DTJ Number 3 September 1987 - Digital Technical Journals

---A Digital Network Architecture OverviewRo ute calculatio n is done in a distributedfashion. Two types of no des are defined in DNA:end nodes, and routing no des. End no des haveonly a single attachment to a network; therefore,they do not need to calculate routes or forwardpackets on behalf of other no des. On theother hand, ro uting no des support multiple linksand forward traffic on behalf of other nodes;therefore, routing no des must calculate routes.Route calculation is performed using three majorcomponents:• An initialization sublayer that determineswhich links interconnect with which no des• A decision process at each routing node thatcalculates routes to all destinations (withinone area for Level I routing or to each areafor Level II routing)• An update process at each routing no de bywhich routing nodes exchange informationabout their ro utesThe routing algorithm runs whenever the initializationsublayer at a routing node detects alo cal to po logy change. It also runs periodicallyto ensure that ro utes throughout the network areco rrect. This routing algorithm, ro bust and selfstabilizing,recovers automatically from corruptionoccurring in routing databases stored inrouting nodes or from any number of simultaneoustopological changes.12The routing layer supports a variety of communicationsfacilities for communicating betweenadjacent nodes. A complete path from a sourcenode to a destination no de can use a mixture oflink types. Three main types are supported: dedicatedlinks using the DDCMP protocol, X.25packet-switched networks, and Ethernet LANs.X.25 packet-switched networks are treated bythe routing layer as a co llection of po int-topointvirtual circuits; hence, these networks functionsimilarly to DDCMP point-to-po int links. Endno des have a particularly simple task on these·types of links since end no des make no decisionwhen sending a packet out; they simply send iton the link to the adjacent node.End no de ro uting is somewhat more complexon Ethernet LANs since each station can sendto any other station on the Ethern et; therefore,the end no des attached to an Ethernet must makea ro uting decision. When sending to no desremote fro m their Ethernet, the end nodes mustsend to a ro uter. When sending to no des on theirEthernet, the end no des send directly to the destinationno de. End no des fo llow a simple procedureto determine which path to follow. If arouter is present and the end no des do no t knowabout a particular destination, they forward theirpackets to the router. If no ro uter exists or if theyknow a particular destination is on the Ethern et,they send their packets directly to the destinationaddress, using 48-bit Ethernet addressesderived fro m the 16 -bit destination no deaddress.End no des learn that particular nodes are ontheir Ethernet by receiving packets directly fromthose no des or by being informed by a router.This approach was chosen to reduce the memoryand overhead in end no des while still permittingmultiple Ethernet LANs to reside in one DNAarea. The alternate approach was to limit eachDNA area to a single Ethernet, which wo uld havelimited the size of Phase IV networks.A DECnet network, like a complex network ofroads, is subject to co ngestion sho uld it be overloaded.The routing layer incorporates severaldesign decisions to reduce the potential for congestion,to prevent lo cal congestion from spreadingglobally, and to minimize the impact of congestionon network performance. To minimizecongestion, traffic sho uld be kept out of congestedportions of the network. To accomplishthat, each node restricts the number of buffersavailable to traffic originating at a no de, therebygiving priority to traffic transiting the no de.Two design decisions help to prevent theglobal spread of local co ngestion. The first decision was to keep ro uting as a function of the network topology, not of the network load. The seconddecision was to handle co ngestio n byallowing packets to be discarded at a no de whenan output queue has filled, instead of slowingdown input to the node. This second decisionminimizes the impact of traffic flowing throughthe node that does no t need the congested link.The discard policy also prevents buffer deadlock,which occurred in early research netwo rks, bypreventing circular buffer waiting conditions.The performance impact of co ng estion is minimizedby this policy fo r limited buffer sharingbetween co ngested links.14Perhaps the most important decision made indesigning the DNA routing layer was to provide a"best-effort" delivery service instead of a "reliable"service. This decision wa s made for a varietyof reasons. First, implementing functions at20Digital TecbnU:aiJournalNo. 3 September 1986