John Ousterhout has been on a mission to fix (TCP) latency in data centers - i personally am unconvined about moving stuff all into NICs - i think it was one of van jacobson's thoughful observations that then you just move the problem to host to NIC latency...and in the meantime, people have done stuff like this: Towards μs tail latency and terabit ethernet: disaggregating the host network stack https://dl.acm.org/doi/abs/10.1145/3544216.3544230 just by being very smart at sofware and understranding where the hardware resources are heading..
indeed, by being cleverer in switches you can do this Re-architecting datacenter networks and stacks for low latency and high performance https://discovery.ucl.ac.uk/id/eprint/10068163/1/ndp.pdf which is actually being built in at least one product... but that doesn't help in wide area, router-to-router multihop control plane traffic management much, as far as I know (hum - maybe it would - maybe we could ask those folks to a side event and have them think about applicability > Jon Crowcroft <[email protected]> wrote: > > Yes, additionsally, though, we can learn lessons from data centers > > where there is a whole body ofwork on dealing with many-to-many > > (e.g. apps using map/reduce platforms and similar) that cause > > congestion in ways that might often be analgous to control plane > > traffic between routers of many kinds - this work post dates most > the > > for instance, I think, this recent talk: > https://netdevconf.info/0x16/session.html?keynote-ousterhout > (I haven't watched it, and I couldn't be at netdev this time) > > I think that netdevconf is a good example of connecting IETF/academia and > the > R-parts of BigTech. > > Also there is: > Semantic Address Routing and Hardware - SARAH > [email protected] > > > > > <<signature.asc>> >
