On Tue, Nov 13, 2012 at 9:49 AM, Christian Huitema <[email protected]> wrote: >> No. Today if you have a set of PA prefixes and your address changes from > one >> to another, your TCP connections all break. > > That's true, has been for quite some time, and yet nobody seems to be doing > much about it. Which makes you wonder how big of a problem that is in > practice. If applications were really hurting, you would hear complaints > from application developers. But you don't. The applications that need > reliable long duration sessions incorporate some trivial checkpoint and > restart mechanism, or some pretty elaborate consistency protocols for big > databases. They probably would do that no matter what the reliability of > TCP, as long as it is not "perfect." And PA renumbering is probably not very > high in their list of "stuff that occasionally break TCP." > Outcome is that the networking people have to build larger L2 DC networks so that applications don't break when a VM is moved between locations - now the DC networks are becoming so large that we need another layer (NVO3) to extend the VLAN range of the already very large L2 DC networks. Life would be so much easier to move a session from one VM to another VM, with the help of a session identifier - instead the packets are encaped with new headers on the "wire" and not at the hosts where the encap/decap of the session layer should occur without changes for the applications. Of course, some application will suffer; if all applications are preserved as such then there is no true separation between identifiers and locators because some of the applications make use of the overload of Loc/ID (also transport protocol needs to evolve in order to make multi-homing more convenient). However, it is a nice starting point to apply the encap of a new header at the wire but it shouldn't be the final goal, the goal should be to have a stack containing a session layer.
Switching topic: Danny and Shane had some thoughts about BGP, perhaps there is room for improvement on inter-domain routing. I agree there might some unfinished topics for us, found an interesting paper on how BGP might be improved. http://www.cs.princeton.edu/~jrex/papers/multipath06.pdf Patrick _______________________________________________ rrg mailing list [email protected] http://www.irtf.org/mailman/listinfo/rrg
