Efficient Signaling Algorithms for ATM Networks

Transcription

1 Efficient Signaling Algorithms for ATM Networks See-Mong Tan Roy H. Campbell Department of Computer Science University of Illinois at Urbana-Champaign 1304 W. Springfield Urbana, IL Abstract ATM provides a connection-oriented model of communication. Each conversation between two end-points occurs over a virtual circuit. All cells for the same conversation flow along the same virtual circuit. For the circuit to be established, a call setup process is initiated by the source host and run over all intermediate switches in the path, as well as the destination. In the current ATM standard, signaling messages are carried between any pair of adjacent switches on a special reserved channel. Switch control software implement call setup protocols and the call setup algorithms defined by the protocols. Various such protocols exist, such as SPANS[5] and Q.2931[2]. We enumerate several properties that all ATM signaling algorithms must satisfy. With current signaling software and protocols, call setup time depends linearly on the number of switches in the setup path. As ATM network diameters grow larger, current standards result in large latencies in call setup time. When the call setup time dominates over the transmit time, the network throughput is low. To enable better signaling performance, we are exploring call setup algorithms in which physically non-adjacent switch controllers may communicate directly. ATM signaling topologies are essentially plastic and do not have to follow physical network topologies. The signaling channels between any two switch controllers may be hardware mapped through any intermediate switches; the intermediate switch controllers are not involved in the transmission of these signaling messages. Any arbitrary signaling topology can be overlaid on the physical network. A signaling overlay can be mapped onto an ATM network in the ATM meta-signaling phase. The simplest interconnection strategy is one where all the switch nodes are fully connected by virtual circuits. In this case, we demonstrate two algorithms which significantly improve performance. Performance is characterized in this paper by the number of messages and the number of rounds. Minimizing the number of messages in a signaling algorithm frees network bandwidth for carrying other data. Our enhanced linear algorithm achieves the minimum number of messages. Minimizing the number of rounds will have the greatest effect on call setup latency. Our multicast call setup algorithm achieves log performance in the number of rounds, where is the number of switches. 1 Introduction ATM provides a connection-oriented model of communication. Each conversation between two endpoints occurs over a virtual circuit. The switch translation hardware performs a VCI translation to correctly switch ATM cells from one port to another. For the circuit to be established, a call setup process is initiated by the source host and run over all intermediate switches in the path, as well as the destination. The intermediate switches must be informed in order to set up the appropriate translation maps in their hardware. The destination must be informed of the incoming call, and it can choose to either accept or 14/1

2 Figure 1: Simple Linear Setup Algorithm reject the call. Similarly, any one of the intermediate switches may reject the call, if, for example, it cannot honor the bandwidth requirements of the new call. If all parties agree, then the virtual circuit is established and data can begin to flow. In the current ATM standard, signaling messages are carried out-of-band. Between any pair of adjacent switches, a special reserved channel is used to carry signaling messages. Signaling in ATM switches is performed by the switch controller. Switch control software implement call setup protocols and the call setup algorithms defined by the protocols. Various such protocols exist, such as SPANS[5] for Fore Systems ATM switches, and Q.2931[2], the standard adopted by the ATM Forum for User-Network Interface signaling. In previous work, we designed the GULP protocol as well as an architecture for signaling software[7][3] for the XUNET (Experimental University Network) project[6]. For signaling to be possible, the switch s hardware cell translation map is set such that cells received on a port along the special reserved signaling channel are sent to the switch controller. The controller treats these cells as signaling messages sent from the adjacent controller or host. Similarly, the controller communicates with other adjacent network nodes by sending messages on the special reserved channels connecting it to other nodes. With current signaling software and protocols, call setup time depends linearly on the number of switches in the setup path. As network diameter grows larger, current standards result in large latencies in call setup time. When the call setup time dominates over the transmit time, the network throughput is low. This is particularly undesirable for short calls. In this paper, we develop metrics for signaling algorithm performance. Then we present two signaling algorithms for ATM networks that reduce the call setup cost. Performance is characterized in this paper by the number of messages and the number of rounds (section 2). Minimizing the number of messages in a signaling algorithm frees network bandwidth for carrying other data. Our enhanced linear algorithm achieves the minimum number of messages. Minimizing the number of rounds will have the greatest effect on call setup latency. Our multicast call setup algorithm achieves log performance in the number of rounds, where is the number of switches. In section 2, we describe two simple metrics used to measure signaling algorithm performance. In section 3, we take a look at signaling topologies and how they may be exploited to achieve better signaling performance. Then in section 4, we give some assumptions and properties that all signaling algorithms should satisfy for correctness. In sections 5 and section 6, we present the two signaling algorithms described above. A graph is presented in section 7 showing the relative cost of each signaling protocol in terms of the number of switches in the route. Related work and some concluding remarks are presented in section 8. 2 Performance Metrics for Signaling Algorithms We first need a way to quantify the performance of a signaling algorithm. Consider the simple linear call setup algorithm. Suppose that host wishes to communicate to host and further that there as are intermediate switches 1, 2,,, with connected to 1 and connected to. Host begins by sending a Setup message to 1. Each switch propagates the Setup message to 1 for 1 1. The final switch sends the Setup message to host. If decides to accept the call, it sends an Accept message to and this message propagates backwards through all switches to host. When a switch receives an Accept message, it sets its translation map to enable the virtual circuit. If, at any point during 14/2

3 the forward propagation of the Setup message, a switch or the host decides to reject a call, it can send a Reject to the previous node. The Reject propagates backwards to host. Thus if receives an Accept, the call was successfully set up, otherwise, if receives a Reject, the call was unsuccessful. The simple linear algorithm is common. SPANS, Q.2931 and GULP are all protocols for the simple linear setup algorithm and all have similar semantics. Suppose there are switches in the call setup route. We see that for successful setup, all linear setup algorithms require: 2 1 messages, with 1 Setup messages and 1 Accept messages, 2 1 rounds, with 1 rounds for forward propagation and the same number for backward propagation. That is, there is one message per round. In this paper, we characterize signaling algorithms by these two basic criteria: messages and rounds. A round is defined as a step in the algorithm where a node 1. receives one message, 2. processes it and 3. sends another message. 1 We justify this metric by noting that the time required for processing a message by the switch controller software is significantly more than the time required for transmitting a message. Minimizing the number of messages in a signaling algorithm frees network bandwidth for carrying other data and is an important feature of the design. Minimizing the number of rounds decreases the setup latency. An efficient algorithm should thus minimize both messages and rounds. 3 Signaling Topologies In the current ATM standard, each adjacent pair of switches share one signaling channel. The signaling topology follows the physical network topology: switches can send signaling messages only to adjacent switches. In this model, the simple linear setup algorithm is the best that one can do in terms of call setup. To enable better signaling performance, we are exploring call setup algorithms in which physically non-adjacent switch controllers may communicate directly. The signaling channels between any two switch controllers may be hardware mapped through any intermediate switches; the intermediate switch controllers are not involved in the transmission of these signaling messages. Note that any arbitrary signaling topology can be overlaid on the physical network. We call such a virtual network a signaling overlay. For simplicity, we consider the case for a fully connected signaling overlay. Many other topologies are possible and we are investigating alternatives. For the fully connected case, all switch controllers may send signaling messages to any other. Thus, for a call setup involving a sequence of switches 1, 2,,, each switch controller for switch can communicate directly with each of the controllers for all,!#" $. Since the controller software for the intermediate switches do not need to handle signaling messages from 1 to, sending a message from 1 to is only trivially more expensive than sending one from 1 to 2. The signaling overlay may be set up through the ATM meta-signaling phase. 4 Properties & Assumptions The network model we consider consist of switches and hosts. We assume hosts can only communicate with the switch to which they are directly connected. Nodes refer to both switches and hosts. Signaling channels carry control messages between switch controllers. These channels are assumed to be reliable. In our model, hosts are directly connected to the ATM switches in an ATM network. 1 For the initiator node, we discount condition 1 and 2, since the initiator begins the algorithm by sending a message without needing to first receive a message. 14/3

4 Figure 2: Enhanced Linear Setup Algorithm Assumption 1 (Host Communication) An end host can only send signaling messages to and receive signaling messages from the control software on the switch to which it is directly connected. This is reflected in the ATM User-Network Interface (UNI) specification recently agreed upon by the ATM Forum. The proposed algorithms in this paper do not modify any semantics of the standard ATM UNI interaction. The following properties apply to any signaling algorithm. We consider the case of host, the initiator, wishing to set up a call to host, the responder, with intermediate switches in the route. Property 1 (Correctness) Before the first switch sends a message to the initiator indicating a successful call setup, it must be that: the destination has accepted the call, and all intermediate switches have accepted the call and have set up their hardware translation maps. This property ensures correctness in the call setup process. When the initiating host receives a positive reply, the circuit must actually have been set up. We constrain our protocols not to be optimistic. Property 2 (Pairwise Negotiation) There must be some pairwise communication between any pair of nodes. Since adjacent nodes must negotiate on a virtual circuit identifier to use, they must exchange messages with one another. For any pair, it is sufficient for one of the pair to unilaterally decide on the VCI and inform the other. 2 5 Enhanced Linear Algorithm We now present an algorithm which is minimal in the number of messages and works in half the number of rounds of the simple linear case. Figure 5 illustrates the enhanced linear algorithm. Setup signaling messages are forward propagated node by node as in the simple linear algorithm. However, when a switch receives a Setup message and can accept the call, it sets up its translation tables first before forwarding the message. The host is signaled by the last switch. If any switch or host along the forward propagation phase cannot accept the call, it sends a Reject message which propagates backwards to the initiator. If the last switch receives an Accept from the destination host, then it sends an Accept directly to the first switch. On receiving an Accept, the first switch can confirm that the call is set up by forwarding the Accept to the initiator. We see that the enhanced linear algorithm satisfies the Correctness Property (Property 1), since 1 indicates a successful call setup to the initiator only after receiving a message from. But this can only happen if each of the destination accepted the call and each of the subsequent switches also accepted 2 Unilateral decisions may result in the channel collision problem. Suppose switches % 1 and % 2 participate in two simultaneous call setups. % 1 may choose to use VCI #6 for call 1, and inform % 2. At the same time, % 2 may choose the same VCI for call 2 and inform % 1. The channel collision problem may be handled by requiring adjacent nodes to allocate VCIs from opposite sides of the VCI space, ie. % 1 allocates from low to high, while % 2 allocates from high to low. The detailed solution to the problem appears in [7]. 14/4

5 the call and have set up their translation maps. In addition, there was also pairwise negotiation of VCIs between each pair of switches in the route. This algorithm uses 4 messages and the same number of rounds. The essential idea is to use the enhanced network to bypass sequential backward propagation. We call this the enhanced linear setup algorithm. This algorithm gives the minimum number of messages and half the number of rounds required of the simple linear algorithm. It presents an improvement over the simple linear algorithm. However, the number of rounds still depends linearly on the number of intermediate switches in the route. 6 A Multicast Algorithm in &(' log )+* Rounds Initiating host S1 Sn Destination host Figure 3: Multicast algorithm: Initiate We now present a multicast signaling algorithm that will terminate in log number of rounds. Many ATM switches support multicasting. This can be done in hardware, thus sending more than one message becomes not more expensive than sending one. Even when hardware multicast facilities are unavailable, the effect may be realized by serially transmitting to each destination in turn. If the time required for processing a message is large compared to the time required for sending one, then this method of multicasting does not become very much more expensive than single transmission. As we choose to exploit multicast, we need to refine our definition of a round. The definition is restated thus: a round is one step in the algorithm where some set of nodes: receive at most one message each process that message, and send one or more messages (perform a multicast). The initiator starts by sending a Setup message to the first switch 1 (figure 3). Figure 4: Multicast algorithm: Multicast 1 consults its routing algorithm and finds the route 2,,. It then performs a multicast to each switch in the route in one round (figure 4). Figure 5: Multicast algorithm: Pairwise negotiate 14/5

6 On receiving the multicast, each switch selects an unused VCI to use with its forward neighbor, and then performs a pairwise negotiation of VCIs for this call with its neighbor in the forward direction; both negotiates a VCI and informs the destination of the call (figure 5). If the destination accepts the Figure 6: Multicast algorithm: Destination response call, it replies by sending an Accept to (figure 6). Figure 7: Multicast algorithm: Collect replies The next phase collects replies from each of the switches (figure 7). Reply collection is essential to ensure correctness (Property 1). The first switch must be sure that each of the subsequent switches as well as the destination must have accepted the call before informing the initiating host of successful call setup. This can occur by having switches 2,, send their replies to 1 directly, but now 1 must handle 1 messages, thus requiring 1 rounds. What is required is a distributed reply collection method where each switch controller only processes one message in each round. S1 S2 S3 S4 Parent Parent Figure 8: First Round of Reply Collection We accomplish this by treating the switches in the path as nodes of a tree. Replies propagate up the tree. In the first round, each odd numbered switch is the parent of its subtree, comprising itself and the next higher (even numbered) switch. Figure 8 illustrates the first round of the reply collection process for a 4 switch call setup path. Child nodes send their acknowledgements to the parent nodes. Thus 1 receives a reply from 2, 3 receives a reply from 4, and so on. Figure 9 illustrates the second round of reply collection. 1 is now the root of the subtree comprising itself and the next 3 higher numbered switches. Replies are sent from 3 to 1. In general, at the th round of the reply collection process, a switch is a parent if!, / 1, log 43. A parent receives replies from child 5 where 6 7! 2. Thus with nodes, log rounds are required for reply collection. When 1 has collected all replies, it knows that all switches as well as the destination have accepted the call, and it can then inform the source that the call has been set up. This algorithm requires 4 messages from and to hosts, 8 1 messages for multicasting, 8 1 messages for pairwise negotiation, and 1 messages for collecting replies, for a total of messages. However, it only requires 5 log rounds. 14/6

7 Parent S1 S2 S3 S4 Figure 9: Second Round of Reply Collection 7 Performance Comparison Figure 10: Multicast algorithm: Reply to source The graph in figure 7 compares the theoretical performance in terms of the number of rounds for the simple linear, enhanced linear and multicast algorithms. For networks with diameters of three or more switches, the multicast algorithm can vastly outperform the standard case. 3 8 Related Work and Concluding Remarks In this paper, we have described a new approach to efficient signaling algorithms. We note that the signaling topology of an ATM network is essentially plastic and does not have to follow the physical network topology. A signaling overlay can be mapped onto an ATM network by setting up signaling channels between physically non-adjacent switches. The simplest interconnection strategy is one where all the switch nodes are fully connected; in this case, we demonstrated two algorithms which significantly improve performance. Performance is characterized in this paper by the number of messages and the number of rounds. Minimizing the number of rounds will have the greatest effect on call setup latency. Our multicast call setup algorithm achieves : log performance in the number of rounds. Although the multicast algorithm achieves sub-linear performance, it consumes a large number of virtual circuits for the signaling overlay alone. We are investigating other interconnection strategies which do not require so many virtual circuits, such as hypercube and mesh topologies. The multicast algorithm is similar to the method used in the Aurora network in that it uses source routing[4]. However, we require no special switch hardware but instead locate our signaling extensions within the software of the switch controller. Other approaches to minimizing call setup time include the Virtual Path approach[1]. Permanent virtual circuits can be trunked together in a virtual path from the source switch to the destination switch. A separate signaling channel is used between the source switch and destination switch. Call setup time is now constant, since signaling communication only occurs between the two switches. In the Virtual Path method, both virtual circuits and resource allocation along the virtual path must be pre-allocated and virtual circuits must be partitioned between low bandwidth data and high bandwidth video calls. As a result, network resources may be over allocated or improperly allocated, for example, more video calls may be requested than there are video circuits pre-allocated. In contrast, our method does not require pre-allocation of either virtual circuits or network resources. Resource and virtual circuit allocation for user calls can be entirely dynamic. 3 Figure 7 assumes that the cost of processing a message in a round is the same in all three algorithms. This assumption may not hold, particularly in the multicast case. The multicast algorithm may require more complex per-message processing. 14/7

8 * n + 1 n log(n) / log(2) Figure 11: Algorithm performance for simple linear (2;7< 1), enhanced linear (;$< 4) and multicast (=?> log ;8@ ) algorithms. 9 Acknowledgements David Putzolu was of immense help in getting this manuscript ready. References [1] Nikos G. Aneroussis and Aurel A. Lazar. Managing Virtual Paths in Xunet III: Architecture, Experimental Platform and Performance. In IFIP/IEEE International Symposium on Integrated Network Managment, Santa Barbara, California, May [2] ATM. BISDN User Network Interface Layer 3 Specification for Basic Call/Bearer Control. ATM Forum/CCITT Draft Recommendation Q.93B, [3] Roy Campbell, Sean Dorward, Anand Iyengar, Chuck Kalmanek, Gary Murakami, Ravi Sethi, Ce- Kuen Shieh, and See-Mong Tan. Control Software for Virtual-Circuit Switches: Call Processing. In Springer-Verlag, volume 653 of Lecture Notes in Computer Science, pages Springer- Verlag, Berlin, [4] B. S. Davie. Milestones in Development and Implementation of the Aurora Gigabit Network Test Bed. Telecommunications, 28(7):45, July [5] Fore. SPANS Protocol Specification Version 2.0. Fore Systems Inc. Publication, [6] A. G. Fraser, C. R. Kalmenek, A. E. Kaplan, W. T. Marshall, and R. C. Restrick. Xunet 2: A nationwide testbed in high-speed networking. INFOCOMM 92, pages , [7] See-Mong Tan. An Architecture for Call Processing. Master s thesis, Department of Computer Science, University of Illinois at Urbana-Champaign, Urbana, Illinois, April /8