UPC Measurements in ATM Networks

Transcription

1 UPC Measurements in ATM Networks Christian Schuler, Vassilios Courzakis, Monika Jaeger, Lutz Mark GMD FOKUS Berlin Hardenbergplatz Berlin schuler@fokus.gmd.de July 1996 Keywords: ATM, UPC, policing, leaky bucket 1 Abstract An important element of the Quality-of-Service concept in ATM networks are the Usage Parameter Control functions (UPC). They enable a future broadband network to control the behaviour of potential customers. The UPC procedures take the necessary actions to enforce the compliance of an ATM connection to a negotiated traffic contract. While the UPC mechanisms are defined in detail in the various ATM and B-ISDN specifications, the application, analysis and control of these complex algorithms are still open issues. In this paper we introduce some methods for the analysis of the UPC functions of an ATM switch, which can be used in a broad range of test configurations. We provide detailed measurement concepts for the transparency of UPC mechanisms and show some interesting results for a reference network setup. 2 Introduction: Who needs UPC and why? In the concept of ATM networks the support of various service qualities is an important issue, because it is one of the advantages of ATM compared to other networking technologies. In order to achieve a guaranteed service quality QoS, the user has to commit to a specific traffic contract during connection setup. The parameters of this traffic contract should be guaranteed by the service provider. They will be one of the main factors a potential customer will be charged for. On the other hand, the service provider needs to control if the customer s cell stream entering the network really behaves compliant to the negotiated traffic contract. If not, the network equipment has to take appropriate measures to protect the network from intentionally or accidentially misbehaving users. These so called policing functions are well defined in the ATM and B-ISDN specifications under the category Usage Parameter Control, abbreviated UPC. The UPC functions must be performed on virtual connection basis at the user s entry point to the network, which will be the port of a public or private ATM switch. UPC functions are required for all service classes (CBR, VBR and ABR) except for Un Bit Rate traffic class UBR. For the latter neither bit rate nor QoS are guaranteed and therefore UPC is not mandatory. It should be mentioned, that this traffic class is the most commonly used class today. This has several reasons: - UBR it is easy to implement, because it needs no UPC or priority functions in the switches - in the LAN sector, most of the traffic today is data. For data the QoS is not as important as for audio or video connections - many of today s ATM networks are operated under very low load compared to the available capacity. In this case QoS is not an important issue - today there is no charging for reserved bandwidth, while most of the existing ATM networks are either testbeds or LAN s. If users will have to pay for the bandwidth they use, they will demand for guarantees If the ATM technology of today will migrate from LAN s to B-ISDN and WAN networks, there will be a stronger need for QoS and thus UPC. Another factor will be the integration of realtime video and audio services, which is as well one of the promises of ATM. In this paper we will first give a short overview of the UPC concepts defined in the related specifications [1],[5]. In section 4 we explain our methodology for the traffic parameter selection, which we used in the experiments described in section of 11 GMD-FOKUS

2 3 Overview UPC Concepts: TM 4. Review In the following experiments we will relate mainly to the ATM Forum Traffic Management 4. Standard [1], which in turn is derived from the ITU I.371 Specification [5]. Most of the concepts are the same as in the UNI 3.1 standard and are transferred to the TM 4. standard in version 4. of the UNI specification. While our main purpose is not to repeat the procedures described in detail in the standard, we will give a short summary of the concept.the ATM network can be used for diverse applications with different traffic characteristics. ATM-Forum s Traffic Management Specification 4. defines five different service categories: Constant Bit Rate (CBR), for applications with a static amount of traffic and tight delay variation constraints. Variable Bit Rate (VBR), for applications with bursty traffic characteristics. In the TM 4. there are distinguished real-time VBR (rt-vbr) for applications with constrained delay and delay variation requirements, and non-real-time VBR (nrt-vbr) for non-real-time applications. Un Bit Rate (UBR), for non-real-time applications which do not have to specify traffic related service requirements. Available Bit Rate (ABR), for applications which can cope with changing network resources during data transfer. Depending on the service category the user describes the traffic charachteristics and the QoS-requirements with different parameters during conneciton setup. UPC has to guarantee a certain QoS and has to control the user s traffic. The UPC function deals with the traffic parameters which are the same for rt-vbr and nrt-vbr. Therefore we do not have to distinguish them and in this study the results are valid for both rt-vbr and nrt-vbr. The UPC functions are located at the public or private UNI, which will be represented by the input port of an ATM switch. The switch needs to determine, if the cells submitted by the user are conforming to the negotiated traffic contract. Attribute CBR rt-vbr nrt-vbr UBR ABR Traffic Parameters PCR, CDVT PCR SCR, MBS, CDVT SCR MCR QoS Parameters p-t-p CDV un un un maxctd un un un CLR un network specific PCR: Peak Cell Rate SCR: Sustainable Cell Rate MBS: Maximum Burst Size CDVT PCR : Cell Delay Variation Tolerance for PCR traffic CDVT SCR : Cell Delay Variation Tolerance for SCR traffic MCR: Minimum Cell Rate MBS: Maximum Burst Size p-t-p CDV: peak-to-peak Cell Delay Variation maxctd: maximum Cell Transfer Delay CLR: Cell Loss Ratio Table 1:ATM Service Category Attributes of 11 GMD-FOKUS

3 The conformance definition for ATM cells is based on one or more instances of the Generic Cell Rate Algorithm GCRA (see figure 1). This algorithm can be implemented either as Virtual Scheduling Algorithm or as Continuous- State Leaky Bucket Algorithm, both implementations must lead to the same results. Each GCRA has two parameters, the Increment I and the Limit L, which correspond loosely spoken to the size and the draining rate of an imaginary leaky bucket. ters. Which parameters are needed and thus how many GCRA Algorithms are needed to support the UPC functions, depends on the service class. Each service class might commit to several different conformance definitions, which are network specific. For some classes the Cell Loss Priority bit in the cell header is to be considered. Cells with CLP= might be tagged, if they are nonconforming. This means that the CLP bit is set to 1 in the output cell stream instead of discarding the cell. Arrival of a cell k at time t a (k) The GCRA parameters are derived from the traffic parameters, which in turn depend on the desired QoS parame- X = X - (t a (k) - LCT) TAT < t a (k) NO? YES TAT =t a (k) X <? YES Non Conforming Cell YES TAT > t a (k) + L? NO X = NO TAT = TAT + I Conforming Cell Non Conforming Cell YES X > L? NO X = X + I LCT = t a (k) Conforming Cell VIRTUAL SCHEDULING ALGORITHM TAT Theoretical Arrival Time t a (k) Time of arrival of a cell CONTINUOUS-STATE LEAKY BUCKET ALGORITHM X Value of Leaky Bucket Counter X auxiliary variable LCT Last Compliance Time I Increment L Limit A algorithm Figure 1: GCR The ABR service category needs the most sophisticated UPC functions, because the service guarantees are based on a closed-loop control mechanism, the Enhanced Proportional Rate Control Algorithm EPRCA. The conformance of an ABR connection can be controlled by a Dynamic Generic Cell Rate Algorithm, which takes into account the response times of the end system. The ABR UPC functions are not subject of this study, because at the time of our experiments there was no implementation of this service available. The traffic contract has some drawbacks to other important ATM functions. During connection setup, which will be initialized by a signalling protocol, the desired traffic contract will be subject to the Connection Admission Control of 11 GMD-FOKUS

4 function CAC. This function will decide, if there are enough resources in the network to supply the desired traffic parameters. E.g. in the case of CBR, the total sum of the peak cell rates must not exceed the link bandwidth. If this would be the case with the new connection, the call has to be rejected. The CAC functions are becoming more complex with the VBR service classes because they incorporate overbooking of a link in order to improve the bandwidth utilization. The overbooking will be based on the mean cell rates of the VBR connections and allows the network provider to exploit the advantages of statistical multiplexing. It is subject to further studies, how large the multiplexing gain will be and what requirements this has to the switch buffers. Nevertheless the UPC functions for VBR traffic already exist and are examined in this paper. 4 Traffic Parameter Selection The traffic parameters, which are used to control the UPC functions, must be negotiated during connection establishment. To evaluate the UPC functions on real network equipment, we need to generate a test signal with known traffic parameters. The first problem we encounter is the mapping of commonly used traffic source parameters to the parameters of a UPC traffic contract. Especially for VBR traffic this usually requires a reliable knowledge of the statistical properties of the source cell stream. Even with this knowledge the proposed algorithms for the parameter mapping are not trivial [1]. A first step is to define the relation between the quality of service parameters and the actual traffic parameters. For the VBR service category the QoS parameters CLR, CDV and CTD will have an influence on the traffic parameter settings. In this context an important aspect of the cell loss definition becomes relevant: depending on a user- or a network perspective different CLR definitions apply. From the network perspective a CLR NP is negotiated for all conforming cells during connection setup. From the user perspective, a VBR cell stream may experience a certain amount of cell loss introduced by the UPC functions. The latter is identical with the nonconformance ratio γ M. γ M is defined as ratio of the number of cells exceeding the traffic contract and the total number of submitted cells. γ M will not contribute to the CLR NP, because it is in fact caused by a traffic contract violation. Nevertheless it is important for the user and will have drawbacks on the UPC parameter selection. Another QoS criterion for VBR cell streams, which might have an influence on the UPC parameter selection, is the probabilistic delay bound α: P( D > D max ) α It s definition is based on the assumption, that cells delayed in excess of D max are useless for the receiving process, e.g. a video codec. The α value will be important for the CDVT parameter determination. A higher delay will result (1) in a higher α and a realtime ATM application will demand for a maximum tolerable CDV. With the previous considerations the nominal cell loss ratio seen by the user is equal to: CLR = γ M + CLP NP + α N nc γ M = N T N T = total number of cells N nc = number of nonconforming cells While the traffic contract parameters are defined in relation to a GCRA algorithm, they are not directly suited for describing a traffic source. The CDVT is not intended to describe the source behaviour and might have different values at each interface along the path of a connection. The CDVT is intended to control the tolerance with respect to the cell delay variation introduced by the network. To determine an exact value for CDVT requires knowledge about the number of networks or switches a cell stream will pass. Some other parameters are used in a different sense to the way, they are used in source descriptions. I.e. the sustainable cell rate SCR of a traffic contract is not the same as the average cell rate of a VBR source. In fact the SCR will have a value somewhere between the peak cell rate and the avarage cell rate. The exact SCR depends on the setting of the MBS parameter. A similar problem occurs with the maximum burst size MBS: it depends implicitly on the PCR and the CDVT. In general we can say, that the selection of the UPC parameters is a minimization problem with up to four variables. On one hand the increase of the rate parameters will reduce the maximum possible link utilization. On the other hand the increase of CDVT and MBS will result in higher switch buffer consumption for the same cell loss probability. For a given cell loss requirement, a tradeoff between buffer cost and network utilization must be made. In addition it must be considered, that the increase of buffering will also increase the cell delay and delay variation. This might be an important issue for time critical realtime services. Recent studies propose a cost function for finding an optimum set of parameters [1]. The necessary calculations for the mapping of the QoS objectives to UPC parameters will have major drawbacks on the achievable network resource utilization. While in normal operation an optimum parameter setting is intended, for our measurements we used a different approach. Our main objective was not to minimize buffer consumption, cell loss or cell delay, but to examine the number of nonconforming cells with a specific traffic parameter setting. Therefore we intentionally generated traffic contract violations and compared the number of tagged or discarded cells to the expected values. In particular the thresholds, at which a switch starts discarding cells, are an (2) (3) of 11 GMD-FOKUS

5 interesting experiment outcome. To make the effects of the source parameters on the UPC functions visible, we used a simulation software with two GCRA processes. Input to the simulation is a previously recorded or generated cell stream. As result of the simulation, we observed the number of conforming and nonconforming cells as well as the fill of the corresponding leaky buckets. In section 5 we will compare the results of the simulations with the measurement results. This corresponds to a minimal interarrival time of 29.9 us. The cell interarrival times are Gaussian distributed with mean 1 us and standard deviation of 2 us, the statistic is shown in figure 3. It should be noted, that the GCRA algorithm is defined in form of a continuous state leaky bukket. This results in the fact, that the graph in figure 2 is not quantized to cell steps on the y axis. 4.1 CDVT Determination: The first example for the leaky bucket simulation addresses the problem of CDVT determination for an arbitrary cell stream. The CDVT is intended to allow the UPC function to control the cell delay variation introduced by the network on a cell stream. The maximum cell delay variation depends on the link rate, which means that the value will be smaller for higher data rates. As the CDVT describes the size of a leaky bucket, it can be understood as the number of cells, which can be sent at full link rate. The CDVT for a given PCR can be calculated from measured or simulated maximum leaky bucket fill fill max in cells by the formula: CDVT = fill max PCR With direct access to the GCRA variables the CDVT can be written as: CDVT = max( TAT t a ( k) ) with TAT = Theoretical Arrival Time, t a (k) = Time of arrival of the kth cell cells 2 1 fill max Figure 2: GCRA Leaky Bucket Fill The simulation in figure 2 shows the fill of the imaginary leaky bucket for 1 cells of a CBR cell stream with random cell delay variation. The traffic contract PCR is 1 cells/s, while the nominal measured PCR is 334 cells/s. (4) (5) Count e-5 8e I.A.T. [s] Figure 3: Interarrival Time Distribution As mentioned above, in addition the simulation tool keeps track of the maximum leaky bucket fill, which is 11 cells for the above example. This gives an CDVT value of 11 us for the given rate of 1 cells/s, if all cells shall be conforming. It is obvious, that the value of the CDVT depends strongly on the cell delay variation introduced by the network, i.e. cell clumping introduced by multiplexing in a switch. The TM 4. standard recommends, that the receiver CDVT should be designed to handle a connection traversing three networks, each having three switches in tandem. At this point it should be mentioned, that the size of the CDVT corresponds to the size of the buffer necessary to eliminate the cell delay variation introduced during the transmission. 4.2 Matching the characteristics of a VBR source: While the determination of the CDVT with a given PCR is a rather simple problem, the matching of VBR source characteristics with the traffic parameters (PCR,SCR,MBS) is a more demanding objective [22]. As mentioned above, many tradeoffs have to be made. Important factors are buffer consumption, bandwidth utilization and QoS requirements like cell delay variation CDV and cell loss ratio CLR [21]. The UPC function for the VBR traffic class can be implemented as two GCRA algorithms: GRCA(T,τ) with T = 1/PCR, τ = CDVT GRCA1(T s,τ s ) with T s = 1/SCR, τ s = burst tolerance BT of 11 GMD-FOKUS

6 The burst tolerance depends on the MBS, the T and the T s and is set equal to: τ s = ( MBS 1) ( T S T ) In general the SCR will have some value between the mean cell rate MCR and the PCR of a source. The closer the SCR is to the PCR, the closer the system is to peak cell rate allocation and the lower is the statistical multiplexing gain. While it is relativly simple to derive the mean cell rate and PCR parameters from a cell stream, the peak-tomean ratio is not sufficient to describe the burstiness of a VBR source [19][2]. Another important factor is the distribution of the burst sizes, which demands for a specific burst tolerance BT value in order to match a user defined cell loss probability. If a higher BT is selected, the SCR can be made smaller to get the same CLR. But on the other hand, with larger buffers the CDV will become larger. Rate [Mb/s] (6) The example in figure 4 shows the leaky bucket fills for 1 cells of a JPEG coded video cell stream. The frames are transmitted in bursts of about 3 cells at an intra-burst cell rate of 24 kcells/s or 1 MBit/s, which can be seen in figure 4 (a). The measured total mean cell rate including the gaps between the frames is 861 cells/s, the PCR derived from the minimum interarrival time is 238 kcells/s. The last value is equal to the full link rate, which means that the cell stream contains back-to-back cells. The UPC parameters are set to PCR1 25 kcells/s, CDVT 1 us, SCR 85 cells/s and MBS 3 cells respectively. The imaginary leaky bucket fills of the two GCRA algorithms are shown in figure 4 (b). The dotted lines are the sizes of the leaky buckets in cells. If one of the two buckets experiences overflow, the cells are discarded as nonconforming. With this parameter setting a total of 66 nonconforming cells are found in the displayed simulation interval. 5 Measurement Results The test equipment for our measurements consists of a time-synchronized cell stream generator and traffic analyzer. The test cell stream is looped back to the receiver module of the BATES test system, where several traffic parameters can be analyzed in realtime. For port overload tests there is a second cell stream generator which does not necessarily have to be time-synchronized with the the analyzer. The traffic of both generators is switched to the same output port. BATES ATM Test System Cell Stream Generator I Analyzer Cell Stream Generator II (a) Bandwidth Limit1 GCRA GCRA1 In Port I Out In Port OutII Port II UPC UPC [cells] (b) Leaky Bucket Fill Figure 4: JPEG Cell Stream Limit ATM Switch Figure 5: Loopback Test Configuration The tools used for the following measurements are a subset of the BATES ATM Test System developed at GMD- FOKUS [6]. This includes in particular the following realtime tools: - test generators for bandwidth ramp, burst and random ATM cell streams - realtime implementation of a GCRA algorithm - time dependent realtime analysis of bandwidth, cell interarrival times, Cell Loss Priority - time dependent cell loss and delay measurements of 11 GMD-FOKUS

7 Time dependent in this context means that the measurement results are recorded in subsequent intervals of 5 ms. This allows us to investigate the interaction of various parameters during the total measurement period with a precision, which could not have been reached by standard software tools. In this context the use of the tagging option at the ATM switch was quite useful to allow a more detailed analysis of the UPC function under the assumption, that the switch behaviour with respect to cell discarding and cell tagging is the same. Because the tagged cells with CLP=1 are still contained in the cell stream, the analysis of the discarded cells is much more detailed than for lost cells. The cell loss can only be detected, if the next cell after the lost cells arrives. The information about the exact arrival time of the discarded cell is lost as well as the timestamp. An interesting method for UPC measurements is the use of extreme parameter settings, i.e. PCR= or CDVT=, where they are allowed by the UPC configuration software. The results of the measurements under these conditions allow some conclusions on the details of the implementation. For the interpretation of the results we had to consider measurement errors due to the following causes: - resolution of the analysis intervals: 5 ms - resolution of the analysis timers and the GCRA calculations: 1 ns - accuracy of the traffic generators: 1-1 ms For all performed measurements our reference ATM switch introduced no additional cell delay for the UPC functions. On the network side in our measurements we used a Fore ASX-2 ATM switch, which implements the traffic contracts listed in table 2. For cbr and vbr a tagging option is available for CLP= cells exceeding the pcr and scr cell rates, respectively. Traffic Contract TM 4. FORE CBR.1 cbr pcr1, cdvt Traffic Parameters cbr pcr, pcr1, cdvt VBR.1 vbr pcr1, scr1, mbs1, cdvt VBR.2 vbr pcr1, scr, mbs, cdvt Table 2:Traffic Contract Parameters When testing UPC functions, a first decision has to be made wether to use random or deterministic test signals: for most of our tests with a single source we used deterministic cell streams like constant rate streams, ramps or burst sequences. For some test scenarios we used random cell streams to generate many independent transitions between conforming and nonconforming cell sequences. Only if the effects of more than one source shall be examined, the use of independent random sources is necessary in order to capture the effects of unsynchronized user cell streams. The TM 4. standard describes a simple non-conformance measurement in form of a 1-point process, which has as a result the non-conformance ratio γ M. The actual policing ratio is defined as γ P. In this context the transparency δ of the UPC mechanism can be defined by the accuracy with which the UPC function approaches the ideal mechanism. N pol γ P = (7) N T δ = γ M γ P N T = total number of submitted cells N pol = number of policed cells (tagged or discarded) A positive δ means, that the UPC function is less strict than the ideal mechanism. A negative δ means, that the UPC implementation discards more cells than necessary, which should never happen. In our measurements we will concentrate on this transparency, especially it s dependence on time and it s relation to the traffic characteristic. The most interesting measurements are performed in conjunction with a realtime implementation of the GCRA algorithm. The γ M value could be determined by comparison of the lost or tagged cells with switch UPC and the detected nonconforming cells without UPC. In extent the UPC transparency could be measured by performing the GCRA a second time on the previously policed cell stream. This cell stream must not contain any nonconforming cells, if the same UPC parameters are used. With the realtime GCRA software arbitrary long measurements of hours or days become possible. 5.1 Cell Rate Measurement For the measurement of the SCR and PCR values we propose two methods: the first one uses ramp shaped test cell streams to determine the clipping treshold. The second one uses bursts with a cell rate much higher than PCR and a very high measurement resolution of 1 ms. To measure the clipping at the PCR, we used the cell loss measurement capabilities of our test equipment. The example in figure 6 shows the clipping for a CBR traffic contract with PCR set to 25 kcells/s. (8) of 11 GMD-FOKUS

8 bandwidth 1 lost cells ky buckets, which can be seen in the different size of the overshots at the clipping treshold for SCR and PCR. The parameters in figure 7 (a) were set to SCR=8 cells/s, MBS=5 cells, CDVT=1 us, tagging enabled. [kcells/s] PCR 8 6 [kcells] 4 2 In figure 7 (b) we see the result of a burst measurement with the same traffic contract. As test cell stream bursts of 6 MBit/s (about 14 kcells/s) and length.9 s were used with a.1 s silent interval between the burst. The latter is necessary to allow the leaky bucket of the GCRA to become empty. The minimum gap is equal to τ S, which is 42 [ms] in our example. 2 4 Figure 6: Peak Cell Rate Measurement As expected, at this rate the first cell loss occures. For higher rates the number of received cells remains constant, while the number of lost cells increases with the source bandwidth. Rate [Mbit/s] PCR - SCR, CLP=1 SCR, CLP= 2 4 (a) Ramp Method with Tagging 5.2 MBS Measurement In the next measurement we determined the MBS parameter of a VBR traffic contract. The UPC parameters were set to PCR=25 cells/s, SCR=8 cells/s, MBS=5 cells and CDVT=1 us. An useful model for evaluating the performance of the GRCA1(T s,τ s ) is the worst-case conforming source [17], which can be described by a full rate on-off source. This type of source generates cells at the PCR during bursts of length B and is silent the rest of the time. The length of the on and off periods is given by B B T on = and T (9) T T off = s T s We generated a worst-case conforming source, which constists of a bursts with PCR of the length MBS. Then we increased the burst size one cell every burst and measured the cell loss on the connection in a loopback configuration. The exact burst size, for which the first cell loss occurs, can be easily determined. Bandwidth [Mbit/s] PCR SCR Loss s (b) Burst Method Figure 7: PCR and SCR Rate Measurement To examine the clipping at the SCR for VBR traffic contracts, we used the tagging option. This setup enabled us to record both clipping points at a time, which can be seen in figure 7 (a). Interesting is also the different size of the lea Burst Size [cells] Figure 8: MBS Measurement Figure 8 is the measurement result for a VBR traffic contract with a MBS parameter of 5 cells. The histogram shows, that the nominal MBS is 512 cells in contrast to the selected parameter. Obviously the used FORE switch implements a GCRA(T s,τ s + τ) for a public UNI interface. This definition is based on the fact, that the cell stream might have experienced a variable delay in a customer pre of 11 GMD-FOKUS

9 mises network. The resulting MBS* can be calculated by the formula τ MBS = MBS (1) T s T While the burst size is incremented 1 cell per burst, the absolute cell loss per burst does not constantly grow as expected. We blame this effect on the granularity of the GCRA algorithm. 5.3 Marginal Values: PCR=, CDVT= The measurement with zero CDVT is particularly interesting for evaluation of the UPC transparency. With a ramp over the full link bandwidth the result shows, that the GCRA algorithm is not able to generate a straight clipping at the PCR like for CDVT > (see figure 6) The cell loss rate increases discontinuously in steps from. to 1/2 and to 2/3, if the cell rate R crosses the PCR and the 1.5 * PCR threshold respectivly. The CLR in relation to the actual GCRA input cell rate R can be determined by CLR = ε , (11) ε + 1 ε = integer T T' ---- with T = 1/PCR, T = 1/R [kcells/s] received cells lost cells (a) Switch UPC cell rates higher than 185 kcells/s the resulting CLR behaves different from the expected function. This might be due to the fact, that for high data rates the UPC function is less accurate because of resolution limits. Another factor could be the limited generator and measurement resolution of 1 ns. In figure 1 we see the measured γ P values in comparison to the theoretical γ M calculated by the simulation tool. For CDVT settings below 1 us there are some differences, but with common values above 1 us γ P and γ M are almost identical. We noticed a strong dependence of the absolute cell rate. For lower cell rates the difference between measured and expected γ P was below the measurement accuracy. γ.4.3 γ M ideal γ P testtool γ P switch 1 2 CDVT [us] Figure 1: γ M measurement for small CDVT values A PCR of is an interesting case for the cbr traffic contract of the FORE switch: it defines a seperate PCR for cells with CLP= and for cells with CLP=1 or. The PCR for the CLP= cells might be set to zero, which means that all cells with CLP= should be discarded. Because the increment I of the GCRA algorithm is the inverse of the PCR, a practical implementation needs to choose some maximum value I limit for I. [kcells/s] nonconf conf [kcells] lost cells (b) Realtime GCRA Figure 9: Bandwidth Ramp with CDVT= In the measurement shown in figure 9 the PCR was set to 1 kcells/s and the CDVT was set to zero. For submitted Figure 11: Switch UPC, CBR, PCR = of 11 GMD-FOKUS

10 We measured I limit by sending an arbitrary cell stream of CLP= cells: the switch lets one cell pass the GCRA every.833 sec regardless of the actual send cell rate. The value can be easily determined by performing a long measurements and dividing the measurement time by the number of arrived test cells. The number of the measured lost cells corresponds to the I limit period of.83 sec. It is about 83 cells for a mean generator cell rate of 1 kcells/s. 6 Conclusion In our experiments we developed some methods for evaluation of the UPC functions of ATM switches. These methods represent an important contribution to troubleshooting and performance analysis of future broadband networks. For conformance testing the implementation of the GCRA algorithm seems to be the most attractive solution. For expected traffic parameter settings the measurement resolution of our test equipment was sufficient. On the other hand most parameters and mechanisms of the UPC functions are accessible by suitable measurement tools, but with this approach it becomes more complex to derive the standard defined conformance measures. The matching of source parameters and the related quality of service seems to be an unresolved problem for the use of VBR traffic contracts. Because only the use of VBR will allow the network provider to exploit of the statistical multiplexing gain of the ATM protocol, there is a need for further research on this topic. This includes the evaluation of the buffer requirements in relation to the traffic contract parameters in order to achieve the desired service quality. Other interesting topics for future research are the UPC functions for ABR service and the performance evaluation of various simultaneous UPC functions for different competing service classes [21],[23]. 7 References [1] The ATM Forum, Traffic Management Specification Version 4., April 96 [2] ITU-T Draft New Recommendation O.191, Equipment to assess ATM layer cell transfer performance, Geneva March 95 [3] ITU-T Recommendation I.356, ATM Layer Cell Transfer Performance, Geneva July 93 [4] ITU-T Recommendation I.361, B-ISDN ATM Layer Specification, Geneva March 93 [5] ITU-T Recommendation I.371, Traffic Control and resource Management in B-ISDN, Geneva July 93 [6] D. Elias, BATES - BERKOM ATM Technology Evaluation System, 7th IEEE Workshop on Local and Metropolitan Area Networks, Marathon 95 [7] A. E. Eckberg, "BISDN/ATM traffic and congestion control," IEEE Network, vol. 6, pp , Sept [8] L. Fratta, L. Musumeci, G. Gallassi, and L. Verri, "Congestion control strategies in ATM networks," ETT (European Transactions on Telecommunications), vol. 3, no. 2, pp ,1992 [9] D. K. Hsing, "Performance study on the ``Leaky Bucket'' Usage Parameter Control mechanism with CLP tagging," in Conference Record of the International Conference on Communications (ICC), vol. 1, (Geneva), pp , 1993 [1] Brian L. Mark and G. Ramamurthy, "Real-time Estimation of UPC Parameters for Arbitrary Traffic Sources in ATM Networks," in Proceedings of the Conference on Computer Communications (IEEE Infocom), (San Fransisco, California), Mar [11] Tadanobu Okada, Hirokazu Ohnishi, and Naotaka Morita, "Traffic Control in Asynchronous Transfer Mode," IEEE Communications Magazine, vol. 29, pp , Sept [12] Naoki Yamanaka, Youichi Sato, and Ken ichi Sato, "Usage parameter control and bandwidth methods for ATM-based BISDN," ACM Computer Communication Review, vol. 22, pp , Mar [13] A. Bohn Nielsen, E. Aarstad, H. Pettersen, T. Renger, R. Elvang, J. Kroeze, and J. Witters, "Experiments on ATM Traffic Control in the EXPLOIT Testbed," tech. rep., RACE Project EXPLOIT, Contact renger@ind.uni-stuttgart.d4.de [14] Arthur W. Berger and A. E. Eckberg, "A BISDN/ ATM traffic descriptor and its use in traffic and congestion controls," in Proceedings of the IEEE Conference on Global Communications (GLOBE- COM), (Phoenix, Arizona), pp (9.4), IEEE, Dec [15] Z. L. Budrikis, G. Mercankosk, M. Blasikiewicz, M. Zukerman, L. Yao, and P. Potter, "A generic flow control protocol for BISDN," in Proceedings of the Conference on Computer Communications (IEEE Infocom), vol. 2, (Florence, Italy), pp (7A.1), IEEE, May 1992 [16] P. E. Boyer, M. J. Servel, and F. P. Guillemin, "The Spacer-controller: An efficient UPC/NPC for ATM networks," in International Switching Symposium, vol. 2, (Yokohama), p. A9.3, Oct [17] B.T. Doshi, Deterministic rule based traffic descriptors for Broadband ISDN: Worst dase behaviour and connection acceptance control, in GlobeCom 93, pp , December of 11 GMD-FOKUS

11 [18] D. Anick, D. Mitra, and M. Sondhi, "Stochastic theory of a data-handling system with multiple sources," The Bell System Tech. J., vol. 61, pp , October 1982 [19] A. E. Eckberg, "A generalization of peakedness to arbitrary arrival processes and service time distributions," in Bell Laboratories Technical Memorandum, 1976 [2] A. E. Eckberg, "Approximations for bursty and smoothed arrival delays based on generalized peakedness," in Proc. 11-th International Teletraffic Congresss (Kyoto, Japan), 1985 [21] H. Kobayashi and Q. Ren, "A diffusion approximation analysis of an ATM statistical multiplexer with multiple types of traffic, Part I: Equilibrium state solutions," in Proc IEEE International Conference on Communications, vol. 2, (Geneva, Switzerland), pp , May 1993 [22] D. Reininger, G. Ramamurthy, and D. Raychaudhuri, "VBR MPEG video coding with dynamic bandwidth renegotiation," in Proc. ICC '95, Seattle, Washington, pp , June 1995 [23] W. Whitt, "Tail probabilities with statistical multiplexing and effective bandwidths in multiclass queues," Telecommunication Systems, vol. 2, pp , of 11 GMD-FOKUS