[ Pobierz całość w formacie PDF ]
.If the route cachealready contains a route to end system S and a packet is received from S with a route that differs from the one in the route cache, the destination can either overwrite the cache entry orkeep it.4.4.4 Route SelectionSuppose that an end station receives multiple copies of an all paths explorer packet.Whatcriteria should the station use to select a route? Again, the standard leaves it up to the end-station implementer.Some possibilities are as follows.1.Select the first packet received, on the theory that it traveled on the fastest path.2.Select the one that indicates a path with the maximum packet size.3.Select the one with the fewest hops.4.Select some combination of the preceding.For example, select the packet thatarrived first, from among the ones with at least a certain packet size; or select the onethat has the fewest hops, from among the ones that arrived within a certain amount oftime.5.Select the most recently received route.(Note: This is a bad idea.Why? Seehomework problem 4.)4.5 Source Routing versus Transparent BridgingSuppose that you were on a standards body evaluating the technical merits of source routingversus transparent bridging or that you were making a decision as to which scheme to adoptin your own network.What sorts of criteria could be used?4.5.1 Bandwidth OverheadThe number of paths in a network is exponential.Theoretically, an all paths explorer packetwill generate a copy of itself for each path through the network.Thus, theoretically, thenumber of copies of one all paths explorer packet that will be generated is exponential.Exponential grows very fast, especially with rich connectivity.Basically, every time there is a choice (if many bridges are on a LAN or a bridge has manyports), the number of packets is multiplied by the number of choices.The maximum routelength is 14 route designators, but there are actually 26 choice points in a path.(Each routedesignator contains both a LAN number and a bridge number, and that is why there are abouttwice as many choice points as route designators.The reason there are not 28 choice pointsin a route is that the first item is the source's LAN, which is fixed.The last item is the bridgeportion of the final route designator, which is blank.) When an all paths explorer packet istransmitted on a LAN, all the bridges on that LAN receive a copy.Each bridge then generatesa copy of its copy for each additional port to which the bridge attaches.The number of packets spawned as a result of a single endnode's attempt to start aconversation with a single other endnode is on the order of the average number of bridges oneach LAN to the 13th power multiplied by the average number of ports (more than two) oneach bridge raised to the 13th power.Even if the destination can be reached with a route oflength 2, after an all paths explorer packet is launched there is no way to stop it until it hasspawned all its copies.In my opinion, this exponential overhead, invoked every time any endnode wishes tocommunicate with any other endnode, is a fatal flaw of source routing.In small topologies thatare simply trees, with bridges that have only two ports, the exponential factor might not bevery serious.But people build transparent bridges with hundreds of ports, so it might not beunreasonable to imagine them building source route bridges with hundreds of ports.If therewere two such bridges in a network, the proliferation of all paths explorer packets would bevery serious.Source routing bridges also use extra bandwidth because the headers must be larger, butthat's hardly worth mentioning given the exponential proliferation of explorer packets.Transparent bridges do not make optimal use of bandwidth either; they waste bandwidth byusing only a subset of the topology, by not using optimal routes, and by forwarding packets needlessly before cache entries are established.But their bandwidth waste is not nearly asspectacular as with source routing.4.5.2 Ease of ConfigurationTransparent bridges are truly plug and play.Although there are plenty of parameters anetwork manager can use to fine-tune performance in a specific situation, there is no need tomodify any parameters unless desired.In contrast, source routing requires that every LAN be assigned a number and that everybridge be configured both with the LAN number for each of its ports and with a 4-bit bridgenumber for each pair of LANs it connects (although establishing restrictions on the topologiesmight enable a bridge to have only a single 4-bit number that would be unique for each pair ofLANs it connects).Misconfiguration can cause problems such as loops or severe duplicationof data packets.4.5.3 UniversalitySource routing requires support in the endnodes.Thus, any endnodes without source routingwould be unable to have their packets routed through an (SR-only) source routing bridge.Such stations would still be able to communicate with other stations on their local LAN butcould not communicate with any stations on remote LANs.On the other hand, transparent bridges are not completely transparent, especially when usedto connect different types of LANs.The packet size problem can be solved if the networkmanager configures all the endnodes to use the smallest maximum packet size allowed byany of the LANs.Alternatively, some sort of clever procedure in the end stations can be usedto determine the best packet size.One such procedure, described in Section 3.6.3, employsthe priority field on FDDI to signal that somewhere along the path between the source and thedestination, the packet is bridged through an 802.3 LAN
[ Pobierz całość w formacie PDF ]
