Full Duplex Switched Ethernet for Next Generation ”1553B”-based Applications · 2013. 7. 19. · 2.2. ”1553B”-based applications require-ments Prior to replace the MIL-STD

Full Duplex Switched Ethernet for Next Generation ”1553B”-based Applications

Ahlem MifdaouiToulouse University

ENSICA1 place Emile Blouin

31056 Toulouse, [email protected]

Fabrice FrancesToulouse University

ENSICA1 place Emile Blouin

31056 Toulouse, [email protected]

Christian FraboulToulouse University

ENSEEIHT/IRIT2 rue Charles Camichel31071 Toulouse, France

[email protected]

Abstract

Over the last thirty years, the MIL-STD 1553B data bushas been used in many embedded systems, like aircrafts,ships, missiles and satellites. However, the increasing num-ber and complexity of interconnected subsystems lead toemerging needs for more communication bandwidth. There-fore, a new interconnection system is needed to overcomethe limitations of the MIL-STD 1553B data bus. Amongseveral high speed networks, Full Duplex Switched Ether-net is put forward here as an attractive candidate to re-place the MIL-STD 1553B data bus. However, the key argu-ment against Switched Ethernet lies in its non-deterministicbehavior that makes it inadequate to deliver hard time-constrained communications. Hence, our primary objectivein this paper is to achieve an accepted QoS level offeredby Switched Ethernet, to support diverse ”1553B”-basedapplications requirements. We evaluate the performance oftraffic shaping techniques on Full Duplex Switched Ethernetwith an adequate choice of service strategy in the switch, toguarantee the real-time constraints required by these spe-cific 1553B-based applications.An analytic study is conducted, using the Network Calcu-lus formalism, to evaluate the deterministic guarantees of-fered by our approach. Theoretical analysis are then inves-tigated in the case of a realistic ”1553B”-based applicationextracted from a real military aircraft network 1. The resultsherein show the ability of profiled Full Duplex SwitchedEthernet to satisfy 1553B-like real-time constraints.

1. Introduction

The MIL-STD 1553B is a widely used avionic data busthat has been deployed in various military applicationsfor decades [11]. It is a 1 Mbps command/response

1This work has been sponsored by a DGA/MRIS program

multiplexed data bus with a centralized system control.However, this traditional data bus is no longer effectivein meeting the communication demands imposed by thenext generation ”1553B”-based applications. In fact, withthis bus approach, communications are closely coupledto the bus controller, which limits system modularityand reconfigurability. Another limitation is its inabilityto handle the heavy throughput demands caused by theincreasing subsystems’ complexity. The need to supportmany simultaneous data flows also renders MIL-STD1553B unsuitable for next generation applications. Thecurrent solution used to handle this problem consists inincreasing the number of MIL-STD 1553B data busesand integrating dedicated data buses with higher ratescompatible with the 1553B interface, like STANAG 3910[8] or even other buses like SCI links [1], to supportthe important amount of exchanged information betweensubsystems. Using these solutions makes global intercon-nection system heterogeneous and complex. Moreover,real-time constraints guarantees are difficult to prove.Clearly, a new interconnection system is needed to fulfillthese requirements.

Among several high speed networks, Switched Ethernetpresents significant interests thanks to its reduced costs,its flexibility and its expandability. In fact, recently, anARINC 664-compliant Full Duplex Switched Ethernet(AFDX) network [5] has been integrated into new gener-ation civil aircrafts like the A380, to replace traditionalARINC-429 data buses. Thanks to policing mechanismsadded in switches, this technology succeeds to supportthe important amount of exchanged data ([2, 6, 9, 10]).Therefore, after this successful civil experience, StandardFull Duplex Switched Ethernet is put forward here as afuture interconnection technology to replace MIL-STD1553B.

However, this COTS technology cannot be directly

Proceedings of the 13th IEEE Real Time andEmbedded Technology and Applications Symposium(RTAS'07)0-7695-2800-7/07 $20.00 © 2007

CORE Metadata, citation and similar papers at core.ac.uk

Provided by Open Archive Toulouse Archive Ouverte

https://core.ac.uk/display/12038924?utm_source=pdf&utm_medium=banner&utm_campaign=pdf-decoration-v1

deployed for time-critical applications where messageshave to be delivered within a determined time limit. Theuse of switches and full duplex links eliminates collisionsbut this is not enough to prove that deadlines constraintsfor each traffic frame are effectively respected . In fact, thecollision problem is shifted to congestions in the switches.This may occur when simultaneous traffic data attemptto share network segments and messages can be lost ifswitch queues overflow. Therefore, in order to use FullDuplex Switched Ethernet in ”1553B”-based applications,an additional mechanism is needed to avoid informationloss and to guarantee a real-time behavior with low latency.Various real-time communication solutions were offeredfor the CSMA/CD Ethernet. These approaches range fromimplementing a master-slave communication model overEthernet (FTT-Ethernet [23]) to token passing (RETHER[26]) or TDMA ([14]) to schedule nodes access. In ourprevious work [18], these solutions were described andtheir adaptation and their applicability over SwitchedEthernet were discussed. After a comparative study of theircapacities to handle real-time constraints, the choice oftraffic shaping approach was justified and first experimentswere conducted.

This paper extends our previous works by giving ageneral analytical study to characterize ”1553B”- basedapplications and to replace the MIL-STD 1553B databus by a reliable Full Duplex Switched Ethernet. First,a general network model is designed to describe trafficshapers implementation at data sources and to integrateswitch service policies among those offered by commonswitch technology (FCFS, SP, WFQ). Then, delay boundanalysis are conducted to investigate the validity of ourgeneral network model. This general analytical way isapplied in the case of a realistic MIL-STD 1553B aircraftnetwork to show our Ethernet proposal’s ability to pro-vide deterministic transmission with bounded delays andguaranteed deadline constraints. Hence, the originality ofthis paper consists in the proposed reliable interconnectiontechnology to support the increasing demands of nextgeneration ”1553B”-based applications with reference toSwitched Ethernet temporal analysis and use of theoreticaltools like the Network Calculus [7, 12].

The paper is organized as follows. An overview of MIL-STD 1553B and a brief examination of data communicationrequirements for ”1553B”-based applications are presentedin section 2. Section 3 surveys other work in the area andrelates it to our work. Section 4 and Section 5 show our net-work model and the delay bounds analysis of traffic shap-ing approach with different service disciplines, respectively.Theoretical results obtained in our case study are then pre-sented in section 6. Section 7 concludes the paper.

2 Background

2.1. An overview of MIL-STD 1553B

The MIL-STD 1553B is a 1 Mbps serial asynchronousdata bus on which the messages are multiplexed amongusers. To initiate all messages, the bus required a central-ized control bus [?]. Figure 1 shows an example of a MIL-STD 1553B data bus system. The Medium Access Con-trol (MAC) mechanism is described as command / responsetime division multiplexing: the Bus Controller (BC) sendscommand messages at predetermined times to Remote Ter-minals (RT), to give them access to the bus. A special RTcalled Monitor (M) can receive and store every circulat-ing message but cannot transmit information on the bus, itsprincipal use is to verify data bus system performance.The BC follows the instructions stored in a memory called

Bus Controller

Remote Terminal

Subsystem

Bus Monitor

Optional redundantbuses

Figure 1. MIL-STD 1553 Architecture

transaction table to control communication and to monitormessage requests. This table complexity depends on thenumber of RTs and the amount of data to be transferred;and several techniques have been developed to determineit. Thus, the BC’s operation is well ordered, pre-establishedand specific for each step necessary to achieve data.The MIL-STD 1553B data bus can support different typesof messages: periodic and sporadic with strict deadlines thatmust be respected in transmission. The polling cycle timemust be carefully chosen since it must efficiently transferall data, thus preventing any RT from being polled eithertoo often or too little. The standard defines different typesof messages formats and each format is divided into datacommunication messages and communication managementmessages. Each message uses only standardized word typesthat are described in [11].Hence, the MIL-STD 1553B is suitable for applica-tions where the generation and transfer of data are pre-established. It is a shared media network that supports pre-cise timeliness in a deterministic way, however the applica-tion flexibility is reduced because a static transaction tableis used to transmit information.


2.2. ”1553B”-based applications require-ments

Prior to replace the MIL-STD 1553B data bus, theapplication requirements and performance criteria must bedefined [24]. These requirements concern both technicalaspects and costs.

Technical requirements are based on the timeliness andthe accuracy of data which is critical for safety concerns.Therefore, the determinism of the data bus and a latencyof information transmission that respects the deadlines con-straints are both very important. Moreover, these applica-tions need several classes of service and priority controlwith guaranteed qualities of service for each traffic class.In fact, in our ”1553B”-based applications, there are fourimportant categories of messages identified by their period-icity and their temporal deadlines:

• Urgent sporadic messages sent by the BC and that haveto be received under a predefined small bounded time,like alarms.

• Periodic messages that depend on the time cyclesdefined by the BC and that are also hardly time-constrained, like sensor data.

• Sporadic messages that have known deadlines to re-spect but without any urgency.

• Sporadic messages that do not have to respect stricttime constraints, like file transfers.

To fulfill real-time guarantees for these applications,a major frame is often defined to transfer all periodicmessages at least once and it shall be no smaller than thebiggest message period. Every major frame consists of afinite number of minor frames to meet the requirements ofthe higher update rate messages: at the beginning of eachnew minor frame, an interrupt occurs and the bus controllerstarts issuing the messages for that minor frame.

Costs requirements are also important since the choiceof the data bus must be made with the goal of meetingthe design requirements for the least amount of money.Therefore, commercial availability, expandability andmaintainability are among the most important characteris-tics that the data bus should have.

Clearly, to design a new data bus for ”1553B”-based ap-plications, we have to guarantee deterministic informationtransmission with bounded latency that respect the real-timeconstraints (technical requirements) and with reduced costs(cost requirements), while the bandwidth may not be theprimary design concern.

3 Related work

3.1. Replacing MIL-STD 1553B

To support the increasing demands of next generation”1553B”-based applications, various solutions were offeredto replace the MIL-STD 1553B data bus. Some of theseinvolve the development of dedicated architectures like theFDDN (Fiber Optic Data Distribution Network) [4] andothers use commercial-off-the-shelf (COTS) technologies.This latter approach is an attractive solution to reduce costsand to add flexibility and expandability. However, amongseveral COTS interconnection technologies, only fewnetwork products are expanding their application domainto replace MIL-STD 1553B data buses.

Parish and Briggs [22] suggested a communicationarchitecture based upon ATM. By using 1553 emulationover ATM network, the coexistence of the two protocols ispossible. However, the use of dedicated entities to connect1553 remote terminals to an ATM network makes thissolution complicated to adapt.

Fiber Channel (FC) [20] is also an interesting COTSproduct nominated by Murdock and Koenig. In their pa-per, they compare FC and MIL-STD 1553B characteristicsand confirm that FC is easier to use since it does not need abus loading analysis, nor a bus controller. However, the FCability to guarantee deterministic transmission and deadlinerespect, required by the ”1553B”-based applications, is stillnot validated.Switched Ethernet is put forward in this paper as an attrac-tive candidate to replace the MIL-STD 1553B data bus. Infact, this COTS technology is indisputably the most cost-effective solution: low price, component maturity and nospecial staff training is needed since all the network engi-neers perfectly know Switched Ethernet. Then, to achievean accepted QoS level offered by Switched Ethernet in or-der to support next generation ”1553B”- based applicationsrequirements, the key of our solution is the use of trafficshapers at data sources to control traffic; and an adequatechoice of switch service policy to guarantee the real timeconstraints of these critical applications. Note that today’sswitches like the recent Cisco [3] ones offer interesting fea-tures like advanced QoS. These switches implement Priorityhandling IEEE 802.1p to classify packets and can supportprocessing up to four queues per output port. In addition,Cisco switches support FCFS, SP and WFQ scheduling.

3.2. Traffic shaping approach

Traffic Shaping approach has been initiated by Kweonand Shin [15] to achieve real-time communication over


Ethernet. The idea is that a smooth traffic, in which mes-sages arrive at a constant rate, suffers less from collisionsthan a bursty traffic. Specifically, a traffic smoother isinstalled in every station between the Ethernet MAC layerand the UDP or TCP/IP layer: first it gives real-time(RT) packets priority over non real-time (NRT) packets,second it smooths the NRT-stream to reduce collision withRT-packets from the other nodes. Real-time traffic is notsmoothed and is sent as soon as it arrives. This approachprovides statistical guarantees which is not sufficient forhard real-time systems like the ”1553B”-based applications.

Unlike this approach where only the non real-time trafficis smoothed, the Traffic Shaping approach developed byLoeser and Haertig [17] is based on the fact that all theincoming traffic has to be controlled in order to guaranteesome deterministic performances of the network. Theyshow in experiments, in the special case of Fast and GigabitSwitched Ethernet, that transmission without packet losscould be guaranteed when using the Traffic Shaping.However, this guarantee do not fulfill the ”1553B”-basedapplications requirements and specially the deadlineconstraints.

This last approach presents a similarity with the one sug-gested in this paper, which consists on the use of trafficshapers at data sources to control traffic and to guaranteethe integrity of the ”1553B”-based applications. However,we will add a priority handling method in data sources andin the switch to assure a good isolation level for urgent mes-sages with hard deadline constraints. Then, an adequatechoice of service policy in the switch is needed to guar-antee low bounded delays. Finally, we show the validityof our proposal to achieve deterministic communications inthe case of a realistic ”1553B”-based applications.

4 Network modeling

4.1. Traffic model

In our model, in order to integrate the different character-istics of the traffic generated by the ”1553B”-based appli-cations, four parameters (T, D, L, P ) are defined for eachtraffic stream:

• The periodicity T : for a periodic message, it is the pe-riod and for a sporadic message, it is low bounded asits minimal inter-arrival time.

• The temporal local deadline D: (the message life du-ration) it is the period for a periodic message and themaximal response time for a sporadic message.

• The length L: the maximal length of a message

• The priority level P : We attribute a priority level, thatis determined by the real-time requirements of the in-formation, to each traffic category described in 2.2 .As a result, we define four priorities: alarms will betagged with the highest priority (4), then periodic traf-fic the medium priority (3), and the lowest priorities (2and 1) for the asynchronous messages of the two lastcategories. Therefore, P is a natural in {1, 2, 3, 4}.

4.2. Traffic shapers implementation

The traffic shaping idea is that reliable transmissionwith bounded delays is possible when there is a trafficcontrol at data sources. Each subsystem has to controlits streams in accordance with their periodicity and theirpacket’s maximal length. The traffic shaper regulates apacket stream using a leaky bucket concept characterizedby a maximal size b and a rate r carefully chosen for eachstream. At the output of the traffic shaper, the maximaltransmitted burst size is less than b and data leave with arate r.

For each traffic shaper, a traffic shaping interval T isdefined: once the bucket gets empty, the next amount ofdata is generated not earlier than T . So, the bucket musthold at least the amount of data r ∗ T that can arrive duringT time units. To guarantee the integrity of the trafficcharacteristics on the MIL-STD 1553B, we consider foreach stream a leaky bucket characterized by its maximalsize L and its rate L/T in order to have one packet of sizeL per period T .

We model a terminal that transmits n flows as shown inFigure 2. Since there are multiple transmitted flows to dif-ferent destinations, we need to put one traffic shaper perflow. Then, these flows are multiplexed inside the terminalbefore sending them on the data bus. The multiplexer char-acteristics in the terminal output determine the processingpolicy and in our approach only FCFS and SP policies areconsidered.

• FCFS policy is used in sources when the schedulingpolicy at the switch is also FCFS.

• SP policy is used to integrate each flow priority level atdata sources, when the scheduling policy at the switchis SP or WFQ.

4.3. Switch policies and model

An Ethernet switch is an active device that identifiesthe destination port of an incoming packet and relays it tothe specific port [19]. If multiple packets have the same


flow1

flow2

flowN

(b1,r1)

(b2,r2)

(bN,rN)

MUX

sources traffic shaper buffers

msgs receptor

Sub

syst

em

Figure 2. Model of traffic shaping in a termi-nal

destination port, buffers are used to resolve the problem ofcollision however frames are lost when buffers overflow.Ethernet switches can be identified by their switchingtechnique and their scheduling policy.

First, two types of switching techniques are currentlyimplemented in Ethernet switches: Cut Through and Storeand Forward. With the first, only the header of each packetis decoded to determine its destination port and the rest isforwarded without any error checking mechanism. Withthe latter, the switch waits until the complete reception ofthe packet and forwards it to the destination port if it issuccessfully verified. In our model, we choose the secondswitching technique for safety reasons since no corruptedpacket will be forwarded. Then comes the schedulingpolicy which will be used to forward packets at the switchoutput port. We have focused on the three most widelyimplemented policies: First Come First Served (FCFS),Static Priority (SP) and Weighted Fair Queuing (WFQ);and this is to guarantee the lowest bounded delays.

• FCFS is the simplest policy. Packets are served in theirarrival order without taking into account their temporalcharacteristics and mainly their deadlines, which cancause real-time constraints violations.

• Using SP, packets are queued and forwarded accordingto their priorities. So, a queue is selected for transmis-sion if all traffic classes queues with higher priority areempty at the time of selection; and for a given queue,the scheduling order is FCFS with a non-preemptivemanner. In our study, the 802.1p priority model [19]which defines a 3 bits priority field in the extended Eth-ernet frame (8 priority levels), is used to manipulate thefour priority classes defined in section 4.1. Starvationfor the lowest priority queues represents the main SPpolicy drawback.

• With WFQ, a fair service is guaranteed for each queue.In fact, a weight is attributed to each queue to deter-mine its associated bandwidth. The WFQ algorithm is

based on the computation of a virtual finish time of ser-vice with the fluid GPS model ([21]), which dependsonly on the service share weight and the packet length.This virtual finish time is tagged into the packet andthe scheduler selects the packet with the lowest finishtag to be transmitted. Delays with WFQ can be longerthan with GPS, however WFQ offers the same proper-ties as GPS, like fairness and flexibility. In our case,weights are associated to every priority class by takinginto account their temporal deadlines. Therefore, fourqueues are defined to serve in switch output ports: onefor each priority class.

The Ethernet switch used for our study is modeled asfollows (figure 3): buffers in the input ports to representthe Store and Forward mode, then a filtering and relayingprocessor, and finally a multiplexer with four queues in eachoutput port to send the multiplexed flow to the destinationwith a determined scheduling strategy.

Inputport

Filtering &relayingprocessor

port

ReceivingBuf

TBuf

Output

Figure 3. Model of frame transmission in anEthernet Switch

5 Delay bound analysis

To investigate the validity of our network model andits efficiency to support the ”1553B”-based applicationsrequirements, the main metric that has been chosen is theend to end delay. To evaluate the QoS level offered by oursuggested network, the maximal end to end delay boundswill be compared to the temporal deadlines of each priorityclass of traffic. To achieve this aim, we have chosen toconduct analytic studies instead of simulations, which arecommonly used to validate models. In fact, simulationscannot cover the entire domain of the model applicabilityand specially rare events that represents worst-case func-tioning. Moreover, these latter are always conducted witha given confidence level always less than 100 percent.So, clearly, simulations cannot provide the deterministicguarantees required by our critical application, where a


failure might have a disastrous consequence on our system.

��

��

�

�

��

Figure 4. Computation of buffer and delaybounds

Our analytic study is based on the use of Network Cal-culus theory, introduced by Cruz [25] and developed in aneater way by Leboudec [16], because it is well adapted tocontrolled traffic sources and provides deterministic end-to-end delay bounds. This formalism differs from traditionalqueuing theories in the model of the traffic entering thenetwork. In fact, instead of using a stochastic process forthe entering traffic, the compliance to some regularity con-straints is enough to model the traffic. These constraintslimit traffic burstiness in the network. Delay and queuesize bounds depend on the traffic arrival described by theso called arrival curve α, and on the availability of the tra-versed node described by a service curve β [16]. As shownin Figure 4, the packet delay D is bounded by the horizon-tal distance between α and β whereas the queue size B isbounded by the vertical distance. The calculation of thesebounds is greatly simplified in the case of affine curves. Infact, the most used curves are: the leaky bucket arrival curveα(t) = b + rt with b the maximal burst and r the rate (wesay that the flow is (b, r)-constrained); and the rate latencyservice curve β(t) = max(0, R(t − T )) with latency Tand rate R. Bounds in this case are simply b

R + T for de-lay bound and b + r ∗ T for queue size bound. Finally,there is an important theorem that describes the burstinessconstraint evolution for each flow: considering a flow thatpasses through a node with an input arrival curve bin anda finite crossing delay D, the output flow is constrained bybout with bout(t) = bin(t + D).

5.1. Definition of the end-to-end delaybound

The end to end delay communication bound of a givenpacket or a traffic class, when crossing one switch, can bedefined as follow (figure 5):

DEED = DSRC + DSW + 2 ∗DPROP

• DSRC is the processing delay for transmission at thesource and it depends in the multiplexer policy in thesource station.

��

��

��

��

��

��

��

Figure 5. Schematic diagram of the end toend communication delay

• DSW is the duration a frame might be delayed in theswitch and is equal to the technological switch relay-ing latency (ts) plus the queuing delay (tq). The lat-ter bound represents the time a queued frame sits inthe queue of the switch output port, including the timeneeded to be emitted on the output cable. DSW de-pends on the scheduling policy used in the switch andan exhaustive calculus of this bound is given later.

• DPROP is the propagation delay needed to propagatethe electrical signal from the source to the switch andthen from the switch to the destination, which is pro-portional to the length of cable connecting the stationto the switch. In our model, this delay is considered asinsignificant.

5.2. Source processing delay

In this part, we explain the source processing delaycalculation DSRC . We consider a subsystem that sends astreams set S = {s1, s2, ..., sn} with an output capacityC. Each stream si is (bi, ri)-constrained thanks to theuse of traffic shapers based on the leaky bucket concept.Moreover,

∑i ri < C is assorted as a stability condition.

In our subsystem model (see figure 2), the multiplexercharacteristics determine the processing policy in the sub-system output. Hence, for a FCFS policy, we choose aFCFS multiplexer and for a SP policy, a 4-FCFS multiplexer(one queue per priority level). To obtain the processing de-lay bounds, we use theorems founded by Cruz ([25]) forthese respective multiplexer behaviors.

• FCFS policy: the following theorem gives an upperbound on delay for an FCFS multiplexer.

Theorem 1 Considering two streams {1, 2} with{C1, C2} the respective input capacities and {α1, α2}the respective arrival curves, the delay of any data bitentering FCFS multiplexer with an output capacity C


from stream 1 is upper bounded by D1, where:

D1 =1C

maxu≤0{α1(u) + α2(u +

L2

C2)− Cu)}

With L2 the maximal packet length in stream 2.

In our case study, all input multiplexer’s links have in-finite transmission rate thanks to the use of buffers inour subsystem model. In fact, according to Cruz ([25]),buffers are characterized by an infinite exit transmis-sion rate . Hence, the term L2

C2is null and the gener-

alization of the precedent theorem for n streams withaffine arrival curves gives an upper bounded delay Dfor any stream si, where:

D =1C

maxu≤0{∑

i

bi + (∑

i

ri − C) ∗ u} (1)

Given that∑

i ri < C, it is easily verified that u = 0is the argument of the maximum in the definitionof D. Consequently, the processing delay for each

stream si has a maximal bound DSRC =∑

ibi

Cthat depends on the burstiness of each stream andthe output capacity. This bound causes a burstinessincrease at the output of the subsystem. In fact, eachstream si is (b∗i , ri)-constrained at the source outputwith b∗i = bi + ri ∗DSRC .

• SP policy: Cruz gives in ([25]) a theorem to calculatethe maximal delay for a non preemptive priority mul-tiplexer.

Theorem 2 Suppose there are n input links to thismultiplexer with α1, α2,...,αn the n respective arrivalcurves for the input traffic. Assume that stream 1 hasthe highest priority and stream n the lowest one andall streams have the same maximal packet length L. Ifeach stream i is (bi, ri)-constrained for i = 1..n, themaximal delay is:

Di =

∑j=1..i bi + L

C −∑j=1..i−1 ri

In our case study, the considered streams set is S ={S1, S2, S3, S4} with Sk the streams subset havingthe priority k. Each priority class k is (Bk, Rk)-constrained where Bk =

∑i∈Sk

bi and Rk =∑i∈Sk

ri and Lk is the maximal packet length in Sk.Hence, the maximal processing delay for each priorityk is:

DkSRC =

∑i≥k Bi + maxi<k Li

C −∑i>k Ri

(2)

5.3. Switch queuing delay

As said before (5.1), DSW = ts + tq , where ts is thetechnological delay and tq the queuing delay. In general,ts is given by the industrials and it depends on the switchcapacity C (at 10 Mbps, ts is about 60µs). In this part, wefocus on the switch queuing delay bound tq that dependsmainly on the scheduling policy. Hence, these bounds foreach flow are analyzed according to whether the policy ofthe node is FCFS, SP or WFQ. First, we determine theservice curve offered to each flow by the part of the switchthat represents the final stage of the forwarding mechanism:the queuing and the multiplexing. Then, giving the servicecurve and the input traffic arrival curve, maximal queuingdelay bounds are calculated using the Network Calculusformalism. The considered streams set destined to a givenswitch output port is E = {e1, e2, ..., en}. Each streamei is (b∗i , ri)-constrained where b∗i = bi + ri ∗ DSRC andDSRC its respective source processing delay.

• First Come First Served policyIn the FCFS policy, the queuing delay depends mainlyon the queue length. The input arrival curve of theglobal traffic is α(t) = σ + ρt where σ =

∑i b∗i and

ρ =∑

i ri and the service curve offered to this ag-gregate traffic is simply β(t) = Ct. Hence, the delaybound is the maximal horizontal deviation between αand β, i.e σ

C .

• Static Priority policyIn the SP policy, the offered guarantees depend on thepriority level. In our case, we have four priorities andthe considered stream set E = {E1, E2, E3, E4} withEk the streams subset having the priority k. Eachtraffic priority class k has an arrival curve αk(t) =σk+ρk∗t, where σk =

∑i∈Ek

b∗i and ρk =∑

i∈Ekri.

The strict priority policy guarantees to a given prioritylevel to be selected before the lower priorities and afterthe higher priorities. However, since the transmissionof a packet on the network cannot be preempted, in theworst case, one packet of maximal length with lowerpriority is served before. The maximum packet lengthbelonging to a priority k will be noted Lk,max.Therefore, the service curve offered to the traffic classwith priority k is given by the rate latency curve

βk(t) = Rk(t− Tk) (3)

Where Rk = C − ∑i≥k ρi the offered rate to the

priority k after serving all the higher priorities; andTk = maxi<kLi,max

C the maximal waiting time whena packet with lower priority is served before. Hence,the maximal horizontal deviation between αk and βk


is σk

Rk+ Tk which is the maximal delay bound for the

priority k.

• Weighted Fair Queuing policyAs said before in 4.3, with WFQ, weights are associ-ated to each priority level and not as known classicallyto each individual stream. In fact, it is more interestingin our case to offer a fair service to each priority levelthat respects its deadline constraints and does not de-pend on the properties of the other priorities. Hence, anappropriate weight φk that respects the stability condi-tion

∑k φk = 1 is considered for each priority level k.

As bellowed, each traffic priority class k has the arrivalcurve αk(t) = σk + ρk ∗ t. To determine the servicecurve offered by the WFQ node for each priority levelk in our case study, the Leboudec’s result [16] con-cerning the service curve offered by a GPS node (thefluid model of WFQ) is used. Assume a (σ, ρ)- con-strained stream with a weight φ entering a GPS nodewith an output capacity C, the associated service curveis β(t) = rt where r = φ ∗ C. In the other hand,the Parekh’s and Gallager’s result [21] on the devia-tion between the delay bound under GPS and the delaybound under WFQ is utilized to make the link betweenthe GPS and the WFQ service curves. This deviationis less than Lmax

C which represents the time to send apacket of maximal length under WFQ.Hence, using these results, we derive the researchedservice curve for each traffic priority class k that is:

βk(t) = ck(t +Lk,max

C) (4)

Where ck = φk ∗ C. The maximal delay bound foreach priority k with WFQ policy is then σk

ck+ Lk,max

Cwhich is the maximal horizontal deviation between αk

and βk.

6 Performance evaluation

6.1. Case study

Our case study is a representative real-time traffic ona MIL-STD 1553B data bus in a modern french militaryaircraft. The selected data bus is the busiest one amongthose aboard this aircraft. Hence, we estimate that it is themost representative one of the MIL-STD 1553B behaviorand its real-time requirements. For these reasons, thissingle case study is considered in this part.

The traffic is circulating between twenty subsystemsconnected to their associated terminals. The transactiontable of the MIL-STD 1553B bus is statically defined insuch a way that time constraints are enforced and terminals

Switch

T1 T2 T3 T4 T5

T6

T7

T8

T9

T10

T11T12T13T14T15

T16

T17

T18

T19

T20

T10T16

UK

QD

YY

JR

TY

JT

HW

UB

J

AP

IKR

SH

2HA

KL

VG

QZJ EQF1 CFV BHT

BYNC WDSQ BHUJ

SXJJ

AFGDHULP

Figure 6. General model based on a real caseof a 1553B network

are polled in a determined sequence. As opposed to thistransmission control approach, Switched Ethernet withtraffic shaping approach is based on the fact that equip-ments can emit their data simultaneously when the trafficis well controlled. Real-time flows are described in tables1 and 2. So, one can see that for periodic messages, thelargest period is about 160 ms and the most frequent valueis 20 ms; and for sporadic messages, there are differentresponse time bounds and the most urgent one is about 3 ms.

As a result, the major frame has a duration of 160 msand minor frames 20 ms, in order to meet the requirementsof the higher update rate messages. For each message,we define a deadline that conforms to our model (see4.1): for periodic messages, the deadline is the periodthat ranges between 20 ms and 160 ms; and for sporadicmessages their maximal response time. We suppose thata subsystem can generate at most one sporadic messageof each type once every minor frame (20 ms). Therefore,the hardest real-time constraint to respect is the maximalresponse time for the most urgent sporadic messages (3 ms).

Table 1. Periodic Flow DescriptionPeriod (ms) Number of flows Lmax(bytes)

20 40 11040 10 7680 15 120

160 100 116

In order to integrate the above traffic in a simple mannerwhen replacing the MIL-STD 1553B data bus with a FullDuplex Switched Ethernet network, a MAC address is


Table 2. Sporadic Flow DescriptionResponse time (ms) Number of flows Lmax(bytes)

3 1 7220 80 120160 90 120

infinity 20 120

attributed to each terminal and the different terminals areconnected to one switch with a 10 Mbps capacity. Every1553 message generated by a 1553 terminal is encapsulatedin an Ethernet frame that respects the minimal framesize (64 bytes) and contains the source and destinationaddresses. Then, this frame is carried over the SwitchedEthernet network. Figure 6 depicts our general model: aswitch in the middle is connected to twenty terminals thattransmit periodic and sporadic messages with determinedcharacteristics.

To validate our proposal, the maximal end to end de-lay bounds are calculated for each traffic class using theanalysis in section 5, and then compared to the associateddeadline. The obtained results are essential to show theefficiency of Switched Ethernet to guarantee deterministictransmission with bounded delays required by ”1553B”-based applications.

6.2. Analytical delay bounds

In order to calculate the analytical maximal end to enddelay bounds, we have developed a Java computing toolbased on our general model. Given the traffic character-istics and the network architecture model in the input, thissoftware determines the arrival curve of each stream at ev-ery point of the network; and the service curve offered byeach network component according to its processing policy.Afterward, it gives the maximal delay bounds that repre-sents the horizontal distance between these two curves. Amaximal end to end delay bound can then be determined.This tool works as described in the followed algorithm 1.

In our case study, given the important number of streamsand the existence of multicast and broadcast transmissionmodes, it was more convenient to give a maximal end-to-end delay bound related to each destination subsystem thanto each individual stream. Hence, initially, the tool iden-tifies the set of received streams at each terminal (line 6).Then, for each stream in the identified set, it determines itsinitial arrival curve (line 9), its associated path (line 10) andservice curves offered by crossed components along thatpath according to their processing policy (line 11). After-ward, the delay bound calculation is propagated from one

Algorithme 1 End to end delay bounds calculus

1: T ← {T1, T2...Tnterminals}

2: S ← {s1, s2...snstreams}3: Policy ∈ {FCFS, SP, WFQ}4: EEDDEST ← NULL-VECTOR(T.length)5: for i = 1 to nterminals do6: R← Vector-rcv-streams(Ti, S)7: EDDstreams ← NULL-VECTOR (R.length)8: for j = 1 to R.length do9: α← Initial-arrival-curve(R(j))

10: Path← Vector-crossed-components(R(j))11: β ← Vector-service-curves(Path, Policy)12: for k = 1 to Path.length do13: D ← Delay-calculus (α, β(k))14: α← Left-shift-curve (α, D)15: EEDstreams(j)← EEDstreams(j) + D16: end for17: end for18: EEDDEST (i)← maxj∈R EEDstreams(j)19: end for

crossed component to another by resolving the burstinessconstraint evolution of each stream. In fact, knowing thearrival curve and service curves, the submitted delay boundis calculated for each stream (line 13) and then its outputarrival curve (line 14). This latter curve will be the inputarrival curve for the next network component and so on un-til the last component. Now, since submitted delay boundsare known for each stream and in each point of the network,a maximal end-to-end delay bound can be determined foreach stream along its path (line 15). Finally, the end to enddelay bound associated to each destination terminal can becalculated as the maximal end to end delay bound amongthose associated to its set of received streams (line 18).

Figure 7. Analytic Maximal End to End Delaybounds with FCFS policy


First, with the FCFS policy, the end-to-end delay boundsfor each stream set, characterized by a destination subsys-tem, are computed and the obtained results are presentedin Figure 7. Clearly, one can see that all end-to-end delaybounds are larger than 3 ms which means that the dead-line constraint for the most urgent sporadic messages is vi-olated. Moreover, traffic streams sent to subsystems 11 and16 have an end to end delay bound larger than 20 ms. Thisvalue represents the smallest period for periodic flows andthe response time of some sporadic flows. Hence, the dead-line constraints associated to these flows are not respected.It is worth to note that this simple policy is the one usedinside the switches of the AFDX network to support civilapplications requirements, where one guaranteed latency toall traffic flows is provided; thus, real-time constraints re-quired by this military application would be violated, de-spite the relative ratio between Switched Ethernet (10Mbps)and MIL-STD 1553B (1Mbps) and the low bus utilization(4%). Hence, the requirements of communication protocolsare distinct in civil and military applications and the AFDXis not designed to support the deadline constraints requiredby the ”1553B”-based applications. Moreover, as opposedto some received ideas, increasing the offered bandwidth isnot sufficient to have a real-time behavior with SwitchedEthernet.

Figure 8. Analytic Maximal End to End Delaybounds with SP policy

In order to achieve the QoS level offered by SwitchedEthernet, several service classes are required to assure agood isolation for urgent messages with hard deadline con-straints. Hence, SP and WFQ policies are selected here toguarantee the priority handling in data sources and in theswitch. The obtained results with SP and WFQ are respec-tively shown in figures 8 and 9. Clearly, the end to enddelay bounds are inherently reduced for the most urgenttraffic class (priority 4) and satisfy the associated deadline

Figure 9. Analytic Maximal End to End Delaybounds with WFQ policy

constraint. Moreover, for each destination subsystem, thepriority classes 3 and 4 have end to end delay bounds lessthan 20 ms which respect the hardest deadline constraint re-quired by these two classes. Obviously, all delay boundscannot improve and as a result low priorities delay boundsgrow compared to the FCFS bounds.To validate our network reliability in a general way, we sug-gest to determine maximal delay bounds in a global man-ner instead of maximal delay bounds related to subsystems.Hence, assume that each priority class k destined to a sub-system i has an end to end delay bound Dk

EED,i. The globalmaximal end to end delay bound associated to each priorityk is then:

DkEED = max

iDk

EED,i (5)

As a result, in our case study, the obtained global boundsare described in table 3 (for each traffic class, only thehardest deadline is considered which represents the smallestdeadline of a traffic class).

Hence, one can see that deadline constraints are violated

Table 3. maximal end to end delay bounds

Priority Hardest Deadline (ms) DEED (ms)FCFS SP WFQ

4 3 21,8 0,63 1,93 20 21,8 9,8 18,62 20 21,8 19,7 191 infinity 21,8 38,8 37,8

with FCFS policy, whereas they are respected with SPand WFQ policies. On the other hand, with the SP policy,the maximal end to end delay bound of high priority isreduced compared to the FCFS policy, with a noticeable


degradation of low priorities. Moreover, with the WeightedFair Queuing policy, a more fair service is offered to allpriority classes with a little amelioration of low prioritiescompared to SP policy. Clearly, this configurable policybridges the gap between the simplest policy FCFS, thatserves packets without taking into account their temporalcharacteristics and causes a real-time constraint violationin our real case application; and the SP policy that makes astrict segregation between priority levels by selecting thehigher priority to be served.

These results show that the priority handling methodcombined with the traffic shaping approach may be a goodmean to improve the Switched Ethernet reliability and toachieve the QoS level required by ”1553B”-based applica-tions.

7 Conclusion

We have shown, through this paper, that Full DuplexSwitched Ethernet is an attractive candidate to replace theMIL-STD 1553B data bus in next generation ”1553B”-based applications. To guarantee an acceptable real-timebehavior of our suggested network, traffic shapers wereintegrated in data sources to assure the traffic control.Then, a priority handling method is added to have a goodisolation level for urgent messages with hard deadlines.This was achieved by using SP and WFQ policies in theswitch. Using the Network Calculus theory, an analyticstudy was conducted to calculate deterministic guaranteesoffered by our proposal in the case of a realistic militaryaircraft network. Obtained results show the efficiency oftraffic shaping with SP and WFQ policies to provide deter-ministic transmission with respected deadline constraints,as required by next generation ”1553B”-based applications,whereas a simpler policy like FCFS failed to provide theseguarantees.

Currently, we are working on the delay bounds op-timization obtained with the traffic shaping approach.Therefore, we are computing WFQ weights associated toeach priority class in order to have better delay boundsfor low priorities than the SP policy. This problem can bemodeled as a multi-objective optimization problem whereweights represent the set of variables; and deadline andpriority constraints the set of constraints to be respected.This assumption will be proved and validated analytically.

The next step will be to evaluate other techniques,like the Time Triggered paradigm [13], over Full DuplexSwitched Ethernet to achieve the real-time behavior re-quired by ”1553B” based applications. Then, their relativemerits will be compared to Traffic Shaping approach

developed in this paper.

References

[1] J. Bogaerts, A. Divia, R. Muller, H. Renardy, and J. Cern.SCI Based Data Acquisition Architectures. IEEE Transac-tions on Nuclear Sciences, 39:85–94, April 1992.

[2] H. Charara, J. Scharbarg, J. Ermont, and C. Fraboul. Meth-ods for bounding end-to end delays on an AFDX network.Dresden, Germany, 2006. Proceedings of the 18th Euromi-cro Conference o n Real-Time Systems (ECRTS06).

[3] Cisco. Cisco switches.[4] M. Cohn. A proposed Local Area Network for next-

generation avionic systems. NAECON, 1988. Proceedingsof Aerospace and Electronics Conference.

[5] A. E. E. Committee. Aircraft Data Network Part 1, SystemsConcepts and Overview, ARINC Specification 664. An-napolis, Maryland, 2002. Aeronautical Radio.

[6] F. Frances, C. Fraboul, and J. Grieu. Using Network Cal-culus to optimize the AFDX Network. Toulouse, 2006.Proceedings of the 3rd European Congress Embedded RealTime Software.

[7] J. Georges, E. Rondeau, and T. Divoux. Evaluation ofSwitched Ethernet in an industrial context by using the Net-work Calculus. Vasteras, Sweeden, 2002. 4th IEEE Interna-tional Workshop on Factory Communication Systems.

[8] A. Gillen and J. Shelton. Introduction of 3910 High SpeedData Bus. San Diego, CA, USA, 1992. Military Communi-cations Conference, MILCOM.

[9] J. Grieu. Analyse et valuation de techniques de commutationEthernet pour l’interconnexion de systemes avioniques. PhDthesis, INP, Toulouse, 2004.

[10] J. Grieu, F. Frances, and C. Fraboul. Preuve de determinismed’un reseau embarque avionique. Montreal, Canada, 2003.Proceedings of Colloque Francophone sur l’Ingenierie desProtocoles.

[11] C. E. Incorporated. MIL-STD-1553 Designer guide. 1982.[12] J. Jasperneite, P. Neumann, M. Theis, and K. Watson. De-

terministic Real-Time Communication with Switched Ether-net. Vasteras, Sweeden, 2002. 4th IEEE International Work-shop on Factory Communication Systems.

[13] H. Kopetz. The Time-Triggered Architecture. In Proceed-ings of the First IEEE International Symposium on Object-Oriented Real-Time Distributed Computing.

[14] H. Kopetz, A. Damn, C. Koza, M. Mulazzani, W. Schwabl,C. Senft, and R. Zainlinger. Distributed Fault-Tolerant Real-Time systems: the Mars Approach. IEEE Micro, february1989.

[15] S.-K. Kweon and K. Shin. Achieving real-time communica-tion over Ethernet with adaptive traffic smoothing. Washing-ton, USA, 2000. IEEE real-time technology and applicationsSymposium.

[16] J. Leboudec and P. Thiran. Network Calculus. SpringerVerlag LNCS volume 2050.

[17] J. Loeser and H. Haertig. Low latency hard real-time com-munication over switched Ethernet. IEEE, Proceedings ofECRTS’04.


[18] A. Mifdaoui, F. Frances, and C. Fraboul. Real-time guaran-tees on Full-Duplex Switched Ethernet for military applica-tions. In Proceedings of the 3rd European Congress Embed-ded Real Time Software, Toulouse, France, 2006.

[19] M. Molle and G. Watson. 100Base-T/IEEE802.12/PacketSwitching. IEEE Communication Magazine, 64-73, August1996.

[20] J. R. Murdock and J. R. Koenig. Open Systems AvionicsNetwork to replace MIL-STD 1553. IEEE AESS Systemsmagazine.

[21] A. Parekh and R. Gallager. A generalized processor sharingapproach to flow control in integrated services networks: thesingle node case. IEEE/ACM transactions on Networking, 1,June 1993.

[22] D. Parish, R. Briggs, D. Chambers, C. Hunter, and N. Kel-sall. 1553Emulation over ATM (Asynchronous TransferMode) A hybrid Avionics Communications Architecture.IEEE AESS Systems magazine.

[23] P. Pedreiras, L. Almeida, and P. Gai. The FTT-Ethernet Pro-tocol : Merging Flexibility, Timeliness and Efficiency . Pro-ceedings of ECRTS, 2002.

[24] L. Pinney. Joint Advanced Strike Technology Program,Avionics Architecture Definition. Arlington, VA, 1994.

[25] R.Cruz. A calculus for network delay, part 1 : network ele-ments in isolation. IEEE transactions on information theory,37, January 1991.

[26] C. Venkatramani and T. Chiueh. Supporting Real-Time Traf-fic on Ethernet. San Juan, Puerto Rico, 1994. IEEE real-timesystems symposium.


Documents

Full Duplex Switched Ethernet for Next Generation ”1553B”-based Applications · 2013. 7. 19. · 2.2. ”1553B”-based applications require-ments Prior to replace the MIL-STD