[as individual] Hi Lingli,
If you mean by „unified deployment” that you need similar AQM schemes with the same parameters/goals, then no; For ECN to achieve its goal (provided the end hosts are reactive), only the marking probability needs to be roughly proportional to the level of congestion. If a path has multiple congestion points at different levels, the ECN marks will indicate the appropriate end-to-end congestion level that path should react upon. So, deploying AQM with entirely different behaviour (ie. one to keep queue sojourn time down, and another that simply tries to keep the queue length slightly below maximum) is just fine. There are other aspects – in the end hosts – that should be improved upon, though. Not all transports feature granular enough feedback mechanisms, and not all transports support granular responses to varying levels of ECN feedback. Having consistency there is important. 3, Deployment considerations, it is somehow related to the first two, do we have to assume a unified deployment of consistent end-to-end ECN behavior in order to achieve the optimal gain? If so, the above issues need to be addressed. Otherwise, how can one choose between end-to-end and hop-by hop? If ECN is ignored by non-supportive systems (as they should), ECN only needs to operate at the bottleneck. This actually makes for a nice deployment situation in principle, given that the bottleneck is often the access link. There are, of course, other deployment issues with ECN, and they are discussed at length in various recent and not-so-recent papers. I think citing them and discussing the current state of things briefly in this draft could be a good idea.
_______________________________________________ aqm mailing list [email protected] https://www.ietf.org/mailman/listinfo/aqm
