[ Pobierz całość w formacie PDF ]
.The registry also defined a way to map globalblocks of IP addresses into global blocks of IPX addresses.The real answer, if a true IPX Internet had taken off, was to force people either to renumberbased on a block of addresses they'd be given by their ISP or use IPX+ (see Section 9.7).9.6.4 Internal IPX Network NumbersWhen people found out I worked at Novell, the second question everyone seemed to ask mewas "What's an internal network number?"First, I'll explain the problem that is solved by internal network numbers.IPX, like IP, assignsan address to each interface.So a server that has two links has two addresses (see Figure9.14).Figure 9.14.Server with two links and two addresses On net 41, S must have an address of the form 41.something.Let's say that S's Ethernetaddress on the top LAN is a, and on the bottom LAN is b.So S's two IPX addresses would be41.a and 79.b.Server S chooses one of its addresses say, 41.a and advertises itself asthat address.The way IPX works, the client, wanting to reach the service at IPX address 41.a, broadcasts amessage on its LAN asking which router can help it reach net 41.A router that is a good pathto that network replies.Both R and S are good paths to 41, so both reply.C chooses the firstreply and uses that router to reach service 41.a.C is as likely to choose R as S.If C choosesR, all its packets to S will be one-hop suboptimal because C will send it to R, which willtransmit it onto the top LAN, where it will reach S.(When told this problem, my immediatereaction was "So?" I never get very excited about suboptimal routes.I think people should begrateful if their packets get there at all.)The solution was to assign S an internal network number say, 92.To the outside world thepicture in Figure 9.14 becomes the one shown in Figure 9.15.Figure 9.15.Server with two links and two addressesThe server advertises the address of its service as 92.something, usually a low number suchas 92.1.Now when C asks who is the best path to net 92, only S replies.This technique solves the one-hop-suboptimality problem.But it eats up network numbers,and there is only a 4-byte network number space.IP has a 4-byte space, but that space mustbe shared between the net and the node.IPX might seem to have a lot more addresses forLANs because the 4-byte network number is only for LANs.But because all the servers windup using network numbers, there really aren't very many more IPX network numbers than IPnetwork numbers.But the real problem is that this approach requires extra configuration.I liked IPX so muchbecause it was plug-and-play for the clients.And I was surprised to find that all servers mustbe assigned an internal network number, even those with only a single link and for whichtherefore there is no one-hop-suboptimality problem.I suggested making configuration of aninternal network number optional, not only for the single-link server case but also whencustomers prefer living with one-hop suboptimality sometimes and saving configuration (andnetwork numbers).But I was told that customers did not mind configuring the internal network number.Loved it, in fact.And that it would be more confusing for them to have to sometimesconfigure it and sometimes not need to configure it.OK.But then in a later release, the implementation gave the option of having the serverchoose its own internal network number.Yikes! It picks one at random, looks at the routingtable, and assumes that if the random number it chose for a network number is not currentlyreachable in the routing table it's safe to use for its own internal network number.So much foraggregatable addresses! So much for routing hierarchy, in which you don't see the addressesreachable in other parts of the network.Why would anyone consider putting in such a featureif customers really enjoyed configuring the internal network number?Anyway, don't use the option of letting the server choose its own internal network number.Incidentally, IP has the problem that nodes having multiple links have multiple addresses.It isa nuisance, not only for the one-hop suboptimality but because also sometimes a servicemight be unreachable at one address but reachable at a different address.So the internalnetwork number is helpful in these cases.CLNP and DECnet do not have the problem (unless a node has links in multiple areas)because level 1 routing in CLNP and DECnet routes to the node and not to the link.Therefore, a node having multiple links can still have a single address.9.7 IPX+IPX+ was intended both as an enhancement to IPX and a migration path to 16-byte IPv6addresses, while keeping the simplicity and plug-and-play nature of IPX.It solved the siterenumbering problem and expanded the size of the address, and it did not require modifyingthe routers in the customer's network.It was approved in about 1994 by the now-defunct IPXadvisory committee, a short-lived open standards body that worked on IPX.Originally the IPXadvisory committee consisted of e-mail lists and occasional short, focused meetings ofimplementers.Then the IPX advisory committee was folded into a larger organization run bymarketeers.Then the length and frequency of meetings, as well as the quality of the food,increased immensely, but the focus and technical content got lost.In addition, the world wasconvinced that what it really wanted was IP, so people decided to stop working on IPX.IPX+ may have been implemented but certainly has not been widely deployed.Still, the ideasare worth learning, and IPv6 is evolving so that one of its modes of operation will resembleIPX+
[ Pobierz całość w formacie PDF ]