On Wed, Mar 25, 2015 at 9:57 AM, Paul Vixie <p...@redbarn.org> wrote: > > > > Tony Finch > Wednesday, March 25, 2015 7:31 AM > > k...@wide.ad.jp <k...@wide.ad.jp> wrote: > > I don't know if such ISPs still exist, but when we started Anycast, it > was reported that some ISPs did per-packet load balancing. > > If they do that they will utterly wreck the performance of unicast TCP, > which generally does not cope well with out-of-order packets. > > > and yet, in a UDP-first UDP-mostly or even UDP-only operating environment, that assumption has been sound. if we're about to make it unsound then we have to say so.
I think thats an interesting, and valid point. We've had the luxury of assuming a very high percentage of all DNS except zone tx is UDP, and packet order issues have largely been ignored because they were lost in noise. Now we're exposed to the consequence of flows, packet trains, and its probably true that TCP out of order delivery is going to hit us. I say this, because I have been exposed to precisely this cost from a DC which is doing what Tony describes, and I can see of the order 10x less throughput from DE to AU compared to DE to US, because of this problem in file transfers, because they use odd BGP routing and load management. split path is hard on TCP if the packet order gets tossed. -G
_______________________________________________ DNSOP mailing list DNSOP@ietf.org https://www.ietf.org/mailman/listinfo/dnsop