A Study of Adaptive Rate Algorithm in ATM Network Flow Control

Transcription

1 A Study of Adaptive Rate Algorithm in ATM Network Flow Control Surasee Prahmkaew and Chanintorn Jittawiriyanukoon Faculty of Science and Technology Assumption University, Bangkok, Thailand Abstract In this paper we present a flow control algorithm designed to improve the performance of ATM network and to gain higher throughput by guaranteeing negotiated cell loss ratio (CLR) for all cell streams passing through the usage parameter control (UPC). In particular, the cases in which a Tahoe, Reno, New Reno, and SACK schemes are applicable in peak-cell-rate (PCR) are discussed. The proposed algorithm improves the performance by adjusting the growth and the declination of window size, rather than using advertised window per se, by means of incoming traffic rate, number of cell drop, current window size and cell delay time. Several simulations are performed to study how different kinds of flow control algorithm behave when congestion existed and compared to our proposed algorithm. By varying advertised windows size in each algorithm, we can obtain the impressive results that support our algorithm. Keywords: Flow control algorithm, cell loss ratio (CLR), usage parameter control (UPC) Tahoe, Reno, New Reno, SACK, sliding window. Introduction The Asynchronous Transfer Mode (ATM) can provide a variety of services using a unified interface and architecture. In ATM networks, information is segmented into fixed-sized blocked called cells and the necessary number of cells is transmitted for the amount of information in the network. Therefore, the traffic control methods (Allman, et al. 1999) used in other existing transfer modes is not applicable to ATM networks. The principle of ATM traffic flow control should be simple. At connection setup, the user specifies both Quality of Service (QoS) requirements and, using a source traffic descriptor (Shioda and Saito 1997), the anticipated traffic characteristic of the connection. Network resources for the connection are assigned on the basis of the source traffic descriptor values and the QoS requirements. If there are not enough network resources, the connection is rejected. If the characteristics do not conform with the source traffic descriptor specified at the connection set up by monitoring the cell stream from the connection. If the characteristics do not conform, a penalty is imposed on the connection, e.g. some cells from the connection may be discarded. To simplify traffic flow control specification base on best QoS requirements and monitoring by the network, the traffic descriptors are required to be observable and easily checked through some algorithm mechanisms. In this sense, the specification of, and the resources allocation based on, the best traffic pattern have recently become a key issue for ATM traffic control. There are many previous studies consideration in flow control algorithms (Pakdeepinit and Jittawiriyanukoon 22); the behavior of each flow control schemes (Rathgeb 11) with varying advertised windows sizes is not mentioned as well. In this paper, we proposed an algorithm that improves the performance of ATM network by adjusting the growth and declination of flow control 69

2 window sizes. Besides our proposed algorithm and four existing flow control schemes that are Tahoe, Reno, New Reno, and SACK (Leerujikul and Jittawiriyanukoon 23) are discussed. Finally, the performance evaluation, especially in term of throughput, number of cells loss, mean time in queue, mean queue length and utilization of ATM link, between different types of these algorithms and our proposed algorithm is compared. Background The existing congestion avoidance algorithms are discussed first. Tahoe algorithm includes Slow Start, Congestion Avoidance, and Fast Retransmit. The Reno implementation is the enhancement to Tahoe by modifying the Fast Retransmit operation to include Fast Recovery. Selective Acknowledgments (SACK) has been proposed to efficiently recover from multiple segment losses. It sends duplicate ACKs (SACKs) acknowledging the out-oforder segments it has received. From these SACKs, the sender can reconstruct information about the segments not received at the destination. Further details can be found in (Matthew, et al. 1996; Kevin and Sally 1996; and Yang 2). Adaptive Rate Control (ARC) and Proposed Algoritm Consider the Tahoe, Reno and New Reno, and SACK when the burst traffic occurred either in a short or long time situations. They started advertised windows size form 1 to maximum windows size. It is increased by 1 (slow-start technique) for very successful transmission. When the window is filled up to the maximum of setting, most of input traffic would be discarded or tagged for the reason of exceeding the capacity of finite window size. When this situation occurred, most of window size will start at congestion window (cwnd) size (recovery technique). Cwnd always set equal to 1 for Tahoe, half of maximum windows size for Reno, New Reno and up to multiple loss for SACK. Unlike the ARC proposed solution in alleviating the number of cells would shrink and expand the window size by automatic adjusting the windows size base on source rate, cells drop, and cells delay. With ARC algorithm, it works like a gate to control the arriving cells. When a cell arrives at the system and if a cell drop is not present, the cell is transmitted without delay. If cell drop is present, the dropped cell will be blocked in a cell queue (Q c ) and waiting to retransmit when ARC finishes adjusting the new windows size in order to overcome the cells drop. At the same time to maintain quality of service (QoS), the maximum cell delay time has been defined as CDVT that means the cells have been waiting in the cell queue longer than CDVT will be discarded finally. Fig. 1 illustrates the above situation according to the cell arriving process, two states of problem for ARC. Source rate ( λa ) Apply flow control (Tahoe, Reno, New Reno, Sack and ARC) Cell queue length (Qc) Non- conforming window size (1- Max_size) Fig. 1. Flow control In case that the arrival traffic (average arrival cell rate or traffic ( λ a ) is less than the cell drop rate (λ p ) and cell drop is not yet present. ARC will set window size to be one (the minimum size). In other hand, if (λ a ) is larger than (λ p ), ARC will adjust the window size regarding to cell drop rate (λ p ) and arrival traffic rate (λ a ). ARC will adjust the window size between one and three (the maximum size based on analytical model shown in Fig. 2). ARC algorithm is described as follows: Leak rate (λp) 7

3 Procedure calculate window size Find out current allocation rate (Ai) Input Traffic UPC Dropped cells Input buffer ATM Frabric Mbps Departing cells Output buffer Interface Find out current window size (Wi) do while cell is transmitting { if cell drop is found then { calculate new allocation rate current allocation rate = new allocation rate Input Traffic Input Traffic Departing cells UPC ARC ATM Frabric Mbps Output buffer Interface Dropped cells Departing cells UPC ATM Mbps Frabric Output buffer Interface Dropped cells Fig. 2. Simulation model calculate new windows size current window size = new window size } else no cell drop is found then { current allocation rate = current allocation rate current window size = current window size } } end_do where : Find current allocate rate (Ai) Source _ rate = WindowSize * Cellsize ( Cell _ drop * Cell _ delay _ time) Cell delay_time = cell traveling time between source to destination Find current window size (Wi) WindowSize = Bandwith* Cell_ delay_ time [8] Simulation Model Fig. 2 shows a simulation model used in this paper. A. Input Traffic The telecommunications traffic can be basically classified into five categories: data, voice, video, image and graphics (Al-Wakeel, et al. 1999). This research confines the discussion to mainly data, voice and video. Data sources are generally bursty in nature whereas, voice and video sources can be continuous or bursty, depending on the compression and coding techniques used. Continuous sources are said to generate constant bit rate (CBR) traffic and bursty sources are said to generate variable bit rate (VBR) (Kouvatsos 1995) traffic. Hence, only VBR traffic will be considered as an input for the study. B. Characteristics of a Queuing Network Model There are three components that have certain characteristics that must be examined before the simulation model is developed. 1. Arrival Characteristic: The pattern of arrival input traffic is mostly characterized to be Poisson arrival processes (Khoshnevis 1994). Like many random events, Poisson arrivals occur such that for each increment of time (T), no matter how large or small, the probability of arrival is independent of any previous history. These events may be individual cells, a burst of cells, cell or packet service completions, or other arbitrary events. The probability of the inter-arrival time between event t, is defined by the inter-arrival time probability density function (pdf). The following formula gives the resulting probability density function (pdf), which the inter-arrival time t is larger than some value x when the average arrival rate is λ events per second: 71

4 e fx( t) =, λt, for for p( t x) = Fx( x) = x t t < e λx x p( t > x) = 1 Fx( x) = λe λ dx = 1 λe λx Queuing theorists call Poisson arrivals a memory-less process, because the probability that the inter-arrival time will be X seconds is independent of the memory of how much time has already expired. The formula of memoryless process is shown accordingly: λs P( x > s + t X > t) = P( X > s) = e, for s t > This fact greatly simplifies the analysis of random processes since no past history, or memory, affects the processes commonly known as Markov processes. The probability that n independent arrivals occur in T seconds is given by the formula Poisson distribution: where P(n, T) = ( λt) n (e - λt )/ n! P(X) = probability of X arrivals, n = number of arrival per unit of time, λ = average arrival rate, E{n T} = λt = expected value of n for a given interval T, and e = The combination of these two thoughts in a commonly used model is called the Markov modulated Poisson process (MMPP) or ON/OFF bursty model. In this paper, the burstiness is varied by altering the T ON and T OFF. 2. Service Facility Characteristics: In this paper, service times are randomly distributed by the exponential probability distribution. This is a mathematically convenient assumption if arrival rates are Poisson distributed. In order to examine the traffic congestion at output of ATM link as STM-4 (622.8 Mbps), the service time in the simulation model is specified by the i speed of output link, giving that a service time is.6815 µs per cell. 3. Source Traffic Descriptor: The source traffic descriptor is the subset of traffic parameters requested by the source (user), which characterizes the traffic that will (or should) be submitted during the connection (Aida and Saito 1995). The relation of each traffic parameters referring to the ATM Forum (ATM Forum 1995) used in the simulation model is defined below. PCR = λ a = 1/T in units of cells/second, where T is the minimum inter-cell spacing in seconds (i.e., the time interval from the first bit of one cell to the first bit of the next cell). This research focuses on four cases as follows: PCR = λ a = Gbps (3, cells/s). Hence, T =.25 µs (1/3,998,956 s) PCR = λ a = 848 Mbps (1,999,732 cells/s). Hence, T =.5 µs (1/1,999,732 s). PCR = λ a = Mbps (1,333,152 cells/s). Hence, T =.75 µs (1/1,333,152 s). CDVT = τ in seconds. This traffic parameter normally cannot be specified by user, but is set instead by the network. Recommendation I.371 defines the minimum CDVT at a public UNI. For LB mechanism, a single bucket depth of CDVT cells and a nominal cell inter arrival spacing T, note that approximately CDVT/T cells can arrive back-toback. Results and Discussion The comparison between existing flow control algorithms (Tahoe, Reno, New Reno, SACK) and ARC is shown in graphs. The experiment has been set the maximum window size (WindowSize) to be 3 so does our proposed algorithm (ARC). With the burst/silence ratio 1:, the average inter-arrival cell rate defines as.25,.5 and.75 µs. Fig. 3 illustrates the throughput against inter-arrival cell rate. Fig. 4 illustrates utilization of link against interarrival cell rate. Fig. 5 illustrates cells drop 72

5 against inter-arrival cell rate. From Figs. 3-5, ARC gives the highest performance, follows by SACK, New Reno, Reno, and Tahoe. Conclusion ARC gives a highest performance in case of high congestion (where the input arrival traffic rate is greater than the ATM fabric service rate), it allows more cells drop to retransmit by adjusting the window size directed to the arrival traffic (λ a ) and number of cells drop. On the other hand, ARC also gives remarkable better performance compared to existing algorithms in the case of noncongestion Throughput Percentage (x1)% Utilization of link Reno New Reno Tahoe SACK ARC Inter arrival time (Microsec) Fig. 4. Utilization of link Number of Cells (x1) Number of Cells drop Reno New Reno Tahoe SACK ARC Inter arrival time (Microsec) Number of Cells drop (x1) Fig. 3. Throughput Reno New Reno Tahoe SACK ARC Inter arrival time (Microsec) Fig. 5. Number of Cells drop 73

6 References Aida, M.; and Saito, H Traffic contract parameters and CAC guaranteeing cell-loss ratio in ATM networks. IEICE Trans. Commu. E78-B(1). Al-Wakeel, S.S.; Rikli, N.E.; and Al-Wehaibi, A.A Evaluation of Fairness Strategies for ATM Congestion Control Policing Mechanisms. IASTED Proc. Int. Conf. Applied Modeling & Simulation (AMP 99). Allman, M.; Paxson, V.; and Stevens, W TCP Congestion Control. RFC2581. ATM Forum User-network interface (UNI) Specification version 3.1, The ATM Forum Technical Committee, Prentice Hall, Englewood Cliff, NJ, USA. Kevin, F.; and Sally, F Simulation-based Comparison of Tahoe, Reno, and SACK TCP. Computer Comm. Rev. 26: Khoshnevis, B Discrete Systems Simulations. McGraw-Hill, New York, NY, USA. Kouvatsos, D.D Performance Modelling and Evaluation of ATM Networks Vol. 1, Chapman & Hall, London, England. Leerujikul, G.; and Jittawiriyanukoon, C. 23. Evaluation of TCP over Satellite Network with the Proposed Sliding Window Algorithm. Proc. InTech 3, pp Low, S. 21.Equilibrium & Dynamic of TCP/AQM. Sigcomm. August 21 /TCPaqm.ppt Matthew, M.; Jamshid, M.; Sally, F.; and Allyn, R TCP Selective Acknowledgment Options. Internetdraft RFC 218. Pakdeepinit, P.; and Jittawiriyanukoon, C. 22. Performance Comparison of ATM Policy Mechanisms with Telecommunications Traffic. Proc. 6th World Multiconference on Systematics, Cybernetics and Informatics, pp Shioda, S.; and Saito, H Connection Admission Control Guaranteeing Negotiated Cell-Loss Ratio of cell Streams Passing through Usage Parameter Control. IEICE Trans. Commu., E8-B (3): Yang, D. 2. TCP performance and flow control. Candle Computer Rpt. 22 (2). 74