A PERFORMANCE ANALYSIS OFTHE IEEE 802.5 TOKEN RING … Bound... · token ring optimizes the token...

Philips J. Res. 43,532-553, 1988 R 1199

A PERFORMANCE ANALYSIS OF THE IEEE 802.5TOKEN RING PROTOCOL

by U. KILLAT·) H.-A. MUSCATE·) and B. WOLFINGER ••)·)Philips GmbH Forschungslaboratorium Hamburg, D-2000 Hamburg 54, Germany

··)Camputer Science Department, University of Hamburg, D-2000 Hamburg 13, Germany

AbstractThis paper presents a performance evaluation for token ring based local-area networks. Performance is measured in terms of the delay-through-put characteristics. The investigation is carried out with a detailed sim-ulation model which accurately describes the access mechanism of the to-ken ring protocol as specified in the IEEE 802.5 draft standard. Inparticular, the impact of different values for the token holding timer andof the 'wait my address' mode are studied. The results reflect the behav-iour of the token access method under different load assumptions in-cluding Poisson-type traffic or, alternatively, a 'session-oriented' traffic.The token ring access control mechanism is compared with token bus,CSMA/CD and distributed polling.

Keywords: collision detection, distributed polling, E-net simulation, LAN,local-area network, protocol, simulation, token bus network,token ring network

1. Introduetion

The communication capabilities provided by local-area networks (LANs)are affected, to a large extent, by the performance of the medium accesscontrol protocol. In geographically concentrated areas, communication sub-networks usually realize a bus or ring topology. The IEEE 802 standardiza-tion body has established o.g. three standards for LAN medium access con-trol protocols: CSMA/CD, token bus and token ring 1). In the present paperwe provide a detailed performance analysis of the IEEE 802 token ring pro-tocol.

An often used way to tackle the problem of performance evaluation of anaccess protocol is to create an analytical model, derive an approximate an-alytical solution and then study the limits of the validity by comparision withsimulation studies. Bux 2), for instance, has chosen this way using a mean

532 Philip' JournnI of Research Vol. 43 Nos 5/6 1988

Philips Journalor Research Vol. 43 Nos 5/6 1988 533

A performance analysis of the IEEE 802.5 token ring protocol

delay approximation in his investigation of token ring networks. As it iscommon the arrival processes he assumed are of Poisson-type, but the sourcesmay have different arrival rates.

In the area of polling systems many queueing models have been studiedwhich are also applicable to token passing schemes 3). Several investigationsof token ring networks made use of these queueing models 2,4.5.6).Token passing schemes were studied under various service disciplines.

Some analyses used exhaustive service strategies 2.7), whereas others inves-tigated a 'round robin' strategy, in which each node having captured the to-ken is permitted to transmit exactly one packet t-ê-").

Some papers included the impact of additional delay caused by a busydestination station in their investigation 8.9). The performance of token ringnetworks in the case of short-packet applications was studied by Loucks etal. 6).Not much work has been done so far in validating the models by perform-

ance comparison with actual networks. The Poisson-type packet arrivalsusually assumed permit closed form analytical solutions but quite often areno good approximation of the load requests observed in a typical imple-mentation of a token ring network 10).

Some approaches studying the performance of access protocols simply re-strict on determination of maximum throughput instead of delay versus of-fered load analysis, because modelling can then be done by using only firstorder statistics. For instance, a study of the characteristics of a backbonenetwork, based on several access methods which are similar to LANs, in-cluding a token passing strategy, can be found in ref. 11. A backbone net-work in that context is defined as a communications structure interconnect-ing various LANs.The major objective of this investigation is to gain a deeper understanding

of the delay-throughput characteristics of the token ring protocol by meansof simulation studies using a quite detailed local-area network model. Inparticular, we have been able to include a timer facility and a 'wait my ad-dress' mode, both specified in ref. 1, in our model. Moreover, different trafficgenerators which allow investigating the behaviour of a token ring networkunder Poisson-type or 'session-oriented' load have been implemented.

Of course, in simulation studies assumptions have to be made on the nu-merical values of the particular parameters concerned. In the following someperformance criteria will be presented as a function of offered load for val-ues of the relevant system parameters. The results of the simulation runs areinterpreted in qualitative and quantitative terms. Though in general thetrends of the effects are not unexpected, the facility to predict the size of the

U. Killat, H.-A. Museate and B. Wolfinger

effects is extremely useful and at least instructive.Section 2 of this paper presents a short description of the IEEE 802 token

ring protocol. In sec. 3 we briefly describe the communication system modelused. Section 4 presents and discusses the results obtained from the simu-lation runs in which the sensitivity of token ring performance for various pa-rameter values is tested. The last section includes a performance compari-son of the token ring protocol with the three competing protocols: Distributedpolling, token bus and CSMA/CD, the latter two being also LAN standardsproposed by IEEE 802.

2. Token ring based local-area networks

The IEEE 802 token ring protocol for local-area networks with a ring to-pology uses a token mechanism to control the medium access. A station ter-minates its own transmission cycle by sending a special bit pattern (token)to its successor on the ring. The token is passed around the ring until itreaches a station with a transmission request. The IEEE 802 specification oftoken ring optimizes the token handling by converting a received token onthe fly to the header of the first packet.

Generally the service discipline of the transmission queues is exhaustivebut in order to prevent a very busy station from holding the token for a verylong time, the IEEE 802 standard supplies a token holding timer (THT),limiting the maximum length of a transmission cycle. This timer facility pro-vides increased fairness (equal access to the ring) between the stations, shar-ing the communication medium. If in the remaining THT timer interval noneof the packets in the transmission queue can get served, the station termi-nates its transmission cycle and releases the token prematurely. The effectof different values for THT timer on some performance characteristics, e.g.on packet waiting times, will be analysed in sec. 4.3.The IEEE 802 token ring protocol standard also introduces a packet-ori-

ented priority scheme. As a major consequence of this, a 'wait my address'operation mode is added to the protocol specification. Although we haveneglected the distinction between priority classes, our investigation studiesthe impact of the 'wait my address' mode in sec. 4.2. When transmission ofthe packets has been completed, the station checks whether at least one framewith the station's address in the 'source address' field has returned. If this isnot the case, the station has to defer the transmission of a token ('wait myaddress') until the required event occurs.

Due to the fact that in local-area networks the influence of transmissionerrors on performance evaluation is marginal, we have neglected in the model

534 Philips Journalor Research Vol.43 Nos 5/6 1988

Phllips Journal of Research Vol. 43 Nos 5/6 1988 535


and simulation studies presented here the corresponding error handling fea-tures specified in the IEEE-standard.

3. Models for token ring networks

In our investigation we used the evaluati.on net (E-nets, see ref. 12) ori-ented modelling system FORCASD (FORtran-based Computer-Aided Sys-tem Design tool, 13)) supporting the description of E-net modules and theirimplementation on computer systems. E-nets are derived from Petri-Nets andintended to supply a universal, illustrative and easy to apply means of sys-tem modelling and simulation.

3.1. Model description based on evaluation nets

A detailed description of E-net system modelling can be obtained fromref. 13. Here we give only a short overview.The evaluation net shown in fig. 1 consists of:- transitions, that represent the activities,- locations, that represent the interconnection elements (token residences)

between the transitions and finally- tokens (information carriers) that represent the dynamic modelobjects

including their information content (attributes).A transition itself is described by:- the transition scheme, that represents the transition logic with respect to

the token flow,- the procedure describing the transition process (optional), representing the

model activities that are executed at this transition in order to affect thetoken attributes,

- the transition time t (t;;. 0), that represents the time being necessary toexecute the model activities associated with that transition,

- a resolution procedure in the case of transitions of type conflict or queue(see below); the resolution procedure permits data dependent decisionswith respect to the token flow according to the net's status or the currentvalues of the token attributes,

- a queue size specification in the case of a queue transition (see below) inorder to limit the maximum number of tokens in a queue.A transition is said to be 'enabled' if necessary conditions at their input

and output locations are fulfilled. If enabled it 'fires' i.e. the activities rep-resented by that transition are executed. This means in particular that afterelapse of the transition time first the tokens are moved from the input lo-cations according to the rule of that transition type. Secondly the attributes

---§@--

U. Kil/at, H.-A. Muscate and B. Wolfinger

transition of typetransit: Tl.T9.ilOselection: T3.TlIdistribution: TS.il2absorber: T6queue: T2. 14. T7. TB

queue 5 izes:T2: 600T4: 300T7: 250TB: 1000

transition time parts:0.0: Tl ..... T4.T6.T7.TB.ill'TRA (transml ss Ion time of a

packet): TS'TC (length of a transmission

£ycle): T9" •'WHA(length of the ".!!.alt !!!Y ~ddress"

phase): TlO'JIRiP (propagation time of the token

to the next station with atransmission request): Tl2

resolution parts (Indicated by a hexagon 0):SELECT: selects packets of the current active

stat IonFC Implements a FCFSservice desclplineRANDOM:Implements a RANDOMservice desclpllneHQ: !_ests, whether lransmission g_ueue T..

is emptyTE: lests, whether all queues are !.lT1lty

locations (Indicated by a directed rectangleD):Input: Input location of the E-netExit: exit location of the E-netlocations representing states of the Token Ringnetwork (compare Fig. 2):Zl, ... ,Z4 and, in addition,ZS (only of auxiliary significance)

process parts:PI: Initllizes the attributes of a new packet In the systemPS: ill1l1ements the premature fire event of transition T9 in the case

that the queue transition T4 is erJ1lty (means "end of transmission cycle")

Fig. 1. E-net defining a token ring network.

of the tokens are (possibly) changed according to the transition process.The communication network model implemented includes two parts:a token ring system model,a traffic generator model.

536 Phlllps Journalof Research Vol. 43 Nos 5/6 1988

Phllips Journal of Research Vol. 43 Nos 5/6 1988 537


3.2. Token ring system model

We have defined the token ring system model by means of E-nets. But aswe assume the reader not necessarily to be familiar with E-nets, we will startthis section with an incomplete but easy to interpret model description interms of a queueing network and a state diagram. This description is fol-lowed by the explanation of the E-net-type model used in our studies.Figure 2 shows the simplified model of the token ring system. New data

packets from the packet generator have to wait at server SI (see fig. 2a). Allpackets of the presently active (token holding) station get to server S3' The

al queueing netr--------..,packets from: SI i r: -,

the traffic ---+-~' II I Igenerator L.. .J

;- III -~DlI l~ JL- ...J

,...--------,I II S4 II :::::m----:-=orpacket exitL J

SI,S2: Serves only packets of the actual active station (if any).Service time = O.

S3: If the model is not in state ZI, packets are switched to S2,otherwise, if S4 is not busy (the queue of S4 is empty), to S4.Service time = O.

54: The max. number of packets at 54 equals 1. Packets never have towait in the queue of 54'Service time = ~TRA.

bl state diagram

states:ZI: active transmission cycle

transitions:TeE: !_ransmission gcle ~nd

MAR:".!!!y2_ddress" r.eceived

F: successor found

Z2: "wait my address"

Z3: determine successor

Z4: wait arrival of newpacket in system

NF: successor .!lot found

NP: arrival of .!lew£_acket in system

Fig. 2. Simplified model of a token ring LAN.

3.3 Traffie generatorPoisson arrival processes are the traffic patterns usually assumed in local-

area network models. We too have implemented a 'Poisson traffic' modelsupplying all stations of the LAN with load generating processes (packetgenerators) that create packets whose interarrival times are exponentiallydistributed.Poisson traffic may be a satisfactory valid approximation for transaction-

oriented applications. However, very often nodes in computer networkscommunicate after having established a common session, e.g. of interactive


system is now in state Z, (for the definition of the states see fig. 2b). In thecase that the timer doesn't force the station to terminate its transmission cycleprematurely, all data packets at S3 are propagated one by one via S4 to theirdestination. Each transmission lasts the (service) time "'IRA per packet. Inthe other case the remaining packets at S3 (after timer THT has expired) arebackscheduled to server S2'

After the end of a transmission cycle, i.e. the time interval spent in stateZl' the system gets in state ~ (see fig. 2b). Having terminated the 'wait myaddress' phase (compare sec. 2) the token ring protocol permits the actualstation to release the token. The system tries to find a successor station. Ifall queues are empty, this proceeding is delayed (state Z4) until the nextpacket arrives from the traffic generator. The time needed by the token totravel from one station to the next is taken into consideration in any case.Figure 1 shows an E-net model of the token ring system. The subnet com-

prising the set of transitions {T1, ... ,T8} models the packet flow; the subnetspanned up by the set of transitions {T9, ... ,Tl2} models the flow of the(LAN) token. The queue transitions TI, T8 and T4 correspond to the queuesof the server stations SI' S2 and S3 respectively (see fig. 2a). The process partPlof transition Tl serves to intialize the attributes of the tokens (packets)generated by the traffic generator. The new packets get into the token ringsystem via the location INPUT and finally leave the model at the absorbertransition T6. T3 schedules the packets from the queues TI respectively T8to transition T4. As long as the model remains in state ZI (see fig. 2b) tran-sition T5 switches the packets to T6 choosing a delay time of "'IRA (see fig.1) per packet. Otherwise the packets are propagated to T8 via TI.If the queue of the actual station is empty, transition T5 (process part)

fires TI in order to terminate the transmission cycle and to change the sys-tem state to ~. At the end of the 'wait my address' phase the state is changedto Z3' After the time TJUMP (see fig. 1) has passed a new transmission cycleis started, if possible, otherwise the system gets into state Z4'

538 Phllips Journol of Research Vol. 43 Nos 5/6 1988

Philips Journalof Research Vol. 43 Nos 5/6 1988 539


or file transfer type. We took this fact into account by providing the 'sessiontraffic' as a second traffic pattern model (compare refs. 14 and 15).

Session starts are generated by a Poisson process. Within a session (theactivity phase of a station) another Poisson generator of different intensitygenerates packets. The length of a session depends on the number of pack-ets per session (NPS). NPS is a uniformly distributed random variable takenfrom the interval (l..NPSMAX).The packet length is determined by a fraction of constant length (NCBY)

and a fraction which is assumed to be exponentially distributed (with meanNMYBY). Common parameters (the values refer to the values used in thisinvestigation) of our traffic generators are: Number of Sources NS=lOO,Number of Constant portion data BYtes per packet NCBY E {0,1O}, Num-ber of variable portion (negative-exponential distribution with mean f.L) dateBYtes NMYBY €{100,1000}. In our investigation we have considered pack-ets with different mean packet lenghts (L): small packets (NCBY = 10,NMYBY = 100) and large packets (NCBY = 0, NMYBY = 1000).

3.4. Verification of the token ring network model

We have tested our simulation model by comparing it with the analyticalmodel used by W. Bux in his LAN performance study ê). Of course, such acomparison requires a simplified configuration of the token ring simulationmodel. This means, in particular, that the 'wait my address' mode and thetimer facility had to be neglected. Furthermore, the comparison includes onlythe mean delay time of a packet and the mean transmission cycle length T.Under these restrictions we have achieved an almost ideal conformity of theresults of both models. The packet delay times evaluated by the analyticalmodel lie clearly within the confidence intervals evaluated by our simulationmodel. The relative error of the mean transmission cycle time T, defined asthe difference between calculated and simulated T divided by calculated T,stays well belowO.S %.In addition, the FORCASD debugging facility turned out to be a powerfultool allowing the test of the model activities in a stepwise execution mode.By means of this tracing facility we were able to test our token ring modelfor the more general case, where THT and WMA were included.

4. Simulation results for token ring networks

The performance of an access protocol is characterized mainly by its av-erage delay - throughput characteristics. Simulation runs have been carriedout to evaluate the dependence of average waiting time on totaloffered load.

U. KWat, H.-A. Museate and B. WoLfinger

0.ü10 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

P -------Fig. 3. Token jing protocol: mean waiting time E (W) versus p, curve-parameter: meanpacket length (L) [byte]; - Poisson traffic; --- session traffic; ..... analytical results; /:i

Poisson traffic, 1 bit latency per station.

The waiting time W is defined as the time interval between packet gen-eration at the traffic generator and the start of the successful transmission.

The offered load p is defined as the aggregated packet lengths per timeinterval, including all control information representing administration over-head, such as source and destination addresses, CRC check sequences etc.,normalized with respect to the channel capacity. Mean waiting time 95%-confidence intervals are indicated in the corresponding figures (figs. 3, 10 to13) as far as they do not lie within the marks of the curves.The values of the four most relevant technical parameters of the token ring

LAN assumed for the simulation experiments are'- Transmission rate: 10Mbit/s corresponding to one of the standard values.- Latency per station: 10 bit.

Latency per station is defined as repeater delay plus signal propagationdelay. For a good performance of the IEEE 802 token ring protocol the ringlatency should be negligible with respect to the mean transmission cyclelength. The ring latency is equal to the number of stations times the latencyper station. For LANs normally a value for latency per station of 1 to 3 bitsis used (see for example refs. 5 and 6), though, from an implementation pointof view, a higher value may be advantageous. A backbone network, inter-

100

IIIE

~UJ

540

10

0.1

~110

Philip. Journalof Research Vol. 43 Nos 5/6 1988

100

A performance analysis of the IEEE 802.5 token ring protocoL

1lilE

I~

10

0.1

0.Q10 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0p----<-

Fig. 4. Token ring protocol: standajd deviation of waiting time (.6.W) versus p, curve para-meter: mean packet length (L) [byte]; --Poisson traffic; --- session traffic.

connecting a set of (heterogeneous) LANs which are locally concentrated,may need 5 bits or even more (see ref. 11). We have chosen 10 bits to coverthe worst case.- Token holding timer (THT) interval length: 10ms which is the value rec-ommended by the IEEE 802 standard') as a default timeout value. This lim-its the transmission cycle here to 105 bits.

Fig. 5. 'Wait my address'-time (WMA) versus p, curve-parameter: mean packet length (i)[byte]-- Poisson traffic; --- session traffic.

50 WMA [us)

40

30

20

10 .L~ _==~~~~~~~~~~- P0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Philips Journalor Research Vol. 43 Nos 5/6 1988 541

U. KWat, H.-A. Muscate and B. Wolfinger

4.1. Experiments for token ring networks with different traffic characteristics

As one might intuitively expect, mean waiting time E (W) increases withoffered load p, as confirmed in fig. 3. The larger the mean packet length Ithe higher are the standard deviations of packet length and packet interar-rival time distributions. Therefore, I affects the delay strongly, especially atlow offered loads, where the influence of the traffic pattern is negligible, upto a load of about 0.4, as fig. 3 shows. The higher the load, however, themore the differences in waiting time for session traffic and Poisson traffic in-crease. This effect is due to the fluctations in the arrivals of session trafficand is still more obvious when the standard deviations is measured, as shownin fig. 4.

4.2. Experiments for token ring networks operated in the 'wait my address'mode

We next study the impact of the 'wait my address' operation mode. Figure5 visualizes, that the WMA-time defined as the delay caused within a stationby this operation mode due to deferring token transmission, mainly dependson the mean packet length (I). Sufficiently high transmission cycle lengthsT compared to ring latency totally prevent any 'wait my address' overhead.Relatively large packets (;;;,:1000 bytes) and/or high queue occupations of thestations avoid significant performance losses. This is in particular true for thehigh station latencies which we have chosen in this investigation and con-firmed in fig. 5.

The highest WMA-times occur in the case of Poisson traffic and shortpackets. Poisson traffic requires many token movements from station to sta-tion. In a first approximation under this traffic pattern assumption eachtransmission cycle consists only of a single packet even for high loads as fig.6 shows. With short packet length most of the time is consumed in waiting

N

65432

~~~~~~~~~~~p0.1 0.2 Q3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Fig. 6. Mean number of served pgckets per transmission cycle (N) versus p, curve-parame-ter: mean packet length (L) [byte); - P.oisson traffic; --- session traffic. .

542 Philips Journalor Research Vol. 43 Nos 5/6 1988

100


1tilEKJ

10

0.1

0.010 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0p---

Fig. 7. Average cycle time eversus p, curve-parameter: mean packet length CL) [byte];- Poisson traffic; --- session traffic.

up for the first returning packet to about p = 0.6. For higher load the WMAtime is reduced because of the increased length of the transmission cycles.Figure 3 also illustrates the negative influence of the 'wait my address' op-eration mode on performance characteristics under these transmission cycleare quite low, a packet waiting in a queue suffers from all WMA-times ofstations that will get the token earlier than the station of this packet.This behaviour can be observed in terms of the cycle time C defined as

the time between subsequent visits of the token to a station queue, shownin fig. 7. At low load the minimum value is found, approximately equal tothe ring latency. Here again the 'wait my address' overhead is reflected in arelatively high average cycle time in the case of Poisson traffic' and shortpackets.We conclude from additional experiments (see fig. 3) that performance loss

due to the 'wait my address' mode is negligible if the latency per station isreduced to one bit.

4.3. Experiments for token ring networks with different timer interval lengtn

We next investigate the influence of the token holding timer (THT) in-terval length on performance characteristics. The IEEE 802 standard rec-ommends a default time-out value of 10ms.

Phllips Journalof Research Vol. 43 Nos 516 1988 543

500,,11

,,//'.2~~=r=-~ -

1oo~-L __L--L __L--L~~-L __ ~~ __ ~0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9_ 1.0

P

/IIII//l

/_JOOD _------

-~---. I

U. Kil/at, H.-A. Museate and B. Wolfinger

3000 TI fJS I

2500

2000

1500

1000

Fig. 8. Mean transmjgsion cycle length Cf) versus p, curve-parameter; mean packet length(L) [byte]; - Poisson traffic; --- session traffic.

Due to the low value of the mean number of served packets per trans-mission cycle eN) in the case of Poisson traffic as shown in fig. 6, in mostsituations a station serves only one packet during a transmission cycle. Inthis case the impact of the timer is negligible because, with the token ringprotocol, a (non-empty) transmission cycle includes at least one completepacket, regardless of its length.Figure 8 shows the mean transmission cycle length Cl) versus offered load.

The time-out value may influence waiting time significantly only if the meantransmission cycle length is not too small. Therefore, session traffic togetherwith large packets is the only critical case. This conclusion is confirmed bythe simulation results shown in fig. 9.For small time-out values and session traffic the performance decreases (see

fig. 9). This may be expected, because the timer causes additional tokenmovements which would not occur without timer. But it may be surprisingthat token ring considering the mean waiting time performs better with a welladjusted timer than with no timer (i.e. infinite time-out value). The expla-nation of this minimum for long and the plateau for short packets in fig. 9is not self-evident. Short packets have a better chance to get served in theremaining timer interval. Therefore, token ring tends to favour small pack-

544 Philip. Journalof Research Vol. 43 Nos 5/6 1988


16000

14000

2- 12000~UJ 10000

8000

6000

0.00

\110\,"'.... .......... _ ..-.-- ..--------

1000

time - out intervallength lusl t

recommended valuenot optimal

Fig. 9. Mean wgjting time E (W) versus time-out intervallength, curve-parameter: meanpacket length (L) [byte]; -- Poisson traffic (p = 0.9); --- session traffic, (mean packet

length = 110 at p = 0.85, mean packet length = 1000 at p = 0.7).

ets. Consider the MIMll single server model. In such a system the best serv-ice discipline (minimizing the mean waiting time) under non-preemptive as-sumptions is SJN (SJN = shortest job next, see for example 16). Strategiesthat favour short packets, in this case token ring with a THT, create smallermean waiting times than service time independent disciplines like token ringwithout a THT do. This effect surpasses the performance loss due to addi-tional token handling overhead for sufficiently high time-out values. Oursimulation experiments have shown, that this conclusion is robust also withrespect to lower latencies per station (e.g. one bit latency per station). Thisresult is not unexspected since timing out enforces additional token passingand a low ring latency reduces token passing overhead. From fig. 9 it canalso be seen that the time-out interval length of 10 ms recommended in thestandard") is not the optimum value.

5. Simulation results comparing token ring with token bus, CSMAlCD anddistributed polling

Though the main subject of this paper is a study of the token ring protocolitself as a function of the various parameters, the E-net simulation can also



be applied to evaluate other medium access methods under the same cir-cumstances. The token ring 17) and CSMA/CD 17) are two other widely ac-cepted methods, similarly standardized in IEEE 802.We assume the reader to be familiar with the token bus and CSMA/CD

protocols, a short description of which can be found in 14).An additional access control mechanism can be derived from a token bus

scheme. As it is well-known token bus represents a system in which eachstation maintains an address .to which the token is passed thus creating anvirtual ring. This virtual ring can be dynamically changed by integrating orexcluding stations. Considering the token bus as a stable ring whose mem-bers include all stations attached to the system, we have another accessstrategy. According to 14) we call this strategy 'distributed polling'.Based on the specifications of ref. 17 FORCASD simulation models al-

ready had been developed 14) for distributed polling, token bus, andCSMA/CD. Simulation runs have been carried out on comparable topolo-gies and with identical traffic assumptions for the models of token ring, dis-tributed polling, token bus, and CSMA/CD based LANs. The performanceof the token ring is contrasted in a number of relevant aspects to the per-formance of the other three.

100

lilE

i ,

J I .Ji !_///

.../ .._ ..-- .-1"-'

.---- •.-.- I,/

10

0.1

0.010 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0p---

Fig. 10. Poisson traffic: mean waiting time E (W) versus p, mean packet length (L): IlObytes token bus protocol; CSMA/CD protocol; _._._ distributed polling; - to-

ken ring protocol; 6. token ring, 1 bit latency per station.

546 Philips Journul of Research Vol. 43 Nos 5/6 1988

Philips journal of Research Vol. 43 Nos 5/6 1988 547

A performance anaLysis of the IEEE 802.5 token ring protocoL

100

10

.l I·I /i , I,../ j.,./. ....>:__ r·..,/......... - /.::::::~.._._..- ~

",';

"K:IJlE

~UJ 0.1

0.Q10 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0p---

Fig. 11. Poisson traffic: mean waiting time E (W) versus p, mean packet length (i): 1000bytes ....... token bus protocol; --- CSMA/CD protocol; _._.- distributed polling; -. - to-

ken ring protocol; l!. token ring, 1 bit latency per station.

Most of our results for token ring LANs are based on the estimate of alObit latency per station. We have complemented these results for token ringby some mean delay values in the corresponding figures, referring to 1bitlatency per station.

Figures 10 to 13 show the mean waiting time for token ring, as in fig. 3,compared now to the mean waiting time for distributed polling, token busand CSMA/CD, all for the same four relevant parameters as above.

5.1. Token ring versus distributed pollingIn spite of the same nature of distributed polling and token ring, figs 10

to 13 show that the latter performs significantly better. Even in the case ofPoisson traffic and short packets, where token ring suffers strongly under thenegative influences of the 'wait my address' operation mode, the result is thesame. This superiority of token ring results from the improved token han-dling of token ring relative to distributed polling.

In particular, the complete bus is blocked when a station is sending thetoken to its successor on the virtual ring in the case of distributed polling.However, in the case of a token ring LAN, signals can be present concur-

100


0.010~--:OL---:O"'=2---='0-:--~----:''':----='''=---:'------:'-----'----'.1 . .3 0.4 0.5 0.6 0.7 0.8 0.9 1.0p---

Fig. 12. Session traffic: mean waiting time E (W) versus p, mean packet length (L): 110bytes token bus protocol; CSMA/CD protocol; _._._ distributed polling; -- to-

ken ring protocol; Ä token ring protocol; 1 bit latency per station.

rentlyon different sectors of the physical ring thus allowing an earlier trans-fer of data. Token ring takes advantage of this fact by converting the tokento the header of the first packet of a new transmission cycle.

IIIE

!LU

~/

/ po' ./., .'."I' ,y'...~/ <..,

I ...-I ....... "._·_.·t·- .....•._._.- / .....

I .I....··.....··r...........I

I

10

0.1

5.2. Token ring versus token bus

In the case of Poisson type traffic the performance of token bus is verypoor (figs 10 and 11). This is not surprising, because under these load con-ditions token bus requires time consuming changes of the virtual ring forevery packet transmission. For session traffic the waiting time characteristicis improved (figs 12 and 13) but token ring still behaves better.

5.3. Token ring versus CSMA/CD

CSMA/CD is another competitor of the token ring protocol. At low loadsthe latter performs worse because mean waiting times depend on the tokenbeing available (figs 10 to 13). The difference is quite small, however, anddecreases with shorter latency. Beyond a load of p = 0.3 the token ring per-forms better because of the effect of collisions in CSMA. In order to explainsome of the differences between token ring and CSMA/CD-LANs mean

548 Philips Journalof Research Vol. 43 Nos 5/6 1988


0.Q10 8 9 00.1 0.2 0.3 0.4 0.5 0.6 0.7 O. 0. 1.p---

Fig. 13. Session traffic: mean waiting time E (W) versus p, mean packet length (I): 1000bytes token bus protocol: --- CSMA/CD protocol; _._._ distributed polling; -- to-

ken ring protocol; ~ token-ring, 1 bit latency per station.

Fig. 14. Difference in mean waiting time illCSMA/CD and token ring protocol versus p,-- P..Qissontraffic, mean packet length (L): 110 bytes; Pojgson traffic, mean packet

length (L): 1000 bytes; session traffic, mean paçket length (L): 110 bytes; _._._ sessiontraffic, mean packet length (L): 1000 bytes.

100

11)

E

§:UJ

E (WCSMAlCD )-E (WTR) [ms]

0.1

/,f" ~,f {'

/ ./ ..,

J"J('"./ .'I <>

_jo .''I.7 ......_..,........ ..,...- .....

I .'~.. '····i..................

0.1

-0.1

Phllips Journol of Research Vol. 43 Nos 5/6 1988 549

100


0.010 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0p---

Fig. IS. Poisson traffic: Stgjrdard deviation of waiting time (Ll.W) versus p, curve-parameter:mean packet length (L) [byte]; --- CSMA/CD protocol; -- token ring protcol.

IIIE

I~

l// ,,-I ,1/I "1./

~"",,"" I .

'"1000 ,.'"

10

0.1

packet waiting times are displayed in detail in fig. 14. For high loads the be-haviour of CSMNCD is dominated by the packet collisions on the bus. Ata critical load waiting time increases asymptotically with offered load.Not only the mean value E (W) but also the standard deviation of waiting

time is of importance. The standard deviation of waiting time as observedin token ring is always smaller than in CSMA (figs 15 and 16). Separate runsshow that under all boundary conditions investigated in our simulations,CSMNCD is superiour to token ring with respect to standard deviations ofpacket waiting times only for offered load values below 0.1.

Under the assumption of one bit latency per station, the performance oftoken ring is improved (see figs 10 to 13). With respect to the mean of wait-ing time, in this case the region where token ring outperforms CSMA/CD isalso extended into the area of low loads (p;;;' 0.2).

6. Conclusions

The objective of this investigation was to present a performance evalua-tion of the IEEE 802 token ring protocol, which illustrates in an easily ac-cessible way in quantitative terms earlier studies of this protocol. The majorconclusions are:

550 Philips Journol of Research Vol. 43 Nos 5/6 1988


0.010 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0p---

Fig. 16. Session traffic: stgndard deviation of waiting time (ÄW) versus p, curve-parameter:mean packet length (L) [byte]; --- CSMA/CD protocol; -- token ring protocol.

a) Token ring performs quite well over the whole range of traffic loads con-sidered. This result is robust with respect to different packet lengths andtraffic patterns. The token ring protocol is less sensitive to variations inoffered traffic than the other methods. From a performance point of viewnone of the protocols CSMA/CD, token bus or distributed polling be-haves as favourable as token ring, except the CSMA/CD protocol underlight load.

b) Due to the 'wait my address' operation mode token ring suffers a signif-icant loss of performance in the case of Poisson traffic, short packets andhigh ring latencies. On more realistic traffic pattern conditions like ses-sion traffic, 'wait my address' can be accepted which, of course, is a pre-requisite to the priority protocol extension. This assertion is valid inde-pendent of the ring latency 1). The 'wait my address' overhead is negligiblein the case of one bit latency per station.

c) The timer (THT) influences the token ring performance significantly onlyfor high session-oriented traffic loads. In the case of session traffic andlarge mean packet length CL) a well adjusted timer permits a moderatebut worth mentioning improvement of the waiting time characteristics.Unlike earlier performance evaluation studies for token ring based LANs,

100

UIE

I~

10

0.1



our model description has not been achieved by means of queueing net-works. (For the use of queueing networks for computer system and com-puter network modelling cf. e.g. Kleinrock 16), Hayes 18) and Schwartz 19)).In this paper a class of extended Petri nets has been applied for modelling.This alternative description method allows a concise and complete specifi-cation of computer network models.

REFERENCES1) IEEE Project 802, Local Area Network Standards Draft E, August 1984.2) W. B ux, IBM Res. Rep. RZ 1199 (1982).3) A.G. Kohnheim and B. Meister, J. ACM, 21, 470 (1974).4) R. Cherukuri, L. Li and L. Louis, Proc. Comp. Netw. Symp., New York, 57-68, 1982.5) W. Bux, IEEE Trans. Commun., COM-29, 1465 (1981).6) W.M. Loucks, U.C. Hamacher, B.R. Preiss and L. Wong, IEEE Trans. Comp.,

C-34,,1006 (1985).7) I. Rubin and L.F. M. DeMorase, IEEE J. Selected Areas Commun., Los Angeles, Cal-

ifornia, 1983, 935-947.S) A.K. Agrawala, J.R. Agre and K.B. Gordon, Proc. COMPSAC 1978, Chicago 1978,

p.674-679.9) A. S. Sethi and T. Saydam, Proc. 9th Conf. on Local Computer Networks, IEEE Cat-

alog No. 84CH2081-8, Silver Spring, MD, 1984, 26-31.10) J. Sventek, W. Greiman, M. O'Del! and A. Jansen, Proc. Computer Networking

Symp., Silver Spring, MD, 1983.11) A. Danthine and E. Vyncke, ESPRIT Techn. Week 1984, Network, Amsterdam, North

Holland, 1985, 337-353.12) G.J. Nutt, Fall Joint Comp. Conf. 1972, AFIPS Conf, Proc. 41, 279 (1972).13) N. Dahmen, Proc. First European Simulation Congress, ECS83, Aachen, Springer Ver-

lag, Berlin, Heidelberg, New York, 1983.14) N.' Dahmen, U. Kil!at and R. Stecher, Proc. Sec. Intern. Symp. Performance of Com-

puter Communication Systems, North Holland, 1984, 79-94.15) B. Wolfinger, Intern. Conf. IEEE INFOCOM'86, Miami 1986,609-615.16) L. Kleinrock, Queueing Systems, Vols. I and I1, J. Wiley and Sons, New York 1975 and

1976.17) IEEE Project 802, Local Area Network Standards Draft 0, December 1982.IS) J. Hayes, Modeling and Analysis of Computer Communications Networks, New York,

Plenum Press, 1984.19) M. Schwartz, Telecommunication Networks - Protocols, Modeling and Analysis, Addi-

sion-Wesley Publ, Comp., Reading Mass., 1987.

AuthorsUlrich Kil!at; Diplom-Physiker, University of Hamburg, Germany, 1969; Ph.D., Universityof Hamburg, Germany, 1973; Philips Forschungslaboratorium Hamburg, 1973- . He worked inthe fields of optical storage, optical communications, distributed switching methods, broadbandswitching, and performance evaluation. Since 1983 he is heading the Telematics Group of thelaboratory.

Hans-Achim Muscate; Diplom-Informatiker, University of Hamburg, 1986; Philips For-schungslaboratorium Hamburg, Germany, 1985- . His work was concerned with modellingtechniques and tools for performance evaluation. Since 1986 he has been working in the fieldsof office automation systems induding document processing systems.

Bernd Wolfinger; 'Maitrise de Mathématiques et Applications Fondamentales' (AppliedMathematics), Université Claude-Bernard, Lyon, France, Diplom-Mathematiker, Universityof Karlsruhe, Germany; Ph.D. (Computer Science), University of Karlsruhe, Germany; Nu-

552 Phillps Journalof Research Vol. 43 Nos 5/6 1988



clear Research Center, Karlsruhe, Germany, 1975-1980; Department of Computer Science,University of Karlsruhe, Germany, 1980; Professor for Computer Science at the University ofHamburg, Germany, 1981- . During 1985, Professor Wolfinger spent a sabbatical with IBM Th.J. Watson Research Center, Yorktown Heights. His primary research interests include the de-sign and the analysis of computer networks, formal descriptions of communication protocols,adaptive routing algorithms, modelling techniques and tools for performance evaluation ofcomputer systems.

A PERFORMANCE ANALYSIS OFTHE IEEE 802.5 TOKEN RING … Bound... · token ring optimizes the token...

Documents

Transcript of A PERFORMANCE ANALYSIS OFTHE IEEE 802.5 TOKEN RING … Bound... · token ring optimizes the token...