Unformatted text preview: ON WORST CASE TRAFFIC IN ATM NETWORKS Bahadir Erimli, John Murphy and Jennifer Murphy Abstract
Asynchronous Transfer Mode ( ATM ) networks, which are the emerging standards for the future broadbandISDN networks, allows for the input tra c from users to vary both from one call to another and within the call. ATM speci es a method for controlling the tra c ow across the user network interface ( UNI ). This involves each user negotiating call parameters with the network. Once these parameters have been decided then a contract is made between the user and network. The network must then enforce the contract in order to guarantee performance and quality of service to other users. What is of interest to the network is given a particular set of users and contracts what is the worst tra c that the users could input to the network while still abiding by their contracts. This type of input tra c would be called the worst case tra c as it would produce the lowest performance in the network. The reason why this is of importance is that for simulating network performance we would like to have the worst case tra c inputs. Furthermore, it's important for the tra c controller, network, to know the possible worst case tra c so that it can assign parameters accordingly. The contract parameters have been decided upon by the standards organisations and what is needed now is to decide what type of tra c can pass these tests and produce the lowest network performance. This problem has been studied in the literature. We give some theoretical background to explain some of the results in the litrature, and we further look at some examples of types of worst case tra c sources. We show that for the two most common types, the greedy on{o and the three state source that either can be worse, depending on the situation. This would imply that there is no general worst case tra c as has been considered for the last number of years. 1 Introduction
The primary role of tra c control and congestion control is to protect the network and the user in order to achieve network performance objectives. An additional role is to optimise the use of network resources. The tra c control and congestion control mechanisms should not rely on other higher layer protocols which are either application or service speci c. However protocols may make use of the information in the ATM layer to increase their e ciency. There are two levels of congestion and control involved with ATM, the call level and the cell level. With ATM connections there is a unidirectional specifying of the Quality of Service 1 Bahadir Erimli, Electrical Engineering, 11681, California Institute of Technology, Pasadena, CA 91125, USA. email: [email protected] 2 John Murphy and Jennifer Murphy, Electronic Engineering, Dublin City University, Dublin, Ireland. email: [email protected] ( QOS ) parameters. These parameters are speci ed at connection setup and are guaranteed by the network. To guarantee the QOS the network must be able to obtain enough information from the user about the connection and be able to ensure that no other connections that share the resources degrade the QOS. A user must enter into a contract with the network about the parameters of the call. Then the network will implement tra c control to avoid problems with degraded QOS before they occur. This includes Network Resource Management ( NRM ), Call Admission Control ( CAC ), Usage Parameter Control ( UPC ), and selective cell discard. The network will also control the case where congestion does occur by implementing Explicit Forward Cell Indication ( EFCI ), selective cell discard and reaction to UPC failure. Within the ATM cell there are a number of bits available for congestion and priority setting. These include the Payload Type Indicator ( PTI ) and the Cell Loss Priority bit ( CLP ) both of which are contained in the header of the ATM cell. The rst bit of the PTI tells that the cell is a user cell and the next bit tells if congestion has been experienced by the cell in the network. The last bit di erentiates between two di erent types of ATMSDU's. The CLP bit is for high and low priority setting of the cells. This can be done by the user and/or by the network. A cell entering the network with low priority is subject to being discarded by the network in times of congestion. Tra c control is necessary to protect the network so that it can achieve the required performance objectives. UPC enforces a contract between the user and the network about the nature of the call. This prevents any one user from causing excessive tra c and hence degrading the quality of service provided to the other users. It is necessary to determine what is the worst tra c a user can in ict on the network while still abiding by UPC. The Leaky Bucket Algorithm is commonly used to implement UPC. 1.1 Asynchronous Transfer Mode Model
A simpli ed ATM switch model consists of N users feeding a nite FIFO bu er ( B places ) and is shown in Figure 1.
Network UPC UPC Users Output FIFO Traffic of Interest . . UPC Figure 1: Simpli ed Model The arrival process from the users is random, but the UPC algorithm for worst case analysis makes the arrival process to the FIFO bu er deterministic. There are two types of arrival patterns that we consider after the UPC and both of them are periodic. The service process is also deterministic ( at rate r ). The worst case tra c is that which creates the highest cell loss for a certain type of UPC, like the leaky bucket. A cell is lost when a cell arrives and the bu er is full. 1.2 Contract and Usage Parameter Control
The tra c contract speci es the negotiated characteristics of the connection. A connection tra c descriptor is the set of tra c parameters in the source tra c descriptor, the cell delay variation tolerance and the conformance de nition. The conformance de nition is used to decide which cells are conforming in the connection. A typical conformance de nition is the leaky bucket 1] or Generic Cell Rate Algorithm ( GCRA ) although many such algorithms may be used in tandem. The CAC will use the connection tra c descriptor to allocate resources and to derive parameters for the UPC. Any connection tra c descriptor must be enforceable by the UPC. Even though a cell is found to be nonconforming that does not mean that the connection is not conforming. The precise de nition of a compliant connection is left to the network operator. However a connection where all the cells are conforming is compliant. The tra c contract consists of the connection tra c descriptor and a requested QOS for each direction of the connection. This includes the de nition of a compliant connection. The private UNI may support a di erent tra c to the public UNI. The contract must contain the Peak Cell Rate ( PCR ) of the source tra c, the cell delay variation and the cell delay variation tolerance. Sustainable cell rate and burst tolerance are optional parameters. For best e ort tra c the only parameter speci ed is the PCR and the network may not reject the call because that bandwidth is not available but it may impose a di erent PCR. CAC is used to decide if a connection should be accepted or continue to be accepted in the network. It is required that the tra c contract be accessible to the CAC. The prime concern is to achieve the required QOS for the new or renegotiated connection as well as to ensure that the connection will maintain the QOS of all the other connections in the network. As well as deciding to accept the connection the CAC must determine the parameters needed by the UPC and route and allocate the resources to the connection. Even if high and low priority are not set the network may set them for nonconforming cells.
Incoming Cells Outgoing Cells Leaky Bucket Figure 2: Leaky Bucket or UPC Algorithm UPC is the set of actions the network take to monitor and control tra c. This includes the validity of the connection. The operation of the UPC shall not violate the QOS objectives of a compliant connection. However the excessive policing actions on a compliant connection are part of the overall network performance degradation and so safety margins should be engineered to limit the e ect of the UPC. The UPC can also fail to take action on a non compliant connection. Policing actions on the non conforming cells are not to be allocated to the network performance degradation of the UPC. At the cell level the UPC may pass a cell, change the priority of the cell or discard the cell. A low priority cell is discarded by the UPC if it is non conforming. Following the UPC shaping may be implemented on the conforming cells to reduce cell clumping. It is optional for the network operator to allow the UPC to initiate the release of a non compliant connection. When two levels of priority are used the UPC may discard high priority cells even though if the UPC were performed on the high priority alone the cells would be conforming. The UPC and CAC are operator speci c and should take into account the tra c contract to operate e ciently. It is speci ed that the signaling should take into account experimental tra c parameters that could be proprietary to either the manufacturer or network operator. It is optional to allow the the operation of these parameters across the UNI by mutual agreement. It is optional for the user to be allowed to mark cells as low and high priority. It is also optional for the network to mark cells as low priority if they are not adhering to the tra c contract. The cell loss ratio for low priority cells must be higher than for high priority. The leaky bucket algorithm is a UPC standardised by the ATM Forum 2] and is shown in Figure 2. The operation of the leaky bucket is that a splash is added to the bucket (counter increment) for each incoming cell when the bucket is not full. When the bucket is full cells cannot pass through to the network unmarked but the bucket leaks away at a constant rate. The important parameters to be de ned in this system are the leak rate of the bucket ( R ), the bucket capacity ( M ) and the peak cell emission ( p ). We assume that if the cells cannot pass the leaky bucked unmarked then they are lost and are not taken into account in the cell loss rate. This is because the network will only give guarantees to the marked cells and the unmarked cells will not interfere with the marked ones. The UPC standard of the ATM Forum 2] is actually a double leakybucket that controls the two di erent cell rates R and p, each with a separate bucket. The peak rate, p, is controled by a bucket of size 1 and leak rate, p, and the second bucket operates as explained above. For a cell to be a conforming cell, it needs to be conforming with both buckets at the time it's transmitted. 1.3 Tra c Types
When considering which type of sources will produce the worst performance in the network while still maintaining the contract the rst point to note is that the type of source will not allow the leaky bucket to ever over ow. In other words the total available number of cells that are allowed to enter the network will enter to produce the lowest performance. Two types of sources that have been proposed that could be the worst types are a two state source and a three state source as shown below in Figure 3.
Greedy OnOff Ton Toff time Ton Three State Toff time Ton T* Toff time Ton T* Toff time Figure 3: Source Types The two state source ( greedy on{o ) emits a burst of cells at the peak cell emission until the bucket is about to over ow and then falls silent waiting for the bucket to empty. This occurs periodically depending on the parameters of the system. The three state source is similar to the greedy on{o source except that it keep emitting cells at the leak rate after a burst. Therefore it's operation is to emits a burst of cells at the peak cell emission rate until the bucket is about to over ow and then emit cells at the leak rate of the leaky bucket for some time and then fall silent to allow the bucket to empty. This would then be repeated again. What can be seen is that the three state source would have a longer period than the greedy on{o for the same system parameters. The general belief is that the greedy on{o source gives rise to the worst case tra c as it would have the largest variance possible. However the three state source has been proposed as producing longer queues and therefore larger loss. 1.4 Finite and In nite Bu ers
A study of the three state source 5] compared it's performance to that of the greedy on{o source, for a number of sources into an in nite bu er. What was found in those simulations is that that the three state sources produced higher bu er occupancy than the greedy ono sources. The survivor function, P Q > q ], was found where Q was the bu er occupancy. This function was then assumed to approximate to cell loss in a nite bu er. However this we believe would only be true for input tra c that would be statistically independent of the queue lengths, which is not the case here as the input tra c is periodic and deterministic. We can demonstrate this fairly simply with a small example as is shown in Figure 4. Here we have the plotted the bu er occupancy for the in nite case of both the greedy on{o and the three state on the same axis. Note that we are considering only a single source here on it's own.
Type Number of Cells in Buffer
4 4 0 Three State Source Greedy onoff Source
0 0 6 12 Cell Time Figure 4: Bu er Occupancy in an In nite Bu er We have chosen the period T of the greedy on{o to be half of the period of the three state source so that we can just examine a single period of the three state for cell loss. There are 2 cells arriving per cell time while the source is emitting at the peak rate p and this continues until the bucket is full to capacity M , which in our example is 4. Then the bucket leaks away at a rate of 1 cell per cell time R for the greedy on{o but for the three state source it emits at this rate R for 6 time units, which is denoted by T and here is equal to the period of the greedy on{o source. Therefore we look over 12 time units as both repeat exactly the same after this. Over this period there are 12 cells inputted to the bu er which is the same as the leak rate R multiplied by the time period. The three state source gives rise to higher bu er occupancy than the two state source as can be seen in the in nite bu er plot. This would lead us to believe because of the higher queue occupancy for the in nite case that we would have higher cell loss in the nite case. However if we now consider a nite bu er of capacity 2, we nd that that the three state source losses 2 cells, one at time period 1 and one at time period 2. However the greedy on{o source losses one at time period 1 and one at time period 2 and again one at time period 7 and one at time period 8. So the cell loss rate for the three state source is 0.166 while the cell loss rate for the greedy on{o is 0.333. The fact of higher bu er occupancy in the in nite case does not necessarily mean higher cell loss rates in the nite bu er when the arrival process is deterministic. For deterministic arrivals the loss in a nite queue is not necessarily related to the bu er occupancy in the in nite case but more on the method of arrivals, or the process of arrivals, past the nite places in the queue. This problem was simulated 4] and these conclusions help to explain the results obtained therein. Furthermore, the survivor function P Q > q ] can provide more con icting results if viewed more carefully. As an example, let's look at a situation with M = 2, p = 1, R = 0:5 and r = 1. For the threestate source, let's leave T as a variable and the overall period of the source be T (T = 6 for the twostate case). Then we can calculate the exact frequency that each possible queue occupancy is going to occur at.
P Q P P P = 0] = (4T ; 7)=T 2 2 2 Q = 1] = (:75T ; 4:5T + 12)=T 2 2 Q = 2] = (:25T ; 3)=T 2 Q = 3] = (:5T ; 2)=T Using these formulas, we can easily calculate the survival function and compare it for di erent values of T . For the probability P Q > 2], starting from T = 6, P Q > 2] for the threestate sources is less than that of the twostate source as seen in Figure 5. On the otherhand, the average queue occupancy of all the threestate sources in this case are greater than that of the twostate source, but the cell loss rate of the twostate source is greater than any of the same threestate sources for a bu er length of 2 cells ( B = 2 ).
1.0 0.8 2state source, T=6 3state source, T=8 3state source, T=24 P[Q>=q] 0.6 0.4 0.2 0.0 0.0 1.0 2.0 3.0 q, buffer occupancy (cells) 4.0 Figure 5: Comparison of Two State and Three State Sources This shows that the survival function for an in nite bu er cannot be used to make conclusions on the nite bu er situation for this deterministic tra c pattern. In fact, if the tailend of the simulation in 5] is calculated, it can be seen that P Q > 898] for the twostate is higher than that of the threestate source, indicating that there is a crossover point in the graph that was unlikely enough not to appear during the simulation process. 1.5 Continuous and Discrete Variables
The problem contains both continuous and discrete variables and therefore it is important to distinguish between the two types. We consider a simpli ed model similar to 3] where there are two identical, independent users feeding a nite bu er. If the users were not independent then the worst case would be the greed on{o source with all sources in phase emitting together. However when we consider that the users or sources are independent then the phase between the sources is random. This then gives rise to the probability of cell loss because there is a probability of phase di erence between the sources. We assume that the two sources are randomly phased, which means we can consider one source as a reference source and the other is than an amount out of phase. The amount out of phase will determine the number of cells lost. The more in phase the more loss we expect. The probability of each of the possible combinations of out of phase is just the reciprocal of the number of the possible combinations. We consider two possibilities, either two greedy on{o sources or two three state type sources. The cell losses for all distinct combinations of the two tra c patterns are calculated and averaged. The cell loss rate for each scenario is compared. There are a number of constraints on the problem due to the discrete nature of the variables. 1. The sources may be out of phase only in one cell time units which means that we cannot use integration as was used in 3]. 2. The on period of the two state source is the time taken for the bu er to just about to over ow whilst emitting at the peak rate and also to emit any cells which arrive during this time. In the discrete case this is equal to dM=p ; Re ; 1. 3. Cells are only lost as units { no fractional cell loss. Therefore it is necessary to ensure that in using general formulae the cell loss is truncated to an integer. We investigate the case of M 1. 4. The discrete nature of the cell also implies that the bu er size must be an integer. 5. The service process can be assumed to be continuous. This means a cell can be served as it arrives. Alternatively, can think of the server waiting until the cell has fully arrived before starting to serve it. 2 New Approach to Cell Loss
The comparison of cell loss from both source types in the two identical user system is revisited assuming the cell arrival and service processes to be discrete. The cell loss rate is computed as the total number of cells lost in one period divided by the total number of cells emitted in one period. The total number of cells lost in any one period is the average of the cells lost by each combination of the two tra c patterns. It is assumed that the minimal phase di erence between two sources is one cell time and there are T ( T is the period ) combinations of the two tra c patterns. The total number of cells emitted in one period T equals 2RT . Furthermore the following stipulations are placed on the parameters : The service rate is at least equal to the peak cell emission rate, p r The leaky bucket rate is less than or equal to half the peak cell rate ( and hence the service rate ), R p=2 Together the peak cell emission and the leaky bucket rate exceed the service rate, p + R r The bu er size must be small enough so that cell loss is guaranteed when both sources are in phase, (2p ; r)Ton B These stipulations ensure that cell loss only occurs when at least one of the sources is emitting at the peak rate. If the service process is assumed to begin as the cell is arriving we call this a fast server. Then the cell loss for two greedy on{o sources, x places out of phase with each other, is denoted by C L (x) and is calculated in Equation 1
CL x ( ) = d(2p ; r)(Ton ; x) ; B e (1) where the symbols have the usual meanings. Similarly the cell loss for the three state source can be represented by Equation 2.
CL x ( ) = b(2p ; r)(Ton ; x) + (p + R ; r)(x) ; B c (2) Alternatively if it is assumed that the server waits until the cell has arrived in the bu er before starting to serve the cell, which we call the slow server, then the expressions for cell loss are modi ed for the greedy on{o source as follows in Equation 3
CL x ( ) = d(2p)(Ton ; x) ; (Ton ; x ; 1=p)r ; B e (3) (4) and for the three state slow server the cell loss will be given in Equation 4.
CL x ( ) = b(2p)(Ton ; x) ; (Ton ; x ; 1=p)r + (p + R ; r)(x) ; B c To calculate the cell loss ratio we allow x to vary over all phase possibilities and then divide by the number of combinations and also divide by the number of cells transmitted by both sources. The number of cells transmitted by both sources will be 2RT and the number of phase combinations will be T , so the cell loss ratio, C LR, is given by Equation 5.
C LR = PT ; x
T 1 =0 C L (x) 2RT (5) 3 Counter Examples
To show that there is no single worst type tra c for two identical sources we present a counter example to the traditional theory of the greedy on{o being the worst case. We show that the three state source can produce higher loss for integer values of variables chosen. As mentioned previously we consider two di erent types of servers, the fast serving server and the slow serving server. 3.1 Fast Serving Server
Assuming that the cell is served as it arrives we have an example showing the three state source to give rise to greater cell loss in a nite bu er than the two state source. The following system parameters are used :
 , peak cell emission = 1 r , service time = 1 R, leaky bucket rate = .5 M , leaky bucket capacity = 8 B , bu er size = 12 T (for three state source) ((2p ; r)Ton ; B )=(p ; R) = 6
p Reference Source Other Source with X = 0 # Cells Lost = 3 time X=1 # Cells Lost = 2 X=2 # Cells Lost = 1 X=3 # Cells Lost = 0 time time time time (A) Reference Source time # Cells Lost = 3 time X=1 # Cells Lost = 2 time X=2 # Cells Lost = 2 time X=3 # Cells Lost = 1 time X=4 # Cells Lost = 1 time X=5 # Cells Lost = 0 time X=6
(B) Other Source with X = 0 # Cells Lost = 0 time Figure 6: Fast Server, (A) Greedy On O Source, (B) Three State Source It takes 8 cell times to ll the bucket up but in that time 4 more cells are allowed through because of the constant leak rate of the bucket and while they try to ll the bucket 2 more arrive and then nally one arrives and the bucket is full. Therefore the amount of time that the source is on and emitting cells at the peak rate is given by Ton = dM=p ; Re ; 1 = 15. When the bucket is full it takes M=R seconds to empty normally, however here one time period has already elapsed so Toff = 15. There are 30 phase combinations for the greedy on{o source patterns, however sources that are too far out of phase do not give rise to cell loss. For the greedy on{o sources by examining Equation 1 we conclude that sources that are 3 or more time units out of phase do not produce any cell loss. Remember that this loss can occur when the secound source is a little advanced from the reference and also when it's so advanced that it is almost back in phase with the reference. This can be seen in Figure 6. The cell loss for the two state sources is in total 9 cells, which is calculated from 3 cells lost when in phase, 2 cells lost when either one out of phase and also 29 out of phase, and 1 cell lost when either two out of phase or 28 out of phase. For all other phaseing's there is no cell loss. Therefore using Equation 5 the cell loss rato can be calculated to be 0.01. For the three state sources cell there are 36 possible phase combinations and by examining Equation 2 loss can occur up to 4 units out of phase. The total number of cell lost over all possible phasings can be seen in Figure 6 and is 15 cells. By using Equation 5 we can calculate the loss to be 0.01157. Therefore the three state source produces higher loss than the greedy on{o source for the fast server. 3.2 Slow Serving Server
Reference Source Other Source with X = 0 # Cells Lost = 2 X=1 # Cells Lost = 1 X=2 # Cells Lost = 0 (A) Reference Source time time time time Other Source with X = 0 # Cells Lost = 2 time X=1 # Cells Lost = 1 X=2 # Cells Lost = 1 X=3 # Cells Lost = 0 X=4 # Cells Lost = 0 time time time time time (B) Figure 7: Slow Server, (A) Greedy On O Source, (B) Three State Source If the service process is assumed to begin serving a cell only after it has fully arrived into the bu er we have a similar counter example. The following system parameters are used :
 , peak cell emission = 1 r , service time = 1 R, leaky bucket rate = .5 M , leaky bucket capacity = 8
p  , bu er size =14 T (for three state source) ((2p ; r)Ton ; B + r=p)=(p ; R) = 4
B Similar to the fast serving server Ton = Toff = 15. For the greedy on{o sources by examining Equation 3 we conclude that sources that are 2 or more time units out of phase do not produce any cell loss. This can be seen in Figure 7. The cell loss for the two state sources is over all passible phase combinations equal to 4 cells and by using Equation 5 we can calculate the cell loss ratio to be 0.00444. For the three state sources cell loss can occur up to 2 time units out of phase. This is concluded by examining Equation 4 for the cell loss for a three state source and this is also seen in Figure 7. Here in total 6 cells are lost over all possible phase combinations and so again by using Equation 5 the cell loss ratio can be calculated to be 0.00519. Therefore the three state source produces higher loss than the greedy on{o source for the slow serving server. Therefore regardless of how the service is achieved in the bu er there is an example of the three state source producing more loss than the greedy on{o source. It is therefore shown that there is no single type of worst case tra c source for the two identical source problem, considering these types of sources. 4 Conclusions
From our analysis it may be concluded that the greedy on{o source does not always give rise to the worst case tra c. The rst point that we make is regarding the in nite versus nite bu er and we show that the occupancy in an in nite bu er will not relate to the loss in a nite bu er when the tra c is determinstic. The second point we make is that some of the variables are discrete and therfore attention must be paid when selecting examples and formula. We then concluded by showing two counter examples against the greedy on{o source and were able to say that there is no single worst case tra c type for the two identical source problem. The problem is far from totally solved, though. The result for this two source case is not binding on an n source situation, indicating one area to work towards. On the other hand, the tra c controller still needs to know in some manner what the worst case, or even a supposed or an asymptotic worst case, should be to be able to make decisions for QOS. This brings up the questions of under what conditions is one case worse than others or if there is a single case that we don't know about that performs universally badly for a given tra c controller. Acknowledgments
We would like to thank Dr. B. T. Doshi for assistance with this problem. We are also grateful to Dr. Mike Mandell for helpful discussions. The rst author acknowledges Prof. R.J. McEliece, California Institute of Technology, and the second and third authors acknowledge Prof. Charles McCorkell, Dublin City University, for their continuing encouragement and support of this work. References
1] S. Akhtar, '"Congestion Control In A Fast Packet Switching Network', Master's thesis, Washington University, 1987. 2] ATM Forum, `ATM User{Network Interface Speci cation', Prentice Hall, 1993. 3] B. T. Doshi, 'Deterministic Rule Based Tra c Descriptors For Broadband ISDN: Worst Case Behaviour And Connection Acceptance Control', International Teletra c Congress, ITC14, Vol 1a, pp 591600, Antibes, France, June 1994. 4] T. Worster, 'Modelling Deterministic Queues: The Leaky Bucket as an Arrival Process', International Teletra c Congress, ITC14, Vol 1a,pp 581590, Antibes, France, June 1994. 5] N. Yamanaka, Y. Sato and K. Sato, 'Performance Limitation of Leaky Bucket Algorithm for Usage Parameter Control and Bandwidth Allocation Methods', IEICE Trans. Commun., Vol E75B, No. 2, February 1992. ...
View
Full
Document
This note was uploaded on 04/29/2010 for the course CS 272 taught by Professor George during the Spring '10 term at BU.
 Spring '10
 George

Click to edit the document details