Watt-Alan@mickey.ycc.yale.edu (06/21/91)
>Date: Thu, 20 Jun 1991 15:45:02 -0700 (PDT) >From: Steve Hubert <hubert@cac.washington.edu> >Sender: Steve Hubert <hubert@kamba.cac.washington.edu> >Subject: backplane backbone, routing or bridging? >To: cisco@spot.colorado.edu > >We are about to connect all N (N ~ 20) of our routers to a single AGS+ in >order to use the AGS+ backplane as our "backbone subnet". There will be point >to point ethernet between each router and the backbone router. We carry only >IP traffic over the backbone. It occurs to us that, instead of assigning a >subnet number to each of these point to point links, we could use the AGS+ as >a bridge and use only one subnet number instead of N. I don't think this has >anything to do with the normal bridge-vs-router debate. This would be an N- >port bridge with a single router attached to each port. Does anyone have any >opinions on advantages/disadvantages of the two approaches? For example, is >one way faster than the other? Thanks. > >Steve Hubert >Networks and Distributed Computing, Univ. of Wash., Seattle >hubert@cac.washington.edu > > You may not be able to do this. You cannot both route and bridge the same protocol in a cisco chassis, even if you only route on some interfaces and only bridge on others. In other words, if you want to route IP in either the AGS+ chassis, you cannot bridge IP in it. I wouldn't do it anyway, as you lose the chance of gathering all that per-protocol information in the AGS+ (such as IP accounting). If you want to bridge, look into getting a genuine multiport bridge (A.K.A. fast packet switch). This would undoubtedly save you money over the AGS+ if that's all you plan to do with it. However you do not have to assign a whole subnet to each point-to-point ethernet. You can have a single subnet which is shared by all the point-to-point ethernets on both ends. Assuming you subnet a class B on an 8-bit boundary, and the "point-to-point cloud" is subnet 100, you could define the subnet mask 255.255.255.252 to allow 6 bits of "sub-subnet" and 2 bits of "host" (0 is reserved, 3 is broadcast, 1 and 2 are the two endpoints. AGS+ intf address Brcast Router intf address ---- ------- ------ ------ ------ ------- ether0 XX.YY.100.1 XX.YY.100.3 AGS 1 ether0 XX.YY.100.2 ether1 XX.YY.100.5 XX.YY.100.7 AGS 2 ether0 XX.YY.100.6 ether2 XX.YY.100.9 XX.YY.100.11 AGS 3 ether0 XX.YY.100.10 ... ether18 XX.YY.100.77 XX.YY.100.79 AGS 20 ether0 XX.YY.100.78 ether19 XX.YY.100.81 XX.YY.100.83 AGS 20 ether0 XX.YY.100.83 There you have your 20 ethernet point to point links between the AGS+ and the 20 other routers, and plenty of address space left in subnet 100 for more. - Alan S. Watt High Speed Networking, Yale University Computing and Information Systems Box 2112 Yale Station New Haven, CT 06520-2112 (203) 432-6600 X394 Watt-Alan@Yale.Edu Disclaimer: "Make Love, Not War -- Be Prepared For Both" - Edelman's Sporting Goods [and Marital Aids?]
Watt-Alan@mickey.ycc.yale.edu (06/22/91)
Steve Hubert replied to my posting of 20-Jun-91: |On Thu, 20 Jun 91 22:11:40 EDT, Watt-Alan@mickey.ycc.yale.edu wrote: | |> Subject: Re: backplane backbone, routing or bridging? |> To: hubert@cac.washington.edu |> cc: cisco@spot.colorado.edu |> |> ... |> | |> However you do not have to assign a whole subnet to each point-to-point |> ethernet. You can have a single subnet which is shared by all the |> point-to-point ethernets on both ends. Assuming you subnet a class B |> on an 8-bit boundary, and the "point-to-point cloud" is subnet 100, |> you could define the subnet mask 255.255.255.252 to allow 6 bits |> of "sub-subnet" and 2 bits of "host" (0 is reserved, 3 is broadcast, |> 1 and 2 are the two endpoints. | |I don't understand how to do this with cisco equipment. My understanding is |that the cisco has a single netmask for a network. How do you get the |variable length mask idea to work? I think my fingers ran amok before my mind was properly engaged. I have actually never tried this and now that I think about it, it is quite likely to get you in trouble. However, I am quite sure that nothing stops you from getting yourself a new class C network and subnetting it as described above. This still saves you from wasting unnecessary subnets of your class B. For all that, it may actually work to sub-subnet your class B. On a typical BSD system, datagrams would be output on the proper interface, but the routing protocol daemons might get hopelessly confused. Perhaps the same thing would happen on a cisco box (or is it "Cisco box" these days?). - Alan S. Watt High Speed Networking, Yale University Computing and Information Systems Box 2112 Yale Station New Haven, CT 06520-2112 (203) 432-6600 X394 Watt-Alan@Yale.Edu Moral: Think twice, hit <ENTER> once. Disclaimer: It is a violation of federal law to use this posting in a manner inconsistent with this disclaimer.
equek@hydro.on.ca (Erone Quek) (06/23/91)
>>> Some previous part has been left out. > > However you do not have to assign a whole subnet to each point-to-point > ethernet. You can have a single subnet which is shared by all the > point-to-point ethernets on both ends. Assuming you subnet a class B > on an 8-bit boundary, and the "point-to-point cloud" is subnet 100, > you could define the subnet mask 255.255.255.252 to allow 6 bits > of "sub-subnet" and 2 bits of "host" (0 is reserved, 3 is broadcast, > 1 and 2 are the two endpoints. > >>> The example/table has been left out. > > There you have your 20 ethernet point to point links between the AGS+ and > the 20 other routers, and plenty of address space left in subnet 100 for > more. > > - Alan S. Watt > High Speed Networking, Yale University > Computing and Information Systems > Box 2112 Yale Station > New Haven, CT 06520-2112 > (203) 432-6600 X394 > Watt-Alan@Yale.Edu > > > Disclaimer: "Make Love, Not War -- Be Prepared For Both" > - Edelman's Sporting Goods [and Marital Aids?] > Alan (& cisco), I thought one can not has multiple subnet masks across the same Class B network, i.e. one cannot have 14 bit subnet mask in certain 'part' of the Class B (in your example: subnet cloud 100) and use 8 bit mask in other 'part' of the Class B network. I was assured by cisco's customer engineers that multiple subnet masks within the same class B is something I DON'T want to get into. As a matter of fact, we were in a very similar situation. We ended up 'wasting' a full (8-bit) subnet between a pair of 'interconnecting' routers. cisco people claim they are working on supporting 'variable subnet masks'. But I am not as optimistic since that would mean a fundamental change in their way of distributing IGRP tables (and may be routing decisions too ?). One more problem, how are you going to tell your unix boxes (Sun for example) about your variable subnet masks (assuming one can NOT use default routing due to security reasons) ? Erone Quek ISD Network Computing Operations Ontario Hydro Tel: +1 416 592 6460 EMail: Erone.Quek@hydro.on.ca or ...!utcsri!ohmg1!equek
Watt-Alan@mickey.ycc.yale.edu (06/25/91)
From Alan Watt's correction to his original mistaken posting: >> likely to get you in trouble. However, I am quite sure that nothing >> stops you from getting yourself a new class C network and subnetting >> it as described above. This still saves you from wasting unnecessary >> subnets of your class B. >> From Mark Tassinari's response to the above: > >Possibly, there is. We just installed routers in the same configuration and >were faced with the same decisions. Page 5-5 of the manual states " ...the >network must be set up such that it does not require traffic between any two >subnets to cross another network." > >For the diagram below this would prohibit the use of a new class C network >number for the point-to-point links. > > net 128.115.1.0 128.115.2.0 > | | > | | > -------- ------- > | cisco | | cisco | > -------- ------- > \ / point-to-point > \ / links > \ / <----- > ---------- > | backbone | > | cisco | > ---------- > >We finally decided to burn a subnet for each link (fortunately there were only >10). I hope we will be into the next generation technology by the time we need >those subnets, and that new technology will involve a topology change (like >FDDI). > I should have just kept my mouth shut. You're right; my "fix" for the original mistake was no better. Forget I said anything. However, *I AM ABSOLUTELY, POSITIVELY, CERTAINLY CONVNICED* there is *NO TECHNICAL REASON* why you can't just get a class A address, and subnet it on a 255.255.255.0 boundary. This will give you 8-bit subnets coming out of your ears. Too bad the NIC won't give them out so freely. Aha! maybe I've finally figured out why we need OSI!! :-). - Alan S. Watt High Speed Networking, Yale University Computing and Information Systems Box 2112 Yale Station New Haven, CT 06520-2112 (203) 432-6600 X394 Watt-Alan@Yale.Edu Moral: Measure *thrice*, cut once. or: It is better to remain silent and be thought a fool than to speak up and remove all doubt.
medin@nsipo.nasa.gov (NASA ARC NSI Project Office) (06/25/91)
I'll point out that you don't need OSI, just OSPF with variable length subnet mask support in the IP forwarder. Several vendors already do this. It's not rocket science, and routing technology has surpassed the point where such restrictions are required. Subnets no longer need to be connected if you do things right. Thanks, Milo
Greg Satz <satz@cisco.com> (06/25/91)
>> >> I'll point out that you don't need OSI, just OSPF with variable length >> subnet mask support in the IP forwarder. Several vendors already do this. >> It's not rocket science, and routing technology has surpassed the >> point where such restrictions are required. Subnets no longer need to >> be connected if you do things right. >> >> Thanks, >> Milo Milo, when is the IETF going to write up how variable length subnet masks should be used? There are a number of situations where unexpected behavior can occur. Is everyone expected to learn this for themselves? Greg
thille@cisco.com (Nick Thille) (06/25/91)
Alan, There is just one problem with getting a class C.... > Return-Path: <cisco-request@spot.Colorado.EDU> > From: Watt-Alan@mickey.ycc.yale.edu () > Date: Fri, 21 Jun 91 22:36:13 EDT > To: hubert@cac.washington.edu > Subject: Re: backplane backbone, routing or bridging? > Cc: cisco@spot.Colorado.EDU > > Steve Hubert replied to my posting of 20-Jun-91: > > |On Thu, 20 Jun 91 22:11:40 EDT, Watt-Alan@mickey.ycc.yale.edu wrote: > | > |> Subject: Re: backplane backbone, routing or bridging? > |> To: hubert@cac.washington.edu > |> cc: cisco@spot.colorado.edu > |> > |> ... > |> > | > |> However you do not have to assign a whole subnet to each point-to-point > |> ethernet. You can have a single subnet which is shared by all the > |> point-to-point ethernets on both ends. Assuming you subnet a class B > |> on an 8-bit boundary, and the "point-to-point cloud" is subnet 100, > |> you could define the subnet mask 255.255.255.252 to allow 6 bits > |> of "sub-subnet" and 2 bits of "host" (0 is reserved, 3 is broadcast, > |> 1 and 2 are the two endpoints. > | > |I don't understand how to do this with cisco equipment. My understanding is > |that the cisco has a single netmask for a network. How do you get the > |variable length mask idea to work? > > I think my fingers ran amok before my mind was properly engaged. I have > actually never tried this and now that I think about it, it is quite > likely to get you in trouble. However, I am quite sure that nothing > stops you from getting yourself a new class C network and subnetting > it as described above. This still saves you from wasting unnecessary > subnets of your class B. The only possible problem with this comes if you want to put your class C subnet in between class B segments. ie: (Disclamer: It is late and my brain is tired. I don't guarantee that the addresses below make sense.) Denver Router New York Router E0: 131.108.19.40 E0: 131.108.29.47 S0: 192.27.38.249 192.27.38.250 Note that this is illegal unless there is some other connection between the two routers that has an address of 131.108.xxx.0. This is because discontiguous subnets are illegal. Best Regards, -Nick > > For all that, it may actually work to sub-subnet your class B. On > a typical BSD system, datagrams would be output on the proper interface, > but the routing protocol daemons might get hopelessly confused. Perhaps > the same thing would happen on a cisco box (or is it "Cisco box" these > days?). > > - Alan S. Watt > High Speed Networking, Yale University > Computing and Information Systems > Box 2112 Yale Station > New Haven, CT 06520-2112 > (203) 432-6600 X394 > Watt-Alan@Yale.Edu > > > Moral: Think twice, hit <ENTER> once. > > Disclaimer: It is a violation of federal law to use this posting > in a manner inconsistent with this disclaimer. >
medin@nsipo.nasa.gov (NASA ARC NSI Project Office) (06/25/91)
Hopefully this will be discussed at the coming IETF meeting. Obviously, most people seem to have a position on how to do it since they are or soon will be shipping it in their production router code. I think most of the vendors are pursuing the approach Van Jacobsen has taken for 4.3 Reno. I know you folks are very sharp, and I'm sure the IETF would love to hear how you would choose to implement it! I have my own views, but I'm not a vendor and don't have a good feel for the support issues. Thanks, Milo
oleary@sura.net (dave o'leary) (06/26/91)
In article <9106250555.AA23713@wolf.cisco.com> satz@cisco.com (Greg Satz) writes: >>> >>> I'll point out that you don't need OSI, just OSPF with variable length >>> subnet mask support in the IP forwarder. Several vendors already do this. >>> It's not rocket science, and routing technology has surpassed the >>> point where such restrictions are required. Subnets no longer need to >>> be connected if you do things right. >>> >>> Thanks, >>> Milo > >Milo, when is the IETF going to write up how variable length subnet masks >should be used? There are a number of situations where unexpected behavior >can occur. Is everyone expected to learn this for themselves? > >Greg Greg et al, I'm working on article for the SURAnet newsletter (which I will finish any month now) on packet forwarding decisions in a subnetted environment. When I have a reasonable draft done, I can make it available for interested parties to check out. In particular I am addressing the issues of limitations of the forwarding table without subnet masks, how to kludge around this under special conditions and why that is dangerous, and how things work in a variably-subnetted environment (using SURAnet, with lots of point to point and a couple of bigger ethernets as an example), and how and why non-connected "subnets" (which aren't really subnets anymore) can still see each other in the new world order without Class A, B, and C. Is there anything else that should be covered to make this comprehensive? I'm more than half done at this point, but some of the hairier points remain to be covered. Is this the kind of document that you are interested in? Thanks, dave
satz@cisco.com (Greg Satz) (06/26/91)
There is one case we came up with at the last IETF where two routers share the same network but with different subnet masks. This is illegal. Then there is the case of multiple IP subnetworks on a single cable with different subnet masks. This is also illegal. Right? Greg PS. I agree that hopefully this should be written up at the next IETF too.
tcs@uunet.UU.NET (Terry Slattery) (06/26/91)
>Then there is the case of multiple IP subnetworks on a single cable with >different subnet masks. This is also illegal. > >Right? > >Greg With OSPF, you could have the different IP subnetworks on the same cable but in different areas. The OSPF area id would keep them distinct. This implies an area border router for each area on the same cable. -tcs
vaf@Valinor.Stanford.EDU (Vince Fuller) (06/26/91)
Greg, I don't think there's anything illegal about having multiple subnets on the same cable with different subnet masks provided that all routers agree on the subnet/mask pairs. OSPF with variable-length subnet masks should be able to handle this. --Vince
Greg Satz <satz@cisco.com> (06/27/91)
>> Greg, >> I don't think there's anything illegal about having multiple subnets on the >> same cable with different subnet masks provided that all routers agree on the >> subnet/mask pairs. OSPF with variable-length subnet masks should be able to >> handle this. >> >> --Vince Exactly why this should be written down somewhere. It is enough of a change to the IP architecture that word of mouth can lead to interesting interopable incompatibilities. Jeff Mogul did a good job of describing how subnetting should work. Who is going to augment it with the variable length changes? Greg
jqj@duff.uoregon.edu (JQ Johnson) (06/29/91)
A couple more minor points on variable subnet masks and OSPF: 1/ multiple different address masks on the same physical cable make handling of ICMP address mask request/reply quite difficult. RFC1122 as written (pp. 45-46) does not allow for this case, and would need to be rewritten to allow multiple masks on the same cable (dropping the "first received" rule, disallowing processing of received mask replies until the host's IP address is known, using the SOURCE of the mask reply to test whether the given reply is appropriate to me, guaranteeing that senders of mask replies always use the appropriate source address, etc.). Worse, it isn't clear to me that a receiver of a mask reply can always determine whether it is appropriate. I believe multiple different address masks for different subnets of the same net on the same cable should be illegal. 2/ the OSPF spec restricts the possible set of subnet masks used in a variably-subnetted network routed by OSPF. P. 17 of RFC1131 states "Subnet masks must be assigned so that the best match for any IP destination is unambiguous". This is a fairly weak restriction and in particular is satisfied by any hierarchical design of subnet masks. 3/ some algorithms for routing to variably-masked subnets are linear cost in the number of different subnet masks on a network. I recommend that in thinking about variable subnet masks people keep the set of subnet masks small (say no more than 3 or 4 canonical sizes of subnet on a class B CAN). Conclusions: (a) don't use subnet masks like: subnet mask (three "important" bits) 128.185.2.0 ffff0300 x10 (x = don't care) 128.185.1.0 ffff0500 0x1 128.185.4.0 ffff0600 10x Even if they work, they'll confuse you. (b) Greg is right. Variable-sized subnets introduce substantial new complexity into the subnetting model, and should be carefully documented before people go out and start implementing them widely. -- JQ Johnson Director of Network Services Internet: jqj@oregon.uoregon.edu University of Oregon voice: (503) 346-1746 250E Computing Center BITNET: jqj@oregon Eugene, OR 97403-1212 fax: (503) 346-4397
medin@nsipo.nasa.gov (NASA ARC NSI Project Office) (06/29/91)
Vince responded to this already, since I was on travel, and I fully agree with his response. I think the routing protocol can support a very large degree of functionality, and the IP forwarder support should support this level of functionality as well. Did you see PT's latest note on a modified patricia approach? Thanks, Milo