Performance Comparison of Congestion Controls ... - Semantic Scholar

Performance Comparison of Congestion Controls for ABR Services in ATM Networks Felix Hartanto , Harsha R. Sirisena and Krzysztof Pawlikowski Department of Electrical and Electronic Engineering Department of Computer Science University of Canterbury, Christchurch, New Zealand An end-to-end adaptive rate-based control scheme is currently being considered by the ATM Forum for supporting Available Bit Rate (ABR) services. The scheme is a reactive control where the source adapts its transmission rate based on feedback from the network. The adaptation can be linear or exponential, while the information received from the network can be based on a forward or backward explicit congestion indication. In this paper, we investigate the network performance and bandwidth allocation fairness of various adaptation alternatives in handling trac traversing dierent numbers of hops. The results show that a new EFCI mechanism with exponential-linear increment and exponential decrement oers the best performance and that the choice of adaptation parameters plays an important role in ensuring good performance and fairness for ABR sources. ABSTRACT:

1 INTRODUCTION B-ISDN is expected to support a wide variety of services (including voice, video and data) with diverse charateristics and performance objectives in an integrated fashion. In order to cater for such a wide spectrum of potential applications, ITUT, formerly CCITT, has selected ATM to provide a uniform information transport technique for BISDN. Additionally, the ATM Forum has also de ned a family of service categories, including Constant Bit Rate (CBR), Variable Bit Rate (VBR), Unspeci ed Bit Rate (UBR) and Available Bit Rate (ABR) [3]. The ABR service category is de ned to oer economical support for applications with vague QoS requirements, such as data applications, which present very stringent data integrity constraints while being insensitive to the transport delay. A user of this service is allocated a relatively small dedicated bandwidth, but is allowed to dynamically share the available resources left over from the other high priority service classes. While this service does not provide any guarantees, it attempts to avoid cell losses inside the network at the expense of delay, because cell losses imply retransmission with consequent waste of bandwidth.

To support ABR trac, the network requires a feedback mechanism in order to inform each source on the amount of available bandwidth or on how much data to send. A number of such mechanisms have been proposed; the two main mechanisms are called credit-based ow control and rate-based ow control. The credit-based ow control operates on a link-by-link basis for each virtual channel (VC). In this mechanism, cells can only be sent by the source after it receives credits from the receiver. The source can evaluate the number of cells it can send on the basis of the credits and knowing the number of cells it has sent since the last credits arrived. The number of credits provided by the receiver can be static or adaptive [7]. On the other hand, the rate-based ow control operates on an edge-by-edge basis. In this mechanism, the rate of transmission depends on the feedback from the network. Initially, the source will transmit at a given rate. If no congestion is detected, the source can increase the rate linearly ( xed step) or exponentially (multiplicatively), otherwise, the rate is decreased either linearly or exponentially. The feedback from the network can be based on either forward (FECI) or backward explicit congestion identi cation (BECI) [3, 12, 9]. A version of the rate-based mechanism that operates on a link-by-link basis, called backpressure [5], and a combination of rate-based and credit-based control [8, 10] have also been proposed. Performance comparison of credit-based and rate-based mechanisms was carried out in [2], of credit-based and backpressure in [1], and of ratebased and backpressure in [5]. The sensitivity to link delays of both credit-based and backpressure and their link-by-link operations, have made them less favoured for ATM networks. In late 1994, the ATM Forum voted for rate-based ow control, but without committing to the details of any particular algorithm [3]. Based on this recent development, in this paper, we will mainly focus our investigation on ratebased control mechanisms. The main components of the mechanism are identi ed and various possible mechanisms based on combinations of these components are discussed. In Section 3, we describe the simulation model used in this study.

The simulation results are presented in Section 4. The analysis of results for existing mechanisms leads us to a new mechanism, as presented in Section 4.2, which is shown to oer better throughput. Conclusions are drawn in Section 5.

2 RATE-BASED MECHANISM The rate-based mechanism is a closed-loop endto-end ow control method, operating per virtual channel and having the following three main components: Feedback Mechanism. The simplest feedback mechanism is based on a single-bit congestion indication (CI), which can be routed in either forward (FECI) or backward (BECI) direction, similar to the one used in the frame relay network [11]. A possible implementation of FECI mechanism with negative feedback option is for the switch to set the CI bit in the header of data cells if its output port is congested. At the destination, the receiver lters these CI bits. If it receives N consecutive cells with their CI bit set, then the forward path is declared congested and the receiver sends a Resource Management (RM) cell with CI bit set to the source. Upon receiving this RM cell, the source decreases its transmission rate. If the source has not received a RM cell after a rate update interval has elapsed, then it will increase the transmission rate. In the case of BECI, the lter is implemented at the switch and an RM cell is generated and routed immediately towards the source [5]. An alternative to this implementation, known as FECI with positive feedback option, is for the source to insert an RM cell in the data stream every NRM data cells. The CI bit of the RM cells is set by switches that are experiencing congestion and the cells are routed towards the destination. The receiver routes the RM cells back to the source. Upon receiving an RM cell with CI bit set, the source reduces its transmission rate. The source also reduces its transmission rate if it does not receive the last RM cell sent out, by the time it nishes transmitting a new one [3]. The source increases its rate only if an RM cell with CI=0 is received. An explicit-rate (ER) feedback mechanism is a more sophisticated alternative than the single-bit feedback. In this option, the network also feeds back the maximum rate at which a source can transmit in the appropriate RM eld [3].

Rate Adaptation. The source adapts its trans-

mission rate depending on the condition of the network. The rate adaptation (decrement/increment) can be either linear or exponential. For example, the source can decrease its rate exponentially upon detecting network congestion and increase its rate linearly when the congestion is cleared [12]. Congestion Measure. The congestion measure is used by the switch to determine whether an output port is congested or not. This can be based on the lling level (Q) of the queue at an output port. The output port is considered congested if Q TH and the congestion state is cleared if Q TL . The hysterisis ensure that there are no oscillations. The discussed components show a diversity of possible rate-based mechanisms. In order to distinguish one mechanism from another and to provide an easy identifaction of its type, a 2-tuple notation is used throughout this paper, where FB (feedback mechanism) = FECI, BECI or ER and INC/DEC (rate increment/decrement) = L(linear) or E(exponential). For example, a mechanism using FECI with linear (additive) increment and exponential (multiplicative) decrement is denoted as .

3 SIMULATION MODEL 1

5

2 3 4

6

Transmit terminals Switch1 D1 D2 1 km 10 km

1 2 3 4 5 6 Receive terminals

Switch2 D3 1km

Figure 1: Simulation model. Figure 1 provides a schematic of the simulation setup employed to investigate the performance of various rate-based control mechanisms. In this setup, the network is modelled as comprising two interconnecting switches. Each switch is modelled as a generic output buered switch. When the queue length of the buer exceeds TH = 300 cells, all cells for all VCs routed through that port are marked as congested. The queue remains in the congested state until the queue length falls below TL = 200 cells.

The link between the switches is 100 km long and has a1 speed of 150 Mbps or equivalently 353770 cps , while the links between the switches and the terminals are 1 km long and also have speeds of 150 Mbps. There are four ABR transmit terminals connected to the rst switch and two to the second switch. Each terminal is assumed to have an in nite supply of cells to transmit. Trac from Terminals 1-4 traverses two hops and is considered transit trac for Switch 2, while trac from Terminals 5-6 traverses one hop. The transmit terminals are rate-controlled. In this initial study, we consider EFCI mechanism with various combinations of adaptation alternatives, namely is an EFCI with linear increment and linear decrement, having the parameters LIR (linear increase to rate) and LDR (linear decrease to rate). is an EFCI with exponential increment and exponential decrement, having the parameters RIF (rate increase factor) and RDF (rate decrease factor). is an EFCI with linear increment and exponential decrement, having the parameters: LIR and RDF [12]. is an EFCI with exponential increment and linear decrement, having the parameters: RIF and LDR. Parameter ACR LCR PCR MCR ICR

NRM LIR LDR RIF RDF

De nition Allowed cell rate Link cell rate Peak cell rate Minimum cell rate Initial cell rate Number of cells/RM Linear increase to rate Linear decrease to rate Rate increase factor Rate decrease factor

Value Variable 353770 cps 353770 cps 3540 cps 353770 cps 32 710 cps 158150 cps 1.010 0.125

Table 1 contains a list of parameters associated with these mechanisms. The parameters (LIR, LDR, RIF, RDF) are chosen so that all mechanisms take approximately the same number of steps to increment ACR (allowed cell rate) from MCR (minimum cell rate) to PCR (peak cell rate) and, conversely, approximately the same number of steps to decrement ACR from PCR to MCR, i.e. n+L n+E 500 and n?L n?E 3, where P CR ? MCR LIR log( P CR=MCR) + nE = log(RIF )

1 cps

= cells/seconds.

P CR ? MCR LDR log( MCR=P CR) n? E = log(RDF ) n? L =

(3) (4)

are the number of steps to decrement ACR from PCR to MCR lineraly and exponentially, respectively. The simulation model was implemented using DESC++ [4]. All simulations were run for a period of 1 second. Collection of statistics was started after 100 miliseconds, which was regarded as the initial transient period.

4 SIMULATION RESULTS 4.1 Homogeneous Environment Initially, we consider the case where all sources use the same mechanism, their respective parameters being as listed in Table 1, and start at time 0. Table 2 shows the throughput for the ABR sources. Mechanism

Terminals 1-4 5-6 28330 cps 38340 cps 22460 cps 24250 cps 14550 cps 29620 cps 10720 cps 11580 cps 37250 cps 41910 cps

Total 189990 cps 138360 cps 117440 cps 66040 cps 232820 cps

Table 2: Throughput for same adaptation parameters for all terminals.

Table 1: Initial parameters.

n+L =

are the number of steps to increment ACR from MCR to PCR linearly and exponentially, respectively, and

(1) (2)

From the results in Table 2, we observe that all mechanisms oer the same throughput for traf c traversing the same number of hops. This indicates the fairness of all the mechanisms. Among the rst four mechanisms, the mechanisms with linear increment, and , oer better performance than the ones with exponential increment. This is due to the slow initial rise of the exponential increment as illustrated in Figure 2. We expect that after a transient period has elapsed, ACR for all sources will be around LCR/4, since we have four homogeneous ABR sources competing for the link capacity LCR. The dashed line indicates this target throughput. The linear mechanisms reach the target throughput very quickly, while the exponential mechanisms take about 300 steps to reach the target throughput; beyond that point the increment towards PCR is much faster . However, by then, congestion creeps in and the mechanism is forced to decrease its rate. This explains why the mechanisms never reach the target throughput as shown in Figure 3.

5

Allowed cell rate

4

Mechanism

x 10

linear exponential exponential−linear

3

Total 234250 cps 146320 cps 228570 cps 94610 cps 300870 cps

Table 3: Throughput for more aggressive increment parameters.

2

1

0 0

100

200

Steps

300

400

500

Figure 2: Steps for incrementing ACR.

140 120 100 80 60 40 20 0 0

100

200

300 400 500 600 700 Simulation times (miliseconds)

800

900

Mechanism


Total 294600 cps 267200 cps 286550 cps 93080 cps 345670

Table 4: Throughput for less aggressive decrement parameters.

160

Allowed cell rate (Kcells/s)


1000

Figure 3: Dynamics of ACR. From this discussion, we can postulate a hypothesis that the faster the system reaches the target throughput, the higher will be the total throughput. Pursuing this idea, in Figure 2, we plot the incremental ACR for a new mechanism, denoted as , where initially ACR is increased exponentially until the target throughput or MERI (maximum exponential rate increase) is reached, beyond that ACR is increased linearly. Choosing 10 steps to reach MERI (= LCR/4 = 88440 cps), we obtain LIR = 550 cps and RIF = 1.4 for this mechanism. From Table 2, we can see that the throughput for this proposed mechanism is higher than for the other mechanisms. The plot of the dynamic ACR for this mechanism demonstrates the fact that this mechanism achieves higher ACR much more often than mechanisms with only exponential or linear increments. Reducing the number of steps for incrementing ACR from MCR to PCR or equivalently choosing more aggressive incremental values, say LIR = 7100 cps and RIF = 1.10, we can see from Table 3 that higher throughputs can be achieved. Moreover, the fairness property seems to be preserved. This increase in throughput can also be achieved

through choosing less aggressive decremental values or equivalently increasing the number of steps for reducing from PCR to MCR (refer to Table 4). Again, we can see that the proposed mechanism oers the best throughput. Fairness is also preserved among the trac that travel the same number of hops, but unfairness is observed as the throughput of transit trac (traf c from Terminals 1-4) is much less than that of the trac that travels only a single hop (trac from Terminals 5-6). This is due to the fact that trac from Terminals 5-6 suers less propagation delay to reach Switch 2 than the transit trac, and hence the terminals also require much less time to adapt to the changing conditions within the network. Moreover, by traversing more hops, the transit trac have more chances of encountering congestion points, with the consequence that their CI bits are set more frequently than for the single hop trac. To improve fairness, it is suggested in [6] that higher priority be given to the transit trac.

4.2 Heterogeneous Environment The examples in the previous section assume that all sources traversing the same number of hops use the same adapatation parameters. With the implementation of the rate-control being de ned by the end-terminals, it is possible that dierent terminals choose dierent adaptation parameters. Table 5 shows the throughput for the case where Terminal 1 has more aggressive parameters than Terminals 2-4, i.e. LIR = 35370 and RIF = 0.1592 for Terminal 1 and LIR = 710 and RIF = 0.125 for Terminals 2-6. From the table we can see that for the rst four mechanisms, Terminal 1 achieves a much better throughput than the others, even Terminals 5-6. This obvi-

Mechanism

1 150320 cps 149780 cps 173050 cps 165990 cps 25830 cps

Terminals 2-4 18340 cps 15950 cps 12270 cps 8960 cps 40700 cps

5-6 21600 cps 16730 cps 14840 cps 9370 cps 46260 cps

Table 5: Throughput when Terminal 1 is more aggressive than the rest. ously indicates an unfair situation. Unlike the rst four mechanisms, with the proposed mechanism , the throughput for Terminal 1 is actually worse than the rest. The reason for this is that during the steady state, each terminal has a target throughput of LCR/4. With more aggressive parameters, Terminal 1 exceeds this target throughput more often and as a consequence, its rate is also forced to decrement more often. This can be observed in Figure 2, where the rate adaptation in is more frequent than for the other mechanism. Overall, this indicates that the mechanism has the advantage of oering some kind of protection against aggressive sources from overwhelming the less aggressive ones. Mechanism

1 28360 cps 22460 cps 14550 cps 10750 cps 37270 cps

Terminals 2-4 28360 cps 22460 cps 14550 cps 10750 cps 37460 cps

5-6 38370 cps 24250 cps 29620 cps 11610 cps 42220 cps

Table 6: Throughput when Terminal 1 starts transmitting 500 s later. Table 6 gives the throughput when Terminal 1 start transmitting 500 seconds later than the rest. As we can see, the throughput for all terminals is approximately the same. This indicated that this fairness property is not limited to the homogeneous case where identical sources start at the same time but is also observed when sources start at dierent times.

5 CONCLUSION In this paper, we have investigated forward explicit congestion indication (EFCI) with various rate adaptation combinations. In general, the results show that the choice of the adaptation parameters plays an important role in ensuring good performance and fairness for ABR sources. In particular, the results show that the proposed mechanism, denoted as , which initially increases its rate exponentially until a target throughput and increases its rate linearly beyond that, oers the best throughput. With the possibility of dierent adaptation parameters being used, it is likely that some ter-

minals with more aggressive parameters achieve much better throughput than the others. However, with the proposed mechanism, it has been shown that the mechanism oers protection against such terminals, since the target throughput oers an explicit rate to limit the maximum cell rate from each terminal. This rate can be updated when a connection is established or terminated by the network. Future study is forseen to investigate the dynamic performance of this mechanism, and to compare the mechanism with the explicit rate (ER) mechanism [3] where the explicit rate is fed back by the network dynamically during the information transfer phase rather than only during connection establishment and termination phases.

6 REFERENCES [1] M. Antico, F. Bernabei, and L. Gratta. Traf c Control Mechanisms Comparison for ABR Trac Transport. In Proc. of the Gigabit Networking Workshop, Boston, MA, April 1995. [2] L. Benmohamed, Y. Chang, N. Golmie, R. Schneeman, and D. Su. Simulation Study of Rate-Based and Credit-Based Trac Management Mechanisms. ATM Forum/94-0381, May, 1994. [3] F. Bonomi and K.W. Fendick. The RateBased Flow Control Framework for the Available Bit Rate ATM Service. IEEE Network Magazine, 9(2):25{39, March 1995. [4] V.F. Hartanto. DESC++: Discrete Event Simulation Library using C++ Language. Technical report, University of Canterbury, New Zealand, January, 1993. [5] A. Kolarov and G. Ramamurthy. Comparison of Congestion Control Schemes for ABR Service in ATM Local Area Networks. In Proc. of IEEE GLOBECOM Conference, pages 913{ 918, San Francisco, CA, December 1994. [6] A. Kolarov and G. Ramamurthy. End-toend Adaptive Rate Based Congestion Control Scheme for ABR Service in Wide Area ATM Networks. In Proc. of IEEE ICC Conference, pages 138{143, Seattle, WA, June 1995. [7] H.T. Kung, T. Blackwell, and A. Chapman. Credit-Based Flow Control for ATM Networks: Credit Update Protocol, Adaptive Credit Allocation, and Statistical Multiplexing. In Proc. of ACM SIGCOMM Conference, pages 101{114, London, U.K., September 1994. [8] H.T. Kung and R. Morris. Credit-Based Flow Controls for ATM Networks. IEEE Network Magazine, 9(2):40{48, March 1995.

[9] P. Newman. Backward Explicit Congestion Noti cation for ATM Local Area Networks. In Proc. of IEEE GLOBECOM Conference, pages 719{723, Houston, TX, November 1993. [10] K.K. Ramakrishnan and P. Newman. Integration of Rate and Credit Schemes for ATM Flow Control. IEEE Network Magazine, 9(2):49{56, March 1995.

[11] W. Stallings. Networking Standards: A Guide to OSI, ISDN, LAN, and MAN Standards. Addison Wesley Publishing Company, Reading, MA., 1993. [12] N. Yin and M. Hluchyj. On Closed-Loop Rate Control for ATM Cell Relay Networks. In Proc. of IEEE INFOCOM Conference, pages 99{108, Toronto, Canada, June 1994.