mckee@MITRE.ARPA (09/06/86)
Gentlemen: I solicit advice and document references on how IP should select a particular Gateway when more than one is available. Background My responsibilities include writing communications and network performance requirements for a large military system called the WWMCCS Information System (WIS). The WIS will (initially) consist of about 40 sites. Each site will have: - a 10 Mbps Applitek LAN - 2 or 3 large host processors - 50 to 150 workstations The workstations and hosts are connected to the LAN through Interface Units (IUs). TCP and IP are performed in the IUs. Each LAN has at least one Gateway IU that connects the LAN to the DDN-RVN (DDN-Red Virtual Network, Class A Net 21). All inter-site communications use net 21. For increased throughput and availability, some sites will have 2 or 3 Gateway IUs. Some Gateways will connect to the DDN-RVN at 56 Kbps, others at 9.6 Kbps. Each LAN has a LAN Control Center/Security Monitor (CC/SM). For security and access control, procedures have been established such that the CC/SM is aware of the status of every connection; every TCP active or passive open/close/abort request is routed through the CC/SM. Further, by means of periodic interrogation, the CC/SM knows the health and traffic load of every local IU, including the local Gateways. Questions When IP in a host or workstation IU receives a datagram destined for a remote site, what procedure should be used to select one of the local Gateways? When IP in a Gateway receives a datagram from the LAN, what procedure should be used to select one of the Gateways at the remote site? Any advice or references would be appreciated. H. Craig McKee 'mckee at mitre' (703)883-5505
haverty@CCV.BBN.COM (09/08/86)
Vint et al, A minor clarification. There is no general algorithm in use which spreads traffic across multiple gateways. There is however a special-purpose table-driven algorithm in the Arpanet/Milnet mailbridges, which attempts to spread traffic amongst the seven parallel mailbridges. It is effective only for hosts which are physically connected to the Arpanet or Milnet, but not gateways, because hosts listen (should at least) to redirects but gateways do not. Fortunately this covers a lot of the traffic; without this mechanism, the mailbridge would be a highly reliable bottleneck (more so than it is now), since only one path would be "best", with six !! idle backup gateways at any time. Unfortunately, with more LANs all the time, this interim approach is clearly limited. There is a feature in design stages now for the C/30 family, which will permit multipath routing to occur -- i.e., between points A and B several parallel paths might exist, and all might be in use to carry traffic. With such a mechanism, for example, a network with only 9.6 kilobit/sec trunks could achieve a host-host throughput greater than 9.6 by using the multiple paths. This is a very hard problem, to achieve stability and efficiency and avoid behavior like the Atlantic two-step. Jack Haverty
JNC@XX.LCS.MIT.EDU.UUCP (09/08/86)
The answer to the first question you ask is contained in two RFC's in the TCP/IP guide; I don't remember which volume they are in in the white books, but they are RFC's 814 and 816, "Names, Addresses, Ports and Routes", and "Fault Isolation and Recovery". I don't want to sound like a broken record, but I wish people would read the documentation provided before asking questions. I am aware that those books contain a lot of material: perhaps the NIC could include a 'Suggested Readings' list in the next printing which would list RFC's 813-817 as 'Required'? The second question is sort of covered by RFC's 904 and 888, which describe the current Internet model for routing packets among gateways. Noel -------