[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: DNS based Destination Selection



On donderdag, jun 19, 2003, at 16:07 Europe/Amsterdam, Jay Ford wrote:

The second of my two above points was that the remote end (which owns the
DNS) would be specifying the preferences of the sender of a packet. I'd
rather decide the preferences at my end for traffic I send. Most net
interactions are bidirectional, so this is admittedly not a single-ended
problem, & the issue of who (sender or receiver) gets to decide path
selection for a packet borders on religious, so getting consensus on this
won't be easy.
This assumes that given the choice between two addresses for the other end, choosing address #1 will make the packet flow over ISP A and choosing address #2 will make the packet flow over ISP B. I see no reason why this assumption would hold true in the majority of cases today, and it certainly doesn't _necessarily_ hold true.

   o  you cover dstaddr selection; what about srcaddr selection?

One thought I have just had (so I have no idea how good or bad this idea
is) is if a source is also multihomed it can perform PMTU discovery for
each of the IPv6 addresses it has in that scope. The source address to
use is determined by which ever returns the best PMTU and RTT.
With hop-by-hop routing you don't get to pick a source address: you must use one compatible with the ISP that is the next hop for the destination address in the packet.

You get to do this with source address dependent routing, though. (This will fix the above problem as well.)

This is a slight waste of bandwidth. But if the "best" possible path is
your ultimate goal there have to be trade offs.

The definition of "best" is tricky. It isn't always just bandwidth & RTT.
Either the preference can be expressed in a policy so ask a policy server, or it can be expressed as "optimize this set of variables" in which case you try the permutations, perform measurements and ultimately decide.

Or make TCP multipath-aware with window and RTT processing per path. Then you get the performance of the combined paths regardless of the features of any particular path.