DTJ Number 3 September 1987 - Digital Technical Journals

New Productsdemands, capacity is expressed in terms of thenumber of users supported. The difficulty withusing this number as the independent variable isthat the designer must account for the resourceconsumption from all layers of the protocol. Thatis generally hard to do .Another difficulty is that the performancerequirements may vary for different higher-levelprotocols. Some may be delay sensitive , Forexample, terminal access protocols that returnechoes end to end are quite sensitive to delay.Other terminal access protocols that allow localediting and echoing are not so delay sensitive.File transfer protocols are not sensitive to thedelay but require high throughput. Therefore, ·to determine the capacity ofan extended LAN,the .designer must investigate both delay andthroughput as applied to the requirements ofparticular protocols and applications that use theLAN.Certain LANs constrain the configuration,owing to either physical layer limitations (suchas the distance over which the line drivers canoperate) or the interaction between the accessmethod of the data link layer and the propagationdelay. For example, Ethernet places a limit onthe maximum number of repeaters betweenany two communicating stations. l7·18 This constraintassures that the propagation delay budget,which is assumed by the access method protocol,will not be exceeded in any configuration.In an extended LAN, the designer may also wishto constrain the configuration based on theperformance expectations of higher-layer protocols. For example, he may require that therebe no more than a certain number of bridgesbetween two stations that use a delay-sensitiveprotocol. In general, the constraints are morecomplex when an extended LAN is configuredwith dissimilar LANs since the individualLANs may provide diffe rent delay /throughputcharacteristics.Another problem when determining capacityis estimating the amounts of traffi c remaininglocal to a subhet and leaving that subnt. Theworst case occurs when all traffic must be forwardedfrom a subnet through one or more levelsof backbone, thus creating the largest demand onthe resources of the backbone. One way thedesigner can handle this situation is to assumethat all the locally generated traffic must also becarried by the backbone. Increasing the load willthen define the system saturation point at which·the resources of the subnets will likely be underutilized.The additional capacity of the subnetcan then be used only for local traffi c. This calculationdefines the limits for the system withrespect to the ratio of local ;traffic to total trafficthat is possible.Using the above principles we developeda capacity-sizing methodplogy for extendedLANs . The diameter of ah extended LAN isIsized in the following fa shion. The averageone-way delay across the l ongest path cannotexceed 10 milliseconds, chosen as the delaybudget based on analyses of higher-layerprotocols that ate delay sensitive. One such protocolis the LAT protocol used for terminalaccess. 1 0A detailed simulation of that protocol was usedto study different configurations and values ofan average delay budget. The 1 0-milliseconddelay budget allowed for ariance in the delayand kept the protocol operation in the normalstates (without timeouts, ec.) . In addition, theoperating point must be sei: so that none of thelinks in the extended LANI run at greater than90 percent utilization . (Note that this utilizationmay occur at an offered lo :i d of much less than90 percent.) On an Ethernet this limit occurs foroffered loads of anywhere from 4 5 percent to90 percent utilization. The difference betweenthe utilization and the offered load is the overheadon the link. On an Ethernet this differenceincludes delays caused by collisions; on tokenrings it includes delays for token passing and thelike.This methodology assures, that the componentlinks in an extended LAN: are all running instable operating regions, and that the delay isIsimilar to that on a single LAN. The fulfillmentof these conditions is impohant so that the per-'formance expectations of hgher-layer protocolsare still met. Depending on the type of LANs usedin an extended LAN , the number of bridgesallowed (in series) will b ' e different. TokenaccessLANs often have higher average delaysthan Ethernet LANs. These delays could consumesome of the delay budget, which averages 1 0milliseconds. In the case of all Ethernet links, thenumber of bridges allowed in series is somewherebetween seven and nine. Further discussionof the performance aspects of extendedLANs may be found in the paper "Performance ·Analysis and Modeling of Digital 's Networking11Architecture. "19·Digital TecbnicalJournalNo. 3 September 198663