tc (three buffers are used) the per-hop latency is always an integral number of cycle times Tc, with a latency variation at the output of the final hop of Tc.

Ingress conditioning may be required if the source of a DetNet flow does not, itself, employ CQF. See Section 7.1.6.

7.1.5. CQF parameterization

The transmit window for a given DetNet transit node running CQF, for example the transmit window for node A in Figure 6, depends on the interference (Int in section Section 7.1.3), the value of Tc, and the dead time required by the following node B. The size of the transmit window determines how many total bits can be reserved per period Tc by DetNet flows.

Part of the dead time derives from delays and delay variations such as output delay (1) and link delay (2) and preemption delay (3) of Figure 1, all of which are known to node A. However, the dead time also depends on the processing delay (4) of node B and the upon whether node B is using 2 or 3 output buffers, which is not necessarily known to node A.

The information in DetNet transit node B necessary to compute the dead time to be observed by transit node A must be known to the entity responsible for making reservation decisions, whether that is node A itself, or a central controller. A decision can be made, by the controller or by the node, whether to use the dead time and two buffers, in order to reduce the per-hop latency by one cycle time, or to use three buffers and eliminate the dead time and increase the total allocable bandwidth.

If the packet sizes of a DetNet flow are variable, or perhaps even unknown beyond the imposition of a maximum size, then some degree of overprovisioning is required. The measurement used to allocate bandwidth to a given DetNet flow is bit times in one cycle time Tc. Therefore, one extra maximum packet time (less one bit) has to be allocated to a flow per cycle time Tc in order to ensure that, no matter what mix of packet sizes are presented, the flow will get its guaranteed latency.

7.1.6. Ingress conditioning for CQF

Assuming that a DetNet domain is using CQF, it is always possible that the previous node (or sender) may not support the queuing method of CQF, or may support CQF but not use the same configuration in accordance of the current DetNet domain. In this case, ingress conditioning is helpful to shape the flow according to the TSPEC in the current DetNet domain and transmission cycle of CQF, thus control the burstiness and reduce overprovisioning.

A DetNet node running CQF satisfies that a maximum number of max_number_packet_per_cycle packets, each with length no larger than max_packet_size, can be transmitted during a CQF cycle Tc. Also, the dead time Dt, during which the former node cannot transmit to the later node (See Section 7.1.1), should be considered. Here we use the notation max_number_packet_per_cycle, max_packet_size to describe the maximum amount of transmitting data during the available transmit window Tc-Dt.

An ingress conditioner typically consists with a FIFO queue with an output regulator. Every incoming packet enters the FIFO queue, which passes it on if and only if the packet conforms to the CQF requirement. For this purpose, two criteria below is suggested to follow.

The output regulator controls the transmitting of packets, with credit function for instance. At the start of a CQF cycle, credit is set to the maximum bits to transmit in a cycle, max_bit_per_cycle = max_number_packet_per_cycle * max_packet_size. When a packet is transmitted from the node, the credit is reduced by length(packet). The operation of the output regulator can be described as below.

     credit = max_bit_per_cycle; % Initial credit
     t = 0; % Initial time offset within a cycle

     if (t>=0 && t< Tc-Dt) % during a transmit window
         if ~isEmpty(queue) % if queue is not empty
             if credit >= length(packet);
                 credit = credit - length(packet);
     elseif (t >= Tc) % when a cycle end, reset t and credit.
         t = t - Tc; % reset t to [0, Tc-Dt)
         credit = max_bit_per_cycle; % credit is refilled

Other instance of output regulator may also meet the CQF requirement.

7.1.7. CQF ingress conditioning timing model

Consider the input traffic conforms to variable bit rate (VBR) constraint (p, M, r, b), which can be modeled as arrival curve:

alpha(t) = min(pt+M, rt+b)

where p is the peak rate, M is the maximum size of a packet, (r, b) stand for token bucket rate and burst. Note that, if N flows enter the ingress conditioner, each flow f_i conforms to token bucket (r_i,b_i), the arrival curve parameters are of the superposition form: r = sum(r_i), b = sum(b_i), i=1,...,N.

Ingress conditioner, which transmit packets as soon as possible if conformance is satisfied, does not increase worst-case queuing latency, if we consider the latency from the input of ingress conditioner to the first output of CQF node. The CQF node, which transmits at most max_bit_per_cycle bits during a cycle Tc, offers a service curve

beta(t) = max_bit_per_cycle * floor(t-Dt) + min(p*mod(t-Dt, Tc), max_bit_per_cycle)

where Dt is the dead time (see Section 7.1.1), p is the peak rate, max_bit_per_cycle = max_number_packet_per_cycle * max_packet_size is the maximum transmitting data in a CQF cycle, floor(x) calculates the nearest integer less than or equal to x, and mod(x,y)=x-floor(x/y).

According to network calculus, the worst-case queuing delay for ingress conditioner and CQF node, denoted delay_bound, is the maximum horizontal distance between arrival curve and service curve.

Specifically, in general assumptions, ingress conditioner's average output rate r2 = max_bit_per_cycle/Tc is always no less than input traffic rate r (or else congestion loss would happen). In such conditions, the queuing delay up-bound for ingress conditioner and first CQF node is derived as ingress_delay_bound = max(t : beta(t) = (pb-rm) / (p-r)). Therefore, considering a path traverses ingress conditioner and H CQF nodes, the total delay up-bound is H*Tc + ingress_delay_bound.

7.2. Time-scheduled queuing

[IEEE8021Q] section specifies a time-aware queue-draining procedure for transmission selection at egress port of a DetNet transit node, which supports up to eight traffic classes. Each traffic class has a separate queue, frame transmission from each queue is allowed or prevented by a time gate. This time gate controlled scheduling allows time-sensitive traffic classes to transmit on dedicate time slots. Within the time slots, the transmitting flows can be granted exclusive use of the transmission medium. Generally, this time-aware scheduling is a layer 2 time division multiplexing (TDM) technique.

Consider the static configuration of a deterministic network. To provide end-to-end latency guaranteed service, network nodes can support time-based behavior, which is determined by gate control list (GCL). GCL defines the gate operation, in open or closed state, with associated timing for each traffic class queue. A time slice with gate state "open" is called transmission window. The time-based traffic scheduling must be coordinated among the DetNet transit nodes along the path from sender to receiver, to control the transmission of time-sensitive traffic.

Ideally all network devices are time synchronized and static GCL configurations on all devices along the routed path are coordinated to ensure that length of transmission window fits the assigned frames, and no two time windows for DetNet traffic on the same port overlap. (DetNet flows' windows can overlap with best-effort windows, so that unused DetNet bandwidth is available to best-effort traffic.) The processing delay, link delay and output delay in transmitting are considered in GCL computation. Transmission window for a certain flow may require that a time offset on consecutive hops be selected to reduce queueing delay as much as possible. In this case, TSN/DetNet frames transmit at the assigned transmission window at every node through the routed path, with zero congestion loss and bounded end-to-end latency. Then, the worst-case end-to-end latency of the flow can be derived from GCL configuration. For a TSN or DetNet frame, denote the transmission window on last hop closes at gate_close_time_last_hop. Assuming talker supports scheduled traffic behavior, it starts the transmission at gate_open_time_on_talker. Then worst case end-to-end delay of this flow is bounded by gate_close_time_last_hop - gate_open_time_on_talker + link_delay_last_hop.

It should be noted that scheduled traffic service relies on a synchronized network and coordinated GCL configuration. Synthesis of GCL on multiple nodes in network is a scheduling problem considering all TSN/DetNet flows traversing the network, which is a non-deterministic polynomial-time hard (NP-hard) problem. Also, at this writing, scheduled traffic service supports no more than eight traffic classes, typically using up to seven priority classes and at least one best effort class.

8. Parameters for the bounded latency model

The use of the TSPEC parameters defined in [RFC2212] and related documents for IntServ are well-established and adequate for DetNet purposes. The parameterization used by [IEEE8021Q] are somewhat different, as discussed above (Section 7.1.6). These parameters are maximum number of frames per interval, interval size, and maximum frame size. They are more suitable for the physical determination of compliance by a sender than for resource reservation purposes.

9. References

9.1. Normative References

[I-D.ietf-detnet-architecture] Finn, N., Thubert, P., Varga, B. and J. Farkas, "Deterministic Networking Architecture", Internet-Draft draft-ietf-detnet-architecture-08, September 2018.
[I-D.ietf-detnet-dp-sol-ip] Korhonen, J. and B. Varga, "DetNet IP Data Plane Encapsulation", Internet-Draft draft-ietf-detnet-dp-sol-ip-00, July 2018.
[I-D.ietf-detnet-dp-sol-mpls] Korhonen, J. and B. Varga, "DetNet MPLS Data Plane Encapsulation", Internet-Draft draft-ietf-detnet-dp-sol-mpls-00, July 2018.
[I-D.ietf-detnet-use-cases] Grossman, E., "Deterministic Networking Use Cases", Internet-Draft draft-ietf-detnet-use-cases-20, December 2018.
[RFC2212] Shenker, S., Partridge, C. and R. Guerin, "Specification of Guaranteed Quality of Service", RFC 2212, DOI 10.17487/RFC2212, September 1997.
[RFC6658] Bryant, S., Martini, L., Swallow, G. and A. Malis, "Packet Pseudowire Encapsulation over an MPLS PSN", RFC 6658, DOI 10.17487/RFC6658, July 2012.
[RFC7806] Baker, F. and R. Pan, "On Queuing, Marking, and Dropping", RFC 7806, DOI 10.17487/RFC7806, April 2016.

9.2. Informative References

[bennett2002delay] J.C.R. Bennett, K. Benson, A. Charny, W.F. Courtney, and J.-Y. Le Boudec, "Delay Jitter Bounds and Packet Scale Rate Guarantee for Expedited Forwarding"
[charny2000delay] A. Charny and J.-Y. Le Boudec, "Delay Bounds in a Network with Aggregate Scheduling"
[IEEE8021Q] IEEE 802.1, "IEEE Std 802.1Q-2018: IEEE Standard for Local and metropolitan area networks - Bridges and Bridged Networks", 2018.
[IEEE8021Qcr] IEEE 802.1, "IEEE P802.1Qcr: IEEE Draft Standard for Local and metropolitan area networks - Bridges and Bridged Networks - Amendment: Asynchronous Traffic Shaping", 2017.
[IEEE8021TSN] IEEE 802.1, "IEEE 802.1 Time-Sensitive Networking (TSN) Task Group"
[IEEE8023] IEEE 802.3, "IEEE Std 802.3-2018: IEEE Standard for Ethernet", 2018.
[le_boudec_theory_2018] J.-Y. Le Boudec, "A Theory of Traffic Regulators for Deterministic Networks with Application to Interleaved Regulators"
[NetCalBook] Le Boudec, Jean-Yves, and Patrick Thiran, "Network calculus: a theory of deterministic queuing systems for the internet", 2001.
[Specht2016UBS] J. Specht and S. Samii, "Urgency-Based Scheduler for Time-Sensitive Switched Ethernet Networks"
[TSNwithATS] E. Mohammadpour, E. Stai, M. Mohiuddin, and J.-Y. Le Boudec, "End-to-end Latency and Backlog Bounds in Time-Sensitive Networking with Credit Based Shapers and Asynchronous Traffic Shaping"

Authors' Addresses

Norman Finn Huawei Technologies Co. Ltd 3101 Rio Way Spring Valley, California 91977 US Phone: +1 925 980 6430 EMail:
Jean-Yves Le Boudec EPFL IC Station 14 Lausanne EPFL, 1015 Switzerland EMail:
Ehsan Mohammadpour EPFL IC Station 14 Lausanne EPFL, 1015 Switzerland EMail:
Jiayi Zhang Huawei Technologies Co. Ltd Q22, No.156 Beiqing Road Beijing, 100095 China EMail:
Balázs Varga Ericsson Konyves Kálmán krt. 11/B Budapest, 1097 Hungary EMail:
János Farkas Ericsson Konyves Kálmán krt. 11/B Budapest, 1097 Hungary EMail: