CERF@A.ISI.EDU (03/24/88)
Alex, "a carefully designed network..." A network? But won't there be thousands of them? Will they ALL be as reliable as you'd like? I'm less persuaded that somehow datagrams are wrong and virtual circuits are right, if that was what you meant in your note, than I am persuaded that we need mechanisms which make apparent memory sharing a reality. Since we need reliability mechanisms to do this SOMEWHERE in the design, I certainly agree that if they aren't in the host, they must be in the off-loaded part. Vint
brian@wb6rqn.UUCP (Brian Lloyd) (03/25/88)
Alex, I think that I will opt for datagrams too. The only part of the network that you can really trust is that which you control. It is the only part of the network that exists to serve YOUR needs. For this reason I am going to want some sort of reliable transport service sitting on top of the network. Sure the network may seem reliable but can you always be sure? At least a reliable transport service can keep trying even though the underlying virtual circuit network has done a reset on your logical channel. Brian Lloyd, President Sirius Systems, Inc. (301) 540-2066 {bellcore, syscad, cp1, irs3, n3dmc}!wb6rqn!brian Share and enjoy!
tsuchiya@GATEWAY.MITRE.ORG (Paul Tsuchiya) (03/25/88)
There are really two issues with regards to this VC/datagram discussion. One is what the network does, and the other is what the host does. It seems to me obvious that the host wants to protect itself and provide itself good service regardless of what the network claims to provide, and so should run a strong transport like TCP. It is silly to argue that one of the virtues of VCs is that it requires less work on the part of the host. Ignoring hosts, however, and just considering the network layer, the discussion is still interesting. I like datagram in the sense that the network isn't obligated to do sequence and guarenteed delivery and so on, and can squash packets if it has to. However, I like some of the "set up" notions of VC. These days, there are many things that one might want to "set up" (or more appropriately, cache) in the gateways along the path. These include routing information, address information (like a Landmark Address, for instance), VISA information. All of these things can be done without destroying the "datagram" aspect of the network. Some people are thinking of more sophisticated network "setup", like rate request and assignment. This is not datagram in the sense that the network tells you roughly how fast you can go, but it is not VC in that the network is happy to squash packets if it needs to. _________________________________________________________________ Paul F. Tsuchiya The MITRE Corp. tsuchiya@gateway.mitre.org 7525 Colshire Dr. 703-883-7352 McLean, VA 22102 USA _________________________________________________________________
jsa@GATEWAY.MITRE.ORG (03/26/88)
Is it really that "obvious that the host wants to protect itself and provide itself good service regardless of what the network claims to provide, and so should run a strong transport like TCP."? If this were true, wouldn't the Europeans be using TP4 over their PTT-provided X.25 nets instead of TP0? _________________________________________________________________ Jim Ackermann The MITRE Corp. jsa@gateway.mitre.org 7525 Colshire Dr. 703-883-5360 McLean, VA 22102 USA _________________________________________________________________
tsuchiya@GATEWAY.MITRE.ORG (Paul Tsuchiya) (03/26/88)
Sorry, what I really should have said is that it is obvious in an Internet environment that one needs a strong Transport. And in fact, the Europeans are quickly becoming TCP users. One can now buy TCP from something like 30 European computer manufacturers, and they have recently comandeered a large chunk of I think Class B space for the European Internet. _________________________________________________________________ Paul F. Tsuchiya The MITRE Corp. tsuchiya@gateway.mitre.org 7525 Colshire Dr. 703-883-7352 McLean, VA 22102 USA _________________________________________________________________
LAWS@rsre.mod.UK (John Laws, on UK.MOD.RSRE) (03/28/88)
Speaking as a European who has access and is active in using both X25 nets and Internet - X25 service within the UK appears to be very satisfactory. That is to say the 'end-to-end' network service provided by BT on PSS is such that I really have no memory of corrupted info, slow response etc. It could be said that I am paying a 'price' for this, but I appear to have no difficulty in seeing the price for this quality of service as other than cost effective. International X25 to Europe again appears very satisfactory. Indeed I am using that as a means of extending the Internet from RSRE to my European partners. Again this appears to be cost effective for the level of our current requirement. Clearly there is a traffic level at which a leased line is more cost effective - what that level is with 'operational' traffic is part of the experiment. I have no direct experience of using X25 to the US (yet) but it is claimed to be less than satisfactory because it is necessary for BT to downgrade its facilities to match the lowest common facilities of US X25 service providers. Fastest line speed (I think) is 9.6kb, window of 2, cant extend packet size. The Internet works very well when lightly used, but standing here in Europe working via a satellite path shows that too many of the Internet solutions are local (CONUS) optimisations. These problems are now well recognised in the specific attention of individuals eg Van J and his TCP work, and a number of TF's. I note with great interest that some folks in BBN and Mitre (who I have great respect for the many scars they have acquired from years of front-line work) have some degree of the so called 'European' view. John Laws Distibuted Information Systems Div RSRE Malvern UK
mckenzie@LABS-N.BBN.COM (Alex McKenzie) (03/28/88)
Paul, I think you are mixing cause and effect. It is only "obvious in an Internet environment that one needs a strong Transport" because of the deliberate, conscious decision made in designing the DoD internet to require that every connected system would use a strong Transport. A different internet would have been built if some other decision had been made about the strength of one's Transport, and then, no doubt, some other conclusion would be obvious. My suggestion was that as we begin to think about the design of the next generation of networks we re-consider this very decision. The suggestion was "inspired" by the recent messages about off loading protocol processing, and about having a communication channel you trust as much as you trust your disk drive channel. I believe that we can choose to decide to require any level of reliability we want (short of absolute perfection) from the next generation internet. Any point in the spectrum of possibilities will have advantages and disadvantages, and will require more or less protocol processing in the connected systems. I do not believe the choice of operating point is strongly related to the number of next-generation networks which constitute the next- generation internet as Vint Cerf suggested; the constituent networks always have to be engineered to play by the rules. Cheers, Alex McKenzie
tsuchiya@GATEWAY.MITRE.ORG (Paul Tsuchiya) (03/28/88)
Alex, I knew I would regret being so arrogant. I don't disagree with you with regards to rethinking how the host views the internet. I especially feel this way with regards to routing paradigms. However, if we make the host think that it sees memory over the wire, then the thing that it is directly over the wire (the first gateway, or the front-end, or whatever), will have to have strong transport. I guess I haven't quite defined what I mean by strong transport here. I don't necessarily mean something exactly like TCP with windowing and all, but something that does essentially the same job. _________________________________________________________________ Paul F. Tsuchiya The MITRE Corp. tsuchiya@gateway.mitre.org 7525 Colshire Dr. 703-883-7352 McLean, VA 22102 USA _________________________________________________________________
CERF@A.ISI.EDU (03/29/88)
Alex, I was trying to react to your single network language "engineer the network" because the internet will involve many nets and administrations. To the extent we are able to produce common engineering objectives and stick with them, this need not be an impediment, but it is almost always harder to engineer a collection of systems than a single one. I also believe that some form of end/end reliability will be needed, whether it is explicitly visible to the host or process software or not. Even if the reliability features are off-loaded on a peripheral controller board, they'd still be part of the architecture. I'm not sure whether we are having a disagreement or not. Vint
philipp@LARRY.MCRCIM.MCGILL.EDU (Philip Prindeville [CC]) (04/04/88)
Ignoring hosts, however, and just considering the network layer, the discussion is still interesting. I like datagram in the sense that the network isn't obligated to do sequence and guarenteed delivery and so on, and can squash packets if it has to. However, I like some of the "set up" notions of VC. These days, there are many things that one might want to "set up" (or more appropriately, cache) in the gateways along the path. These include routing information, address information (like a Landmark Address, for instance), VISA information. All of these things can be done without destroying the "datagram" aspect of the network. Paul, Sounds to me like you want a frame-relay network. Sort of a hybrid... You set up a "channel" by specifying an end-point, type of service, etc. All packets associated with that channel will have those properties. The subnet maintains state at the sender's DCE only. -Philip