Deterministic Real-Time Communication with Switched Ethernet

J rgen Jasperneite Peter Neumann Michael Theis and Kym Watson u , ,

Abstract
In this paper we show, under which conditions switched ethernet can be called deterministic.The inuence of the service strategies in switched ethernet is investigated by simulation and compared with analytical bounds on the delays.

cient to assess the characteristics of distributed real-time systems. This paper therefore evaluates the real-time characteristics by looking at distributions and upper bounds on transaction times. Simulation techniques and the analytical method Network Calculus (originating from the engineering of Internet networks) are both applied. The results yielded by these methods are discussed and compared.

2 1 Introduction
Ethernet is an internationally standardized local area network technology [IEE00]. Although characterized as having a behavior unsuitable for control level network applications under moderate to heavy trafc loads, much of ethernets non-deterministic reputation derives from the stochastic nature of its collision recovery mechanism. In todays factory communication systems ethernet is used for non-time-sensitive applications. A major step toward deterministic behavior in Ethernet networks is to eliminate the random CSMA/CD bus arbitration. This can be achieved by using the latest Ethernet switch technology instead of a hub-based infrastructure and by directly connecting systems to full duplex ethernet switches. Switch technology divides collision domains into simple point-topoint connections between network components and stations. Collisions no longer occur and the random backoff algorithm is no longer required. The concept of switching or Media Access Control (MAC) bridging, which was introduced in standard IEEE 802.1 in 1993, was expanded upon in 1998 by the denition of additional capabilities in bridged LANs. The aim was to provide additional trafc capabilities so as to support the transmission of time-critical information in a LAN environment [IEE98a], [IEE98b]. By using simulation, [JN01] showed the quantitative inuence of system parameters like scheduling strategy and thinking time within the devices on the typical time behavior in a switched Ethernet. Mean values are not suf J rgen Jasperneite is with Phoenix Contact GmbH, System Develu opment Department, Flachsmarktstr. 8-28, 32823 Blomberg, Germany, Fax:+49-5235-3-39999, e-mail: [email protected] Peter Neumann is with IFAK Institute for Automation and Communication Magdeburg, Steinfeldstr. 3, 39179 Barleben, Germany, Fax:+49-39203-81100, e-mail: [email protected] Michael Theis and Kym Watson are with Fraunhofer Institute IITB, Fraunhoferstr. 1, 76131 Karlsruhe, Germany, Fax:+49-721-6091-413, e-mail: {the, wat}@iitb.fhg.de

Case Study

In order to meet the requirements for small, medium and large congurations the station number N is taken from the set {10, 50, 100} for all scenarios. The simulation model was realized with OPNET Modeler 7.0B from OPNET Technologies, Inc., by extending the standard models for switches and MAC to incorporate priority queuing. In the simulation experiments only full duplex transmission channels (FDX) with a transmission capacity of C = 100M bps are used to guarantee a collisionfree system. In the switches and devices the scheduling strategies FCFS (First Come, First Served) and PQ (Priority Queuing) are considered. Packet processing is asumed to be store and forward (i.e. a packet must be completely received before the transmission to subsequent switch or station may begin.) In addition, we assume that packet processing times in a switch or station are negligible. 2.1 Topology Switches enable free network topology conguration. Of the numerous types of topologies only the star and line topology are considered in the scenarios here. The simplest topology is the star topology because it only has one switch. However, the star topology requires maximum cabling which is of concern to automation technology. Originally, eldbus systems were developed in order to replace parallel cabling of sensors and actuators. For this reason we look at a line topology in which every switch is assigned to one device. This creates a eldbus-compliant cabling concept. In future devices this switch function may be part of the device itself. This will reduce costs for cabling and infrastructure. Due to the large number of switches the line topology is the most unfavorable topology for real-time behavior. 2.2 Load Parameters In order to create a data trafc mix the following four message types (M Ti ), which can be found in every indus1

SW1

Star-Topology

Feature Transaction type

ST1 ST2 ST3 ST4 STN-1 STN

SW1

..

SWk ..

Switch Fullduplex Connection

Line Topology

ST1

ST..

STN

Station

Device with integrated Switch

Figure 1. Considered topologies

trial application, are dened from the application viewpoint. There are cyclic as well as acyclic event types. When evaluating the performance of a network topology we also have to consider the information ow distribution dened by the application. The information ow distribution describes every communication relationship between the devices within a message type. As we can see in actual congurations the central information ow distribution (C) is very important in the eld of automation technology. This is the case if there is one special device in the system which is involved in all communication relationships, either as source or sink. This form applies to the application (part) near the process in which a central control system is used. In the simulation experiments we assume a uniform distribution of the destination addresses. 2.3 Scenarios In these synthetic scenarios a typical automation application is used which consists of a central control system and distributed intelligent eld devices. Within the M T1 message type the cyclic process data exchange forms the major part of the data with its short user data lengths typical of eld communication. In addition, events, such as for example alarms from the eld devices to the central control system, are transmitted within the M T2 message type. The scenarios also include acyclic parameterization and network management (M T3 ) as well as transfer of large data amounts (M T4 ) from the distributed eld devices to the central control system.
Type M T1 M T2 M T3 M T4 Description Cyclic send and request of process data objects with response Transmission of events, e.g., alarms Acyclic transmission for, e.g. network monitoring, diagnostics or device conguration Acyclic transmission of datagrams for le transfer or IP-based applications

Distribution information ow Percentage of packets uM ti Distribution packet generation Distribution destination addresses Length MAC payload [Bit] Distribution Length MAC payload user priority

MT1 Conrmed Client: ST 1 Server: ST 2 ..ST N Central (ST 1 ) 80%

Message Type MT2 MT3 Unconrmed Conrmed Publisher: Client: ST 2 ..ST N ST 1 Subscriber: Server: ST 1 ST 2 ..ST N Central Central (ST 1 ) 10% (ST 1 ) 9.5%

MT4 Conrmed Client: ST 1 Server: ST 2 ..ST N Central (ST 1 ) 0.5%

D D (cyclic) Lreq : 368 MT

1

R D Lreq : 368 MT
2

R R Lreq : 1024 MT
3

R R Lreq : 1024 MT
4

Lres : M T1 D

368 D Medium (6)

Lres : M T3 D High (7)

1024

Lres : M T4 D Low (4)

12000

Standard (5)

Table 2. Dening the workload

Table 1. Denition of message types

The distributions used for the packet interarrival times and the packet destination addresses are either deterministic (D) or random (R) (cf. Table 2). The metrics are shown as a function of the system load. The system load is dened as the ratio of arrival rate to service rate of a service system, whereby the service rate is the link capacity C. Since a switched Ethernet comprises numerous segments with individual dedicated service systems we want to focus on the segment with the highest load in the sense of a bottleneck analysis. Due to the central information ow distribution of the selected workload, the physical link between station ST 1 and the port of the corresponding switch has the highest load. The service system at the port of the switch to which the central station is connected has the highest load due to the asymmetrical workload. Therefore, it is considered as reference for the system load from now on. The relative system load is within the range 0 < < 1. As stated in [Jas02] the generation of transactions of a given type on a given connection is constrained by a leaky bucket regulator (refer to the denition below). In the scenarios we assume that with increasing arrival rate of the sub-application M Ti the burstiness also increases. Besides the ow rate rM Ti , the bucket size bM Ti of the leaky bucket arrival curves is increased with increasing load . Based on the analysis of communication requirements [JN00] the bucket size is chosen between bmin = 2 packets for = 0.05 and bmax = 12 packets for = 1.2. We rst describe the analytical approach using Network Calculus and subsequently consider the results using simulation. 2

3 Network Calculus
3.1 The principles of Network Calculus Network Calculus is a widely applicable technique to assess the real-time performance of communication networks and is based on the fundamental work of [Cru99a], [Cru99b] and [GP93], [GP94]. For a full introduction the reader is recommended to consult [BT01]. Network Calculus basically considers networks of service nodes and packet ows between the nodes. Whereas traditional queuing theory deals with stochastic processes and probability distributions, Network Calculus involves bounding constraints on packet arrival and service. These constraints allow bounds on the packet delays and work backlogs to be derived, which can be immediately used to quantify the real-time network behavior. Traditional queuing theory, on the other hand, normally yields mean values and perhaps quantiles of distributions. The derivation is often difcult and upper bounds on end-to-end delays may not exist or be computable. The packet arrival process in Network Calculus is described with the aid of so-called arrival curves which quantify constraints on the number of packets or the number of bits of a packet ow in a time interval at a service node. Typical packet ows of interest are, for example, the packets of a given message type entering or departing from a service node. Let F be a packet ow and let R(t) be the number of packets of F arriving in the time interval [0, t]. We say the ow is constrained by or has arrival curve (t) if, for all 0 s t R(t) R(s) (t s) whereby (t) is a non-negative, non-decreasing function. Although one can consider packet arrival constraints, it is mostly more convenient in the context of this paper to consider the corresponding work arrival curve. We say that a packet P requires W [sec] work in the service node N , if N needs this long to process P . Typically, W can be calculated from the packet length (including overhead), the transmission rate and interframe gap on the medium. Let Rw (t) be the amount of work due to ow F which arrives [0, t]. Then F is constrained by a work arrival curve w (t) if, for all 0 s t Rw (t) Rw (s) w (t s) A service curve describes the service of ow F in node N . Let Rw (t) be the amount of work completed in the time interval [0, t]. Then the non-negative, non-decreasing function (t) is called a (minimal) work service curve for F in N if, for all t 0, Rw (t) = Rw (t) inf {Rw (s) + (t s) : 0 s t}
s

where denotes the deconvolution operator. In a packet network it is important to consider the packetization of the output. If packets ow from N1 to N2 then work arrives at N2 for further service only when the complete packet has been received. If (t) is a work service curve of a packet ow and max is the maximum service time of packets of the ow, then 1 (t) = ([t max ]+ ) is a work service curve in N1 for the ow including packetization of the output. The constraints given by an arrival and service curve for a ow sufce to calculate an upper bound on the delay of a packet in the service node and on the work backlog. The packet delay is bounded by the horizontal distance between w and , whereas the work backlog is bounded by the vertical distance: delay sup{inf{s 0 such that (t) (t + s)}} (1)
t0

work backlog sup{(t) (t)} (2)

t0

Computations with arrival and service curves can be greatly facilitated if concave arrival and convex service curves can be applied. The most important representatives are the leaky bucket arrival curve (t) = rt+b with rate r and burst or bucket size b and the rate latency service curve (t) = max(0, (tT )R) with latency T and rate R. Both are determined by just two parameters. The above delay bound in this case is simply T + b/R (if r R), which manifests the dependence on the service latency and the burst size. The output ow is constrained by the leaky bucket arrival curve (t) = (t + T ). 3.2 Results for FCFS and PQ service nodes We give two specic results on the Network Calculus for service nodes serving two or more ows according to a FCFS or PQ service strategy [refer to [BT01], Corollary 6.2.1 and Corollary 6.4.1, and [Wat02] for these and more general results]. Proposition 3.1 [FCFS Split] Consider a service node N serving two packet ows F1 and F2 in FCFS order. Assume that N gives the aggregate ow a rate latency work service curve (t) = max(0, (t T )R) and suppose that F2 is constrained by the leaky bucket work arrival curve 2 (t) = r2 t+b2 . Let = T +b2 /R. Dene 1 as follows: 1 (t) = [(t) 2 (t )]+ f or t = 0 f or 0 t <

Then 1 is a rate latency work service curve for F1 with rate R r2 and latency .

where denotes the convolution operator. The work output ow is itself constrained by the arrival curve w (t) = sup{w (t + s) (s) : s 0}
s

Proposition 3.2 [PQ Split] A service node N gives prioritized packet ows Fi (i = 1, , P ; Fi has higher priority than Fj whenever i > j) non-preemptive priority 3

service. Let i be the service time of packets of ow Fi for i = 1, , P . Suppose that each Fi is constrained by the leaky bucket work arrival curve i (t) = ri t + bi . For j = 1, , P dene Lj
H rj

= =

max{i : i < j} ri
{i:i>j}

bH j

=
{i:i>j}

bi
H max{t rj t Lj bH , 0} j

j (t) =

H If rj < 1, then j is a rate latency work service curve for Fj .

One could in principle use the above results directly to do a network calculus of the star and line topologies considered in this paper. However, the iterative application of the FCFS and PQ split calculations along a chain of service nodes yields unacceptably high delay bounds due to the rapid increase in burst and latency values. The way around this problem is to consider the service curve offered to a ow in a complete subsystem. If ow F with work arrival curve is given a work service curve 1 in node subsystem S1 before entering node subsystem S2 where it is given work service curve 2 , then F is given the work service curve = 1 2 in the combined system S = S1 S2 . The Network Calculus can then be done using the input arrival curve and the service curve in S. Compare the pay bursts only once principle in [BT01], 1.5.3. 3.3 Network Calculus for the line topology In the case of the line topology we sketch the procedure used in [Wat02] to nd suitable ows and subsystems. For the line topology the packet ow Fk,i in a message type (class) k with conrmed service is for i = 2, ..., N ST 1 (request) SW 1 SW i ST i ST i (response) SW i SW 1 ST 1 and for unconrmed service ST i (data) SW i SW 1 ST 1
Let Fki = j>i Fk,j be the aggregate ow of packets in class k which are transmitted from SW i to SW i+1 for 1 < i < N. Let Si denote the subsystem ST 1 (request) SW 1 SW i for i = 1, , N . Suppose that we have derived a service curve for the conrmed service ows Fk,i and Fk,i in Si1 . Note that Fk,i1 has a different output port in SW i1 and is served in parallel. We rst compute a work arrival curve for Fk,i at SW i by considering the packetized output of this ow from Si1 . We

then apply the PQ and FCFS splits at SW i to derive ser vice curves for Fk,i+1 and Fk,i+1 , rstly in SW i , and then in Si by concatenation. By iteration we thus nd service curves for the Fki in the subsystem Si . Using the arrival curve for this ow at ST 1 , we can then compute a delay bound for the ow Fki on the path ST 1 SW i as well as an arrival curve for Fki at ST i where the response packets are generated. The unconrmed packet ows have to be added at ST i . We can compute work arrival curves Aw (Fki , ST i ) at ST i by considering the packet output of Si . For the highest priority packets (or for all packets in case of FCFS), we can calculate the delay bound on the outbound path from ST 1 to ST N as follows. Let r[packet/s] and b[packet] be the packet rate and burst parameters of a leaky bucket arrival curve constraining the total highest priority ow of request packets at ST 1 . Let req and rsp be the service times of its request and response packets respectively. The delay of highest priority packets from ST 1 to ST N is bounded by
out out DH = L + breq + N max

where L is the maximum service time in ST 1 of a lower out priority packet (0 if there are none) and max is the maximum service time of an outbound request packet (cf. [Wat02]). On the inbound path from the ST i back to ST 1 it is advantageous to aggregate the higher priorities. Let p(k) denote the priority of class k and let Fki = {Fjm : p(j) p(k), m i} be the aggregate of all ows of priority p(k) through SW i . For i = 1, , N dene the subsystem
in Si = {ST j : j i} {SW j : j > i}

The output of this subsystem is the input for SW i . The in input of ow Fki to Si is constrained by the work arrival curve
in Aw (Fki , Si ) = {j,m:p(j)p(k),mi}

Aw (Fjm , ST m )

For a class k let k be the maximum service time of an inbound packet of priority at least p(k), and let k be the maximum service time of an inbound packet of priority less than p(k). We set k = 0 if there are no lower priority packets. Dene the service curve w (k, t) = [t k k ]+ Then w (k) is a packetized work service curve for Fki in SW i . It is shown in [Wat02] that the aggregate ow Fki in in Si is given the packetized work service curve
in Bw (Fki , Si ) = w (k) w (k) (N i+1)f old

 The input of ow Fki to SW i is constrained by the work arrival curve

10 transaction time [ms]

MT1 STAR FCFS N=50

2

MT2 STAR FCFS N=50

10 transaction time [ms]
2

in Aw (Fki , SW i ) = Aw (Fki , Si )

in Bw (Fki , Si )

14.36 10
1

101 2 100 101 102

12.93

3.07 100 101 102 0.05 0.50 rel. load p 0.95

Let m be the class with the largest priority less than p(k) and let Gmi = {Fmj : j i} be the aggregate ow of packets of priority m which traverse SW i for 1 i N . We rst apply the PQ split at ST N to calculate a service curve for GmN :
in Bw (GmN , SW N , t) = [t m Aw (FkN , SN )(t)]+

0.05

0.50 rel. load p

0.95

MT3 STAR FCFS N=50

10 transaction time [ms]
2

MT4 STAR FCFS N=50

10 transaction time [ms]
2

We then apply the PQ split at the service node SW i to derive a service curve for Gmi for 1 i N : Bw (Gmi , SW i , t) = [t m Aw (Fki , SW i )(t)]+ The concatenation of these service curves yields a service in curve for Gm1 in S1 SW 1 , from which we can compute a delay bound for the inbound path. The sum of the outbound and inbound delay bounds for each conrmed priority class gives the desired end-to-end delay bound. It is shown in [Wat02] that the end-to-end delay of the highest priority packets (or for all packets in case of FCFS) is bounded by
out out DH = DH + (b + rDH )rsp + N rsp + (N + 1)

10

13.47 3.9

101

10.35

14.7

100 101 102 0.05 0.50 rel. load p 0.95

100

101 0.05 0.50 rel. load p 0.95

Figure 2. Box plots of the time distribution for star topology and FCFS

where is the maximum service time of an inbound packet of lower priority (0 if there are none).

4 Real-Time Evaluation
4.1 Simulation Results For evaluating the real-time behavior of switched Ethernet, mean values are not sufcient because they only allow a statement of the typical time response. Therefore, the distributions of the absolute transaction times recorded by simulation are considered rst. Fig. 2 shows the absolute transaction times (globally1 ) determined by simulation as a function of selected relative system loads in a so-called box plot for the star topology with a medium device number (N = 50) and FCFS scheduling strategy. Load and system parameters shown in Table 2 apply. The processing time within the stations and the network delay in the switches are set to TT H = TS = 0. Box plots present a compact graphical visualization of the frequency distribution of data records on the basis of a 5-number summary (pentagram) [Sch98]. The rectan gle (box) includes the rst quartile X0.25 up to the third 0.75 and thus comprises 50% of all data in the quartile X center of a distribution. The horizontal line in the rect angle locates the median X (second quartile). The upper and lower horizontal line represent the maximum and minimum values of the distribution. The additional circle is the mean value X. Furthermore, the values for the 95% percentile (dot-and-dash line) and the maximum (dashed 5

line) are shown for a load of = 0.8. Generally, all distributions in Fig. 2 have a positive skew. The collapsed rectangle for the message types M T1 M T3 clearly shows that for low system loads the transaction times come close to the mean value. For real-time evaluation the upper extreme values are of vital interest. As an example a deterministic upper limit (i.e. absolute upper bound) and a statistical upper limit with ZM ax = 0.95 (i.e. distribution percentile) are shown for a system load of = 0.8. For the chosen load of = 0.8 the result for all message types and the star topology is that a deterministic upper limit less than 15 ms can be guaranteed for a single transaction. For the statistical time limit the value for M T1 M T3 is less than 4 ms in 95% of all cases and equal to 10.36 ms for M T4 . For the line topology (Fig. 3) the central area of the transaction times has a much greater distance to the lower extreme value than in the star topology (Fig. 2). This wider distribution leads to accordingly higher mean values of the transaction times. Due to the high number of switches the jitter is already at a high level even if system loads are small. In the line topology an upper deterministic limit less than 19 ms can be kept for all message types at a load of = 0.8. For the statistical time limit the 95% percentile for M T1 M T3 is less than 6 ms and equal to 13.03 ms for M T4 . Attention should be drawn to the fact that there is a big difference between the corresponding mean values for both topologies and a comparatively small difference for the extreme values determined for a medium load. In comparison, Figs. 4 and 5 show the box
1 In this context globally means that the diagrams comprise the transaction times of all links within one message type.

MT1 LINE FCFS N=50

102 transaction time [ms] 101 100 101 102 0.05 0.50 rel. load p 0.95 transaction time [ms] 18.57 5.43 102

MT2 LINE FCFS N=50

16.71 101 100 101 102 0.05 0.50 rel. load p 0.95 4.81

MT3 LINE FCFS N=50

10 transaction time [ms] 10
2

MT4 LINE FCFS N=50

10 transaction time [ms]
2

18.26
1

5.71

13.03 101

18.59

100 101 102 0.05 0.50 rel. load p 0.95

100

101 0.05 0.50 rel. load p 0.95

Figure 3. Box plots of the time distribution for line topology and FCFS

and M T1 also have a better time treatment with regard to the absolute delay time and the jitter. This is clearly shown with the smaller rectangles and the maximum values. The distance between the 95% percentile and the absolute upper limit for = 0.8 and above indicates that there is only a small number of large values in the distribution (long tail). However, the message type M T4 is given a much less favorable service than in the FCFS scenario beginning from a system load of 50%. This behavior is not critical because this message type has a throughput-oriented denition and does not include time-critical applications (best effort). The priority settings effectively isolate the inuence of the long data packets in M T4 on the time-sensitive transactions of the message types M T1 to M T3 with their short data packets. When applying PQ in line topologies the maximum value for the high-priority message type M T3 is only weakly dependent on the load. This effect becomes less as the priority drops. As expected the jitter increases due to the numerous switches to be traversed. This leads to the wider distribution of transaction times, which we already know from FCFS and the line topology.
MT1 Line PQ4 N=50
102 transaction time [ms] 101 100 101 102 transaction time [ms] 102 10.19 5.04 101 3.92 100 101 102 0.05 0.50 rel. load p 0.95 0.05 0.50 rel. load p 0.95

MT2 Line PQ4 N=50

MT1 STAR PQ4 N=50

102 transaction time [ms] 101 10
0

MT2 STAR PQ4 N=50

102

6.49

9.32

transaction time [ms]

101 100 101 102 0.08

3.01

0.16 101 102 0.05 0.50 rel. load p 0.95

0.05

0.50 rel. load p

0.95

MT3 Line PQ4 N=50

102 transaction time [ms] 101 4.22 100 101 102 0.05 0.50 rel. load p 0.95 transaction time [ms] 102

MT4 Line PQ4 N=50

49.38 62.56

MT3 STAR PQ4 N=50

102 transaction time [ms] 101 10
0

MT4 STAR PQ4 N=50

102 transaction time [ms] 56.86 75.98

6.61

101

101

100

0.8 0.11

101 102

100

101 0.05 0.50 rel. load p 0.95

101 0.05 0.50 rel. load p 0.95 0.05 0.50 rel. load p 0.95

Figure 5. Box plots of the time distribution for line topology and PQ4

Figure 4. Box plots of the time distribution for star topology and PQ4 4.2 Analytical Delay Bounds Especially when considering the rare upper extreme values relevant for real-time processing, the question arises as to whether the absolute upper bound can be determined with simulation experiments. Without doubt the maximum value determined by simulation can occur. However, we have to take into account that even higher values may occur when the simulation runs are longer or when the start conditions are changed. In any case 6

plots when the PQ scheduling strategy with four trafc classes (P Q4) is applied. The preferential service given to the high priority message type M T3 in the star topology (Fig. 4) is a typical feature. The rectangles of the box plots have collapsed in the entire load range selected. Even the maximum values are much smaller than in the FCFS scenario. In the selected load prole the message types M T2

Line PQ4 N=50

max. transaction time [ms] max. transaction time [ms] 104 10
3

Star PQ4 N=50

103
MT1 MT2 MT3 MT4

P[t0,t]
Packet Generation

MT1 MT2 MT3 MT4

10

102 101 100 0.0 0.2 0.4 0.6 0.8 1.0

t0
10
1

t
WS

time t

100 0.0 0.2 0.4 0.6 0.8 1.0

rel. load p

rel. load p

Packet Service

min

ps

S idle S busy

Line FCFS N=50

max. transaction time [ms] max. transaction time [ms] 103
MT1 MT2 MT3 MT4

Star FCFS N=50

103
MT1 MT2 MT3 MT4

t0

t1

ts

time t

Figure 7. Bad case scenario The delay of pS from its generation until reaching SW is at least D0 (pS ) = tS t . Therefore the delay of pS until leaving SW is at least

10

10

101

101

100 0.0 0.2 0.4 0.6 0.8 1.0

100 0.0 0.2 0.4 0.6 0.8 1.0

D(pS ) D0 (pS ) + D(pS , S) WS + t1 t Now t1 t0 + min , where min is the minimum time required by a packet to reach SW. Depending on the network topology and service strategies, it is in general possible to construct packet generation scenarios with larger values for t1 . We conclude that D(pS ) WS + t0 t + min In order to obtain the highest possible delay of a data packet (bad case), scenarios must be found which maximize the right hand side part of the above inequality. The analytical method assumes that on all connections (even within a message type) a burst can be generated at the same time. This results in a high work WS in the interval. However, in the simulation model the connections within a message type are not congured independently. There is always a nite time t > 0 between the bursts on the different connections within a message type M Ti . In the simulation method the work WS is therefore generated within a larger time interval t t0 than in the analytical method and thus leads to smaller time bounds D(pS ).

rel. load p

rel. load p

Figure 6. Upper bounds on the transaction times determined by analysis

the simulation results allow the statement that an upper time limit set by the application is exceeded if its value is smaller than the maximum value determined by simulation. The analytical method presented above was applied to derive sure upper bounds on the transaction times. Fig. 6 shows the analytical results for the transaction time bounds as a function of the relative system load (and implicitly the burst value) for the star and line topologies as well as for the P Q4 and FCFS scheduling strategies. For a load of = 0.8 the analytically determined values are 2 to 40 times greater than the maximum transaction times determined by simulation. Let us discuss the reasons for these deviations by means of a so-called bad case (but possible) scenario for the FCFS scheduling strategy: Consider a set P of packets generated in the time interval [t0 , t ] which are to be served in the FCFS switch SW, (SW 1 for the line topology). The packet paths before reaching SW may be arbitrary. Let WS be the work in the Service Node SW. We will consider the delay of the packets in the system until leaving SW. Now let t1 be the rst arrival of a packet of P at SW. Note that this packet need not be the rst generated. Let tS be the last arrival of a packet of P at SW. Note that this packet need not be the last generated. Denote it by pS . Cf. Fig. 7. In the interval [t1 , tS ], work amounting to at least WS arrives at SW. In the same interval, the amount of work executed is at most tS t1 . Hence the work remaining at time tS is at least WS (tS t1 ). Since SW is FCFS, the delay of pS in SW is at least D(pS , S) = WS (tS t1 ). 7

5 Summary
The analytical method presented in this paper used Network Calculus techniques to derive deterministic guarantees for bounds on transaction times, if the worst case source behavior can be characterized. The simulation results conrmed the applicability of these bounds for scenarios with equivalent source behavior. The presented results show that, with regard to maximum transaction times the results of the analytical method applied here show noticeable deviations from the results of the simulation method. We therefore have to be careful if simulation with the necessary use of random generators is the only method applied for real-time evaluation. We

can only be certain that a limit value set by a technical process is exceeded if the maximum value determined by simulation is higher than this limit value. An open issue in general is the derivation of a quality metric for the upper bounds determined by the Network Calculus approach. However, in some situations it is possible to calculate bad cases which show that the Network Calculus upper bound is only moderately higher than the worst case. Simulation is a useful approach to generate bad cases. It was also shown that using more than one trafc class with user priorities leads to a good isolation of timesensitive and time-uncritical data. In this case the upper time bound of high-priority applications can be reduced by a value comparable with FCFS scheduling, but only at a the cost of the low-priority application. The question is still open as to whether switched Ethernet has real-time capabilities or not. This primarily depends on the requirements of the technical process and cannot be answered in general. But it was shown that in a typical application with N = 50 devices maximum transaction times on the MAC level within milliseconds could be guaranteed for real-time services.

[IEE00]

IEEE, New York. Carrier sense multiple access with collision detection (CSMA/CD) access method and physical layer specications, 2000. ANSI/IEEE Std 802.3-2000 Edition (ISO/IEC 8802-3:2000(E)). J. Jasperneite. Performance Evaluation of a Class-of-service based Local Area Network for using at the Field device level. Internal Research Report, Phoenix Contact, 2002. J. Jasperneite and P. Neumann. Measurement, Analysis and Modeling of Real-Time Source Data Trafc in Factory Communication Systems. In 2000 IEEE International Workshop on Factory Communication Systems, pages 327 334, Porto, Portugal, Sept. 2000. J. Jasperneite and P. Neumann. Switched Ethernet for Factory Automation. In 8th IEEE International Conference on Emerging Technologies and Factory Automation, pages 205212, Antibes - Juan les Pins, France, Oct. 2001. R. Schlittgen. Einf hrung in die Statistik u Analyse und Modellierung von Daten. Oldenbourg Verlag, M nchen/Wien, 1998. u

[Jas02]

[JN00]

[JN01]

[Sch98]

References
[BT01] Le Boudec and P. Thiran. Network calculus: a theory of deterministic queuing systems for the Internet. In Lecture Notes in computer science, volume 2050. Springer Verlag, 2001.

[Wat02] K. Watson. Network Calculus in Star and Line Networks with centralized communication. Technical report, Fraunhofer-Institut f r u Informations- und Datenverarbeitung (IITB), Karlsruhe, April 2002. Bericht-Nr. 10573.

[Cru99a] Rene L. Cruz. A Calculus for Network Delay, Part I: Network Elements in Isolation. IEEE Transactions on Information Theory, 37:114 131, Jan. 1999. [Cru99b] Rene L. Cruz. A Calculus for Network Delay, Part II: Network Analysis. IEEE Transactions on Information Theory, 37:132141, Jan. 1999. [GP93] R. G. Gallager and A. K. J. Parekh. A generalized processor sharing Approach to ow control in integrated services networks: the singlenode case. In IEEE/ACM Transaction on Networking, volume 1, pages 344357, June 1993. R. G. Gallager and A. K. J Parekh. A generalized processor sharing Approach to ow control in integrated services networks: the multiple-node case. In IEEE/ACM Transaction on Networking, volume 2, pages 137150, April 1994.

[GP94]

[IEE98a] IEEE, New York. Media Access Control (MAC) Bridges, 1998. ANSI/IEEE Std 802.1D-1998. [IEE98b] IEEE, New York. Virtual Bridged Local Area Networks, 1998. ANSI/IEEE Std 802.1Q-1998. 8