Fairness in bandwidth allocation for ABR congestion avoidance algorithms

Transcription

1 Fairness in bandwidth allocation for ABR congestion avoidance algorithms Bradley Williams, Neco Ventura Dept of Electrical Engineering, University of Cape Town, Private Bag, Rondebosch, South Africa {bwillia, Abstract-To determine the effectiveness of traffic management algorithms, a performance analysis of the algorithms is performed. The analysis should show how well the algorithm maximises available resources under various traffic conditions. The correctness of the performance analysis depends on the selection of appropriate performance metrics to quantify algorithm performance. The traffic management algorithms being tested support Explicit Rate congestion control schemes for the Available Bit Rate service. The algorithms under study are: ERICA, which is a standard ER congestion avoidance algorithm, and two extensions to ERICA, namely ERICA+, which considers the effect of length of the queue that buffers the ABR cells, and ERICA+ with MCR guarantees, which assigns bandwidth to contending connections proportional to their Minimum Cell Rates. The performance analysis of the congestion avoidance algorithm was carried out on a testbed created by the authors to allow for the introduction of congestion avoidance algorithms to traffic in an ATM network. The authors show all the ER algorithms under test effectively allocates bandwidth among contending connections according to the fairness policy employed and exhibit similar transient response times. I. INTRODUCTION The aim of traffic management [] for digital networks is to ensure that resources available are used efficiently without overloading the network. This is achieved by the introduction of traffic management algorithms, which continually monitor the effects of traffic on resources in the network. The effectiveness of traffic management algorithms employed will determine how well the network can handle increased traffic loads and how quickly the network can recover from fault conditions such as congestion. For the Available Bit Rate service of Asynchronous Transfer Mode, a class of traffic management algorithms called Explicit Rate (ER) congestion avoidance algorithms is used to proactively limit the aggregate ABR rate that can be handled by the network switches at any specific time to avoid the occurrence of congestion. The authors have developed an architecture [] to introduce and test the performance of ER algorithms in an ATM network. Determining the performance of ER algorithms is based on the selection of performance metrics to test the performance against. The purpose of this paper is to show how a performance metric, fairness in bandwidth allocation [], was used to perform a comparative evaluation of ER algorithms. The rest of paper is arranged as follows. Section II provides background and performance criteria of ER algorithms. Section III explains the performance metric, fairness in bandwidth allocation, for ABR traffic. Section IV explains how traffic data is analysed in terms of fairness in bandwidth allocation. Section V presents the findings of our experiments in the comparative performance analysis of ER algorithms, and section VI presents the conclusion of our work. II. ABR AND EXPLICIT RATE ALGORITHMS Of the five classes of service [] that ATM offers, ABR is designed for data traffic with bursty unpredictable behaviour. The ABR service uses a feedback flow control mechanism that allows for traffic sources to adapt to the bandwidth dynamically available in the network and attempts to avoid congestion, thereby minimising the possibility of data loss in the network. The ABR feedback control loop as shown in figure requires the participation of sources, switches, and destinations. The source has to send at a rate that all the network switches can support. The destination has to initiate the feedback path. Each switch has to correctly report the source sending rates it can handle. The authors would like to thank Telkom SA, Siemens, National Research Foundation (NRF) and The Department of Trade and Industry (DTI) for supporting this research project.

2 Source FRMs Switch BRMs Figure : ABR flow control Destination For the source to receive this feedback from the network, the source has to initially send in-band resource management (RM) cells in the forward direction, called Forward Resource Management (FRM) cells. The RM cell contains a number of fields including the explicit rate (ER) field. The ER field indicates to the source the rate the network can support at an instant in time. When the RM cell traverses the switch, an ER algorithm on every ATM switch in the network path computes a fair share for the connection (i.e. it provides information on the congestion status of the network) and inserts it into the ER field of the RM cell, which now travels in the reverse direction back to the source as backward RM (BRM) cells. The ER field contains the sending rate of an ABR connection, in cells per second, which the particular switch can support at that stage. Every switch performs this calculation for every contending ABR connection across the switch. Upon receiving the BRM cells the source acts appropriately to control the Allowed Cell Rate (ACR) of each ABR connection, by setting the Allowed Cell Rate to the ER value received in the BRM cell. Because of the bursty nature of ABR, it is important for a network carrying ABR traffic to dynamically adjust its resource allocation to deal with the varying ABR traffic. ABR-capable networks need to prioritise traffic streams, as ABR traffic is of a lower priority than CBR and VBR. Therefore, ABR traffic should not be allowed to use resources requested by CBR and VBR sources at the Call Setup stage. Since ABR is a lower priority service class, ABR congestion avoidance algorithms should be relatively inexpensive to implement in terms of processing capabilities of network switches. This is the design philosophy followed by ER congestion avoidance algorithms. The three ER congestion avoidance algorithms are: A. ERICA (Explicit Rate Indication for Congestion Avoidance) The ERICA algorithm [] is designed to achieve efficiency, fairness, controlled queuing delays and a fast transient response. The algorithm aims to achieve an efficient and fair allocation of the available bandwidth, for the ABR traffic, to all the competing sources in the network. ERICA has become an ATM Forum standard []. As the name suggests, a switch using the ERICA scheme periodically calculates the Explicit Rate that each ABR source can send at, while avoiding congestion at any point in the link. In ERICA, ABR traffic is not allowed to occupy the Total ABR Capacity, but only a large fraction of it, called the Target ABR capacity. The Total ABR Capacity is the fluctuating bandwidth left over for ABR traffic after higher priority traffic has been accommodated. ERICA avoids congestion by reserving a small fraction (Total ABR Capacity Target ABR Capacity) of the Total ABR capacity to deal with ABR traffic suddenly sending above the Target ABR capacity, without congesting the link. The Target ABR Capacity has been found in simulation studies [] to be optimal at approximately 9% for LANs and 9% for WANs, of the Total ABR Capacity. In calculating the Explicit Rate for each ABR connection, the algorithm calculates a FairShare for each connection, which is the share of the available bandwidth to be allocated to that connection. The FairShare is the ratio of the Target ABR capacity to the number of active VCs. From this, it can be seen that irrespective of the size of each ABR connection, the algorithm partitions bandwidth equally among contending connections. This fairness policy is known as max-min fairness. The bandwidth assigned to each ABR connection is the amount of bandwidth available at the most congested point in the network, divided by the number of contending ABR connections. B. ERICA+ In ERICA+ [], the Target ABR Capacity is not a constant fraction of the Total ABR Capacity, but a varying quantity dependent on the length of the ABR output queue of the switch. If the queues were to fill up quickly, more space would be allocated for preventing cell loss due to overflowing queues, and vice versa. This approach achieves a higher network utilisation, (%) than the original ERICA algorithm. This is because ERICA+ does not restrict ABR traffic to a Target ABR Capacity to avoid congestion. The main difference between ERICA and ERICA+ is that, while with ERICA allowance is made for queue growth when nearing congestion, with ERICA+, the output queue length is measured directly, giving a better indication of the amount that the ABR connection rates should be reduced by. C. ERICA+ with MCR guarantees It can be argued that the max-min fairness policy, although a good starting point in considering bandwidth partitioning among contending ABR connections, is not the best fairness policy to employ. ABR connections can specify a Minimum Cell Rate (MCR), which represents the minimum bandwidth that the network should reserve for that connection. Maxmin fairness does not take note of MCR, which can cause a network to under allocate bandwidth to an ABR connection. III. FAIRNESS IN BANDWIDTH ALLOCATION Max-min fairness and MCR guarantees are two examples of fairness policies used in ATM networks. To illustrate the

3 difference between these two policies, an example is provided. Consider the situation in which two ABR connections, VC with a MCR of cells per second (cps) and VC with MCR of cps, are contending for network resources. The customers of the connections expect to get a sustained connection rate of at least MCR. There is a point of impending congestion with the ATM switch there, which can support an Available ABR capacity of 6 cps. With max-min fairness, the switch would aim to assign bandwidth to each connection of 8 cps, which would not satisfy the MCR of VC of cps. With a fairness policy that takes note of the MCR of each ABR connection, the bandwidth could have been partitioned to satisfy both MCR conditions. ERICA+ with MCR uses a fairness policy of MCR + proportional weighting to MCR. This means that the ER algorithm will divide the Available ABR capacity calculated by a switch by first assigning each connection its MCR, and then assign the left over ABR capacity to the connections, proportionally to its MCR. Using the example, VC would receive bandwidth (in cps) of its MCR,, plus its proportional weighting, /(+) = /, multiplied by the left over bandwidth, 6 ( + ). So bandwidth of VC is + /* = cps. VC would receive 6 = cps. Using this fairness criterion for two connections, one notes that the ratio of current cell rates is equal to the ratio of minimum cell rates, i.e. (CCR of VC) / (CCR of VC) = (MCR of VC) / (MCR of VC). The choice of fairness policy used is a decision made by the network administrator. It impacts on how difference priority traffic classes are allowed to use the network, and how customers are billed for network usage. With max-min fairness, bandwidth unused by higher priority traffic can be used more efficiently used by ABR connections, as any bandwidth assigned to ABR connections can be taken away when needed. With MCR guarantees, a user of ABR traffic is assured that he/she will get at least a known minimum bandwidth share, and might be willing to pay for that assurance. IV. ANALYSING ABR TRAFFIC DATA To evaluate the fairness in bandwidth allocation of the ER algorithms that are being studied, the following procedure will be followed: A. Configure an ATM network to handle various ABR traffic scenarios. B. Decide on different ABR traffic scenarios as input to the network. C. Implement an ER algorithm onto all switches in the network. D. Choose the traffic data that will provide the information to determine the fairness in bandwidth allocation. E. Quantify the degree of fairness achieved with each algorithm and display the fairness graphically. F. Draw comparisons from the graphical representation. A. Configuring ATM network The configuration to be used in this study is a single-pointcongested link, with one source generating two ABR connections, and one destination. This configuration is adequate to study the basic performance features of ER algorithms, and can be scaled to more complex configurations. A number of virtual channels from the source are carrying ABR traffic towards the destination, through the switch, on the same virtual path. B. Traffic scenarios The authors have identified four different traffic scenarios that will be used to test the algorithm. Scenario : ABR sources send traffic at the target ABR capacity, the point at which the algorithm starts limiting the ABR sending rates. This is the ideal traffic operating point, and the authors will note whether the algorithm would keep the network at this point. Scenario : Same as scenario, but second connection starts sending seconds after first connection, so the investigator can note the effect of changing bandwidth requirements. Scenario : This scenario follows the same configuration as scenario, but with background VBR traffic. The VBR traffic is modelled as a Poisson process. Scenario : Scenario is the same configuration as scenario, but with VBR traffic. C. Implementing ER algorithms The ER algorithms were implemented [] in software on the ATM switch. In the case of ERICA, only the information contained in the RM cells are used to calculate the ER of an ABR connection, while with ERICA+, the effect of queue growth in the switch is also considered. An output queue was simulated, as the sending rates generated was too low to cause real queue growth. D. Choosing traffic data

4 To determine how fairly an ER algorithm allocates bandwidth among contending connections, the traffic data required is the MCR of each connection, and, for every measurement interval, the Current Cell Rate (CCR) of each connection. The CCR of each connection is continually monitored to see whether the bandwidth share of the connection is in accordance with the fairness policy used. Fairness Index: ERICA The MCR is used in the case of the MCR-dependent fairness policy to determine whether the CCR of a connection is at least equal to its MCR and whether the ratio of CCRs is as the MCR policy states. Figure : Scenario - ERICA E. Present results graphically Fairness Index: ERICA+ The degree of fairness can be quantified as the degree of error from the ideal conditions, and displayed on a graph. For max-min fairness, the CCRs of the connections should agree, so the ideal condition would be where the ratio of CCRs is equal to one. For MCR guarantees, the ideal ratio would be proportional to their MCRs. The authors have decided to have the ideal ratio equal to (MCR of VC ) / (MCR of VC ). F. Drawing comparisons By inspecting the results graphs, we can determine which algorithm gives the best fairness in bandwidth allocation performance. The next section presents a selection of the resultant graphs. V. RESULTS OF EXPERIMENTS The experimental results are presented as a set of graphs of the degree of fairness in bandwidth allocation for each of the three ER algorithms, over a time period. Each ER algorithm operated on four different traffic scenarios. In each case, the MCR of VC is equal to cps and the MCR for VC is equal to cps, with (MCR of VC)/(MCR of VC) equal to.. The measurement interval is the time elapsed between measurements taken of traffic data. In the experiments, the measurement interval is ms..... Figure : Scenario - ERICA+ 6 9 As can be seen from figures and, the fairness index settles to after measurement intervals, or after * ms =.6 seconds, which is the ideal fairness point for max-min fairness. ERICA and ERICA+ yielded the same results for all configurations Fairness Index: ERICA+ MCR Figure : Scenario with ERICA+ - MCR guarantees Figure shows that ERICA+ with MCR proportional weighting fairness settles at a fairness index of., which is the ideal point with the MCRs as stated above.

5 Fairness Index: ERICA+ - Staggered Fairness Index: ERICA+ with VBR Figure : Scenario - ERICA+ Figure shows a fairness index increasing from after measurement intervals, which corresponds to, seconds after start. This is the time when the second VC starts having traffic. It takes measurement intervals for the fairness index to reach a steady state of. Figure : Scenario - ERICA+ Figure is the first graph with the introduction of VBR traffic. As in Scenario with ERICA+, it takes measurement intervals before the fairness index settles at..... Fairness Index: ERICA+ MCR - Staggered 9 9 Fairness Index: ERICA+ MCR with VBR Figure 8: Scenario with ERICA+ - MCR guarantee With the introduction of VBR traffic, ERICA+ with MCRrelated fairness shows stabilisation at. with oscillations around that point, after measurement intervals. Figure 6: Scenario with ERICA+ - MCR guarantee In figure 6, the profile of the graph is similar to that in figure, but with a peak at.9 in figure 6, against a peak at 6.6 in figure. Fairness Index: ERICA+ with VBR: Staggered Figure 9: Scenario - ERICA+ The fairness index rises from zero after measurement intervals in figure 9. Afterwards, it oscillates around with a variance of about..

6 Fairness Index: ERICA+ MCR with VBR - Staggered Traffic Management in ATM Networks, 99. Website of R. Jain Figure : Scenario ERICA+ - MCR-related fairness In the final figure, figure, the introduction of VBR causes a sharp rise of the fairness index to, after which it settles at. with some oscillations with a variance of.. VI. PERFORMANCE ANALYSIS AND CONCLUSION ERICA and ERICA+ exhibited the same behaviour on network traffic for all four traffic configurations. This result suggests that the explicit interrogation of queue length might only be necessary for sending rates much faster than was achieved in the testbed. It took three measurement intervals for the fairness index to settle after a sudden change in traffic conditions. This figure holds for all ER algorithms. This suggests that at the sending rates used in the experiments, all the ER algorithms have the same transient response time. The introduction of VBR in the network did not have much affect on the stability of the ER algorithms in terms of fairness in bandwidth allocation. For the traffic and network configurations used in these experiments, most noted having two ABR connections at sending rates of around cps, all the ER algorithms, ERICA, ERICA+ and ERICA+ with MCR proportional guarantees, have similar capabilities in terms of fairness in bandwidth allocation. BIBLIOGRAPHY [] The ATM Forum Technical Committee, Traffic Management Specification, Version., ATM Forum, March 999. [] B. Williams, G. Mountain and M. Ventura, Experimental Evaluation of ABR Explicit-Rate congestion avoidance schemes in a Gigabit ATM network, SATNAC. [] S. Kalyanaraman, Traffic Management for the Available Bit Rate (ABR) Service in Asynchronous Transfer Mode (ATM) Networks, Ph.D. dissertation, The Ohio State University, 99. [] S. Kalyanaraman, R. Jain, S. Fahmy, R. Goyal and B. Vandalore, The ERICA Switch Algorithm for ABR