On Sun, Jun 24, 2012 at 10:29 PM, Eric Dumazet <[email protected]> wrote: > On Sun, 2012-06-24 at 22:00 -0700, Dave Täht wrote: >> From: Dave Taht <[email protected]> >> >> ECN was not part of the original codel design and adding support >> for it revealed problems in two areas. 1) ECN can be gamed. >> 2) Dropping packets under overload more rapidly frees up >> bandwidth than marking packets. >> >> Two possible scenarios of use - on egress from a network, >> ecn_target could be set low, to drop more often, to >> ensure lowest latency for other packets. >> >> On ingress, it could be set high to mark packets more often, >> to lower data loss while still signalling the end application >> that bandwidth is a problem. >> >> ecn_target is not engaged until after codel enters a dropping >> state overall. > > I would suggest 'drop_above' instead of ecn_target, since its quite > different than the 'target'
OK. I will respin this after resolving the issue I raised in the previous patch comment (just switch ecn to being an int?), and do some testing. As noted I was (de)impressed at the amount of ecn drops I had with it set to 2*target on pure ecn enabled streams. In one test it was 95% drops... And I'm still open to saner approaches. Another one might be to at graduated levels drop big packets down to smaller ones. > (And dont display/output it if ecn is not set, no need to confuse users > who didn't enable ecn on CoDel) > > > > _______________________________________________ > Codel mailing list > [email protected] > https://lists.bufferbloat.net/listinfo/codel -- Dave Täht http://www.bufferbloat.net/projects/cerowrt/wiki - "3.3.8-6 is out with fq_codel!" _______________________________________________ Codel mailing list [email protected] https://lists.bufferbloat.net/listinfo/codel
