On 08/10/2010 02:31 EDT, Tony Li wrote: > I'd recommend that we also consider locator selection as another > degree of freedom. I agree with Joel that flow control belongs to > the transport layer, and that path selection is part of the network > layer. > > Is locator selection a transport function? It would seem that having > it be a unilaterally network layer function is over constraining the > architecture. Yet, transport cannot perform intelligent locator > selection without network layer input.
The problem is that the layer-to-layer API is constrained, i.e. there almost isn't any! It's like in many protocols: you want the other end (of a protocol, an API, any communication) to do something, but you have no way of telling him what you want. However, you know that if you tweak the information he gets that he will probably do what you want, so you overload an existing mechanism. We've done it for years with BGP for example e.g. the way AS paths are used. So yes I agree completely that a higher layer shouldn't be communicating its path wishes using src/dst pairs but it has no other globally implemented mechanism ... and I don't think we're going to get one for years. One of the points of MIF is to get L3 and below to be more predictable in their behavior, which will make L4's problem easier and lessen the need to do anything else. On 08/10/2010 17:25 EDT, Brian E Carpenter wrote: > I can't speak for MPTCP, but is seems to me that MPTCP is based on the > assumption that it *can* perform intelligent locator selection based on > feedback from transport layer ACKs (or their absence). That's if your > definition of intelligence is 'get the most out of the network layer > without caring about path selection.' It treats the network layer as a black box: if it gives L3 input X (source and destination), a particular outcome happens and it does it more, or less, depending on whether it likes that outcome. On 08/10/2010 17:37 EDT, Tony Li wrote: > At least as I understood it, MPTCP can consider the full cross > product of locators and make use of dynamic feedback for adjustments. > The question I'm raising is whether that's sufficient. > > For example, if transport got feedback about path selection and > congestion from the network layer, then perhaps transport could be > more intelligent. For example, if multiple locator pairs all result > in paths that coincide on a particular congested link, then the > benefits of multi-pathing could be extremely limited. > > More information is available, if transport is willing to increase > its interaction with the network layer. All true. Scott _______________________________________________ rrg mailing list [email protected] http://www.irtf.org/mailman/listinfo/rrg
