© Michel Kadoch 1 Attributs et Classification des commutateurs ATM

© Michel Kadoch 2 Caractéristiques Architecturales des commutateurs ATM Les caractéristiques architecturales des commutateurs ATM sont évaluées pour leur pertinence à rencontrer les exigences de performance des protocoles qui supporteront les services à offrir dans B-ISDN. Un commutateur ATM comprend un ensemble de N ports d’entrées, et N ports de sorties, un switch fabric, et un processeur de controle de gestion - management control processor (MCP). Switch Fabric MCP N N Input ControllersOutput Controllers

© Michel Kadoch 3 Contrôleurs I/O contrôleurs de port d’entrées fournie: buffering; cell replication (copying) pour multicast; traitement des cellules; translation VPI/VCI; multiplexage du trafic provenant de plusieurs equippements à basse vitesses; demandes de connexion de chemin et réservations à travers le switch fabric.

© Michel Kadoch 4 Switch Fabric

© Michel Kadoch 5 Management and Control Processor (MCP)

© Michel Kadoch 6 Buffering

© Michel Kadoch 7 Classification de Switch Fabrics ATM ATM Switch Fabrics Time division Shared memory Shared medium BusRing Space division Single path MatrixBanyanSorted Banyan Delta Multiple path Augmented Banyan Parallel planes Load sharing Recirculation

© Michel Kadoch 8 Commutation Time Division

© Michel Kadoch 9 Space Division Switching Dans la méthodologie “space division”, le switch fabric peut supporter plusieurs connexions au même moment. Les connexions sont basées sur la disponibilité de chemins physiques du switch fabric qui ne sont pas en conflit. 2 x 2 switching element a0a0 a1a1 b0b0 b1b1

© Michel Kadoch 10 Switching Elements a0a0 a1a1 b0b0 b1b1 a0a0 a1a1 b0b0 b1b1 a0a0 a1a1 b0b0 b1b1 a0a0 a1a1 b0b0 b1b1 Straight through Lower broadcast exchange Upper broadcast

© Michel Kadoch 11 Routing Routage de cellules à travers un « space-division switch » peut être accomplit par: self routing label routing self routing compte sur les interconnexions régulières du switching elements dans le fabric. label routing, le champ VCI dans l’entête est utilisé par chaque switching element pour prendre les décision de lien de sortie.

© Michel Kadoch 12 Representative Architecture Matrix and Fully-Connected Switches Also referred to as crossbar switches. Provide N 2 paths between N input and N output ports. A single contact pair for each input-ouput combination. inputsoutputs Complexity grows with the number of links, i.e., O(N 2 )

© Michel Kadoch 13 Knockout Switch A knockout switch is a matrix architecture. Any input can transmit to any output. The switch is self routing. input 0 1 N-1 output Knockout concentrator filter RR 0N-1 Filtering is based on the packet destination address The concentrator with R buffers (R<N) implements a selection algorithm for selecting R cells out of the maximum N possible cells arriving at that port. Worse case:N-R cells lost R out of N selection algorithm is used by the concentrator bus bus drivers Separate output buffers can be maintained and a shifter can be used to allocate cells to the buffers in order to ensure proper sequencing of transmitted cells on each virtual connection in an ATM Network.

© Michel Kadoch 14 Multistage Switches Two of the basic switching fabrics used for multistage switching are the Banyan and the Delta networks. Delta Networks N input port switch composed of b x b simpler switching elements. There are K stages, where N = b K (K=log b N) and N/b switches in each stage (column).

© Michel Kadoch 15 Example of Delta network 8 x 8 Delta 2 network, where b = 2, N = 8, K = log 2 8 = 3 columns (switches in a column) K N one half of the deck the other half of the deck a perfect shuffle

© Michel Kadoch 16 Multiple connections The bits of the destination address provide the required routing tags. The digits in the destination address are used to set the state of the stages Perfect shuffle Stage 1 Stage 2Stage destination port address white bit controls switch setting in each stage

© Michel Kadoch 17 Internal blocking Internal link blocking as well as output blocking can happen in a Delta network. The following example illustrates an internal blocking for connections of input 0 to output 3 and input 4 to output Perfect shuffle Stage 1 Stage 2Stage 3 blocking link ???

© Michel Kadoch 18 Output Blocking The following example illustrates output blocking for the connections between input 1 and output 6, and input 3 and output Perfect shuffle Stage 1 Stage 2Stage output blocking

© Michel Kadoch 19 Blocking Reduction Means to reduce the probability of blocking or increase throughput: installation of a distribution (randomizing) network in front of the switch to reduce or eliminate the possibility of internal blocking. Recirculate packets that cause blocking to the input ports, including them in the next cycle. Provide a contention resolution phase among the input ports. Subject the traffic between the stages of the network to a handshake protocol that synchronizes transmission and reception across the stages. Add extra stages to the multistage network and produce extra paths. Provide a number of buffers on each link or in the switching elements. Increase the bandwidth of the internal links relative to the input links. Increase network throughput by using several networks in parallel.

© Michel Kadoch 20 The Banyan Network The banyan network is another self-routing switching fabric, similar in topology to the delta network Stage 1 Stage 2Stage

© Michel Kadoch 21 Adding Load Distribution Because of internal blocking, input traffic is sorted on the destination ports. The network is known as Batcher sorter Perfect shuffle Stage 1 Stage 2Stage 3 B A A B A = Connection to output port 3 (011) B = Connection to output port 2 (010)

© Michel Kadoch 22 Output blocking To eliminate output contention if there are multiple packets with the same destination: Remove the duplicate-destination packets from the group and include them in the next batch for sorting and routing; submit only requests (into the group to be sorted) that have no destination in common.

© Michel Kadoch 23 Adding Recirculation Recirculating networks can redistribute, to the input ports, packets that were not successfully delivered in a given cycle. Switch fabric recirculation buffer This technique deals with output or internal blocking.

© Michel Kadoch 24 Starlight Switch Structure The Trap network examines the output of the sort network and removes all packets with the same destination address. The duplicates are routed back to the sort network for the next cycle with a higher priority. Concentrator Sort network Trap buffer Routing Banyan Network Input Output

© Michel Kadoch 25 Adding a contention resolution phase Three Phase Algorithm Phase 1 The input ports that have traffic, submit a transmission request to the sorting network. The sorter sorts the requests based on the destination. One of the requests that contests the access to the destination is selected; the others with the same destination address are purged. Phase 2 The sorter sends an acknowledgement to the input ports from which requests have been received and selected. Phase 3 The ports selected transmit their packets. This method can have some blocking problems, and phase 1 and 2 are overhead.

© Michel Kadoch 26 Adding Extra Stages One way to augment a Banyan network is to add extra stages to increase the number of paths from input ports to output ports. Each additional stage doubles the number of paths between any input-output pair. A better load distribution can increase throughput, alleviate blocking, and make the fabric less sensitive to the load on input ports.

© Michel Kadoch 27 Adding Internal Buffers Adding buffers also augments a Banyan network. The resulting configurations are known as buffered Banyans. The Integrated Service Packet Network (ISPN) described by Turner 1. The switch is constructed from 2 x 2 switches in which each link buffers one packet. The fabric connects 1024 ports using 5120 switching elements arranged in 10 stages. The internal links operate at eight times the speed of the external ports. So, for 100% load, an input port results in only 12.5% load on the internal links. A back pressure flow mechanism prevents contention at the internal links. When a packet arrives at a free link, the packet is not buffered. 1. J. Turner, « Design of an Integrated Service Packet Network », Proceedings Ninth ACM Data Comm. Symposium, ACM, New York, 1985,pp

© Michel Kadoch 28 Adding Parallel Switching Planes One way to increase throughput is to connect the input and output ports to multiple planes of the switch fabric. Each input port can distribute its traffic to the multiple homogeneous fabric planes. Similarly, each output port can be fed from multiple fabric planes.

© Michel Kadoch 29 Putting it all together The Sunshine switch is an example that incorporates: Batcher Banyan; Internal and output buffering; parallel fabrics; recirculation with dedicated input ports; self routing; level of service priority; trunk groups between internal nodes. Batcher sort 1 N Trap Network Concentrator Banyan 1 Banyan K 1 N Delay TT Select at most K cells per destination. Separate selected cells from the rest. Selector Input port Output port

© Michel Kadoch 30 Commercial Systems ATM strategy for major vendor classes and some commercially available ATM switches. Hub and Router Vendors Uses two stage strategy for introducing ATM. Phase 1 Add interface board to connect hubs or routers to an ATM switch over SONET (OC-3) or T3. Phase 2 The next step will be to incorporate an ATM switching module (switch fabric) into the hub or router. Several vendors are considering attaching to an ATM adaptor box.

© Michel Kadoch 31 Multiplexer and WAN Vendors They are working on or have already released ATM switches or adapters for the LAN. Examples: ATMX from Adaptive, Newbridge’s 361X ATM switch, the Synchrony product from Ascom Timplex, Stratacom’s BPX, GTE Government System’s ATM switch.

© Michel Kadoch 32 Computer Systems Vendors IBM, NEC, DEC, Sun Microsystems, HP, and AT&T are developing intelligent hubs that offer ATM support.