Quantitative Performance Comparison of Various

IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 19, NO.
3, SEPTEMBER 2022 2899
Quantitative Performance Comparison of Various

Traffic Shapers in Time-Sensitive Networking
Luxi Zhao , Paul Pop , Member, IEEE, and Sebastian Steinhorst , Senior Member, IEEE
Abstract—Owning to the sub-standards being developed by depend on the real-time capabilities of their communica-
IEEE Time-Sensitive Networking (TSN) Task Group, the tra- tion networks. Time-Sensitive Networking (TSN) [1] enhances
ditional IEEE 802.1 Ethernet is enhanced to support real-time standard Ethernet [2], aiming at providing deterministic com-
dependable communications for future time- and safety-critical
applications. Several sub-standards have been recently proposed munication for real-time traffic. Over the recent years, TSN
that introduce various traffic shapers (e.g., Time-Aware Shaper has become a high-profile and active standardization effort
(TAS), Asynchronous Traffic Shaper (ATS), Credit-Based Shaper with a strong research community both in academia and in the
(CBS), Strict Priority (SP)) for flow control mechanisms of queu- industry. Several companies, such as Belden, Cisco Systems,
ing and scheduling, targeting different application requirements. Intel Corporation, NXP Semiconductors, Siemens, TTTech
These shapers can be used in isolation or combination and there
is limited work that analyzes, evaluates, and compares their Computertechnik and Huawei Technologies are developing
performance, which makes it challenging for end-users to choose TSN switches with various capabilities. The Avnu Alliance
the right combination for their applications. This paper aims consortium has been established to evaluate the interoperabil-
at (i) quantitatively comparing various traffic shapers and their ity and conformance of such products to the TSN standards.
combinations, (ii) summarizing, classifying, and extending the TSN integrates multiple traffic types implemented by different
architectures of individual and combined traffic shapers and
their Network calculus (NC)-based performance analysis meth- scheduling mechanisms (traffic shapers), such as the Time-
ods, and (iii) filling the gap in the timing analysis research on Aware Shaper (TAS) standardized by IEEE 802.1Qbv [4],
handling ATS and CBS used for different priority queues, and the Asynchronous Traffic Shaper (ATS) standardized by IEEE
two novel hybrid architectures of combined traffic shapers, i.e., 802.1Qcr [5], the Credit-Based Shaper (CBS) standardized by
TAS+ATS+SP and TAS+ATS+CBS when ATS and CBS used IEEE 802.1Qav [7]. These shapers can be used separately
at the same queue. A large number of experiments, using both
synthetic and realistic test cases, are carried out for quantitative or in several combinations. TAS is based on a global clock
performance comparisons of various individual and combined synchronization (via IEEE 802.1AS [3]) implementing the
traffic shapers, from the perspective of upper bounds of delay, time-triggered traffic to guarantee deterministic transmission.
backlog, and jitter. To the best of our knowledge, we are the first ATS avoids using the global clock synchronization, but it is
to quantitatively compare the performance of the main traffic still able to provide real-time guarantees by reshaping traf-
shapers in TSN. The paper aims at supporting the researchers
and practitioners in the selection of suitable TSN sub-protocols fic flows per hop to reduce the burstiness of traffic. CBS is
for their use cases. an asynchronous traffic shaper that implements a bandwidth
reservation mechanism.
Index Terms—TSN, traffic shapers, combinations, real-time
performance, worst-case, comparison. Many related works have already been proposed for the
schedulability analysis and configuration for different traffic
shapers. For TAS, which relies on global clock synchroniza-
I. I NTRODUCTION tion, the scheduling synthesis for time-triggered (TT) traffic,
OWADAYS, modern cyber-physical and embedded which is also called scheduled traffic (ST), has been studied
N systems, including systems in the automotive, indus-
trial automation, avionics and aerospace domain increasingly
in [10]–[14] using different implementation methods to syn-
thesize Gate Control Lists (GCLs). Vlk et al. [14] increase
the schedulability and throughput of TT by proposing a sim-
Manuscript received 25 March 2021; revised 15 November 2021, ple hardware enhancement of a switch. Oliver et al. [15] relax
24 March 2022, and 27 May 2022; accepted 29 May 2022. Date of pub- the constraints of the scheduling model to increase the solution
lication 3 June 2022; date of current version 12 October 2022. This work
was supported by the National Natural Science Foundation of China under space at the expense of the deterministic scheduling of TAS.
Grant 62071023. The associate editor coordinating the review of this arti- A more flexible class-based (i.e., window-based) TAS model
cle and approving it for publication was V. Fodor. (Corresponding author: is proposed in [16], [17], which does not require strict flow
Luxi Zhao.)
Luxi Zhao is with the Department of Electronic and Information isolation in queues and supports unscheduled end systems.
Engineering, Beihang University, Beijing 100190, China (e-mail: Reusch et al. [40] propose the class-based schedule synthe-
zhaoluxi@buaa.edu.cn). sis for 802.1Qbv. Craciunas and Oliver [9] give an overview
Paul Pop is with the Department of Applied Mathematics and Computer
Science, Technical University of Denmark, 2800 Lyngby, Denmark (e-mail: of the comparison of scheduling mechanisms for TAS in TSN
paupo@dtu.dk). networks and time-triggered scheduling in TTEthernet. In [19],
Sebastian Steinhorst is with the Department of Electrical and Computer researchers solved the stability-aware integrated scheduling
Engineering, Technical University of Munich, 80333 Munich, Germany
(e-mail: sebastian.steinhorst@tum.de). and routing problem for networked cyber-physical systems
Digital Object Identifier 10.1109/TNSM.2022.3180160 based on the 802.1Qbv TSN standard. ATS is developed from
1932-4537
c 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
licensed use limited to: Alvas Institute of Engineering and Technology Department of Library and Information Centre. Downloaded on December 10,2023 at 06:34:01 UTC from IEEE Xplore. Restricti
2900 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 19, NO. 3, SEPTEMBER 2022
the urgency-based scheduler (UBS) proposed by Specht and on these novel combinations. We consider the coexistence
Samii [20] and aims at achieving low latency without design- between time-triggered shapers (TAS) and various event-
ing time schedules harmonized among all end systems and triggered shapers (ATS, CBS, SP). Our findings will support
switches based on global time synchronization. The same researchers and practitioners in understanding the performance
authors [21] propose the synthesis of queues and priority characteristics and mutual effects of different traffic shapers.
assignment for ATS. Zhou et al. [22], [23] present the simu- The main contributions of the paper are as follows,
lation model of ATS implemented in the Riverbed simulator. • We summarize the architectures of the main traffic
Reference [24] proves that ATS will not introduce extra over- shapers and their combinations in TSN. In order to
heads to the worst-case delay of the FIFO system. For CBS, perform a fair comparison, we use the same method
several methods related to performance and schedulability (Network Calculus, NC) to evaluate the performance of
analysis have been proposed in [25]–[29]. each shaper. Based on our and other researchers’ exist-
The above studies all assume the use of a single traffic ing NC-based analysis work for different traffic shapers
shaper. There are also some limited studies on the combination in TSN, we summarize and classify them. The existing
of different traffic shapers. An overview of the combined usage work for the NC-based analysis includes: ATS, CBS, SP
of TAS and CBS in controlling flows in in-vehicle networks individually used, and TAS+SP, TAS+CBS used in com-
was presented in [30]. A simulation study of the coexistence of bination. We complete the general uniform formula for
TAS and CBS is presented in [31]. Zhao et al. [33] propose the timing analysis of the arbitrary number of AVB classes
performance analysis of Audio Video Bridging (AVB) traffic when CBS is used individually.
under the coexistence of CBS and TAS. The same authors [34] • The NC-based performance analysis approach is extended
extend the timing analysis for the arbitrary number of AVB to the combination used of ATS and CBS for differ-
classes under the same architecture of TAS+CBS, consider- ent priority queues, including two cases of ATS+CBS
ing both standard credit behavior and more generally assumed (ATS for high priority queues and CBS for low priority
credit behavior but deviating from the standard 802.1Q [1]. queues) and CBS+ATS (CBS for high priority queues
Mohammadpour et al. [36] consider the combination of non- and ATS for low priority queues). Two novel hybrid
time-triggered control-data traffic (CDT), CBS and ATS, and architectures of traffic shapers, i.e., TAS+ATS+SP (com-
give the latency and backlog bounds for the traffic of CBS pared with TAS+SP) and TAS+ATS+CBS (compared
affected by ATS. However, the CDT model is not a stan- with TAS+CBS) are proposed to understand the impact
dard model required by the TSN standard. In [37], researchers of ATS reshaping on the combined architectures, where
present a simulation model of combined CBS and ATS within SP and CBS are used in the same queue. Even though
the OMNeT++ simulator. ATS+CBS on the same queue is not supported by the
With the increasing number of sub-standards for TSN standards, their combination is still worthwhile to be
networks, there have been several literature reviews related investigated from a research perspective. The NC-based
to TSN networks. Researchers [38] have given a comprehen- timing analysis method is extended to analyze the real-
sive survey on TSN networks, from TSN sub-standards to the time performance of traffic in these combinations. The
existing research of TSN before 2018. Maile et al. [39] provide combinations have been selected to provide comprehen-
an overview of the existing publications that use a Network sive coverage of possible combined traffic shapers in TSN
Calculus approach in the timing analysis for TSN networks. networks, supported by their corresponding NC-based
Researchers in [18] make a comparison between flow-based performance analysis.
(i.e., frame-based) and class-based TAS, which concludes that • A large number of experiments, using both synthetic and
class-based scheduling is easy to plan but loses the advantages realistic test cases, for quantitative performance compar-
of extremely low latency and jitter compared with the flow- isons of various individual and combined traffic shapers
based TAS. Nasrallah et al. [41] presented the performance are carried out, from the perspective of upper bounds of
comparison of class-based TAS and ATS based on simula- delay, backlog and jitter. Especially with ATS shaping,
tions. Nevertheless, the ATS architecture they considered does we highlight interesting results that do not always show
not exactly match the general model of ATS [5], [20]. They the superiority of ATS compared with other shapers, in
apply the ATS shaper at the ingress port of the switch instead, isolation or combination. Moreover, we compare the NC-
and consider another extra urgent queue with the highest prior- based analysis with the closed-form formula proposed
ity before ST traffic. Thus, as stated above, there are currently in [20] for the ATS shaper used individually. We also
no comprehensive and systematic guidelines on the quantita- show the positive function of ATS on the cyclic depen-
tive performance comparison of different traffic shapers, and dencies. We aim at providing a basic reference for the
their further coexistence possibilities and interactions in TSN selection of suitable TSN sub-protocols for researchers
networks. and practitioners.
This paper aims at (i) quantitatively comparing various The remainder of the paper is organized as follows.
traffic shapers, i.e., TAS, ATS, CBS, strict priority (SP) Section II gives the overview of performance metrics for the
scheduling and their combinations; (ii) summarizing, classify- TSN traffic evaluation. Section III summarizes and supple-
ing and extending the architectures of individual and combined ments the worst-case performance analysis for traffic transmis-
traffic shapers and their performance analysis methods; and sion with individual TAS, ATS and CBS shapers. Section IV
(iii) filling the gap in the timing analysis research handling presents novel combined architectures of shapers, and extends
ZHAO et al.: QUANTITATIVE PERFORMANCE COMPARISON OF VARIOUS TRAFFIC SHAPERS 2901
TABLE I
the NC-based analysis. The evaluation of our performance S UMMARY OF N OTATION
comparison of individual traffic shapers and their combinations
is provided in Section V. Section VI concludes of the paper.
The background of the NC method used is briefly introduced
in Appendix A.
II. OVERVIEW OF P ERFORMANCE M ETRICS

IN TSN E VALUATION
In this paper, we will compare the quality of service for each
individual and combined traffic shapers from the perspective
of upper bounds of end-to-end latency, backlog and end-to-
end jitter. The end-to-end latency is the time a frame uses to
traverse the network from the sending node to the receiving
node along its route. The latency upper bound is a significant
QoS metric for real-time applications, which is used to check
if a message meets its deadline. The backlog is defined as the
number of bits waiting in the queue to be served at any time,
and the backlog upper bound can be used to determine the
buffer size needed to avoid frame loss. The jitter represents
the variation in the latency of a flow. High amounts of jitter
indicate poor network performance.
In this paper, flows manipulated by time-triggered shapers
(TAS) can only be periodic flows, and flows handled by the
event-triggered shapers (ATS, CBS, SP) can be periodic or
aperiodic flows. For a periodic flow, we know its frame size,
period and priority, i.e., lf , Tf , Pf . For a sporadic flow, we
assume as the related work that the flow is regulated by a leaky
bucket model (bf , rf ) before entering the network [20], where
bf and rf are the burst and rate of the leaky bucket, respec-
tively. Thus, for a sporadic flow we know its bf , rf , lf , Pf .
In this paper, the traffic class (priority) Pf for the flow f
remains the same on all nodes along its path. TSN supports
at most eight different priorities (0 of lowest - 7 of highest
priority). Tables I and II respectively summarize the notations
and acronyms used in this paper.
A. End-to-End Latency Upper Bounds

Considering a flow f, its source of delay consists of:
which is also the upper bound of delay for each flow f in
(i) Propagation delay Dpro , which is tightly related to the
Q. Then, the upper bound of end-to-end delay of the flow f
physical medium, and considered constant in this paper;
is obtained by the sum of delays from the source ES to the
(ii) Forwarding delay Dfwd on the switch, which is the time
destination ES along its route Rf ,
interval from the time after the frame being fully received,
to the time it arrives at the buffer located after the switching Q
fabric. It is also generally considered constant; (iii) Queuing DE2E,f = Df + Dpro + Dfwd − Dfwd . (2)
delay dQ (t) in the FIFO queue Q for the egress port, which is Q∈Rf
a time-variant depending on the flows’ contention on the port.
The upper bound of queuing delay DQ can be calculated based
on the Network Calculus theory [42], see Appendix A. By B. Backlog Upper Bounds
constructing the input arrival curve αQ (t) of aggregate flows According to the Network Calculus theory, the upper bound
before Q, which represents the upper envelope of flows arrival of backlog in a queue Q is given by the maximum vertical
in any time interval, and the service curve βQ (t), which repre- deviation between the arrival curve αQ (t) of aggregate flows
sents the service guarantee for these flows, the upper bound of before the queue Q and the service curve βQ (t) offered to
queuing latency of any flows in the queue Q can be calculated flows waiting in the queue Q,
by the maximum horizontal deviation of αQ (t) and βQ (t),

DQ,f = DQ = h αQ (t), βQ (t) , (1) BQ = v αQ (t), βQ (t) . (3)
TABLE II
S UMMARY OF ACRONYM
Fig. 1. TAS Architecture.
systems (ESes) and switches (SWs), in the network support

this traffic shaper.
A. Time-Aware-Shaper (TAS)
Relying on the global network clock (IEEE 802.1ASrev [3]),
IEEE 802.1Qbv [4] defines the Time-Aware Shaper (TAS)
used to control a gate for each queue of the output port to
enable time-triggered communication, enabling the determin-
istic transmission of extremely low latency and jitter using
Gate Control Lists (GCLs). In this paper, we consider the flow-
based TAS [11]–[14], which is a widely used model compared
to class-based TAS [16], [17]. Researchers in [18] have con-
C. End-to-End Jitter Upper Bounds cluded that it has much better performance in terms of latency
Jitter refers to the delay variation, i.e., the difference in and jitter compared with the class-based TAS. Moreover, we
end-to-end latency between any selected frames in a flow consider the case that both ESes and SWs can be sched-
transmitting over a network. Then, the upper bound of jitter of uled [11], [12]. With a scheduled ES, the task sending the
a flow f is calculated by the difference between the maximum message and the communication schedule on the ES egress
and minimum bounds of end-to-end latency of the flow f. The port are synchronized.
upper bound of end-to-end latency DE2E,f has been discussed Fig. 1 depicts a TAS architecture in an egress port of a node
previously in Eq. (2). The lower bound of end-to-end latency supporting 802.1Qbv. The switching fabric forwards input
dE2E,f of f is the sum of transmission delays along its route flows to the corresponding output port according to their rout-
without the interference from other flows, which can be given ing information. The traffic class filtering (TCF) dispatches
as follows, input frames to the corresponding queue of the output port
according to their traffic class. For each egress port, there are
dE2E,f = lf /C + Dpro + Dfwd − Dfwd . (4) eight queues, where there may be multiple queues used for
Q∈Rf TT traffic to achieve completely deterministic transmission,
depending on the TT traffic load and construction of GCLs.
Thus the upper bound of jitter for the flow f is,
Frames waiting in a queue are eligible for transmission only
JE2E,f = DE2E,f − dE2E,f . (5) when the corresponding queue gate is open. The TAS control
is implemented based on GCLs which dictate the state of the
As shown above, in order to obtain the performance metrics gates. The open and closed states are represented by 1 and
for different traffic shapers in TSN, the main objective is to 0 respectively in GCLs, as shown with the GCL table beside
construct the arrival curve αQ (t) and service curve βQ (t) for the TAS architecture in Fig. 1. For example, at time t1 , the
the corresponding traffic shapers. gate for the queue Q2 is open (1) while all the rest are closed
(0). Full control of frames can be implemented by mutually
III. P ERFORMANCE A NALYSIS OF I NDIVIDUAL exclusive opening queues.
T RAFFIC S HAPERS Currently, the flow-based TAS can only support periodic
In the following, the performance analyses for each indi- traffic scheduling. The problem of GCL synthesis is to find
vidual traffic shaper, including Time-Aware Shaper (TAS), feasible and optimized offsets and queue allocations for peri-
Asynchronous Traffic Shaper (ATS) and Credit-Based Shaper odic traffic. A frame of a TT flow f on a link (egress
(CBS), are summarized and extended from the state-of-the- port) h = [va , vb ] is defined by the tuple φhf , lf /C ,
art, and specified using a citation after the subtitle of each of which φhf and lf /C respectively denote the start time
performance analysis section to indicate where the relevant (i.e., offset) of transmission and transmission duration of
previous work. Their quantitative performance comparison in the frame on the respective link. Flow f repeatedly sends
Section V-A is based on these analyses. When discussing a cer- frames at times φhf , φhf + Tf , φhf + 2 · Tf , φhf + 3 · Tf , . . .
tain traffic shaper, it is assumed that all nodes, including end Fig. 2 shows an example of a GCL using a Gantt chart,
Fig. 3. ATS Architecture.
Jitter bounds - TAS: Real-time communications are typi-

cally sensitive to jitter. The flow-based TAS model [11]–[14]
that implements a completely deterministic transmission leads
Fig. 2. Example GCL synthesis for ST.
to zero jitter, i.e.,
TAS
JE2E,f = 0. (8)
describing the transmission of two TT flows f1 and f2 , with
the routes r1 = [[ES1 , SW1 ], [SW1 , SW2 ], [SW2 , ES3 ]] and
B. Asynchronous Traffic Shaper (ATS)
r2 = [[ES2 , SW1 ], [SW1 , SW2 ], [SW2 , ES3 ]], respectively.
The x-axis represents the time dimension, and the y-axis lists Asynchronous Traffic Shaping (ATS) is another real-time
the output ports. The rectangles represent the TT frames’ trans- traffic shaper, standardized by IEEE P802.1Qcr [5]. It uses
mission, whose length equals to lf /C . The left side of the asynchronous transmission and, although it does not require
rectangle is the start time of the transmitted frame, which a global clock, it uses the local clocks to reshape traffic in
equals to φhf . The thin shaded row labeled Qi below the link each node. The ATS architecture is shown in Fig. 3. Note that
represents the waiting time of the frame in the corresponding although ATS was originally proposed in [20] for strict priority
queue Qi . queues, the essence of ATS is an interleaved regulator per input
It can be seen that the transmission time of TT traffic is port and per traffic class placed after the class-based FIFO
scheduled in advance. Thus, the performance metrics can be system of an upstream node [24]. It is theoretically possible
obtained together with the GCLs synthesis without the need that ATS can be connected to any class-based FIFO system of
for complex performance analysis methods. The performance upstream nodes, for example in combination used with CBS,
analysis for TAS was first discussed in [11], and we conclude which will be discussed in Section IV-C. It contains two levels
in the following subsection. of queues, (i) shaped queue q and (ii) shared queue Q, and
1) Performance Analysis–TAS [11]: the ATS shaping algorithm is located between them. Shaped
End-to-End Latency Bounds - TAS: A flow f using flow- queues are used to pre-store frames, which are waiting to be
based TAS has a completely deterministic end-to-end latency. reshaped into the leaky bucket model by the ATS shaping
When the GCL is constructed, its end-to-end delay is known algorithm. Shared queues are used for different priority traffic,
by the time duration between the sending time φhf 0 on the and there are at most 8 shared queues. The shared queues
h− follow the class-based scheduling mechanism. ATS has been
source ES h0 and the reception time φf n + lf /C on the des- proposed with the goal of avoiding burstiness cascades. Which
tination ES hn , where hn− is the node before hn . During the shaped queue the frames should enter depends on the queuing
h
design phase of determining the offset φf j (hj ∈ [h0 , hn− ]), schemes as follows [20], [23],
the following inherent delays are considered: propagation • QAR1: frames from different senders (input ports) should
delay Dpro , forwarding delay Dfwd , network precision δ due not be assigned to the same shaped queue in the receiver;
to the time-synchronization, and store-and-forward (transmis- • QAR2: frames from the same sender but with different
sion) delay lf /C , which enforces that a frame is forwarded priority levels are not allowed to be assigned to the same
by a node only after it has been fully received at the node. shaped queue;
Then, the end-to-end latency for the TT flow f is given by, • QAR3: frames are not allowed to be stored in the same
h− shaped queue if the frames sent to receivers are in
TAS
DE2E,f = φf n + lf /C − φhf 0 . (6)
different priority levels.
Backlog Bounds - TAS: To fully control each frame trans- The number of shaped queues in the receiver is related to the
mission, researchers have proposed [11], [12] to isolate the number of senders and the number of priorities assigned to
frames in queues, i.e., at most one flow occupies a queue at a the traffic from the sender to the receiver. Since, in the paper,
time, preventing the frame transmission ordering from being traffic priority for a flow remains the same at all nodes along
disrupted. We depict the queue occupancy with thin shaded its path, the number of shaped queues is only related to the
rows in Fig. 2. Then, the backlog bounds in the queue Q is number of senders (used input ports).
the maximum frame assigned to such a queue, The ATS shaping algorithm (See [5, Sec. 8.6.11.3]) is
TAS derived from the Token Bucket Emulation (TBE) algorithm,
BQ = max{lf }. (7)
f ∈Q which implements the committed transmission rate rf and the
committed burst size bf for each flow by calculating an eligible

time for frame transmission. Note that ATS is not implemented
with per-flow queuing, but the ATS shaping algorithm needs
to record state per flow in order to reshape each flow with
the respective constraint [24]. It also means that the shaping
parameters in the ATS shaping algorithm [5, Sec. 8.6.11.3]
are for per-flow but not for per-queue. Frames waiting in each
shaped queue are forwarded into the shared queue in FIFO
order, following their respective eligible transmission times.
According to the proof in [24] that ATS will not intro-
duce extra overheads to the worst-case delay of the FIFO
system, and inspired by [36] of the combined ATS and CBS Fig. 4. Latency illustration for ATS Architecture.
performance analysis based on the NC method, we summarize,
for the first time, an NC approach to analyze the performance
of ATS used individually as follows in Section III-B1, which shared queue Qi is the sum of output arrival curves from all
forms the basis of supporting combinations of ATS with other shaped queues q connected to Qi ,

shapers. Moreover, we compare the NC-based method with ATS
αQ i
(t) = αq∗ (t), (11)
the closed-form formula proposed by [20] in the experiment q
in Section V-A1.
1) Performance Analysis–ATS [24]: where αq∗ (t) is from Eq. (10). Note that frames in all shaped
Service Curve β ATSQ i (t) - ATS - Shared Queue: The ser-
queues q connected to the same shared queue Qi have the
vice for the traffic in the shared queues obeys strict priority same priority.
scheduling, i.e., flows with low priority can obtain service only By applying αQ ATS and β ATS (t) into Eq. (1) and Eq. (3),
i Qi
the upper bound of latency DQ ATS and backlog B ATS for SP
when the queues of higher priority traffic are empty. Then, by i Qi
ATS reshaping, the service curve for SP traffic with prior- flows passing through the shared queue Qi can be given.
ity i (i ∈ [1, nSP ]) in the corresponding shared queue Qi is Service Curve β ATSq (t) - ATS - Shaped Queue: As proved
given by, by [24, Th. 5], the ATS shaper is a kind of minimal interleaved
⎡ ⎤+ regulator, which has the characteristic that placing it at the
i−1 ATS
αQj (t) l max back-end of the FIFO system will not introduce extra worst-
j =1 >i
βQATS
(t) = C ⎣t − − ⎦ , (9) case delay, i.e., the worst-case delay in the combined system
i C C
of the front-end FIFO system and the ATS shaper is the same
as the worst-case delay of the front-end FIFO system alone.
where [x ]+ = max{0, x }, αQ ATS (t) (Eq. (11)) is the aggregate
j Obviously, the shared queue is served in a FIFO manner. Thus,
arrival curve of SP flows after ATS reshaping with the priority
a flow fed to the shaped queue q on the subsequent node will
j >i {lQj , lQBE }
j higher than the priority i, and l>imax = max max max
not increase the upper bound of the delay for the flow waiting
that is the maximum frame size of traffic with the priority
in the combined element of shared queue Qi− on the preceding
lower than priority i. ATS (t) ≤
node and the shaped queue q, i.e., d ATS − (t) + dq
Input Arrival Curve α ATS Q i (t) - ATS - Shared Queue: The Qi
input arrival curve αQi (t) of aggregate SP flows with priority
ATS D ATS ATS
− , where D − is the latency upper bound of SP flows
Qi Qi
i before the shared queue Qi is related to the total output with priority i waiting in the preceding shared queue Qi− and
arrival curves αq∗ (t) of individual flows from each previously can be calculated by applying αATS (t) (Eq. (11)) and β ATS (t)
Qi− Qi−
shaped queue q. As mentioned, each flow is reshaped into the (Eq. (9)) to Eq. (1), as shown for example in Fig. 4 with two
leaky bucket model before entering the shared queue. Then, switches under ATS architecture in sequence. In the figure,
the output arrival curve of an individual flow f from the shaped Dpro and Dfwd are respectively propagation and forwarding
queue satisfies rf · t + bf , where rf and bf are the committed delays, which are considered as constant in the paper. Then,
transmission rate and burst size for the flow f implemented for those flows transmitting from Qi− to q, as their lower
by the ATS shaping algorithm, respectively. Note that, in this
bound of the delay in the shared queue Qi− is lqmin /C , the
paper, if flow f is aperiodic, bf and rf are set to the same
maximum latency DqATS of SP flows waiting in the shaped q
leaky bucket parameters as before f entered the network, and
can be given by,
if f is periodic, we have bf = lf and rf = lf /Tf . The output
arrival curve of aggregate flows from the shaped queue q is DqATS = DQ
ATS min
− − lq /C . (12)
the sum of the output arrival curves of individual flows in q, i
Note that not all the flows in the shared queue Qi− will enter
αq∗ (t) = rf · t + bf . (10)
into the same shaped queue q, as they may be forwarded to the
f ∈q
other egress port of the subsequent node. Moreover, according
Moreover, according to the queuing schemes, there will be one to the ATS queuing schemes QAR1 and QAR2, flows queuing
or more shaped queues connected to the shared queue. Thus, in the shaped queue q can only come from the same preceding
the input arrival curve αQ
ATS (t) of aggregate flows before the
i
shared queue Qi− .
Then, the service curve βqATS (t) for aggregate flows in the
shaped queue q can be given by means of the pure-delay
function [36]
q
βqATS (t) = δD (t) (13)
q
where δD (t) is the pure-delay function [42] which equals to
0 if t ≤ D and +∞ otherwise, while D = DqATS (Eq. (12))
is the delay upper bound of flows in the shaped queue q.
q (t) - ATS - Shaped Queue: The
Input Arrival Curve α ATS
Fig. 5. CBS Architecture.
input arrival curve αqATS (t) of aggregate flows before reach-
ing the corresponding shaped queue q is related to the output
arrival curve α∗ − (t) when these flows depart the preceding queues are implemented in a single physical queue. It should
Qi
shared queue Qi− , and can be given by, also note that the maximum backlogs in the shared queue
BQATS and shaped queues B ATS cannot be reached at the same
Q− Q− q
∗ i
αQ − (t) = αf i (t) δD i (t), (14) time.
i
f ∈[Qi− ,q]
C. Credit-Based Shaper (CBS)
Q−
where αf i (t) = rf · t + bf is the input arrival curve of the The Credit-Based Shaper (CBS) [7] is another queuing and
Q−
flow f before the shared queue Qi− , δD i (t) is the pure-delay forwarding rule proposed for the bandwidth reservation for
function of the delay bound D = D ATS for aggregate SP Audio-Video Bridging (AVB) traffic. Fig. 5(a) depicts a CBS
−
Qi architecture model of an output port of a node. Currently,
flows of priority i in the preceding shared queue Qi− . Note AVB classes (i.e., Stream Reservation (SR) classes) Mi are
that we use f ∈ [Qi− , q] instead of f ∈ Qi− to emphasize expanded from two to more (a maximum of seven, i ≤ 7) pri-
that not all flows queuing in the shared queue Qi− will be orities supported by TSN [1]. Each AVB class corresponds to a
forwarded to the same shaped queue q. FIFO queue, and has its credit value for the CBS shaper, which
Moreover, different flows sharing the same shaped queue is used to control the transmission of AVB frames. For each
q cannot arrive on the shared queue Qi at the same time, AVB Class Mi , the CBS algorithm has a credit value manip-
because flows sharing a common link are serialized. Thus, by ulated by two different parameters called “idleSlope” (idSli )
taking all flows from Qi− to q as a group, their output arrival and “sendSlope” (sdSli = idSli − C ). For the AVB traffic of
curve from Qi− can be refined by considering the constraint Class Mi , idSli decides its maximum guaranteed bandwidth
from the physical link with the shaping curve σ link (t) = C ·t. reservation, of which the minimum value is set according to
According to the ATS queuing schemes QAR1 and QAR2, all the actual bandwidth usage of AVB Class Mi traffic.
flows in the shaped queue q must be from the same preceding The frame transmission based on the CBS functionality is
shared queue Qi− . Then, the input arrival curve αqATS (t) of shown using an example in Fig. 5(b). The credit is initialized
aggregate flows before the shaped queue q is given by, to zero and is increasing with the idleSlope (idSli ) when AVB
frames are waiting to be transmitted due to other higher pri-
∗
αqATS (t) = αQ − (t) ∧ σ
link max
(t) + lQ − , (15) ority AVB frames or due to the negative credit and decreasing
i i
where α∗ − (t) is given by Eq. (14), x ∧ y = min{x , y}, and with the sendSlope (sdSli ) during the transmission of an AVB
Qi frame. If the credit is positive and no frames are waiting in the
l max is the maximum frame size in the queue Qi− , which corresponding queue, then the credit is set to zero. However,
Qi−
needs to be taken into account since the frame is packetized if no frames are waiting in the queue, but the credit of the
at the switch input. corresponding queue is negative, it will increase with the idle
By applying αqATS (t) and βqATS (t) into Eq. (3), the upper slope until zero.
bound of backlog BqATS in the shaped queue q can be given. Since the standard 802.1Q [1] now supports a multiple num-
Remark: As discussed above, the ATS shaper only plays a ber of AVB classes, we extend the previous analysis work of
role in reshaping the flows, and will not increase the worst-case supporting two classes [27] and three classes [28] to an arbi-
delay of the flow in the node transmission. Thus, the end-to- trary number of AVB classes, as summarized in Section III-C1.
end latency bound for the flow can be obtained by summing The extension proof of credit bounds for an arbitrary num-
up latency bounds DQ ATS (t) only for each shared queue along ber of AVB classes can be found in our previous work [34].
i
its path. The backlogs for ATS are the sum of backlogs for the Although [34] is the NC-based performance analysis for com-
shared queue and the corresponding shaped queue connected bined TAS and CBS, and TT transmission delays AVB traffic,
to the shared queue, AVB credits will not be affected by TT traffic if credit is frozen
during the TT window and the protection interval (“guard
ATS ATS
BQ = BQ i
+ BqATS , (16) band” (GB)), which is one of the cases discussed in [34].
q→Qi 1) Performance Analysis–CBS [27]:
where q → Qi represents all the shaped queues q connected Q i (t) - CBS: The service for AVB traffic
Service Curve β CBS
to the shared queue Qi . This due to that shaped and shared in the individual CBS architecture depends only on the credit
state controlled by CBS. As opposed SP traffic, AVB traffic Appendix A. Finally, the input arrival curve αQ CBS (t) of
i
with low priority can obtain the service even if the queues of aggregate AVB flows of Class Mi before Qi is given by,
AVB traffic of higher priority are not empty. This is because
CBS
∗
link max

CBS max

αQ i (t) = α (t) ∧ σ (t) + l ∧ σ (t) + l − .
the AVB traffic cannot transmit if the CBS credit of its corre- Qi− Qi− Qi− Qi
Qi−
sponding class is negative. The guaranteed service for multiple
numbers of AVB classes Mi (i ∈ [1, nCBS ]) [27], (22)
cimax + The use of the term l max

− is because the frame is packetized
CBS Qi
βQ (t) = idSl i t − , (17)
i idSli at the switch input.1
By applying αQCBS (t) and β CBS (t) into Eq. (1) and Eq. (3),
where cimax is the credit upper bound for AVB Class Mi , i Qi
CBS and backlog B CBS for
the upper bound of latency DQ i Qi
i−1 min
j =1 cj − l>i
max AVB flows of Class Mi passing through the queue Qi can
cimax = idSli · i−1
, (18) be calculated.
j =1 idSlj − C
j >i {lQj , lQBE } is the maximum frame

where l>imax = max max max IV. P ERFORMANCE A NALYSIS OF
size in the traffic with the priority lower than priority Mi , and C OMBINED T RAFFIC S HAPERS
cimin is the lower bound of the credit of AVB Class Mi , In this section, we discuss the combination of different
max
lQ traffic shapers. We will first show the combination of ATS
cimin = sdSli · i
. (19) and CBS used in different queues. Moreover, as we will
C show in Section V-A, from the perspective of latency, jitter
Input Arrival Curve α CBS Q i (t) - CBS: The input arrival
and backlog, TAS outperforms ATS, CBS and SP. However,
curve αQi (t) of aggregate AVB flows of Class Mi before
CBS TAS requires the synthesis of optimized GCLs, which does
entering the corresponding queue Qi of the intermediate node not scale to large networks with many flows. This problem
is related to the output arrival curve α∗ − (t) of these flows can be mitigated by combining different traffic shapers in
Qi the same switch architecture to reduce the number of flows
departing the corresponding preceding queues Qi− connected handled by TAS. Therefore, we believe that the coexistence
to Qi . The output arrival curve of aggregate flows from a of time-triggered shapers (TAS) and various event-triggered
preceding queue Qi− is, shapers (ATS, CBS, SP) will be a promising approach in
Qi− Q−
the time-critical and real-time communication networks of
∗
αQ − (t) = αf (t) δD i (t), (20) the future, to support different performance quality require-
i
f ∈[Qi− ,Qi ] ments of applications, see also the discussion in [44]. Three
combined traffic shapers investigated in the literature are
Q− non-time-triggered-CDT+ATS+CBS [36], TAS+SP inherited
where αf i (t) is the input arrival curve of the AVB flow
from TTEthernet [32] and TAS+CBS [33], [34]. However,
f before Qi− , which needs to be iteratively calculated from
CDT is not a time-triggered traffic type, and the CDT model is
the node before Qi− by Eq. (66) until the source node ES is
Q−
not a standard model required by the TSN. Similarly, we add a
reached, δD i (t) is the pure-delay function of the delay upper citation after the subtitles of the performance analysis sections
bound D = D CBS − for aggregate AVB flows of Class Mi in the for TAS+SP and TAS+CBS. Inspired by the high performance
Qi
preceding queue Qi− . Note that we use f ∈ [Qi− , Qi ] instead of TAS, the ATS reshaping function and the existing com-
of f ∈ Qi− to emphasize that not all flows queuing in Qi− are bined traffic shapers, we are interested in understanding
forwarded to Qi . the impact of ATS reshaping on the combined architectures
Similarly, all AVB flows from the same preceding queue that include TAS. Thus, in the following, we propose addi-
Qi− are regarded as a group. On one hand, due to the phys- tional three architectures of combined traffic shapers, i.e.,
TAS+SP, TAS+ATS+SP, TAS+ATS+CBS. We extend the
ical link constraint σ link (t), they cannot arrive on Qi at the
NC approach to analyze the worst-case performance of traf-
same time. On the other hand, such a group of flows is also
fic under these architectures for the quantitative performance
constrained by the shaping curve σ CBS (t) of CBS, indicat-
Qi− comparison in Section V-B.
ing the effect of CBS on the output of AVB traffic. The CBS
shaping curve σQ CBS (t) is a non-greedy shaping curve, which
i A. ATS+CBS/CBS+ATS
is constructed by the upper envelope of output accumulated
bits of AVB Class Mi from Qi in any time interval, In this section, ATS and CBS used for different queues at
the same egress port are considered. The working mechanism
c max − c min
CBS of ATS and CBS used in combination follows their separate
σQ (t) = idSli t + i i , (21)
i idSli scheduling ways. The only thing that needs to be additionally
1 The detailed derivations [27] of CBS service curve in Eq. (17) and CBS
where cimax and cimin are upper and lower credit bounds
shaping curve in Eq. (21) are not the contributions of the paper. But they
respectively given by Eq. (18) and Eq. (19). The distinc- are concluded in the supplementary document with the uniformly symbols:
tion between greedy and non-greedy shaping is clarified in https://zenodo.org/record/6378112#.YjqQReeZNPY.
where cjmin is the lower bound of credit of AVB Class

Mj given by Eq. (19), q → Qi represents all the shaped
queues q connected to the shared queue Qi , bf and rf
are respectively the burst and rate after reshaped by ATS,
j >i {lQj , lQBE } is the maximum frame size in
max = max
l>i max max
the traffic with the priority lower than priority Mi .

Proof: It is assumed that RQ CBS (t) (resp. R ATS (t)) and
i Qi
∗ CBS ∗ ATS
RQi (t) (resp. RQi (t)) are the arrival and departure
processes of AVB flows of Class Mi (i ∈ [1, nCBS ]) (resp.
SP flows of priority i (i ∈ [1, nSP ]) reshaped by ATS) cross-
ing through an egress port. Let t ∈ R+ be a time point
when the queue Qi of AVB Class Mi is backlogged, i.e.,
RQ∗ CBS (t) < R CBS (t). Then let us define s = sup{u ≤
i Qi
t ci (u) = 0, RQ ∗ CBS (u) = R CBS (u)}, where c (t) is the
i Qi i
credit value of AVB Class Mi at time t. This implies that
∀u ∈ (s, t], the queue Qi is non-empty or ci (u) < 0.
Otherwise, we can always find another s < s’ ≤ t that satisfies
ci (s ) = 0 and RQ ∗ CBS (s ) = R CBS (s ). Therefore, we define
i Qi
the duration (s, t] the busy period of AVB Class Mi .
According to the working mechanism of CBS under
ATS+CBS, the interval Δt = t − s can be decomposed by
Δt = Δt − + Δt + ,
Fig. 6. ATS+CBS/CBS+ATS Combined Shaper Architecture (H/L: high/low
priority).
where Δt − (resp. Δt + ) is the accumulated length of all
periods where the credit is decreasing (resp. increasing).
considered is to set the priority for ATS and CBS queues. In Δt − = ΔtQ CBS represents the transmission duration of AVB
i
this paper, it is assumed two cases: ATS and CBS respectively, Class Mi , and Δt + = ΔtQ CBS is the waiting time duration
j
have the higher priority, of which architectures are respectively of AVB Class Mi due to the transmission for higher priority
shown in Fig. 6(a) and (b). And there are nSP queues and SP flows reshaped by ATS or the transmission for other AVB
nCBS queues. We presume in this paper the non-preemptive Classes or due to the credit ci (t) < 0. And we have
integration mode for different traffic types. The frame trans-
mission based on the CBS functionality is shown with an Δt + = Δt − ΔtQ
CBS
i
.
example in Fig. 5(c), (d). The credit behavior is very simi-
lar to CBS individually used. Note that when CBS is used Due to the definition of s, for ∀u ∈ (s, t], the queue Qi
for low priority traffic, the credit of all AVB classes will be for AVB traffic is not possible to temporarily empty when
increased when the high priority SP frame is on transmis- ci (u) > 0. Thus, there is no case where the credit of AVB
sion. In the following, the NC approach is expanded to the Class Mi is reduced from some positive value P to 0 due to
performance analysis of ATS and AVB traffic in the ATS+CBS resets. Then the variation of credit for AVB Class Mi during
and CBS+ATS architectures. the time interval Δt = t − s satisfies,
1) Performance Analysis–ATS+CBS: Since the flows ci (t) − ci (s) = ci (t)
shaped by ATS have the highest priority, and there will be
= Δt + · idSli + Δt − · sdSli
at most one non-preemptive AVB frame that can impact their
CBS CBS
transmission. Therefore, the performance analysis for ATS is = Δt − ΔtQ i
· idSli + ΔtQ i
· sdSli
completely the same as ATS individually used, as discussed in CBS
max in the = Δt · idSli − ΔtQ · C,
Section III-B1. Note that the only difference is that l>i i
service curve βQi (Eq. (9)) should additionally consider the

ATS
Thus, it is obtained the expression of service times for AVB
maximum frame size of AVB traffic with low priority. Then, Class Mi in any interval Δt,
in this section, we only focus on CBS for low-priority traffic.
Service Curve β CBS Δt · idSli − ci (t)
Q i (t) - CBS:
CBS
ΔtQ = .
Corollary 1: With the impact of high priority SP traffic i C
shaped by ATS, the service curve βQ CBS (t) for low priority
i
Since the service could only be supplied for AVB traffic Mi
AVB traffic of Class Mi (i ∈ [1, nCBS ]) is same to Eq. (17), during ΔtQ CBS , i.e., the descent time Δt − of the credit, then
i
where the credit upper bound cimax for AVB Class Mi is over the interval (s, t], we have
given by,
∗ CBS ∗ CBS CBS
i−1 min SP max
RQ (t) − RQ (s) = C · ΔtQ ≥ Δt · idSli − cimax ,
j =1 cj − n k =1 f ∈q bf − l>i
i i i
cimax = idSli ·
q→Qk
i−1 nSP , ∗ CBS (t) −
j =1 idSlj + k =1 q→Qk f ∈q rf − C
where ci (t) ≤ cimax . Since we have also RQ i
∗ CBS (s) = R ∗ CBS (t) − R CBS (s) ≥ 0 and R ∗ CBS (t) is
RQ
(23) i Qi Qi Qi
CBS = t − s − Δt
ATS − ΔtQi − ΔtLP and P ≥ 0,
a wide-sense increasing function, from which we derive Since Δt<i CBS
the above expression is modified into,
∗ CBS CBS cimax +
RQ (t) ≥ RQi (s) + idSl i t − s − CBS
i idSli c<i (t) − c<i (s) ≤ ΔtATS + ΔtQ + Δt LP · C
⎛
i
⎞
c max +
≥ inf RQCBS
(s) + idSli t − s − i .
i−1
0≤s≤t i idSli + (t − s) · ⎝ idSlj − C ⎠. (25)
j =1
Therefore, the service curve βQ CBS (t) for AVB Class M is
i i
c max For ΔtATS , we have,
βQi (t) = idSli [t − idSli ] , which is same to the expression
CBS i +
nSP
nSP
in Eq. (17). ATS ATS ATS
ΔtATS · C = RQ (t) − R (s) ≤ αQ (t − s),
In the following, we will derive the credit upper bound cimax k Qk k
k =1 k =1
for each AVB Class Mi (i ∈ [1, nCBS ]). Let t ∈ R+ be a time
point when AVB gates are in the open state, and ci (t) > 0. where αQ ATS (t) is the arrival curve of aggregate SP flows of
k
Then let us define s = sup{u ≤ t ∀Qj ∈ Q≤i CBS , c (u) ≤ 0},
j priority k before the shared queue of ATS, given by Eq. (11).
CBS
where Q≤i are the queues with priority no lower than AVB Thus, the above equation can be written,
Class Mi . This implies that ∀u ∈ (s, t], ∃Qj ∈ Q≤i CBS ,

nSP
cj (u) > 0, i.e., there always exists at least one queue in Q≤i CBS ΔtATS · C ≤ (rf (t − s) + bf ),
with some frame to send. Otherwise, we can always find k =1 q→Qk f ∈q
another s < s ≤ t that satisfies ∀Qj ∈ Q≤i CBS , c (s ) ≤ 0.
j where q → Qi represents all the shaped queues q con-
Consider the evolution of the credit value ci (t) of AVB nected to the shared queue Qi . Moreover, since ΔtLP ·
Class Mi between s and t. The credit ci (t) 1) decreases at C ≤ maxj ∈[i+1,nCBS ] {lQ
max , l max } = l max ,Eq. (25) can be
BE >i
speed sdSli = idSli − C when the frame in the queue Qi is rewritten to,
j
on transmission (during ΔtQ CBS ); 2) increases at speed idSl

i CBS nSP max
i c<i (t) − c<i (s) − ΔtQ ·C − k =1 q→Qk f ∈q bf − l>i
when the frame in the queue Qi is waiting when SP flows t −s ≤ i−1
i
nSP .
j =1 idSlj + k =1 f ∈q rf − C
shaped by ATS is on transmission (during ΔtATS ) or when q→Qk
AVB flows with higher priority than Mi is on transmission (26)

(during Δt<i CBS ) or a non-preemptive lower priority frame
By applying Eq. (26) into Eq. (24), it is obtained that
than Mi is on transmission (during ΔtLP ); 3) and may be
reduced from some positive value P to 0 due to resets. Thus ci (t) − ci (s)
i nSP
j =1 idSlj + f ∈q rf − C
the variation of ci (t) during (s, t] is, k =1 q→Qk CBS
≤ − i−1 nSP · ΔtQ i
·C
ci (t) − ci (s) j =1 idSl j + k =1 q→Qk f ∈q f r − C
SP max
CBS
= ΔtQ CBS
· sdSli + ΔtATS + Δt<i + ΔtLP · idSli − P . c<i (t) − c<i (s) − n k =1 q→Qk f ∈q bf − l>i
i + i−1 nSP · idSli
j =1 idSlj + k =1 q→Qk f ∈q rf − C
CBS + Δt
Since ΔtATS + Δt<i LP = t − s − ΔtQi
CBS and P ≥ 0, SP max
c<i (t) − c<i (s) − n k =1 q→Qk f ∈q bf − l>i
The above equation is modified into, ≤ i−1 nSP · idSli .
CBS j =1 idSlj + k =1 q→Qk f ∈q rf − C
ci (t) − ci (s) ≤ −ΔtQ · C + (t − s) · idSli . (24)
i
By definition of s, we have c<i (s) ≤ 0, ci (s) ≤ 0,
i−1 nSP
Let c<i (t) = ji−1=1 cj (t) denote the sum of credits of AVB j =1 idSlj + k =1 q→Qk f ∈q rf −C < 0, and c<i (t) ≥
traffic with a priority higher than Class Mi . At any instant i−1 min
j =1 cj , it is concluded that,
between s and t the sum of credits c<i (t) 1) increases at i−1 SP
min max
most at speed ji−1 =1 idSlj when a frame of SP traffic shaped j =1 cj − n k =1 q→Qk f ∈q bf − l>i
ci (t) ≤ idSli · i−1 nSP .
by ATS is on transmission (during ΔtATS ), or a frame of
CBS ), or a low priority j =1 idSlj + k =1 q→Qk f ∈q rf − C
Class Mi uses the link (during ΔtQ i
frame blocks the link (during ΔtLP ); 2) decreases at least at where cjmin is the lower bound of credit of AVB Class Mj .
speed ji−1 =1 idSlj −C (all the classes from Q<i gain credit,
CBS Since an AVB frame cannot start to be forwarded if the credit
except one which loses credit) when a frame from class with is lower than 0, the minimum credit bound is reached when
higher priority than Class Mi is being sent (during Δt<i CBS ); the size of the frame is maximized, which is the same as
3) be reduced from some positive value P to 0 due to a set of Eq. (19).
resets. Then the variation of c<i (t) between s and t is, Input Arrival Curve α CBS
Q i (t) - CBS:
i−1 Corollary 2: With the impact of high priority SP traffic
CBS
c<i (t) − c<i (s) = ΔtATS + ΔtQ + Δt · idSlj shaped by ATS, the input arrival curve αQCBS (t) for low pri-
LP i
i
j =1 ority AVB traffic of Class Mi (i ∈ [1, nCBS ]) is the same to
⎛ ⎞ Eq. (22), where the CBS shaping curve σQCBS (t) is given from

i−1 max
Eq. (21) by replacing ci with Eq. (23).
i
CBS ⎝
+ Δt<i · idSlj − C ⎠ − P
The proof is the same as the proof of σQ CBS (t) when CBS
j =1 i
individually used.
2) Performance Analysis–CBS+ATS: Similar to the

ATS+CBS in conversely used, flows passing through CBS
have the highest priority and their performance analyses
are the same as the analysis (Section III-C1) when CBS
individually used. It should also be noted that l>i max in the
credit upper bound ci max should take the maximum frame
size of flows reshaped by ATS into account. Thus, in this
section, we only focus on ATS for low-priority traffic.
Q i (t) - ATS - Shared Queue:
Service Curve β ATS
Corollary 3: With the impact of high priority AVB traffic,
the service curve βQ
ATS (t) for low priority SP traffic reshaped
i
by ATS in the corresponding shared queue Qi (i ∈ [1, nSP ])
is given by,
nCBS i−1 +
k =1 αCBS
Q (t) + j =1 αATS
Qj (t)
max
l>i
ATS
βQ =C t− − ,
k
i
(t)
C C
(27)
where αQ CBS (t) (Eq. (22)) is the arrival curve of aggre-

k
gate AVB flows of Class Mk (k ∈ [1, nCBS ]), αQ ATS (t)
j
(Eq. (11)) is the arrival curve of SP flows after ATS
reshaping with the priority j higher than the priority i, Fig. 7. TAS+SP/CBS Combined Shaper Architecture.
j ∈[i+1,nSP ] {lQj , lQBE } that is the max-
and l>imax = max max max
imum frame size of traffic with the priority lower than into,
priority i. n
CBS
Proof: Let RQ ATS (t) (resp. R CBS (t)) and R ∗ ATS (t) (resp. ∗ ATS ∗ ATS CBS
i Qi Qi RQ (t) − RQ (s) ≥ C (t − s) − αQ (t − s)
∗ CBS
RQi (t)) are the arrival and departure processes of SP flows
i i k
k =1
of priority i (i ∈ [1, nSP ]) reshaped by ATS (resp. AVB flows
i−1
of Class Mi (i ∈ [1, nCBS ])). It is assumed that t ∈ R+ be − ATS
αQ j
max
(t − s) − l>i ,
a time point when the queue Qi of SP traffic of priority i j =1
reshaped by ATS is backlogged, i.e., RQ ∗ ATS (t) < R ATS (t).
i Qi ∗ ATS (t) − R ∗ ATS (s) = R ∗ ATS (t) −
Since we have also RQ
Then, let s = sup{u ≤ t ∀k ∈ [1, nCBS ], RQ ∗ CBS (u) = i Qi Qi
k
RQATS (s) ≥ 0 and R ∗ ATS (t) is a wide-sense increasing
RQCBS (u); ∀j ∈ [1, i ], R ∗ ATS (u) = R ATS (u)}. i Q i
k Qj Qj function, from which we derive
During the time interval (s, t], flows of priority i in the n
CBS
shared queue Qi will obtain the service only when the queue ∗ ATS ATS CBS
for higher priority traffic is empty. Moreover, at most one non- RQi (t) ≥ RQi (s) + C (t − s) − αQ k
(t − s)
preemptable frame with lower priority is transmitted within k =1
⎤+
(s, t]. Thus, we have
i−1
ATS max ⎦
− αQ j
(t − s) − l>i

nCBS
j =1
∗ ATS ∗ ATS ∗ CBS ∗ CBS
RQ i
(t) − RQ i
(s) = C (t − s) − RQ k
(t) − RQ k
(s)
ATS ATS
k =1 ≥ inf RQ i
(s) + βQ i
(t − s) .
i−1
0≤s≤t
∗ ATS ∗ ATS
− RQ (t) − RQ (s) − ΔtLP ,
where the service curve βQ ATS (t − s) = [C (t − s) −
j j
j =1 i
nCBS CBS i−1 ATS
k =1 αQk (t − s) − j =1 αQj (t − s) − l>i ] .
(28) max +
Especially, for the input arrival curve αQi (t) in the

ATS
Since ΔtLP · C ≤ maxj ∈[i+1,nSP ] {lQ
max , l max } = l max , and
>i
j BE shared queue, and the service curve βqATS (t) and the arrival
∗ CBS ∗ CBS ∗ CBS CBS
curve αqATS (t) in the shaped queue, are the same with ones
RQ k
(t) − RQ k
(s) = RQ k
(t) − RQ k
(s) (Eq. (11), Eq. (13), Eq. (15)) discussed in Section III-B1.
CBS CBS CBS
≤ RQ k
(t) − RQ k
(s) ≤ αQ k
(t − s)
B. TAS+SP/TAS+CBS
where αQ CBS (t) given by Eq. (22) is the arrival curve of
k
∗ ATS (t) −
With the combination of the completely deterministic trans-
aggregate AVB flows of Class Mk . Similarly, RQ j mission of TAS, we first address the architecture without
RQ∗ ATS (s) ≤ αATS (t − s), where αATS (t) given by
j Qj Qj ATS. One possible combination is that of the Time-Aware
Eq. (11) is the input arrival curve of aggregate flows before Shaper (TAS) and Strict Priority (SP) queuing (i.e., TAS+SP),
the shared queue Qj . Then, Eq. (28) can be changed shown in Fig. 7(a). The combined scheduling mechanisms
traffic shapers. Here, the lower priority traffic (SP/AVB) can

only obtain the remaining service after TT frames are for-
warded. Therefore, for the TAS+SP/CBS combined traffic
shaper, it is necessary to first calculate the arrival curve related
to TT traffic, i.e., the maximum accumulated bits within any
Fig. 8. Non-preemption integration modes of TT and SP/AVB Traffic. time interval that could not be served for lower priority traffic
(SP/AVB), which was first proposed in [32] for TTEthernet
and extended to TSN [33].
of TAS+SP are inherited from TTEthernet which supports Arrival Curve α TASI (t) - TAS [32], [33]: The construction
Time-Triggered (TT) traffic and Rate-Constrained (RC) com- of the arrival curve αITAS (t) for TT traffic needs to be consid-
munication with a strict priority allocation. The difference in ered under two situations. The subscript I = {TT, GB + TT}
TSN is how the TT frames are transmitted. In TTEthernet, is used for distinguishing whether the credit for AVB traffic
TT communication is implemented by directly controlling is frozen or not during GB. If the credit is non-frozen during
the temporal behavior of each individual frame of the TT GB [34] inconsistency with the standard [1], the credit will
flows [45]. However, in TSN, TT communication depends only be frozen during the time slots (windows), occupied by
on the gate control for corresponding TT queues in egress TT frames. Then I = TT. If the credit is frozen during GB
ports, which requires flow or frame isolation constraints in as most literature assumes [33], [35], [36], the GB duration
order to achieve completely deterministic transmission [11], and the corresponding TT time slot can be taken as a whole
[12]. Another possible combination is that of TAS and the to represent the credit frozen duration. Then I = GB + TT.
Credit-Based Shaper (CBS) (TAS+CBS) [34]. The TAS+CBS Moreover, whether the GB and the corresponding TT time slot
architecture is presented in Fig. 7(b). In any combination, TT are taken as a whole to construct the arrival curve αITAS (t)
traffic implemented by TAS always has the highest priority. also depends on the selection of the scheduling architec-
Thus, it will have the same high real-time performance as ture (TAS+SP or TAS+CBS), which will be respectively in
when individually used. SP/AVB traffic has the secondary pri- Section IV-B1 and Section IV-B2.
ority. Best-Effort (BE) traffic has the lowest priority without The time slots for TT traffic are given by GCLs for each
timing guarantee requirements. Different from SP scheduling, egress port, as in the example in Fig. 2. They appear period-
which handles the flows based on their priorities, CBS enforces ically according to the GCL period TGCL , and the number
a bandwidth reservation for multiple priorities of AVB traf- NTT of TT time slots within the GCL period is finite. It is
fic. CBS is used to prevent the starvation of lower-priority assumed that the n th (n ∈ [0, NTT − 1]) time slot occupied
AVB traffic, and can tolerate a certain degree of degradation by TT traffic starts at time onTT and has the time duration
in real-time performance for high-priority traffic. LTT
n . Then, the relative offset between the starting times of
With TSN, there is a gate for each queue of egress ports. the n th and m th (m ∈ [n, (n +NTT −1)%NTT ]) time slots is
TT = o TT −o TT , if m ≥ n, and o TT = o TT +T
GCL −on
Only when the gate is open, the frames in the correspond- om,n TT
m n m,n m
ing queue can be forwarded. If more than one gate opens otherwise. Consequently, the arrival curve αI (t) related to
TAS
at the same time, the frame transmission is based on their TT traffic can be given for all t ∈ R+ [34],
priority. In order to keep the completely deterministic trans- ⎧ ⎫
mission for TT traffic, when an associated gate for TT traffic ⎨(n+NTT
−1)%NTT TAS
t − oI,m,n ⎬
αTAS
I (t) = max TAS
lI,m ,
is open, the remaining gates for other traffic types (SP, AVB, 0≤n≤NTT −1⎩ TGCL ⎭
m=n
etc.) are closed, and vice versa. Thus, lower priority traffic can
(29)
be prevented from occupying the time slots reserved for TT
frames. In this paper, we consider the non-preemption inte- where
gration mode [4] to solve the issue when a SP/AVB frame TT
is already in transmission at the beginning of the time slot TAS
lI,m = mTT· C , GB
L I = TT
reserved for TT traffic, as shown in Fig. 8. The non-preemption Lm + Lm · C , I = GB + TT,
mode introduces a “guard band” (GB) interval before the TT
time slot to ensure no additional delay and jitter for TT traffic. and
TT
The frame is prevented from initiating transmission if there is TAS om,n , I = TT
not sufficient time for the whole frame transmission before the oI,m,n = TT + LGB − LGB ,
om,n n m I = GB + TT .
gate is closed. For SP/AVB traffic, the maximum GB (LGB ) is
related to the transmission time of a maximum SP/AVB frame Note that LGB
m is the minimum value of the maximum frame
waiting in the corresponding queue. of lower priority than the flow of interest with priority i and
For the combined traffic shapers, there is no need to maximum idle time slot between two consecutive TT time
re-analyze TT traffic shaped by the TAS shaper, as TT slots,
traffic is scheduled within pre-allocated time slots and is
not interfered with by other traffic types. However, for LGB max TT
m = min l>i , om,m−1 + Lm−1 ,
TT
lower priority SP/AVB traffic in the combined traffic shaper
TAS+SP/TAS+CBS, the real-time performance is different where l>imax is the maximum frame size in traffic with the
from the one in the corresponding individual (SP or CBS) priority lower than the priority i traffic.
1) Performance Analysis–TAS+SP [32]:

Service Curve β SPQ i (t) - SP: SP traffic of different pri-
orities competes for the leftover bandwidth after serving TT
traffic. Moreover, the service SP traffic obtains also depends
on the integration mode selected. Since we consider the non-
preemption mode, there will be a GB before each TT window
to prevent an SP frame already in transmission from interfering
with TT traffic. Then, in the worst case, the time slot that SP
traffic cannot occupy will be enlarged to GB + TT. SP traf-
fic with low priority can obtain the service only when the
queues of SP traffic of higher priority are empty. Then, the
service curve for SP traffic with priority i (i ∈ [1, nSP ]) in the Fig. 9. CBS forwarding frames under the impact of TAS.
corresponding queue Qi can be given as follows,
for AVB traffic. An example of the CBS working mecha-
i−1 SP + nism under the non-preemption integration mode with different
TAS max
SP αGB + TT (t) j =1 αQj (t) l>i
βQ i
(t) = C t − − − , assumptions on the variation of credit during GB is shown in
C C C
↑ Fig. 9. The service curve for AVB Class Mi (i ∈ [1, nCBS ])
(30) in the corresponding queue Qi is given by [34],
max +
where [f (t)]+↑ = max0≤s≤t {f (s), 0}, αGB + TT (t) is from
TAS α[TAS
M ] (t) ci[ M]
CBS
Eq. (29) with I = GB + TT, αQj (t) (Eq. (31)) is the arrival
SP βQi [ M ] (t) = idSli t − − , (32)
C idSli
↑
curve of aggregate SP flows with the priority j higher than the
max is the maximum frame size in traffic with
priority i, and l>i where M ∈ {F, NF} representing the choice of the credit state
the priority lower than the priority i. during GB (F — frozen credit during GB; NF — non-frozen
Input Arrival Curve α SP Q i (t) - SP: The input arrival curve M ] (t) = {αTT (t), αTT + GB (t)} from
credit during GB), α[TAS TAS TAS
αQi (t) of aggregate SP flows with the priority i before enter-

SP Eq. (29), and the credit upper bound ci[ M ] = {cimax , c max
max
i }
ing the corresponding queue Qi of the intermediate node is of AVB Class Mi . Here cimax is the credit upper bound when
related to the total output arrival curve α∗ − (t) of these flows credit is considered frozen during GB, which equals to the
Qi
credit upper bound (Eq. (18)) of CBS used individually, and
departing the corresponding preceding queues Qi− connected
c max is the credit upper bound when considering the non-
to Qi and the shaping curve σ link (t) of the physical link by i
frozen credit during GB,
taking all SP flows from Qi− to Qi as a group. The calculation
of α∗ − (t) can be done considering Eq. (20), by substituting i−1 min
j =1 cj − l>i
max − σ GB
i
Qi
Qi−
ci (t) ≤ idSli · i−1
= c max
i . (33)
the delay bound in δD (t) with the delay upper bound D SP− ρGB
i + j =1 idSl j − C
Qi
of aggregate SP flows with priority i at the preceding queue where cjmin is the lower bound of credit of AVB Class Mj
Qi− . Then, αQ
SP (t) can be given by,
i (Eq. (19)), and σiGB and ρGB i are parameters of the linear
upper envelope related to GB duration and satisfy ∀s, t ∈
SP ∗ link max
αQ (t) = αQ − (t) ∧ σ (t) + lQ . (31) R+ , s ≤ t, C ·ΔtGB (s, t) ≤ σiGB +ρGB i ·(t −s −ΔtTT (s, t)),
i −
i i
Qi− where ΔtTT (s, t) and ΔtGB (s, t) respectively represent dura-
tion of TT frames emission and of accumulative guard bands
By applying αQ SP (t) and β SP (t) in Eq. (1) and Eq. (3), the
i Qi during the interval (s, t].
upper bound of latency DQ SP and backlog B SP for SP flows of
i Qi The choice of expressions of α[TAS max
M ] (t) and ci[ M ] depends
priority i passing through the queue Qi under the architecture on the credit state during GB, as follows.
TAS+SP can be determined. • M=F: There will be a GB before each TT window to
2) Performance Analysis–TAS+CBS [34]: prevent an AVB frame already in transmission interfering
Service Curve β CBSQ i [M] (t) - CBS: The service for AVB with TT traffic. Then, the time slot that AVB traffic can-
traffic in the TAS+CBS architecture depends not only on the not occupy will be enlarged to GB+TT. Moreover, since
leftover service after serving TT traffic, but also on the credit credit is always frozen during GB+TT, the maximum
state controlled by CBS. AVB traffic with different classes credit value will not be affected by GB+TT time slots and
competes for the remaining bandwidth. When the gate for the equals the credit upper bound when using CBS individu-
AVB queue is open, the variation of associated credit is the ally. Here we take GB and TT slots as a whole and then
same as in the case CBS is used individually, see Section III-C. M ] (t) = αTT + GB (t) (Eq. (29)), ci[ M ] = ci
have α[TAS TAS max max
When the gate for the AVB queue is closed, i.e., during TT (Eq. (18)).
transmission, the credit is frozen. In particular, during GB, the • M=NF: Although there is a GB before each TT window,
gates for all AVB queues are open without any frame transmis- and no AVB traffic class can transmit during GB+TT, the
sion, however. Then, the variation of credit during GB has two credit of the corresponding AVB class will be increased
cases, frozen and non-frozen, which will impact the service during GB, however. Therefore, when deriving the service
curve for AVB traffic, GB and TT slots cannot be taken as

a whole as the maximum credit value will be affected by
GB duration, which is not equal to the one in individually
using CBS anymore. Here we have α[TAS M ] (t) = αTT (t)
TAS
max max
(Eq. (29)), ci[ M ] = c i (Eq. (33)).
Input Arrival Curve α Q [M] (t) - CBS: Similar to the
CBS
i
CBS used individually, the input arrival curve αQ CBS (t) of
i[ M ]
aggregate AVB flows with priority Mi before entering the cor-
responding queue Qi of the intermediate node is related to
the total output arrival curve α∗ − (t) of these flows in the
Qi [ M ]
preceding queues Qi− connected to Qi , to the shaping curve
σ link (t) of the physical link by taking all AVB flows from
Qi− to Qi as a group, and to the shaping curve σQ CBS (t)
i[ M ]
(Eq. (35)) of CBS with the consideration of TAS influence,
CBS
∗

link max

αQ i[ M ]
(t) = αQ −
[M]
(t) ∧ σ (t) + l Q −
i i
Qi−

CBS max
∧ σQ −
[M]
(t) + lQ − . (34)
i i
where the calculation of α∗ (t) can refer to Eq. (20),

Qi− [ M ]
Qi−
and the delay in δD (t) is the delay upper bound D CBS
Qi− [ M ]
−
of AVB Class Mi traffic at queue Qi .
The CBS shaping curve σQ CBS (t) is also a non-greedy
i[ M ]
shaping curve, which is constructed as the upper envelope of
output accumulated bits of AVB Class i from Qi in any time Fig. 10. TAS+ATS+SP/CBS Combined Shaper Architecture.
interval. Its expression depends on the choice of the credit CBS for AVB flows of Class M passing through the queue
BQ i
state during GB and is given by,2 i[ M ]
Qi under the architecture TAS+CBS can be calculated for two
max − c min +
βTT
TAS (t) ci[ M] i
cases of credit during GB, respectively.
CBS
σQi [ M ] (t) = idSli t − + , (35)
C idSli
↑ C. TAS+ATS+SP/TAS+ATS+CBS
where M ∈ {F, NF} represents the choice of the credit state An ATS shaper is a type of minimal interleaved regula-
during GB, ci[ max = {c max , c max } with c max from Eq. (18)
M] i i i tor [24], used to reshape traffic before entering into the queue
and c max
i from Eq. (33), of which the expression selection for each egress port of the middle node in the network. In
depends on the credit state during GB, and βTTTAS (t) represents this section, the hybrid architectures TAS+ATS+SP/CBS are
the minimum amount of service obtained by TT traffic in any presented, aiming to evaluate the reshaping influence of ATS
interval, and is given as follows, on the real-time performance of other event-triggered shapers
n+N −1 under the effect of the time-triggered shaper (TAS). Note that
TAS
TT
βTT (t) = min βTDMA (t + t0 , LTT

m ) , the combination of ATS+CBS on the same queue is not sup-
0≤n≤NTT −1
m=n ported by the standards, since the TSN standards allow a queue
(36) to have only one transmission selection algorithm, either CBS
or ATS. However, the combination of ATS and CBS used for
where
the same queue is also worthwhile to be investigated, since it
βTDMA (t, L) = C · max TGCL
t
L, t − t
TGCL
(TGCL − L) , could be relevant for industrial use cases. In Section V-A, we
will find that ATS used alone is not always better than SP and
and CBS. But the advantage of the reshaping effect of ATS for
t0 = TGCL − LTT TT TT TT lower priority traffic under the hybrid architecture is greater
m − om + on−1 + Ln−1 .
than that of using ATS alone, as will be shown in Section V.
By applying αQCBS (t) and β CBS (t) into Eq. (1) and
Different from the architecture of CDT+ATS+CBS
i[ M ] Qi [ M ]
Eq. (3), the upper bound of latency DQ CBS proposed by [36], which assumes that CDT with the high-
[ M ] and backlog i
est priority satisfies the leaky bucket model, here we consider
2 The detailed derivations [34] of CBS service curve in Eq. (32) under the more general model for TAS, i.e., satisfying arbitrary
TAS+CBS, credit upper bounds cimax and c max i for arbitrary num- time-triggered slots. For this mode, the arrival curve of TT
ber of AVB classes, expressions for σiGB and ρGB i , and CBS shap- traffic satisfies the non-linear staircase function from Eq. (29).
ing curve in Eq. (35) are not the contributions of the paper. But they
are concluded in the supplementary document with the uniform symbols: The architectures of TAS+ATS+SP and TAS+ATS+CBS are
https://zenodo.org/record/6378112#.YjqQReeZNPY. shown in Fig. 10(a) and (b), respectively. Compared with the
i.e., s = sup{u ≤ t RTT ∗ (u) = R ∗

TAS+SP and TAS+CBS architectures in Fig. 7, there are addi- TT (u), RGB (u) =
tional shaped queues and the ATS shaping algorithm is used ∗
RGB (u), RQj (u) = RQj (u), ∀j ∈ [1, i ]}. Since SP traffic in
to reshape SP/AVB flows before admitting them into their Qi cannot be served during the TT/GB time slots, and when
corresponding priority queues (shared queues). The queuing queues for other SP traffic with higher priority are non-empty
schemes for frames entering the shaped queues and the ATS or a non-preemptive lower priority frame lL as well, then over
shaping algorithm are the same as the ATS used individually. the interval Δt = t − s, we have
Moreover, the gate operation is the same as in the architecture ∗ ∗ ∗ ∗
of TAS+SP/CBS without ATS, i.e., TT traffic has the exclusive RQ i
(t) − RQ i
(s) = C (t − s) − (RTT (t) − RTT (s))
∗ ∗
gate opening. SP/AVB frames are allowed to be transmitted − (RGB (t) − RGB (s))
only when the TT gate is closed and their corresponding gates i−1

∗ ∗
are open. As earlier, we use the non-preemption integration − RQ j
(t) − R Q j
(s) − lL . (39)
mode with the GB duration. Since the TT traffic controlled j =1
by the TAS has the highest priority, the traffic reshaping for
With the TAS+ATS+SP architecture, ATS is only used to
SP/AVB flows by ATS will not affect the transmission of TT
reshape SP traffic but without the influence on TT traffic type.
traffic.
Moreover, the integration modes between SP and TT traffic are
In the following, we extend the NC approach to
not influenced by ATS. In this paper, in order to compare with
TAS+ATS+SP/CBS architectures for the quantitative
the existing work, we also assume the non-preemption integra-
performance comparison in Section V-B. ∗ (t)−R ∗ (s))
tion mode. Therefore, the upper bounds of (RTT TT
1) Performance Analysis–TAS+ATS+SP: ∗ ∗
+SP and (RGB (t) − RGB (s)) are the same as the case under the
Input Arrival Curve α ATSQi (t) - SP - Shared Queue:
TAS+SP architecture, and can be taken as a whole, i.e.,
As the output of each SP flow f departing the shaped queue q
∗ ∗ ∗ ∗
is constrained by the committed transmission rate rf and the (RTT (t) − RTT (s)) + (RGB (t) − RGB (s))
committed burst size bf , the output arrival curve of f from ≤ (RTT (t) − RTT (s)) + (RGB (t) − RGB (s))
the shaped queue q satisfies rf · t + bf . It is also the input TAS
≤ αGB + TT (t), (40)
arrival curve of f before entering into the shared queue Qi .
ATS + SP
Therefore, the input arrival curve αQ (t) of aggregate where αGB
TAS
i + TT (t) is the arrival curve related to TT traffic
SP flows with priority i before the shared queue Qi is the sum given by Eq. (29). The non-preemption lower priority frame
of output arrival curves from all the previous shaped queues is maximized by
q connected to Qi , which is the same as the situation of ATS
used individually (Eq. 11), lL ≤ max ljmax = l>i
max
, (41)
i+1≤j ≤nSP ,
ATS + SP Qj ∈QBE
αQ i
(t) = r f · t + bf , (37)
q→Qi f ∈q and is not affected by ATS either.
where q → Qi represents all the shaped queues connected to However, since each SP flow before entering the shared
the shared queue Qi . queue Qj is reshaped to the committed transmission rate
Service Curve β ATS +SP
(t) - SP - Shared Queue: rf and the committed burst size bf , the upper bound of
Q i−1 ∗ ∗
j =1 (RQj (t) − RQj (s)) satisfies,
i
ATS + SP
Corollary 4: The service curve βQ i
(t) for SP traffic
with priority i (i ∈ [1, nSP ]) in the shared queue Qi under the i−1
i−1
∗ ∗
TAS+ATS+SP architecture is RQ j
(t) − RQ j
(s) ≤ R Q j
(t) − R Q j
(s)
⎡ i−1 ⎤+
αTAS αATS + SP
(t) max j =1 j =1
ATS + SP GB + TT (t) j =1 Q l>i
= C ⎣t − ⎦ ,
j
βQ (t) − −
i C C C
i−1
i−1
ATS + SP
↑
≤ (rf · t + bf ) = αQ j
(t). (42)
(38) j =1 q→Qj f ∈q j =1
ATS + SP
where αQ j
(t) is the arrival curve of aggregate SP flows This is different from the case under the TAS+SP architecture.
from Eq. (37), αGB TAS
+ TT (t) is the arrival curve related to TT Then, by inducing Eq. (40), Eq. (41) and Eq. (42) into Eq. (39),
max is the maximum frame size
traffic give by Eq. (29), and l>i we can obtain the service curve (Eq. (38)) for SP traffic in the
with priority lower than i. shared queue under the TAS+ATS+SP architecture.
Proof: Let RQi (t) and RQ ∗ (t) be respectively the input
i
Therefore, even under the architecture of TAS+ATS+SP,
and output cumulative function for SP traffic with priority i the service curve for SP traffic in the shared queue is similar
(i ∈ [1, nSP ]) before and after the shared queue Qi . Moreover, to the one (Eq. (30)) under the TAS+SP in Section IV-B1.
let RTT (t) (resp. RGB (t)) and RTT ∗ (t) (resp. R ∗ (t)) be the
GB The only difference is that, due to the ATS shaper, the rate and
arrival and departure processes of TT traffic (resp. occupation burst of SP traffic are reshaped and restricted before entering
of guard bands) due to GCL implementation by TAS. the shared queue.
It is assumed that t ∈ R+ is a time point when the By applying αQ ATS + SP
(t) and βQATS + SP
(t) into Eq. (62)
queue Qi is non-empty, i.e., RQ ∗ (t) < R (t). Let s be i i
i Qi and Eq. (63), we can determine the upper bound of latency
the start of the last busy period for the traffic with a pri-
ATS + SP ATS + SP ATS + SP
ority higher than or equal to the priority i of SP traffic, DQ i
= h α Q i
(t), β Q i
(t) , (43)
and backlog d ATS + SP

(t)+dqATS + SP (t) ≤ D ATS + SP
. Here, D ATS + SP
Qi− Qi− Qi−
is the latency bound of SP flows with priority i waiting in
ATS + SP ATS + SP ATS + SP
BQ = v αQ (t), βQ (t) (44)
i i i the preceding shared queue Qi− and can be calculated from
for SP flows of priority i passing through the shared queue Qi Eq. (43)). Since the lower bound of the delay in the shared
under the architecture TAS+ATS+SP. queue Qi− for all flows traversing through Qi− → q is lqmin /C ,
Input Arrival Curve α ATS q
+SP (t) - SP - Shaped Queue: the maximum latency DqATS of SP flows waiting in the shaped
Corollary 5: The input arrival curve αqATS + SP (t) of q is given by DqATS + SP = D ATS + SP
− lqmin /C . Then, we
Qi−
aggregate SP flows before entering the shaped queue q is, can conclude the corollary.
2) Performance Analysis–TAS+ATS+CBS:
∗
αqATS + SP (t) = αQ − (t) ∧ σ
link max
(t) + lQ − . (45) Input Arrival Curve α ATS +CBS
(t) - CBS - Shared
i i Qi
Queue: Correspondingly, by ATS reshaping, the input arrival
where the output arrival curve α∗ − (t) is from Eq. (14), ATS + CBS
Qi curve αQ (t) of aggregate AVB flows with priority i
Q− i
before the shared queue Qi is the sum of output arrival curves
by substituting delay upper bound D in δD i (t) with D =
D ATS + SP
= h(αATS + SP
(t), β ATS + SP
(t)) of SP traffic in satisfying rf · t + bf from all the previous shaped queues q,
− − −
Qi Qi Qi
the preceding shared queue Qi− , in which αATS + SP
(t) is αQ ATS + CBS
(t) = rf · t + bf . (47)
Qi− i
q→Qi f ∈q
determined by Eq. (37) and β ATS + SP
(t) is from Eq. (38).
Qi−
+CBS
Proof: In the TAS+ATS+SP architecture, compared with Service Curve β ATS
Q i [M] (t) - CBS - Shared Queue:
the ATS individually used, SP flows have additional wait- Corollary 7: The service curve for AVB traffic of Class
ing time for higher priority traffic, which causes the max- Mi (i ∈ [1, nCBS ]) in the shared queue Qi under the
imum latency in the preceding shared queue Qi− changed TAS+ATS+CBS architecture is same to the service curve for
to D ATS
−
+ SP
= h(αATS −
+ SP
(t), β ATS
−
+ SP
(t)), where AVB traffic under the TAS+CBS architecture, i.e.,
Qi Qi Qi
αATS + SP
(t) and β ATS + SP
(t) are respectively the input max +
Qi− Qi− α[TAS
M ] (t) ci[ M]
ATS + CBS
arrival (Eq. (37)) and service curves (Eq. (38)) for SP flows in βQ [ M ] (t) = idSli t − − , (48)
i C idSli
the shaped queue Qi− . Therefore, according to the basic NC ↑
∗Q −
theory, the output arrival curve αf i (t) of a SP flow f depar- where M ∈ {F, NF} representing the choice of the credit state
Q− Q− during GB (F — frozen credit during GB; NF — non-frozen
ture from the preceding shared queue Qi− is αf i (t)δD i (t), credit during GB), α[TAS
M ] (t) = {αTT (t), αTT + GB (t)} is the
TAS TAS
where D = D ATS + SP
. Then the aggregate arrival curve of max =
arrival curve related to TT traffic, given by Eq. (29); ci[
Qi− M]
all SP flows departure from Qi− to q can be represented by cimax (Eq. (18)) is the credit upper bound of AVB Class Mi
∗Q − if credit is considered frozen during GB, and ci[ max = c max
α∗ − (t) = f ∈Q − , αf i (t), where D = D ATS + SP
. M] i
Qi i Qi− (Eq. (33)) is the credit upper bound if credit is considered
−
Qi →q non-frozen during GB.
Moreover, before a shaped queue q, the phenomenon of ∗ (t) be respectively the input and
Proof: Let RQi (t) and RQ
serialization of flows from the same shared queues Qi− still i
output cumulative function for AVB traffic of Class Mi (i ∈
exists. And according to the ATS queuing schemes QAR1 and
[1, nCBS ]) before and after the shared queue Qi . Moreover,
QAR2, flows in the shaped queue q are all from the same ∗ (t) (resp. R ∗ (t)) be the
let RTT (t) (resp. RGB (t)) and RTT
preceding shared queue Qi− . These characteristics will not be GB
arrival and departure processes of TT traffic (resp. occupation
impacted by the integration of TAS. Then we can conclude
of guard bands) due to GCL implementation by TAS.
the corollary.
It is assumed that t ∈ R+ is a time point when the queue Qi
Service Curve β ATS
q
+SP (t) - SP - Shaped Queue:
∗ (t) < R (t). Then let s = sup{u ≤
is non-empty, i.e., RQ Qi
Corollary 6: The service curve βqATS + SP (t) for aggregate ∗
i
∗ (u) = R ∗
t RTT (u) = RTT (u), RGB GB (u), RQi (u) =
SP flows in the shaped queue q is,
RQi (u), ci (u) = 0}. Since AVB traffic in Qi obtained ser-
q
βqATS + SP (t) = δD (t), (46) vice during Δt − with the associated credit decreased, and
q
be blocked during Δt + and Δt 0 with the associated credit
where δD (t) is the pure-delay function with D = increased and frozen respectively. Then over the interval
DqATS + SP = D ATS + SP
− lqmin /C , and D ATS + SP
is from Δt = t − s, we have
Qi− Qi−
Eq. (43).
Proof: Since the shared queue for each SP priority in the ci (t) − ci (s) = Δt − Δt − − Δt 0 idSli + Δt − sdSli .
TAS+ATS+SP architecture is served in a FIFO manner, the
Then, since idSli − sdSli = C , and ci (s) = 0,
ATS shaper will not introduce extra overheads to the worst-

case delay of such a FIFO system [24]. ∗ ∗
RQ i
(t) − RQ i
(s) = t − Δt 0 idSli − ci (t). (49)
Thus, an SP flow fed to the shaped queue q on the sub-
sequent node will not increase the upper bound of the delay Since
for the flow waiting in the combined element of the shared
queue Qi− on the preceding node and the shaped queue q, i.e., Δt 0 ≤ α[TAS
M ] (t)/C , (50)
ATS + CBS
in which α[TAS
M ] (t) (Eq. (29)) is only related to the TT traffic curve αq[ M]
(t) of aggregate AVB flows before entering
and the credit behavior during GB, Δt 0 is not affected by ATS the shaped queue q is also related to the CBS shaping curve
reshaping. σ ATS + CBS
(t).
Qi− [ M ]
Moreover, ci (t) in Eq. (49) can be lower and upper As can be seen from the CBS shaping curve under the
bounded by, TAS+CBS architecture in Eq. (35), it is related to, on the
one hand, the service curve βTT TAS (t) (Eq. (36)) supplied to
cimin ≤ ci (t) ≤ ci[
max
M ], (51)
TT traffic, which is only related to the time slots reserved for
where cimin is given by Eq. (19), which is only related to the TT frames. On the other hand, the credit lower bound cimin
max = c max is given by Eq. (18) (Eq. (19)) and upper bound ci[ max = {c max , c max } for AVB
maximum frame size in Qi . ci[ M] i M] i i
for the case of credit frozen during GB, which is associated flows in Qi− , which have been discussed in Corollary 7, are not
with the credit lower bound cjmin of AVB traffic with the pri- related to the arrival pattern of AVB flows before the shared
ority Mj higher than Mi , and the maximum frame size l>i max
queue Qi− . Since ATS only has an impact on the flows’ arrival
of AVB traffic with the priority lower than Mi . ci[ M ] = c max
max
i pattern before shared queues Qi− , and shaped queue q, the
is given by Eq. (33) for the case of credit non-frozen during CBS shaping curve σ ATS −
+ CBS
(t) in the TAS+ATS+CBS
Qi [ M ]
GB, which besides the above-mentioned parameters in cimax ,
architecture is the same as the CBS shaping curve σ CBS (t)
is also related to the linear upper envelope for GB duration. Qi− [ M ]
As can be found, ci (t) ≤ ci[max is not related to flows arrival in the TAS+CBS architecture.
The discussion for the aggregate arrival curve α∗ − (t) of
M]
pattern of AVB flows before the share queue Qi . Qi
By the analysis of Eq. (50) and Eq. (51), it is known AVB flows departure from the shared queue Qi− to the shaped
that ATS reshaping on AVB traffic does not change the abil- queue q is similar to the one in Corollary 5.
+CBS
ity to serve AVB traffic of different priorities in the shared Service Curve β ATS
q [M]
(t) - CBS - Shaped Queue:
queue, compared with the service capability for AVB traffic ATS + CBS
Corollary 9: The service curve βq[ M]
(t) for aggre-
under the TAS+CBS architecture. Then, by inducing Eq. (50)
gate AVB flows in the shaped queue q is
and Eq. (51) into Eq. (49), we can obtain the service curve
(Eq. (48)) for AVB traffic in the shared queue under the ATS + CBS q
βq[ M]
(t) = δD (t), (55)
TAS+ATS+CBS architecture, the same as the one (Eq. (32)
q
in Appendix A-B2) under the TAS+AVB architecture. where δD (t) is the pure-delay function with D =
ATS + CBS ATS + CBS
By applying αQ (t) and βQ (t) into ATS + CBS
Dq[ M ] = D ATS + CBS
− lqmin /C , and D ATS + CBS
is
i[ M ]
Qi− [ M ] Qi− [ M ]
i
Eq. (62) and Eq. (63), we can determine the upper bound
from Eq. (52).
of latency
Proof: The discussion is similar to Corollary 6, by consid-
ATS + CBS
DQ ATS + CBS
= h(αQ ATS + CBS
(t), βQ (t)), (52) ering the latency upper bound D ATS−
+ CBS
in Eq. (52) instead
[M]
i i [M] i Qi [ M ]
and backlog of D ATS
−
+ SP
in Eq. (43).
Qi
ATS + CBS ATS + CBS ATS + CBS
BQ = v (αQ (t), βQ (t)) (53)
i[ M ] i i[ M ]
V. P ERFORMANCE C OMPARISON E VALUATION
for SP flows of priority Mi passing through the shared queue
In this section, in order to compare the performance evalu-
Qi under TAS+ATS+CBS.
+CBS ation of individual traffic shapers and their combinations, we
Input Arrival Curve α ATS q [M]
(t) - CBS - Shaped
use a large set of synthetic test cases 3 with different topolo-
Queue:
ATS + CBS gies and a realistic test case, i.e., the Orion Crew Exploration
Corollary 8: The input arrival curve αq[ M]
(t) of
Vehicle (CEV) from NASA [49].
aggregate AVB flows before the shaped queue q is,

ATS + CBS ∗ link max
αq[ M]
(t) = α Q − (t) ∧ σ (t) + l −
Qi
A. Individual Traffic Shapers
i 1) Comparison of NC and Non-NC Approaches for ATS
ATS + CBS max
∧ σ − (t) + lQ − . (54) Evaluation: Before discussing the various traffic shapers, we
Qi [ M ] i
first compare the two different methods used for ATS eval-
where σ ATS + CBS
(t) is same to Eq. (35) of the case with- uation, i.e., the Network Calculus (NC) used in this article
Qi− [ M ]
out ATS, and the output arrival curve α∗ − (t) is from and a non-NC approach proposed in [20]. By comparing the
Qi upper bound of the delay obtained by the two methods, the
Q−
Eq. (14), by replacing the delay bound D in δD i (t) with latency bound for a flow f in an egress port calculated by NC
D = D ATS −
+ CBS
= h(αATS−
+ CBS
(t), β ATS
−
+ CBS
(t)) of is ΔDfQ more pessimistic than the result calculated by the
Qi [ M ] Qi Qi [ M ]
non-NC approach [20],
AVB traffic in the preceding shared queue Qi− , in which

αATS + CBS
(t) and β ATS + CBS
(t) are respectively from lf lf
Qi− Qi− [ M ] Q
ΔDf = max − , (56)
Eq. (47) and Eq. (48). ∀f ∈Q(f ) C − f ∈QH rf C
Proof: In the TAS+ATS+CBS architecture, in addition to
the shaping curve σ link (t) of the physical link due to seri- 3 Details of flows, routes and GCLs for all the test cases can be downloaded
alization of all AVB flows from Qi− to q, the input arrival from https://zenodo.org/record/6378112#.YjqQReeZNPY.
TABLE III
S TATISTICAL H OPS AND T RAFFIC L OAD FOR 100 T EST C ASES
1 and 2, the evaluation results calculated by NC are 0.6% and

1.4% slightly more pessimistic on average than the results by
the non-NC approach, respectively.
However, the non-NC approach proposed by [20] is focused
on ATS in isolation, and thus it is not applicable and cannot
be extended to combinations of traffic shapers. Hence, in this
Fig. 11. Network topologies of synthetic test cases. paper, we consider the Network Calculus approach for eval-
uating various traffic shapers and their combinations. All the
following evaluation results are based on the Network Calculus
approach introduced in this article.
2) Performance Comparison Among TAS, ATS, SP and
CBS: In the first set of experiments, we are interested in
comparing the performance from the perspective of the upper
bounds of end-to-end latency, jitter and backlog without
the frame loss for each individual traffic shapers (including
TAS, ATS, CBS and SP) under different network topolo-
gies. The network topologies are respectively small ring &
mesh (SRM), medium ring (MR), medium mesh (MM), small
tree-depth 1 (ST) and medium tree-depth 2 (MT) which are
inspired by industrial application requirements [45], as shown
in Fig. 11. There are 100 test cases (TCs) randomly gener-
ated. For each test case, there are 15 flows. The frame size
Fig. 12. Comparison of NC and non-NC methods for ATS evaluation.
lf of each flow is randomly chosen between the minimum
(64 bytes) and the maximum (1,522 bytes) Ethernet frame
size, and flows can be periodic or sporadic.4 For the peri-
where lf is the frame size of a flow with the same priority odic flow, the periods are uniformly selected from the set
level as the flow f of interest, lf is the frame size of flows Tf = {1 000, 2 000, 5 000, 10 000} μs. For the sporadic flows,
with a higher priority than f, and rf is the committed burst it is assumed that each flow satisfies the leaky bucket model
size of the flow f supported by ATS. with the burst bf = lf and rate rf = lf /Tf , where Tf is
For the evaluation of the two approaches, we use a synthetic the minimum interval between two consecutive frames. Since
test case where the topology is a medium mesh (Fig. 11), TT flows manipulated by the TAS have no priority division,
including 45 flows with 3 priorities. ATS is applied to all the it is assumed that all flows are assigned to the same priority
priorities. The average traffic load is around 70%, and the level for each use case in this experiment. The GCLs for TAS
physical link rate is set to 100 Mb/s. We show a comparison are generated according to [12]. All the test cases are applied
of the two methods in Fig. 12, where the value on the x-axis to the above five topologies, respectively. The routes of flows
represents the identifiers of each flow, and the y-axis shows are generated according to the routing optimization strategy
the upper bound of end-to-end latency in microseconds. The proposed for TT traffic [12]. The idle slope for AVB traffic is
obtained results are grouped by priority, denoted by vertical set to the default value of 75%. For each test case, we con-
dotted lines, and sorted in increasing order by results within sidered the average hops of flow and the average traffic load
each priority. As we can see from Fig. 12, the performance under each topology. Table III gives the statistics over the 100
evaluation by the NC analysis of ATS is very close to the anal- test cases under each topology. The physical link rate is set to
ysis from [20], which is as expected according to Eq. (56). C = 100 Mb/s.
We note that with a decrease in priority, the gap between For each test case under a given topology, we evaluate the
the two will increase slightly. This is because the denomi- quality of service for different individual traffic shapers, i.e.,
nator of the first term of Eq. (56) is related to the sum of TAS, ATS, CBS and SP, respectively. By applying the NC
the rates of all high-priority flows. The lower the priority of approach, we can get the two evaluation metrics of each flow
the flow of interest, the greater the rate accumulated by the under different individual traffic shapers, i.e., the upper bound
high-priority flow. Thus, the first term of Eq. (56) is becoming
larger. For example, for the highest priority flows, the results 4 Flows served by the TAS are periodic, and ATS and AVB support both
from the two approaches are the same. For the lower priority periodic and sporadic flows.
TABLE IV
D IFFERENCE R ATIO ON M ETRICS OF T WO I NDIVIDUAL T RAFFIC S HAPERS
the upper bounds of worst-case backlog (WCB) in bytes and

jitter (WCJ) in microseconds, respectively.
As can be seen in the figure, TAS performs best with the
lowest latency and backlog, and provides zero-jitter. Such
performance is in line with expectations. Since TAS realizes a
completely deterministic time-triggered transmission through
flow-based scheduling, it avoids the collision of frames of
its own traffic type and avoids the collision with frames of
other traffic types as well. Thus, flows shaped by TAS can
achieve ultra-low latency, backlog and jitter. For the other three
traffic shapers, i.e., ATS, SP and CBS, the overall trend of
their performance comparison can be inferred from the figure.
However, it is difficult to see their comparison on an individ-
ual test case from the figure. Thus, we additionally use the
difference ratio

Xi = XiY1 − XiY2 /XiY2 (57)
to capture the performance comparison of traffic shapers Y1
and Y2 (Y ∈ {ATS, CBS, SP}), where Xi can represent the
end-to-end latency bound or jitter bound of the flow fi or the
upper bound of the backlog for the queue Qi at the egress
Fig. 13. Comparison of different individual traffic shapers under different port. Then, we take the average value X = average(Xi ) as
topologies. the comparison result of two traffic shapers under the current
use case. The comparison results are shown in Table IV.
Regarding the delay and jitter upper bound, SP performs
better than CBS, as shown in Fig. 13(a) and (c), and the
of end-to-end delay and jitter; and one evaluation metric for “Average WCD” and “Average WCJ” of the third column in
each egress port, i.e., the upper bound of the backlog without Table IV. This is because CBS has a bandwidth reservation
the frame loss. Fig. 13 presents the performance evaluation mechanism for traffic. In our case, only 75% of the bandwidth
of the network from the perspective of the upper bounds of is available for AVB traffic. Therefore, compared to SP, the
end-to-end delay, backlog and jitter of different individual traf- service bandwidth obtained for flows shaped by CBS is lower.
fic shapers under different topologies. For each test case, we Concerning the backlog bounds, CBS may perform better than
use the average value of the corresponding evaluation metric SP, as presented in Fig. 13(b) and the “Average WCB” of the
of all flows to represent the metric value under the current third column in Table IV. This is because although the waiting
test case. Therefore, for each individual traffic shaper under time of the flow in the corresponding priority queue has been
each topology, we obtain 100 values of the corresponding met- prolonged by CBS through controlling the credit, it neverthe-
ric, and we use box plots to present these results. Fig. 13(a) less reduces the long-term rate of arrival of flows. Thus, the
shows the evaluation of the end-to-end latency bounds. The backlog upper bounds of AVB traffic are probably lower than
x-axis represents different topologies, and the y-axis shows for SP traffic.
the upper bound of the end-to-end latency (WCD) in microsec- What is more interesting is that, for all the current use cases,
onds. Fig. 13(b) and Fig. 13(c) show the evaluation results on the performance of ATS is not better than that of SP and
CBS, a finding that has not been reported in the related work. experiments, we chose the same topology (MM) to study the
Due to the reshaping function of ATS, which avoids burstiness comparative performance between ATS and SP mechanisms
cascades, we initially hypothesized that ATS would improve when the average traffic load is increasing from 10% to 90%.
the performance of flows. However, as it can be observed in We used 20 random synthetic test cases under each traffic load.
Fig. 13(a), (b), (c) and the fourth and fifth columns in Table IV, The frame sizes and intervals are selected in the same way as
in terms of upper bounds of delay, backlog or jitter, ATS did for the use cases in the previous section. As the comparison
not perform superiorly. trend of individual shapers on the upper bounds of latency and
In the following, we will only focus on the comparison jitter is very similar, we will only show the comparison of the
of ATS and SP, as ATS compared with CBS can be inferred upper bounds of delay and backlog in the following.
from the comparison of SP and CBS discussed above. From We still use the difference ratio from Eq. (57) to represent
the current results, we can draw a conclusion and a hypothe- the performance comparison result of ATS and SP under the
sis. The conclusion is that the advantages of ATS are getting current use case, where Xi can represent the upper bound of
worse with the increase in the concentration of flows transmis- the end-to-end latency of flow fi or the upper bound of the
sion and the number of hops (under the same concentration backlog for the queue Qi at the egress port.
of flows transmission). This is because the more concentrated In order to test the performance of ATS in isolation, we first
the flows are in the network, the more obvious is the serializa- assume that all flows have the same priority, and that there is
tion of flows in transmission, which increases the determinism no interference from other traffic types. This situation can also
of traffic transmission at subsequent nodes. Moreover, in the be similar to the network adopting the preemption integration
case of the constant concentration of flows transmission, as mode. The comparison results of ATS versus SP are shown
the number of hops increases, serialization leads to the same in Fig. 14(a) and (b). As it can be seen, for the upper bound
determinism of traffic transmission, but the time cost of ATS of end-to-end delay (Fig. 14(a)), with the increase of average
reshaping increases accordingly. The hypothesis is that as the traffic load, the performance relationship between ATS and SP
load increases, the advantages of ATS will also increase, which changes. When the average traffic load is lower than 70%, SP
will be discussed in the next section. The reasoning behind this performs better than ATS. ATS shows its superiority only when
hypothesis is that the time cost of ATS reshaping traffic can- the average traffic load increases by more than 70%. For the
not offset the queuing time of traffic with the burst cascade upper bound of the backlog (Fig. 14(b)), the difference ratio
without ATS. Thus, ATS has a negative effect when the traffic between ATS and SP does not change significantly, and when
load has not reached a certain level. Therefore, it is reasonable considering ATS used individually, the backlog performance is
to infer that when the traffic load increases, the impact of ATS always worse than that of SP. This is due to the fact that ATS
will gradually turn positive. is a kind of non-work conserving scheduler that implements
Next, we analyze the results on column 4 in Table IV. Since reshaping through frame waiting. Moreover, as noted, shaped
each test case is applied under five topologies, for some use queues and the shared queue are implemented in a single
cases, the five topologies have the same flows, but just with physical queue, and the backlog for ATS are the sum of back-
different routes. Moreover, the traffic load in Table III reflects logs in shaped and shared queues. It is furthermore assumed
the degree of dispersion of flows transmission in the network. that all flows still have the same priority, but there exists the
For example, the traffic load under the MM topology is the interference of BE traffic with the maximum Ethernet frame
lowest. This is because the MM has the largest number of size of 1,522 bytes. If the preemption integration mode is con-
selectable paths, and thus the flows are more dispersed than sidered, the compared results will be similar to the discussion
the traffic in other topologies. Hence, although the traffic load above (Fig. 14(a), (b)). If the non-preemption integration mode
in MM is lower than that in SRM, the performance comparison is taken into account (as considered in this paper), there will
between ATS and SP is not much different from that in SRM. be at most one BE frame interference when the flow of interest
Compared with MR, the traffic load in MM is close to that of obtains the service. The results with and without ATS reshap-
MR, but the transmission of traffic in MR is more concentrated ing are shown in Fig. 14(c) and (d). As can be observed from
than in MM. Therefore, the advantage of ATS under the MM the figure, low-priority non-preemptible frames have no sig-
topology is higher than under the MR topology. For the ST nificant impact on the performance with or without ATS from
topology, although its traffic load is close to the traffic load in the perspective of upper bounds of end-to-end delays. Latency
SRM, its number of hops is higher, which leads to more times performance comparison results still mainly depend on the
of ATS reshaping along the flow’s route. Thus, compared with traffic load. However, the performance with ATS and without
SRM, the performance of ATS under the ST topology is far ATS from the perspective of backlog upper bounds is signif-
worse than SP. A similar explanation can be extended to the icantly changed, even though ATS still performs worse than
results for the MT topology. without ATS. This implies that the non-preemptible frames
3) Comparison Between ATS and SP Under Changing from low priority traffic have a larger impact on the backlog
Traffic Load: In the previous section, the traffic load of all performance of egress ports without ATS reshaping compared
test cases is in the range from 6% to 47%, and in all these to ports with ATS reshaping. From here, we can also see that
test cases, ATS does not show its advantages in real-time the backlog bounds comparison between individual ATS and
performance compared with TAS, CBS and SP. In order to SP (one priority) is more related to the frame length.
rule out the influence of traffic dispersion degree and num- Moreover, we assign high and low priorities to the traf-
ber of hops under different topologies, for the second set of fic, and the high and low priority traffic each account for
Fig. 14. Comparison of ATS and SP under different traffic load.
50% of the overall traffic load. The comparison results of priority and are interfered by a BE frame. Moreover, the max-
ATS vs SP in terms of upper bound of the delays and back- imum average load of the high-priority traffic is only 45%.
logs for the high and low priority traffic, with and without Nevertheless, for low priority, ATS shows its performance
reshaping by ATS, are shown in Fig. 14 (e) and (f), respec- advantage from the perspective of latency bounds when the
tively, using light blue and dark blue box plots. It can be seen overall average traffic load reaches 30% while low priority
from the figure that for the high priority traffic, the results average traffic load reaches 15%. From the perspective of
are similar to the top 40% compared results in Fig. 14 (c) backlog, no matter for the high-priority or the low-priority
and (d), and similarly ATS does not show its superiority for traffic, the backlogs of ATS are still worse than that with-
high-priority traffic. This is because the transmission of the out ATS when the overall average traffic load is increasing
high priority traffic will only be interfered by at most one from 10% to 80% (the average traffic load for low priority is
low priority frame whose frame length ranges from the mini- from 5% to 40% accordingly). The ATS backlogs of a few
mum (64 bytes) to the maximum (1,522 bytes) Ethernet frame cases shows its superiority when the average load is as high
size, which is same as the case if all flows have the same as 90%.
Therefore, when the ATS is used individually, its positive

effect on latency upper bounds can be highlighted only when
the average traffic load in the network reaches a certain high
level, around 70% to 80%, as in Fig. 14(a) and (c). Moreover,
it always shows a negative impact on the upper limit of the
backlog, as in Fig. 14(b) and (d). When ATS is used for traffic
of multiple priorities, ATS shows a positive impact on latency
upper bounds for low-priority traffic even if the average low-
priority traffic load is not high. However, its impact on the
high priority traffic is still negative, see Fig. 14(e). However,
ATS still shows a negative impact on the backlog bound on
both high and low priority, unless the average traffic load for
high priority is reached at least 90%, as shown in Fig. 14(f).
Based on the above findings, this is why we believe that the
combined use of ATS with TAS will make ATS play a more
active role. At the same time, TAS will perfectly maintain its
advantages of ultra-low latency and jitter.
B. Combined Traffic Shapers

1) Evaluation on Synthetic Test Cases: In this section,
we are interested in the influence of ATS on the real-time
performance of combined traffic shapers.
We first show the evaluation performance of the combi-
nation use of ATS and CBS in different queues, as per the
architecture discussed in Section IV-A1. We will use the
synthetic test cases adapted from Section V-A3 under MM
topology. The average traffic load is increasing from 10% to
60%, including five test cases per load case. For each test case,
flows are divided into four sets, two of which are configured
as high- and low-priority SP traffic served by ATS, and two of
which are high- and low-priority AVB traffic served by CBS.
The idle slopes for each AVB Class are respectively 45% and
30%. We discuss the mutual influence of ATS and CBS under
the two architectures ATS+CBS (ATS has higher priority than
CBS, named “ATS(H)+CBS(L)”) and CBS+ATS (CBS has
higher priority than ATS, named “CBS(H)+ATS(L)”). It is
difficult to say which architecture is better, but we compare
and explain some phenomena using the following Fig. 15.
Fig. 15(a) and (b) respectively present the average worst-
case end-to-end delays of flows scheduled via ATS and CBS,
when ATS and CBS are respectively treated as high priority.
The results for ATS are depicted with yellow triangles and
for CBS are with blue circles, and identifiers on the x-axis
are divided by traffic load. As can be seen from the figure,
for these sets of test cases, no matter when ATS is used for
high- or low-priority, its delay is always better than the delay
of traffic scheduled by CBS when CBS is at the same priority.
This is different from the results of ATS and CBS individ-
ually used under MM topology in Fig. 13. It is due that all Fig. 15. Comparison of ATS(H)+CBS(L) and CBS(H)+ATS(L) under
flows in Fig. 13 are assumed to have the same priority. So the different traffic load.
bandwidth utilization (75%) of CBS is all used for AVB flows
of one priority, which means that the speed of credit recovery
is fast. Moreover, the average traffic load in the Fig. 13 is service time interval that can be assigned to each AVB class
relatively low. Thus, the cost of ATS shaping approximately lengthened. This is because the speed of credit recovery slows
offsets the delay drag caused by the non-full bandwidth usage down. Another interesting finding is that the average worst-
of CBS. However, in this experiment, the bandwidth utiliza- case end-to-end delays of all the flows (including both AVB
tion (75%) is assigned to two AVB classes, which makes the flows and SP flows shaped by ATS) under the architecture
Fig. 16. Comparison of TAS+ATS+SP and TAS+SP under different traffic load.
CBS(H)+ATS(L) may be better than under the architecture by ATS have the same priority and with no interference from
ATS(H)+CBS(L), especially when the average traffic load is other traffic types of lower priority (BE for example).
high. The compared results are shown in Fig. 15(c), in which As can be seen from Fig. 16(a) and (b), similar to ATS
for ATS(H)+CBS(L) are depicted with blue circles and for used individually, with the increase of average traffic load, the
CBS(H)+ATS(L) are with yellow triangles. The reason is that performance of traffic shaped by ATS is gradually improved.
the variation in worst-case end-to-end latencies of ATS from However, with the influence of TT traffic implemented by
high priority to low priority is not as large as that of CBS. TAS, it is obvious that the optimization effect on latency
From the previous experimental results, we conclude that, on bounds of ATS reshaping is better than that of the individ-
the one hand, when the average traffic load is low, the positive ual ATS traffic shaper. For the end-to-end latency bounds,
effect of ATS is not prominent, but only when the traffic load the ATS traffic shaper used individually performs better only
reaches a certain level, its advantages will become more and when the average traffic load is above 70%. However, in the
more obvious (Fig. 14(a), (c)). On the other hand, ATS has a combined traffic shaper TAS+ATS+SP, as long as the aver-
greater positive effect on low-priority traffic (Fig. 14(e)). age traffic load shaped by ATS reaches 20% (overall average
In the leftover section, we will see the impact of ATS on load is 40%), the ATS traffic shaper is superior. Nevertheless,
the combined traffic shapers (TAS+SP or TAS+CBS) under for the backlog bounds, ATS does not show its superiority
different levels of traffic load, as discussed for individual traf- in the combined architecture of TAS+ATS+SP either, due to
fic shapers. The architectures of TAS+ATS+SP and TAS+SP the usage of shaped queues. The performance advantage of
without ATS are first compared. We still use the same topology ATS in combination with TAS is similar to the performance
(MM) as in the Section V-A3. In the first set of experiments, advantage of reshaping for low priority traffic when ATS is
it is assumed that there is 20% of TT traffic achieving deter- used individually. This is because TT traffic based on TAS
ministic transmission based on the TAS, and 10% to 70% of has fixed transmission time slots and has the highest priority,
traffic is SP. Therefore, the overall average traffic load on the and other traffic types cannot use the time slots allocated to
network is 30%-90%. For each traffic load, we have used 20 TT traffic. The existence of TT traffic will greatly increase the
randomly generated test cases. In order to fairly compare with possibility of mutual interference and backlog of other types
ATS used individually, we still assume that all flows reshaped of traffic. Then, the burst cascade of the flow on its route will
Fig. 17. Comparison of TAS+ATS+CBS and TAS+CBS under different traffic load.
be increased. Therefore, the time used by the ATS reshaping

is lower than the waiting time caused by the bursty traffic.
Furthermore, we increase the average load of TT traffic to
30%, while there is 10% - 60% of SP traffic, so the overall
average traffic load in the network is 40% to 90%. Similarly,
for each traffic load, there are 20 test cases generated ran-
domly. Compared results of TAS+ATS+SP and TAS+SP are
shown in Fig. 16(c) and (d), respectively. By comparing with
20% of the TT traffic load, it is found that with the increasing
TT traffic load, the positive impact of ATS is not increasing
significantly.
Fig. 18. Topology with flows that have routes with cyclic dependencies.
Then, we will compare the performance of
TAS+ATS+CBS and TAS+CBS without ATS. The TT
traffic load is still 20%. In addition, we consider the average
load of AVB traffic is from 10% to 40%, as only 75% of Analysis, cannot be used directly for the analysis of cyclic
bandwidth is reserved for AVB traffic, which also includes the dependencies [46], [47]. In the following, we will respectively
bandwidth occupied by TT traffic. In addition, it is also nec- test TAS+SP, TAS+ATS+SP, TAS+CBS, TAS+ATS+CBS
essary to ensure that the maximum link load on the network on a MM topology with routes that have cyclic dependencies,
does not exceed 75%. Similarly, we have used 20 randomly which are shown in Fig. 18. As can be seen in the fig-
generated test cases for each traffic load. The comparison ure, R1=[[ES1,SW1], [SW1,SW2], [SW2,SW3], [SW3,ES8]],
results are shown in Fig. 17. Concerning the upper bounds of R2=[[ES5,SW2], [SW2,SW3], [SW3,SW1], [SW1,ES3]] and
end-to-end latency, ATS also shows its superiority as long as R3=[[ES7,SW3], [SW3,SW1], [SW1,SW2], [SW2,ES6]] are
the average load of AVB traffic is above 20%, and concerning mutually cyclically dependent. The end-to-end latency bound
the backlog bounds, ATS performs worse than the backlog (WCD) for each flow is shown in Table V. As can be seen
performance of the architecture without ATS, which is similar from the results, for the architectures TAS+SP and TAS+CBS
to the case with TAS+ATS+SP, as shown in Fig. 16. But without ATS reshaping the flows in middle nodes, it is not
the difference is that with the AVB traffic load increasing, possible to calculate WCDs using the classical NC approach.
although the optimization of the delay performance with the However, with the ATS reshaping function in the architectures
ATS in the architecture TAS+ATS+CBS is also increasing, TAS+ATS+SP and TAS+ATS+CBS, it is easy to break the
it is not as obvious as that in TAS+ATS+SP. For the upper cyclic dependency, and the WCDs for flows can be calcu-
bounds of backlog, the optimization effect of ATS does not lated using the classical NC method. It has been shown that
increase with the traffic load. This may be because CBS is a cyclic dependencies can be eliminated by reshaping flows in
bandwidth reservation service. CBS itself implements a fairer the network [48]. In conclusion from the above experiment,
service by controlling credit, thereby reducing the long-term we can find that although from the perspective of worst-case
rate of arrival of flows. Therefore, the optimization effect of latency and backlog, ATS does not always perform better than
ATS itself in the TAS+ATS+CBS architecture is weakened. the other traffic shapers, ATS can always handle the cyclic
All the synthetic test cases above are without cyclic depen- dependency of flows. We can say that no matter for CBS,
dencies. Next, we are interested in seeing the ability of ATS SP etc., the combination use of ATS before them will break
to remove the cyclic-dependency of flows in a topology. It the cyclic dependency and make the traditional network cal-
is known that the classical NC approach, e.g., Total Flow culus analysis method feasible. Therefore, if there are cyclic
Fig. 19. Comparison of combined traffic shapers under the Orion CEV TCOrion 1 (TT - 1.5%, SP/AVB with 1 priority).
TABLE V
P ERFORMANCE A NALYSIS OF D IFFERENT T RAFFIC S HAPERS
ON C YCLIC D EPENDENCY T OPOLOGY
Fig. 20. Network topology of Orion CEV.
dependencies between flows in the network, it would be sug- Exploration Vehicle (CEV) from NASA [49]. The test case
gested to use the ATS on node ports along the path of the topology is shown in Fig. 20. The Orion CEV case has 31
cyclic-dependency flow to break the cycles to guarantee the ESes, 15 SWs, 188 dataflow routes connected by a physical
deterministic transmission for the flows. link transmitting at 100 Mbps. In the last set of experi-
2) Evaluation on the Realistic Test Case: In the last ments, we are interested to see the effect of ATS on the
experiment, we use the realistic case of the Orion Crew three novel hybrid architectures of combined traffic shapers
Fig. 21. Comparison of combined traffic shapers under the Orion CEV TCOrion 2 (TT - 15%, SP/AVB with 4 priorities).
TAS+ATS+SP, TAS+ATS+CBS and ATS+CBS. We have is relatively low. The results for Orion CEV conform to the
run the NC-based performance analysis method for both com- outcomes shown in Fig. 16(a), (b) and Fig. 17.
binations TAS+SP and TAS+ATS+SP (resp. TAS+CBS and Then we increase the traffic load in Orion CEV by rais-
TAS+ATS+CBS, ATS(H)+CBS(L) and CBS(H)+ATS(L)) on ing the rate and keeping the frame size of the flow (called
the Orion CEV case, and obtained for each combined traf- TCOrion 2 ). The average network load (resp. maximum link
fic shaper the upper bounds of the maximum latency (WCD) load) for TT traffic is increased to 15% (resp. 54%). The over-
for each flow and the upper bounds of the maximum backlog all traffic load on the network is 25% on average and 69% on
(WCB) for each priority queue in an egress port. maximum. Moreover, we are interested in taking a look at
We first use the original traffic parameters [50] of the ATS’s effect on multiple priorities, and thus we classify the
Orion CEV (TCOrion 1 ), including 99 TT flows and 87 Rate SP/AVB traffic into four priorities. There are 25 flows of pri-
Constraint (RC) flows of the same priority. RC flows are con- ority P1 , 25 of priority P2 , 25 of priority P3 and 12 of priority
sidered as SP and AVB flows under respective combinations P4 . For the AVB traffic, due to the increased traffic load and
of traffic shapers in this paper. The idle slope for AVB traf- the uneven load for each traffic type on each link, it is dif-
fic is set to 75%. The average network load (resp. maximum ficult to assign a fixed idleSlope for each AVB traffic class
link load) for TT traffic is around 1.5% (resp. 5.5%), and the across the entire network. Thus, we calculate the idle slope
overall traffic load on the network is 3.5% on average and of AVB Class Mi for each egress port according to the actual
10% on maximum. The results are shown in Fig. 19, where bandwidth utilization [1, Sec. 8.6.8.2], i.e.,
the upper bounds are normalized to 100 × ln(X ) with X = OperCycleTime
{WCD, WCB}. The obtained results are sorted in increasing idSli = operIdleSlope(Mi ) · ,
GateOpenTime
order by results. As can be seen from the figure, ATS does
not show a positive impact on both the upper bounds of end- where operIdleSlope(Mi ) [1, Sec. 34.3] is the actual band-
to-end latencies of the flows and the backlog upper bounds of width that is currently reserved for the AVB class Mi for each
egress ports under the architecture TAS+ATS+SP/AVB. This port, and OperCycleTime
GateOpenTime is the fraction of effective time that
is because the average traffic load for both TT and SP/AVB the gate is open for AVB traffic. Similarly, the results are
VI. C ONCLUSION
This paper has studied the qualitative performance compar-
ison of the various individual traffic shapers and their possible
combinations. The analysis is based on the Network Calculus
(NC) approach. The paper summarized the existing NC-based
analysis for ATS, CBS, SP individually used, and TAS+SP,
TAS+CBS used in combination, and extended the NC-based
analysis for some other traffic shapers used in combination
which have not been discussed, including ATS and CBS used
in combination for different priority queues, the novel archi-
tectures TAS+ATS+SP and TAS+ATS+CBS when ATS and
CBS are used for the same queue. From plenty of test cases,
we can conclude that SP and CBS have advantages in that SP is
more beneficial to the transmission delay of high-priority traf-
Fig. 22. Comparison of ATS(H)+CBS(L) and CBS(H)+ATS(L) under the
Orion CEV TCOrion 3 . fic. While CBS can specify bandwidth reservations for each
priority of traffic. In addition, due to the credit controlled
by CBS, the long-term rate of traffic arrival is reduced, so
the backlog upper bounds of AVB traffic are probably lower
shown in Fig. 21. The obtained results are grouped by differ- than SP traffic. Compared with SP and CBS, ATS has limited
ent priorities with vertical dotted lines and, respectively, sorted advantages for high-priority traffic. Only when the average
in increasing order by results within each priority. From the traffic load of high-priority traffic in the network reaches
figure, we can find that with the increasing traffic load, the around 80%, ATS shows its superiority. The positive effect of
combination shaper TAS+ATS+SP (resp. TAS+ATS+CBS) ATS on low-priority traffic is more obvious. When the average
outperforms TAS+SP (resp. TAS+CBS) from the latency traffic load of low-priority traffic on the entire network reaches
upper bounds. For the backlog bounds, TAS+ATS+SP does about 20% (overall load 40%), the positive effect of ATS
not perform better than TAS+SP as conforming to results in on the upper bound of delay performance begins to become
Fig. 16. However, more than half of the backlog bounds under prominent.
the TAS+ATS+CBS are better than under the TAS+CBS. For the combined traffic shapers, we first investigate the
This is due to the idle slope of CBS, which is related to performance of ATS and CBS used for different queues.
the serviceability and is set according to the actual band- Whether to use ATS for high priority or CBS for high
width of AVB traffic. The relative load of AVB is very priority is not a definite conclusion, as each has its advan-
large. What is more interesting is that, with the combina- tages. Compared with all the above traffic shapers, TT traffic
tion of ATS, the performance of SP and CBS gets close to implemented with flow-based scheduling by TAS has the high-
each other, but CBS allows the bandwidth reservation for the est performance, with ultra-low latency, jitter and backlog.
traffic. However, it is well known that TAS requires the synthesis
In the last experiment, we are interested in testing combina- of optimized GCLs, which do not scale to large networks
tions ATS(H)+CBS(L) and CBS(H)+ATS(L) using the case with many flows. This problem can be mitigated by combin-
adapted from Orion CEV, that has 25 flows of priority P1 (the ing different traffic shapers in the same switch architecture, to
highest priority), 25 of priority P2 , 25 of priority P3 and 12 reduce the number of flows handled by TAS. Moreover, the
of priority P4 (called TCOrion 3 ). The average network load combined use of ATS with TAS will make ATS play a more
is 10%. One case is ATS used for flows with priority P0 and active role, of which the effect is similar to the reshaping
P1 , and CBS used for flows with priority P1 and P2 . In the impact of ATS used individually on low priority traffic, and at
other case, CBS(H)+ATS(L) is the other way round. The idle the same time, TAS will maintain unchanged its advantages
slopes for each AVB Class are, respectively, 45% and 30%. of ultra-low latency and jitter. Additionally, we have shown
The results are shown in Fig. 22, where the x-axis refers to that even though, from the perspective of worst-case latency
the flow ID, and the y-axis presents the WCD in microsec- and backlog, ATS does not always perform better than the
onds. It can be found that when both ATS and CBS are for other traffic shapers, adding the ATS model before other traf-
high-priority or low-priority traffic, most ATS results are better fic shapers (for example, SP and CBS) can always break the
than CBS. The average WCDs for ATS and CBS classes are cyclic dependency of flows and make the traditional network
given by horizontal lines. As can be seen from the figure, the calculus analysis method feasible.
average WCD of ATS is smaller than that of CBS under both
the CBS(H)+ATS(L) (purple solid line) and ATS(H)+CBS(L)
(green dotted line) architectures, which is conformed to the A PPENDIX A
results shown in Fig. 15(a), (b). Finally, the overall flows’ aver- N ETWORK C ALCULUS T HEORY
age WCDs under the ATS(H)+CBS(L) and CBS(H)+ATS(L) Network Calculus [42], [43] is a system theory proposed
are respectively 3091.7μs and 2899.4μs. This is due to the for analyzing performance guarantees in communication
relatively low overall traffic load of 10%, as can also be seen networks. By constructing arrival curve and service curve
in Fig. 15(c). models, the maximum amount of flow data entered into
network nodes and the minimum service offered by network the maximum horizontal deviation between the graphs of two
nodes can be obtained. Network Calculus is build on min-plus curves α(t) and β(t),
algebra, which includes two basic operators on non-decreasing
D = h(α, β) = sup{inf{τ ≥ 0 | α(s) ≤ β(s + τ )}}. (62)
functions: F↑ = {f : R+ → R|x1 < x2 ⇒ f (x1 ) ≤ f (x2 )}. s≥0
One is the convolution operator ⊗,
The backlog of the flow in the network node is bounded by
(f ⊗ g)(t) = inf {f (t − s) + g(s)}, the maximum vertical deviation between the graphs of two
0≤s≤t
curves α(t) and β(t),
and the other is the deconvolution operator ,
B = v(α, β) = sup{α(s) − β(s)}. (63)
(f g)(t) = sup{f (t + s) − g(s)}, s≥0
The output arrival curve for the output flow R ∗ (t) is bounded
s≥0
where inf means infimum and sup means supremum. by the arrival curve α∗ (t),
The arrival and service curves are defined by means of
the min-plus convolution. An arrival curve α(t) is a model α∗ (t) = α β(t) = sup{α(t + s) − β(s)}, (64)
s≥0
constraining the arrival process R(t) of a flow, where R(t) rep-
resents the input cumulative function counting the total data With the effect of packetization, the output arrival curve in
bits of the flow that has arrived on the network node up to Eq. (64) is changed into,
time t. We say that R(t) is constrained by α(t) iff, α∗ (t) = α (β − l max )(t), (65)
R(t) ≤ inf {R(s) + α(t − s)} = (R ⊗ α)(t). (58) which is proved by [43, Corollary 8.3], where l max is the max-
0≤s≤t
imum frame size of the flow. With the known latency upper
Note that an arrival curve α(t) should be a non-negative bound of the flow, the output arrival curve of the flow can also
wide-increasing function. A typical example of an arrival curve be given by,
is the “leaky bucket” constraint satisfying α(t) = b + r · t for
t > 0 and α(0) = 0, with the maximum burst tolerance b and α∗ (t) = α(t) δD (t), (66)
long-term rate r of the flow. where δD (t) is the pure-delay function which equals to 0 if
A service curve β(t) models the processing capability of the t ≤ D and +∞ otherwise. The output arrival curve can also be
available resource for the network node. Assume that R ∗ (t) is regarded as the input arrival curve of the flow before reaching
the departure process, which is the output cumulative function the next node. Note that in this paper, all the NC-based method
that counts the total data bits of the flow departure from the is based on the Total Flow Analysis (TFA). For TFA, using
network node up to time t. There are several definitions for the the delay bound D calculated from aggregate flows is always
service curve. We say that the network node offers the min- better than using the delay bound Df = h(αf , βf ) for the
plus minimal service curve β(t) (considered in this paper) for flow of interest f [43]. Moreover, the calculation of βf when
the flow iff considering FIFO multiplexing depends on the choice of an
R ∗ (t) ≥ inf {R(s) + β(t − s)} = (R ⊗ β)(t), (59) efficient θ, which is still an open research issue [51], [52],
0≤s≤t ⎡ ⎤+

and offers the strict service curve β(t) iff βf (t) = ⎣β − αj ⊗ δθ ⎦ ∧ δθ , ∀θ ∈ R+ . (67)
R ∗ (t + Δt) − R ∗ (t) ≥ β(Δt), (60) j =f
during any backlog period [t, t + Δt). Note that a service ACKNOWLEDGMENT
curve β(t) should be a non-negative wide-increasing function. The authors would like to thank Marc Boyer from ONERA
A shaping curve σ(t) characterizes the maximum number for the valuable suggestions on the review version of this
of bits that are served during any period of time Δt, which paper.
means that the departure process R ∗ (t) from the server is
always constrained by the shaping curve. A server offers a
R EFERENCES
shaping curve σ(t) iff,
[1] IEEE. “802.1Q—IEEE standard for local and metropolitan area
R ∗ (t + Δt) − R ∗ (t) ≤ σ(Δt), (61) networks—Bridges and bridged networks.” 2018. [Online]. Available:
https://standards.ieee.org/standard/802_1Q-2018.html
i.e., R ∗ (t) ≤ R ∗ ⊗ σ(t). Note that the shaping curve σ(t) in [2] 802.3 Standard for Ethernet, IEEE, Piscataway, NJ, USA, 2015.
[3] IEEE. “802.1AS-Rev—Timing and synchronization for time-sensitive
Eq. (61) is non-greedy shaping, which is different from the applications.” 2016. [Online]. Available: http://www.ieee802.org/1/
definition of greedy shaper satisfying R ∗ (t) = R ⊗ σ(t). The pages/802.1AS-rev.html
shaping curve can be used to additionally constrain the output [4] IEEE. “802.1Qbv—Enhancements for scheduled traffic.” 2016. [Online].
Available: http://www.ieee802.org/1/pages/802.1bv.html
arrival curve by the minimum operation. [5] IEEE. “802.1Qcr—IEEE standard for local and metropolitan area
Three basic results of network calculus are given as follows. networks—Bridges and bridged networks amendment: Asynchronous
If the flow R(t) constrained by the arrival curve α(t) traverses traffic shaping.” 2018. [Online]. Available: https://1.ieee802.org/tsn/
802-1qcr/
the network node offering the service curve β(t), the latency [6] IEEE. “802.1BA—Audio video bridging (AVB) systems.” 2011.
experienced by the flow in the network node is bounded by [Online]. Available: http://www.ieee802.org/1/pages/802.1ba.html
[7] IEEE. “802.1Qav—Forwarding and queuing enhancements for time- [31] P. Meyer, T. Steinbach, F. Korf, and T. C. Schmidt, “Extending IEEE
sensitive streams.” 2009. [Online]. Available: https://www.ieee802.org/ 802.1 AVB with time-triggered scheduling: A simulation study of the
1/pages/802.1av.html coexistence of synchronous and asynchronous traffic,” in Proc. IEEE
[8] IEEE. “802.1Qbu–Frame preemption.” 2015. [Online]. Available: http:/ Veh. Netw. Conf., 2013, pp. 47–54.
/www:ieee802:org/1/pages/802.1bu.html [32] L. Zhao, P. Pop, Q. Li, J. Chen, and H. Xiong, “Timing analysis of rate-
[9] S. S. Craciunas and R. S. Oliver, “An overview of scheduling mecha- constrained traffic in TTEthernet using network calculus,” Real Time
nisms for time-sensitive networks,” in Proc. Real Time Summer School Syst. vol. 53, no. 2, pp. 254–287, 2017.
ETR, 2017, pp. 1–9. [33] L. Zhao, P. Pop, Z. Zheng, and Q. Li, “Timing analysis of AVB traffic
[10] F. Dürr and N. G. Nayak, “No-wait packet scheduling for IEEE time- in TSN networks using network calculus,” in Proc. IEEE Real Time
sensitive networks (TSN),” in Proc. 24th Int. Conf. Real Time Netw. Embedded Technol. Appl. Symp. (RTAS), 2018, pp. 25–36.
Syst., 2016, pp. 203–212. [34] L. Zhao, P. Pop, Z. Zheng, H. Daigmorte, and M. Boyer, “Latency
[11] S. S. Craciunas, R. S. Oliver, M. Chmelik, and W. Steiner, “Scheduling analysis of multiple classes of AVB traffic in TSN with standard credit
real-time communication in IEEE 802.1 Qbv time sensitive networks,” behavior using network calculus,” IEEE Trans. Ind. Electron., vol. 68,
in Proc. ACM 24th Int. Conf. Real Time Netw. Syst., 2016, pp. 183–192. no. 10, pp. 10291–10302, Nov. 2021.
[12] P. Pop, M. L. Raagaard, S. S. Craciunas, and W. Steiner, “Design optimi- [35] M. Ashjaei, G. Patti, M. Behnam, T. Nolte, G. Alderisi, and L. L. Bello,
sation of cyber-physical distributed systems using IEEE time-sensitive “Schedulability analysis of Ethernet audio video bridging networks with
networks,” IET Cyber Phys. Syst. Theory Appl., vol. 1, no. 1, pp. 86–94, scheduled traffic support,” Real Time Syst., vol. 53, no. 4, pp. 526–577,
2016. 2017.
[13] M. L. Raagaard, “Algorithms for the optimization of safety-critical [36] E. Mohammadpour, E. Stai, M. Mohiuddin, and J.-Y. Le Boudec, “End-
networks,” Ph.D. dissertation, Dept. Comput. Sci., Delhi Technol. Univ., to-end latency and backlog bounds in time-sensitive networking with
New Delhi, India, 2017. credit based shapers and asynchronous traffic shaping,” in Proc. 30th
[14] M. Vlk, Z. Hanzálek, K. Brejchová, S. Tang, S. Bhattacharjee, and Int. Teletraffic Congr., 2018, pp. 1–9.
S. Fu, “Enhancing schedulability and throughput of time-triggered traffic [37] B. Fang, L. Qiao, Z. Gong, and H. Xiong, “Simulative assessments of
in IEEE 802.1 QBV time-sensitive networks,” IEEE Trans. Commun., credit-based shaping and asynchronous traffic shaping in time-sensitive
vol. 68, no. 11, pp. 7023–7038, Nov. 2020. networking,” in Proc. 12th Int. Conf. Adv. Infocomm Technol., 2020,
[15] R. S. Oliver, S. S. Craciunas, and W. Steiner, “IEEE 802.1 Qbv gate pp. 111–118.
control list synthesis using array theory encoding,” in Proc. IEEE Real [38] A. Nasrallah et al., “Ultra-low latency (ULL) networks: The IEEE
Time Embedded Technol. Appl. Symp. (RTAS), 2018, pp. 1–8. TSN and IETF DetNet standards and related 5G ULL research,” IEEE
[16] L. Zhao, P. Paul, and S. S. Craciunas, “Worst-case latency analysis for Commun. Surveys Tuts., vol. 21, no. 1, pp. 88–145, 1st Quart., 2018.
IEEE 802.1 Qbv time sensitive networks using network calculus,” IEEE [39] L. Maile, K. S. Hielscher, and R. German, “Network calculus results for
Access, vol. 6, pp. 41803–41815, 2018. TSN: An introduction,” in Proc. Inf. Commun. Technol. Conf. (ICTC),
[17] L. Zhao, P. Pop, Z. Gong, and B. Fang, “Improving latency analysis for 2020, pp. 131–140.
flexible window-based gcl scheduling in TSN networks by integration [40] N. Reusch, L. Zhao, S. S. Craciunas, and P. Pop, “Window-based
of consecutive nodes offsets,” IEEE Internet Things J., vol. 7, no. 7, schedule synthesis for industrial IEEE 802.1 Qbv TSN networks,” in
pp. 5574–5584, Apr. 2021. Proc. 16th IEEE Int. Conf. Factory Commun. Syst. (WFCS), 2020,
pp. 1–4.
[18] D. Hellmanns, J. Falk, A. Glavackij, R. Hummen, S. Kehrer, and F. Durr,
[41] A. Nasrallah et al., “Performance comparison of IEEE 802.1 TSN
“On the performance of stream-based, class-based time-aware shaping
time aware shaper (TAS) and asynchronous traffic shaper (ATS),” IEEE
and frame preemption in TSN,” in Proc. IEEE Int. Conf. Ind. Technol.
Access, vol. 7, pp. 44165–44181, 2019.
(ICIT), 2020, pp. 298–303.
[42] J. Y. Le Boudec and P. Thiran, Network Calculus: A Theory of
[19] R. Mahfouzi, A. Aminifar, S. Samii, A. Rezine, P. Eles, and Z. Peng,
Deterministic Queuing Systems for the Internet (Lecture Notes on
“Stability-aware integrated routing and scheduling for control applica-
Computer Science), 5th ed. Heidelberg, Germany: Springer-Verlag,
tions in Ethernet networks,” in Proc. IEEE Design Autom. Test Europe
2001.
Conf. Exhibit. (DATE), 2018, pp. 682–687.
[43] A. Bouillard, M. Boyer, and E. Le Corronc, Deterministic Network
[20] J. Specht and S. Samii, “Urgency-based scheduler for time-sensitive Calculus: From Theory to Practical Implementation. Hoboken, NJ,
switched Ethernet networks,” in Proc. IEEE Euromicro Conf. Real Time USA: Wiley, 2018.
Syst. (ECRTS), 2016, pp. 75–85. [44] V. Gavriluţ and P. Pop, “Traffic-type assignment for TSN-based mixed-
[21] J. Specht and S. Samii, “Synthesis of queue and priority assignment criticality cyber-physical systems,” ACM Trans. Cyber Phys. Syst., vol. 4,
for asynchronous traffic shaping in switched Ethernet,” in Proc. IEEE no. 2, pp. 1–27, 2020.
Real-Time Syst. Symp. (RTSS), 2017, pp. 178–187. [45] S. S. Craciunas and R. S. Oliver, “Combined task-and network-level
[22] Z. Zhou, Y. Yan, M. Berger, and S. Ruepp, “Analysis and modeling of scheduling for distributed time-triggered systems,” Real Time Syst.,
asynchronous traffic shaping in time sensitive networks,” in Proc. 14th vol. 52, no. 2, pp. 161–200, 2016.
IEEE Int. Workshop Factory Commun. Syst., 2018, pp. 1–8. [46] A. Amari and A. Mifdaoui, “Worst-case timing analysis of ring networks
[23] Z. Zhou, M. S. Berger, S. R. Ruepp, and Y. Yan, “Insight into the IEEE with cyclic dependencies using network calculus,” in Proc. IEEE 23rd
802.1 Qcr asynchronous traffic shaping in time sensitive network,” Adv. Int. Conf. Embedded Real Time Comput. Syst. Appl. (RTCSA), 2017,
Sci. Technol. Eng. Syst. J., vol. 4, no. 1, pp. 292–301, 2019. pp. 1–10.
[24] J. Y. Le Boudec, “A theory of traffic regulators for deterministic [47] A. Finzi and S. S. Craciunas, “Breaking vs. solving: Analysis and
networks with application to interleaved regulators,” IEEE/ACM Trans. routing of real-time networks with cyclic dependencies using network
Netw., vol. 26, no. 6, pp. 2721–2733, Dec. 2018. calculus,” in Proc. 27th Int. Conf. Real Time Netw. Syst., 2019,
[25] J. Diemer, J. Rox, and R. Ernst, “Modeling of Ethernet AVB networks pp. 101–111.
for worst-case timing analysis,” IFAC Proc. Vol., vol. 45, no. 2, [48] L. Thomas, J. Y. L. Boudec, and A. Mifdaoui, “On cyclic dependencies
pp. 848–853, 2012. and regulators in time-sensitive networks,” in Proc. IEEE Real-Time
[26] U. D. Bordoloi, A. Aminifar, P. Eles, and Z. Peng, “Schedulability anal- Syst. Symp. (RTSS), 2019, pp. 299–311.
ysis of Ethernet AVB switches,” in Proc. IEEE 20th Int. Conf. Embedded [49] D. Tamas-Selicean, “Design of mixed-criticality applications on dis-
Real Time Comput. Syst. Appl., 2014, pp. 1–10. tributed real-time systems,” Ph.D. dissertation, Dept. Comput. Sci., Tech.
[27] J. A. R. De Azua and M. Boyer, “Complete modelling of AVB in Univ. Denmark, Kongens Lyngby, Denmark, 2014.
network calculus framework,” in Proc. 22nd Int. Conf. Real Time Netw. [50] M. Paulitsch, E. Schmidt, B. Gstottenbauer, C. Scherrer, and H. Kantz,
Syst., 2014, p. 55. “Time-triggered communication (industrial applications),” in Time-
[28] L. Zhao, F. He, and E. Li, “Improving worst-case delay analysis for Triggered Communication. Boca Raton, FL, USA: CRC Press, 2011,
traffic of additional stream reservation class in Ethernet-AVB Network,” pp. 121–152.
Sensors, vol. 18, no. 11, p. 3849, 2018. [51] M. Boyer, A. Graillat, B. D. De Dinechin, and J. Migge, “Bounding
[29] J. Cao, P. J. L. Cuijpers, R. J. Bril, and J. J. Lukkien, “Tight worst-case the delays of the MPPA network-on-chip with network calculus:
response-time analysis for Ethernet AVB using eligible intervals,” in Models and benchmarks,” Perform. Eval., vol. 143, Nov. 2020,
Proc. IEEE World Conf. Factory Commun. Syst. (WFCS), 2016, pp. 1–8. Art. no. 102124.
[30] L. L. Bello, “Novel trends in automotive networks: A perspective on [52] A. Scheffler and S. Bondorf, “Network calculus for bounding delays in
Ethernet and the IEEE Audio Video Bridging,” in Proc. IEEE Emerg. feedforward networks of FIFO queueing systems,” in Proc. Int. Conf.
Technol. Factory Autom. (ETFA), 2014, pp. 1–8. Quant. Eval. Syst., 2021, pp. 149–167.
Luxi Zhao received the Ph.D. degree in com- Sebastian Steinhorst (Senior Member, IEEE)
munication and information system from Beihang received the M.Sc. (Dipl.-Inf.) and Ph.D. (Dr. phil.
University, Beijing, China, in 2017, where she is nat.) degrees in computer science from Goethe
currently an Associate Professor of Communication University, Frankfurt, Germany, in 2005 and 2011,
and Information System with the Department of respectively. He is currently an Associate Professor
Electronic and Information Engineering. She was a with the Technical University of Munich, Germany,
Postdoctoral Fellow with DTU Compute, Technical where he leads the Embedded Systems and Internet
University of Denmark from 2017 to 2019. She of Things Group, Department of Electrical and
has been a Marie-Curie Research Fellow with the Computer Engineering. He was also a Co-Program
Department of Electrical and Computer Engineering, PI with the Electrification Suite and Test Laboratory,
Technical University of Munich. Her main research Research Center TUMCREATE, Singapore. His
interest concerns worst-case analysis and performance evaluation of determin- research interests include design methodology and hardware/software archi-
istic real-time and safety-critical networks. tecture co-design of secure distributed embedded systems for use in IoT,
automotive, and smart energy applications.
Paul Pop (Member, IEEE) received the Ph.D. degree

in computer systems from Linköping University
in 2003. He is a Professor of Cyber–Physical
Systems with DTU Compute, Technical University
of Denmark. His research is focused on developing
methods and tools for the analysis and optimization
of networked dependable cyber–physical systems. In
this area, he has published over 130 peer-reviewed
papers, three books, and seven book chapters. He
has received the Best Paper Award at DATE 2005,
RTIS 2007, CASES 2009, MECO 2013, DSD 2016,
and ETFA 2020. He has served as a Technical Program Committee Member
on several conferences, such as DATE and ESWEEK. He is the Chairman of
the IEEE Danish Chapter on Embedded Systems. He is the Coordinator of
the Nordic University Hub on Industrial IoT and of the European Training
Network on Fog Computing for Robotics and Industrial Automation.

Quantitative Performance Comparison of Various

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Quantitative Performance Comparison of Various

Uploaded by

Copyright:

Available Formats

IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 19, NO.

3, SEPTEMBER 2022 2899

Quantitative Performance Comparison of Various

II. OVERVIEW OF P ERFORMANCE M ETRICS

A. End-to-End Latency Upper Bounds

Fig. 1. TAS Architecture.

systems (ESes) and switches (SWs), in the network support

Fig. 3. ATS Architecture.

Jitter bounds - TAS: Real-time communications are typi-

committed burst size bf for each flow by calculating an eligible

cimax + The use of the term l max

j >i {lQj , lQBE } is the maximum frame

where cjmin is the lower bound of credit of AVB Class

the traffic with the priority lower than priority Mi .

service curve βQi (Eq. (9)) should additionally consider the

on transmission (during ΔtQ CBS ); 2) increases at speed idSl

AVB flows with higher priority than Mi is on transmission (26)

2) Performance Analysis–CBS+ATS: Similar to the

where αQ CBS (t) (Eq. (22)) is the arrival curve of aggre-

Especially, for the input arrival curve αQi (t) in the

traffic shapers. Here, the lower priority traffic (SP/AVB) can

1) Performance Analysis–TAS+SP [32]:

αQi (t) of aggregate SP flows with the priority i before enter-

curve for AVB traffic, GB and TT slots cannot be taken as

where the calculation of α∗ (t) can refer to Eq. (20),

βTT (t) = min βTDMA (t + t0 , LTT

i.e., s = sup{u ≤ t RTT ∗ (u) = R ∗

and backlog d ATS + SP

1 and 2, the evaluation results calculated by NC are 0.6% and

the upper bounds of worst-case backlog (WCB) in bytes and

Fig. 14. Comparison of ATS and SP under different traffic load.

Therefore, when the ATS is used individually, its positive

B. Combined Traffic Shapers

be increased. Therefore, the time used by the ATS reshaping

Fig. 20. Network topology of Orion CEV.

Paul Pop (Member, IEEE) received the Ph.D. degree

You might also like