kwe@bu-cs.BU.EDU (kwe@bu-it.bu.edu (Kent W. England)) (04/14/89)
In article <825@oregon.uoregon.edu> dsmith@oregon.uoregon.edu (Dale Smith) writes: >FDDI has an 11dB loss budget. If you want to survive a failed >station that has gone into bypass, then you need to figure you'll lose >4dB through the bypassed station, giving you 7dB. But, you have to spit >the 7dB between runs on each side of the failed station, giving you a >3.5dB budget between any two stations. Figuring 1dB per inactive Now this is interesting. This is the second person I've heard talk about optical bypass and FDDI. I don't understand why you want or need optical bypass for FDDI. With the dual counter-rotating ring you have the secondary ring in standby in case of node failure. If a node on the ring fails, then the adjacent nodes "wrap" and use the primary and secondary paths to create a new ring that bypasses the failed node. Why do you want optical bypass fault tolerance as well?
vjs@rhyolite.SGI.COM (Vernon Schryver) (04/14/89)
In article <29548@bu-cs.BU.EDU>, kwe@bu-cs.BU.EDU (kwe@bu-it.bu.edu (Kent W. England)) writes: > With the dual counter-rotating ring you have the secondary > ring in standby in case of node failure. If a node on the ring fails, > then the adjacent nodes "wrap" and use the primary and secondary paths > to create a new ring that bypasses the failed node. Without an optical bypass, how do you survive > 1 fault? If you have a nontrivial number of stations on the ring (i.e. hundreds), you are certain to have > 1 broken computer. (Take a reasonable MTBF, and remember the birthday paradox.) Yes, an optical bypass is expensive. So are FDDI transcievers, chipsets, and so on. If you are playing with FDDI, you don't care about $, at least not for a while. Since the vast majority of individual pairs or "work groups" of computers running operating systems released in Q1 1989 are unable to eat an entire ethernet, you are using FDDI for reasons other than simple, end-to-end performance, but that's another paradox. Vernon Schryver Silicon Graphics vjs@sgi.com
goodloe@ingr.com (Tony Goodloe) (04/15/89)
In article <29548@bu-cs.BU.EDU>, kwe@bu-cs.BU.EDU (kwe@bu-it.bu.edu (Kent W. England)) writes: > Now this is interesting. This is the second person I've heard > talk about optical bypass and FDDI. I don't understand why you want > or need optical bypass for FDDI. > > With the dual counter-rotating ring you have the secondary > ring in standby in case of node failure. > > Why do you want optical bypass fault tolerance as well? With an optical bypass you won't need to do something as drastic as wrapping the ring for something as minor as, say, a powered-down station. Also, with bypass and wrapping you could sustain two faults, for example, a dead node and a cable break, and still be operating hunky-dory (sp?). We were asking ourselves this same question several weeks ago, and have convinced myself that bypasses should be on all stations. My $.02. Tony Goodloe
bobk@boulder.Colorado.EDU (Bob Kinne) (04/15/89)
In article <29548@bu-cs.BU.EDU> kwe@buit13.bu.edu (Kent England) writes: > With the dual counter-rotating ring you have the secondary >ring in standby in case of node failure. If a node on the ring fails, >then the adjacent nodes "wrap" and use the primary and secondary paths >to create a new ring that bypasses the failed node. > > Why do you want optical bypass fault tolerance as well? The purpose of the dual ring is to recover from a cut or damaged connection between nodes. If a node fails or powers down, the optical bypass permits the ring to continue to function in normal, non-wrap mode, bypassing the failed or removed station. Any ring architecture needs to provide this bypass functionality. In some architectures the bypass may be initiated remotely from a ring master in the event the node fails in a manner that disrupts the network.
dsmith@oregon.uoregon.edu (Dale Smith) (04/15/89)
In article <29548@bu-cs.BU.EDU>, kwe@bu-cs.BU.EDU (kwe@bu-it.bu.edu (Kent W. England)) writes: > In article <825@oregon.uoregon.edu> dsmith@oregon.uoregon.edu (Dale Smith) writes: >>FDDI has an 11dB loss budget. If you want to survive a failed >>station that has gone into bypass, then you need to figure you'll lose >>4dB through the bypassed station, giving you 7dB. But, you have to spit >>the 7dB between runs on each side of the failed station, giving you a >>3.5dB budget between any two stations. Figuring 1dB per inactive > > Now this is interesting. This is the second person I've heard > talk about optical bypass and FDDI. I don't understand why you want > or need optical bypass for FDDI. > > With the dual counter-rotating ring you have the secondary > ring in standby in case of node failure. If a node on the ring fails, > then the adjacent nodes "wrap" and use the primary and secondary paths > to create a new ring that bypasses the failed node. > > Why do you want optical bypass fault tolerance as well? A basic issue in ring design is that of fault tolerance and reliability. There are two basic techniques used in FDDI for a ring to survive a station failure. The first, and most desirable, technique is that of bypass. If a station detects that it is no longer functioning properly or if it loses power, it should go into bypass. This means that the ring does not reconfigure, but rather that the node is no longer receiving and retransmitting the optical signals. If the station fails and this failure is not detected internally, then the ring must be reconfigured for it to survive. In this case, the adjancent nodes "wrap" to eliminate the failed station. This "wrapping" is also the technique used to survive a failed link. As vjs@rhyolite.SGI.COM (Vernon Schryver) of Silicon Graphics points out in article <30661@sgi.SGI.COM>, you must have bypass or you will never be able to survive the loss of two or more stations. This will become extremely important when you start buying your SGI (or SUN or DEC) workstations with an FDDI board rather than an ethernet board. (How would you like it if your ethernet died when two workstations were powered off :-).) Dale Smith Internet: dsmith@oregon.uoregon.edu University of Oregon BITNET: dsmith@oregon.bitnet Computing Center UUCP: ...hp-pcd!uoregon!dsmith Eugene, OR 97403-1212 Voice: (503)686-4394
kwe@bu-cs.BU.EDU (kwe@bu-it.bu.edu (Kent W. England)) (04/16/89)
In article <867@oregon.uoregon.edu> dsmith@oregon.uoregon.edu (Dale Smith) writes: > >There are two basic techniques used in FDDI for a ring to survive a >station failure. The first, and most desirable, technique is that of >bypass. If a station detects that it is no longer functioning properly >or if it loses power, it should go into bypass. > >As vjs@rhyolite.SGI.COM (Vernon Schryver) of Silicon Graphics points out >in article <30661@sgi.SGI.COM>, you must have bypass or you will never >be able to survive the loss of two or more stations. This will become >extremely important when you start buying your SGI (or SUN or DEC) >workstations with an FDDI board rather than an ethernet board. FDDI contains specs for ring concentrators, dual and single ring modes, and the SMT (soon) to make all these options and modes work together. Sorry, I still don't think optical bypass is useful. I think the typical backbone will be dual ring with wrap mode as the dominant fault tolerance approach. You might convince me that optical bypass would be useful here, but in my two years' experience with Pronet-80 I don't miss optical bypass. I haven't heard anyone else with Pronet-80 asking for bypass. I think the typical workstation ring will be single ring, because people can't afford and don't need dual ring on the workstation nets. Just like the IBM-PC token rings, the workstation nets will be star wired and the concentrator will take stations in and out of the ring, electrically I imagine. If the workstations have to have optical bypass to handle faults that the concentrators don't handle (or if you don't use concentrators), then how does FDDI deal with an unplugged optical connector? I don't think FDDI deals with optical bypass in the FDDI connector. You know, the IBM token ring shorts out when the station is unplugged and a lot of people have found that to be troublesome. But that's simple compared to optical bypass in the FDDI connector.
goodloe@ingr.com (Tony Goodloe) (04/18/89)
In article <29650@bu-cs.BU.EDU>, kwe@bu-cs.BU.EDU (kwe@bu-it.bu.edu (Kent W. England)) writes: > I think the typical workstation ring will be single ring, > because people can't afford and don't need dual ring on the > workstation nets. Just like the IBM-PC token rings, the workstation > nets will be star wired and the concentrator will take stations in and > out of the ring, electrically I imagine. I agree here, that if you have a single ring subnet coming off of a concentrator, that you don't need bypass on those stations. Any comments I made were thinking about a dual-attached station. Does anyone have any figures based on data that would suggest what portion of stations will be single-attached, and double-attached? tony goodloe
vjs@rhyolite.SGI.COM (Vernon Schryver) (04/19/89)
In article <29650@bu-cs.BU.EDU>, kwe@bu-cs.BU.EDU (kwe@bu-it.bu.edu (Kent W. England)) writes: > > FDDI contains specs for ring concentrators, dual and single > ring modes, and the SMT (soon) to make all these options and modes work > together.... I thought single attach rings were declared illegal last year. My reading of section 6 of X3T9.5/84-49, Update 2-1-89, finds only dual-attach A/B and and Master/Slave connections. Notice that the new CMT diagrams have boxes labeled "optical bypass", and that PC-Reset has words about when you should or should not fiddle with your bypass. If you have only a few stations on the ring, and they are well maintained exemplars of mature products, then one clearly can avoid bypasses (e.g. Proteon). If you like single attach rings, how can you not like optical bypasses? With single attach, aren't bypasses absolutely required? A lot of people talk about FDDI concentrators. I can't make the $'s come out in a way that makes concentrators plausible, assuming everything is dual attach or M/S. What am I missing? Consider: a concentrator on the dual ring and 5 slave stations: I count a total of (12*big $) 12 transceivers (the concentrator has 2 on the ring, and 5 masters, while the 5 stations have 5 slaves) (cheap) <= 6 MAC's (?) one dual-attach optical bypass (or two single-attach bypasses), unless you don't by bypasses. (big $) 1 concentrator, with at least enough smarts, power supply, etc. to do things like PC-Reset -- 5 workstations, each with SMT, OSI, NFS, and maybe a real application if it is small 5 stations on the ring: I count (10*big $) 10 transceivers (cheap) 5 to 10 MAC's. (5*?) 5 dual-attach optical bypass (or 10 single-attach bypasses) -- 5 workstations, each with SMT, ... No matter how deep the tree, there are always > 2 pairs of transceivers per station if you use concentrators, but exactly 2 pairs if you get on the ring. MAC's are (or will be) cheap. As I understand things, it is the transceivers that are not coming down in price. If you don't believe in bypasses, putting everything on the dual ring looks cheaper. Vernon Schryver Silicon Graphics vjs@sgi.com
rpw3@amdcad.AMD.COM (Rob Warnock) (04/19/89)
In article <30878@sgi.SGI.COM> vjs@rhyolite.SGI.COM (Vernon Schryver) writes: +--------------- | No matter how deep the tree, there are always > 2 pairs of transceivers per | station if you use concentrators, but exactly 2 pairs if you get on the | ring. MAC's are (or will be) cheap. As I understand things, it is the | transceivers that are not coming down in price. If you don't believe in | bypasses, putting everything on the dual ring looks cheaper. | Vernon Schryver | Silicon Graphics | vjs@sgi.com +--------------- This is quite true... if you believe that you *have* to use the expensive transceivers (optics) for your Master/Slave connections. The fact is, there are order-of-magnitude cheaper optics currently available that will cheerfully do "FDDI" (in quotes 'cause if's not "real" FDDI then they probably don't like you to call it FDDI), but at 820 nanometers and "only" up to 700 meters or so (instead of 2 km). So I venture to say that some workstation manufacturer might decide to introduce "cheaper-FDDI", with concentrators that have "real" FDDI on the backbone and "cheaper-FDDI" to the single-attach stations. And you can even do cheaper still: Noticing that, as Vernon points out, concentrators already have to have a *lot* of smarts in them, why not make the concentrator really be a multi-port bridge instead, and run something *really* cheap between the bridge and the workstations -- that is, a non-FDDI protocol [that still uses FDDI addressing, &c.]. I've been playing around lately with some designs for a super-simple "external serial backplane" based on point-to-point links [a fiber pair is a easier to handle than a SCSI cable, and a lot faster] using, for example, AMD's "TAXI" chips with (say) an H-P plastic-package 125 Mbaud fiber link. Parts cost to get a full-duplex 100 Mbit/sec "UART"-like interface is less than than $100 in volume ($40 + $43 + glue). Some packet buffer SRAM, some counters, and some PALs would bring that up to, say, a total of $150-200. [Yeah, cost, not retail. But that's still about one-tenth the parts cost of FDDI.] Now I was looking to make multi-master I/O "busses" ("Cheap-O-Bus"?) out of this stuff, with simple star repeater hubs [and CSMA/CD, which limits it to about 250 meters diameter, to reduce collisions], but the same hardware and bus protocol could be used to make a link from a workstation to an FDDI bridge. In fact, if you already had such a fiber I/O bus, an FDDI bridge is just another "shared peripheral". In that configuration, the economics shift considerably back towards using a hub/bridge/concentrator, and to using FDDI for a "backbone", where its extreme reliability features (SMT/CMT/etc.) and optical bypasses have a clear need. Besides, just like people sometimes use Ethernet transceiver multi- plexers [DEC "DELNI", and equiv.] to make standalone "networks" that have no yellow trunk cable, just so could a multi-master "external serial backplane" be used as a limited-distance [250m diameter] LAN. [Disclosure/disclaimer: While I am a sometime consultant to AMD, the mention of specific AMD or H-P products above was just to show that cheap 100Mb/s links *can* be made today (I've seen 'em run). There are sure to be other vendors and other ways to do it -- it's "steamboat time".] Rob Warnock Systems Architecture Consultant UUCP: {amdcad,fortune,sun}!redwood!rpw3 DDD: (415)572-2607 USPS: 627 26th Ave, San Mateo, CA 94403