Real-time Performance Evaluation of Line Topology Switched Ethernet

Transcription

1 International Journal of Automation and Computing 05(4), October 2008, DOI: /s Real-time Performance Evaluation of Line Topology Switched Ethernet Fan Cen Tao Xing Ke-Tong Wu Institute of Acoustics, the Chinese Academy of Sciences, Beijing , PRC Abstract: Recently, switched Ethernet has become an active area of research because of its wide uses in industry. However, its uses have various real-time constraints on data communications. This paper analyzes the performance of the line topology switched Ethernet as a data acquisition network. Network calculus theory, which has been successfully applied to assess the real-time performance of packet-switched networks, is used to analyze the networks. To properly describe the activity of switches, a novel approach of modeling data flows into or out of switches is addressed. Based on our model, a concisely analytical expression of the maximal end-to-end delay in line topology switched Ethernet is derived. Finally, the relative simulation results are demonstrated. These results agree well with the analytical results, and thus they validate the data flow modeling techniques. Keywords: Switched Ethernet, network calculus, end-to-end delay, line topology, real-time network. 1 Introduction Industrial communications are currently based on several kinds of specific networks, namely fieldbus, such as CAN, Profibuses, and DeviceNet. They interconnect industrial devices in order to exchange data for the purpose of monitoring, controlling, and other industrial processes. A large number of these applications are time-constrained, and hence, the main purpose of a fieldbus is to ensure that the end-to-end delays of messages are bounded and remain limited compared with the time-constraints of an application. The fieldbus has solved many communication problems in industrial processes; however, there is no universally accepted standard that supports communication among different fieldbus networks. Limited bandwidth is another disadvantage of the fieldbus. An existing and widely acceptable network technology has been found to be able to replace fieldbus; Ethernet is one of the most popular local area network (LAN) communication technologies. Several years ago, because of its random CSMA/CD bus arbitration, Ethernet could only be used for non-time-sensitive applications. At present, in most newly built Ethernets, the latest Ethernet switch technology is being used instead of the hub-based infrastructure. In other words, every device directly connects to a high-performance switch through a full duplex port such that every conflicting domain has only one device. In such point-to-point architectures, collisions cease to occur and the random back-off algorithm is no longer required. This provides an additional capability for Ethernet to support the transmission of time-critical information [1 7]. This aspect has received considerable attention for some time; current research mainly focuses on methods to determine the maximal delay of a network, in which field Network Calculus plays an important role [8 11]. Many researchers have attempted to develop a general model to represent switched Ethernet; however, their models cannot obtain a concisely analytical result that shows the maximal Manuscript received July 26, 2007; revised March 19, 2008 *Corresponding author. address: cenfan@gmail.com delay, and the key factors impact the delay. A consensus has been reached that due to the diversity in industrial communications, it is difficult to determine whether or not Ethernet is real-time capable. It depends on the limit value set for the industrial application. To derive the maximal delay of a given network, complex models are constructed. A phenomenon, called pay burst only once, is known to offer a closer upper bound of the delay. However, it is difficult to analyze the models, and most results are still hard to compute and are unintuitive. The objective of this paper is to estimate the delay bounds of a data acquisition system based on the line topology switched Ethernet. The reason for selecting line topology is as follows: firstly, line topology is basic and common in factory use; secondly, its performance is usually worse than that of other topologies because of the cascade switches, and therefore, it could act as a reference; lastly, it is easy to be described. The main aim of this paper is to derive a rather concisely analytical result that would facilitate estimation of the realtime performance of line topology networks. We also introduce an approach of applying network calculus to switched ethernet to avoid the pay burst only once phenomenon and thereby simplify the calculation. 2 Network calculus 2.1 Overview Network calculus is a newly developed theory based on min-plus algebra. It is now widely used to assess the realtime performance of communication networks. It is based on the fundamental work of [12, 13] and has been fully developed in [14]. This theory is tailored to the analysis of switched networks used in industrial scenarios. On one hand, this theory models all network elements as nodes and data flows. Networks on buses, such as conventional Ethernet, are difficult to describe using these elements. However, switched networks are suitable in this case. On the other hand, from certain reasonable assump-

2 F. Cen et al. / Real-time Performance Evaluation of Line Topology Switched Ethernet 377 tions about input data flows and service node activities, network calculus provides deterministic results on network delays, backlogs, and throughputs; it is different from the conventional methodology that uses stochastic process theory. This property is very important in industrial usage, which requires a strong QoS guarantee. 2.2 Basic concepts and results The basic concepts of network calculus include the arrival curve, service curve, and min-plus convolution/deconvolution. Here, only some basic and related concepts are covered to clarify the notations below; readers may consult [14] for a full introduction. Arrival curves quantify constraints on the number of packets or the number of bits that flow in a time interval at a service node. Let F be a data flow and R(t) bethe number of packets or bits of F arriving in the time interval [0,t]. We say that the flow is constrained or has an arrival curve α(t) if for all 0 s t, R(t) R(s) α(t s) (1) where α(t) is a non-negative and non-decreasing function. A service curve describes the service of flow F in a node, which is a work station or a switch. Let R (t) be the amount of data output in a time interval [0,t]. Then, the nonnegative, and non-decreasing function β(t) is called the service curve for F in this node if for all t 0, R (t) R β(t) = (2) inf 0 s t {R(s)β(t s)} where denotes the convolution operator. The output flow is constrained by the arrival curve α β(t) =sup 0 s t {α(t s) β(t)} (3) where denotes the deconvolution operator. The constraints given by the arrival and service curve for a flow suffice to calculate an upper bound on the delay of a packet or bit in the service node. The delay is bounded by the horizontal distance between α and β: delay sup t 0 {inf {s 0:α(t) β(t s)}}. (4) There is a specific result in network calculus on a service node serving two or more flows according to a first come first serve (FCFS) strategy. Consider a lossless node serving two flows in the FCFS order. Denote these two flows by F 1andF2, with arrival curves α1 andα2, respectively. Assume that the node guarantees a service curve of β to the aggregate of the two flows. Thus, F 1 is guaranteed by the service curve { [β(t) α 2(t θ)] for t>θ β 1(t) = (5) 0 for 0 t θ where [x] denotes (x x )/2. Conventionally, this property is called FCFS splitting. The widely used arrival curve for describing a data flow is the leaky bucket arrival curve. It is defined by α(t) = rt b for t>0, and 0 otherwise. Having a leaky bucket arrival curve allows a source to send b bits at once, but without exceeding r bit/s over the long run. Service curves of switches are mostly defined by β(t) =C [t T ],which is called the rate-latency service curve. Data is delayed by a fixed time T and then routed out at a rate C. Inreality,this service curve only represents one output port of a switch. These two curves are popular because of their simplicity. We will also apply these in our analysis. 2.3 Pay burst only once principle Consider two systems with service curve β 1 and β 2.What happens when a flow transverses the two systems in sequence? Boudec and Thiran [14] showed that the concatenation of the two systems offered an effective service curve of β 1 β 2 to the flow. Consequently, there are two ways to obtain the delay bounds: 1) by applying the network service curve; 2) by iteratively applying the individual bounds on every node. The result obtained by the latter approach is always greater than that of the former. This is because the delay due to the burstiness of the input flow is calculated twice; this phenomenon is called pay burst only once. While using network calculus to obtain the end-to-end delay of a large system, it is important to apply the pay burst only once principle. However, the use of this principle has two disadvantages. First, there is no universal approach to tell us how to apply this principle. People have to try their best to use this principle wherever they can. Second, the analysis of a large network becomes very difficult and always yields intricate results [15,16]. In the next section, a novel approach is raised to analyze switched Ethernet without the interference of the pay burst only once phenomenon. 3 Switched Ethernet analysis 3.1 State of art in Ethernet switches A switch can be functionally considered as a multiport bridge. However, it is more powerful than a conventional bridge due to its application specific integrated circuit (ASIC) based architecture and its ultra-rapid simultaneous multiple access memory. At present, highperformance switches are increasingly replacing conventional hubs. Commercially available switches from various manufacturers have the following technical properties in common. Store-and-forward. There are mainly three switch fabrics. They have shared memory, matrix, and bus. Shared memory is the most popular architecture; it can avoid the head of line (HOL) blocking. By using shared-memory, packets are always processed in the store-and-forward mode. Work-conserving. Switches will route out the packets received as soon as possible. They make full use of their output ports until their buffers are empty. First come first serve (FCFS) and priority queueing (PQ) strategies. When there are two flows originating from different ports but requiring the same output

3 378 International Journal of Automation and Computing 05(4), October 2008 port, the former strategy will serve the packet that arrives earlier, while the latter strategy would check and assign a non-preemptive priority to one flow. Wire-speed. All recent Ethernet switches are announced to operate with wire-speed and non-blocking. Wire-speed means that all ports of a switch can simultaneously transmit or receive at their full bit rates. It requires that the switch fabric can operate at a bit rate equal to the aggregate speeds of all the ports. A wire-speed switch does not require a back-to-back test and large memory because it can output the data that comes in from all ports in time. 4 Case study 4.1 Network configuration We assume the following application scenario: the Ethernet analyzed is used in a data acquisition system in which dozens or hundreds of identical sensors transmit cyclic messages to a central processing node. As mentioned before, we assume that it is a line topology network. There are N switches in a cascade and each of them is attached to M sensors; thus, there are M N sensors in total. The switches observe the FCFS strategy. Fig. 1 illustrates the network. 3.2 Procedure description As mentioned earlier, in switched Ethernet, every node is directly connected to a switch; thus we can describe an entire network if and only if we know all the data flows into and out of every switch. To obtain the end-to-end delay of a certain path in the network, we follow the procedure below. 1) Select the nodes through which the researched data flow F passes to form a node set S = {source node, switch F 1,switch F 2,,switch F N F, destination node}. Here, assume there are N F switches between the source and destination that relay the messages. 2) Choose bit granularity. The burstiness of all arrival curves, or α(0 ), should be the maximum packet length. In the store-and-forward operation, a packet will not be processed until all its bits are received; therefore, assume that they arrive at once. This burstiness constraint concerning the arrival curves is a basic assumption. 3) All data coming from an input port should be constrained by a single arrival curve. Since only one packet can arrive at a time in the physical channel, the arrival curve satisfying the above burstiness constraint must exist. 4) Each output port of each switch in S has its own service curve; this is obvious since all ports work simultaneously. 5) Use (3) to calculate the output flow, and then use a greedy shaper to reshape it. A greedy shaper is a shaper that delays the input bits in a buffer, whenever the sending of a bit would violate the shaping curve. Therefore, we should output them as soon as possible. Here, the shaping curve is α(0 )C t, whereα(0 ) is the burstiness of the input flow and C is the channel rate. The greedy shaper follows work-conserving and wire-speed properties and ensures that the output flow observes the burstiness constraint. Boudec and Thiran [14] showed that greedy shapers did not increase the delay or buffer requirements in this case. 6) Compute delays for interesting flow F by (4) and (5) at all switches in S. Sum up all delays to obtain the end-to-end delay. This procedure considers the characteristics of popular switching techniques and physical channels. Because burstiness is constrained to one packet length, the pay burst only once phenomenon no longer exists, allowing us to focus on the delay on every switch. Fig. 1 Network topology 4.2 Analytical delay bounds Each sensor ij, for1 i N and 1 j M, generates a data flow constrained by the arrival curve α ij(t) =r ijtb ij. For switch i, assume that it has an identical service curve β(t) =C (t T ). Delay T is due to the multiplexing of packets and other internal management functions in the switch. Consider the farthest node, sensor NM; its packets go through all switches, i.e., S = {sensir NM, switch N,switch N 1,,switch 1, processing node}. Atswitch N, we use (5) to obtain the service curve: β N (t) =C (t T ) [ r Nj(t θ) b Nj ] { (C β N (t) = rnj) (t θ) for t θ (6) 0 for 0 t θ where θ = T bnj/c. Using (4), we obtain the delay at switch N : delay N = T bnj b NM C C. (7) rnj The integrated output flow in which sensor NM sdatais taken has the arrival curve as follows: M M α N (t) = r Njt b Nj. (8) This is still a leaky bucket arrival curve. Before flowing to switch N 1, it must be reshaped by the greedy shaper to eliminate the redundant burstiness. After reshaping, the flow will have a traffic specification (T-SPEC) arrival curve: α reshaped N (t) = min ( M M ) r Njt b Nj, Ct b N (9)

4 F. Cen et al. / Real-time Performance Evaluation of Line Topology Switched Ethernet 379 where b N = max(b Nj), for j = 1,,M. By repeating the above calculation and using the results from the former step, we eventually obtain the end-to-end delay delay =NT N 1 i=1 bnj b NM C C rnj ( M bij b ) i C C bnm M rij C (10) where b i =max(b ij, b i1), for i =1,,N 1andj = 1,,M. The last term is due to the path from sensor NM to switch N. The calculation is easy because every arrival curve needed is either leaky bucket or T-SPEC constrained. In many data acquisition applications, each sensor acquires data at the same rate and all the inbound packets have identical lengths. Hence, assuming b ij = b and r ij = r, we have delay = N(T Mb (N 1)b ) C C Mr b C (M 1)r. (11) Since T is usually negligibly small, the maximal end-toend delay approximately increases with N (the number of switches) and b (the packet length). 5 Simulation Fig. 2 End-to-end delay obtained by network calculus and simulation The simulation model was realized with Network Simulator The number of switches was selected from the set {3, 6, 9, 12, 15}. Packet length was selected from {64, 128, 256, 512, 1024, 1518}, which covers the shortest and the longest Ethernet packets in IEEE All sensors transmit packets simultaneously every 12.5ms with a 2% jitter. Fig. 2 shows the predicted end-to-end delay using network calculus and the simulation results. In our simulation, the average bit rate r does not vary, so the backbone bandwidth is linearly proportional to the number of switches and the packet length. As a result, the packets delay also increases with the backbone bandwidth utility. The maximal backbone bandwidth utility varies from 7.4% to 87.4% and the two surfaces agree well. In addition, Fig. 3 illustrates the percentage difference between the calculation and simulation results. We can see that the calculation results are always larger than the simulation results because network calculus theory always considers the theoretical worst-case scenario. Moreover, the error is less than 20% at all points considered. Thus, the procedure introduced above is useful in estimating real-time network performance. Under the heaviest bandwidth load, the maximal end-toend delay exceeds 13 ms, which is greater than the packetsending period. This may violate the real-time constraints in some factory applications. Fig. 3 Difference between calculation and simulation results 6 Conclusions In this paper, we proposed a new procedure to calculate the end-to-end delay in switched Ethernet using network calculus. Additionally, we applied this procedure to assess the real-time performance of line topology switched Ethernet, which only has aggregate flows. The analytical results presented in this paper using network calculus show two points: first, they agree well with the simulation results, showing that the proposed procedure is applicable in this circumstance; second, the maximal end-to-end delay in

5 380 International Journal of Automation and Computing 05(4), October 2008 the network that we researched linearly increases with the packet length and switch number, which are the key factors impacting the delay. It can guarantee deterministic bounds on transaction times. Depending on various requirements and appropriate network design, the switched Ethernet is able to satisfy the real-time constraints in industrial applications. In real industrial applications, the networks always carry many kinds of data. Hybrid data flows that contain cyclic and acyclic data with different priorities will be studied later. Acknowledgement The authors would like to thank the anonymous reviewers for their advice. References [1] A. Pandey, H. M. Alnuweiri. Quality of Service Support over Switched Ethernet. In Proceedings of IEEE Pacific Rim Conference on Communications, Computers and Signal Processing, IEEE Press, Victoria, BC, Canada, pp , [2] Y. Q. Song. Time Constrained Communication over Switched Ethernet. In Proceedings of the 4th IFAC International Conference on Fieldbus Systems and Their Applications-FeT, Nancy, France, pp , [3] K. C. Lee, S. Lee. Performance Evaluation of Switched Ethernet for Real-time Industrial Communications. Computer Standards & Interfaces, vol. 24, no. 5, pp , [4] Z. Wang, Y. Song, J. M. Chen, Y. X. Sun. Real Time Characteristics of Ethernet and Its Improvement. In Proceedings of the 4th World Congress on Intelligent Control and Automation, IEEE Press, Shanghai, PRC, vol. 2, pp , [5] A. Jacobs, J. Wernicke, S. Oral, B. Gordon, A. George. Experimental Characterization of QoS in Commercial Ethernet Switches for Statistically Bounded Latency in Aircraft Netwoks. In Proceedings of the 29th Annual IEEE International Conference on Local Computer Networks, pp , [6] J. D. Decotignie. Ethernet-based Real-time and Industrial Communications. Proceedings of the IEEE, vol. 93, no. 6, pp , [7] K. C. Lee, S. Lee, M. H. Lee. Worst Case Communication Delay of Real-time Industrial Switched Ethernet with Multiple Levels. IEEE Transactions on Industrial Electronics, vol. 53, no. 5, pp , [8] J. Jasperneite, P. Neumann, M. Theis, K. Watson. Deterministic Real-time Communication with Switched Ethernet. In Proceedings of the 4th IEEE International Workshop on Factory Communication Systems, IEEE Press, Västeräs, Sweden, pp , [9] J. P. Georges, E. Rondeau, T. Divoux. Evaluation of Switched Ethernet in an Industrial Context by Using Network Calculus. In Proceedings of the 4th IEEE International Workshop on Factory Communication Systems, IEEE Press, Västeräs, Sweden, pp , [10] M. Bertoluzzo, G. Buja, S. Vitturi. Ethernet Networks for Factory Automation. In Proceedings of the IEEE International Symposium on Industrial Electronics, IEEE Press, vol. 1, pp , [11] J. P. Georges. T. Divoux, E. Rondeau. Comparison of Switched Ethernet Architectures Models. In Procceedings of the Conference on Emerging Technologies and Factory Automation, IEEE Press, vol. 1, pp , [12] R. L. Cruz. A Calculus for Network Delay, Part I: Network Elements in Isolation. IEEE Transactions on Information Theory, vol. 37, no. 1, pp , [13] R. L. Cruz. A Calculus for Network Delay, Part II: Network Analysis. IEEE Transactions on Information Theory, vol. 37, no. 1, pp , [14] J. L. Boudec, P. Thiran. Network Calculus: A Theory of Deteministic Queuing Systems for the Internet, Springer- Verlag, Berlin, [15] K. S. Watson. Network Calculus in Star and Line Networks with Centralized Communication, Technical Report IITB 10573, Fraunhofer-Institut für Informations-und Datenverarbeitung, Karlsruhe, [16] M. Fidler. Extending the Network Calculus Pay Bursts Only Once Principle to Aggregate Scheduling. In Proceedings of the Second International Workshop on Quality of Service in Multiservice IP Networks, Lecture Notes in Computer Science, Springer-Verlag, vol. 2601, pp , works. Fan Cen received the B. Sc. degree in electronic engineering from the Tsinghua University, PRC, in He is currently a Ph. D. candidate in the Institute of Acoustics, Chinese Academy of Sciences (IA- CAS). He received the Outstanding Student Award from the Graduate School of Chinese Academy of Sciences in His research interests include embedded system design and interconnection net- Tao Xing graduated from Tsinghua University, PRC, in He received the Ph. D. degree from the Institute of Acoustics, Chinese Academy of Sciences (IA- CAS), in He is currently an associate professor at the Sonar Engineering Laboratory, IACAS. His research interests include circuit and system, high performance signal processing, and data acquisition. Ke-Tong Wu received the B. Sc. degree in electronic science & engineering from Nanjing University, PRC, in He is currently a Ph. D. candidate at IACAS. His research interests include statistical and array signal processing, machine learning and their applications.