On Tuesday, 5 April 2016, Rong Pan (ropan) wrote: > > Rasool, > > Thanks for your detailed review. Please see the comments below and a new > draft has been posted. > > Regards, > > Rong > > >For new drop_prob calculation, how can drop_prob_ be just a division of > >queue lengths? > > drop_prob_(i)=queue_.byte_length(i)/max(queue_.byte_length(i)) > > > >>>>>>>>>>>>>>>>>>>>>>>>>RP: Let¹s say there are two flows that are > >>>>>>>>>>>>>>>>>>>>>>>>>queued into > two sfq queues. One flow is an elephant flow and one flow is mice flow. Total > latency is caused by the elephant flow. The above equation makes sure that > the mice flow (whose queue length is zero, or very small amount) won¹t be > dropped. > Besides, since it is mice flow, it would be hard to measure its departure rate > as the queue would never build up.qdelay >
Thank you for the explanation. I found section 4.6 SFQ-PIE in [*] is clearer than in the draft. I hope SFQ paragraph of PIE draft becomes a little clearer. [*] http://www.cablelabs.com/wp-content/uploads/2014/06/DOCSIS-AQM_May2014.pdf > > > >5) In pages 23 and 24, > > > >> //If the queue has been idle for a while, turn off PIE > >> //reset counters when accessing the queue after some idle > >> //period if PIE was active before > >> if ( PIE->drop_prob_ == 0 && PIE->qdelay_old_ == 0 > >> && queue_.byte_length() == 0) { > >> PIE->active_ = INACTIVE; > > > >According to section 5.3 (Turning PIE on and off), the condition of > >deactivating PIE should be: > > > >if (PIE->drop_prob_ == 0 and current_qdelay < QDELAY_REF/2 and > > PIE->qdelay_old_ < QDELAY_REF/2) { > > > >>>>>>>>>>>>>>>>>>>>>>>>>>RP: Fixed. > I am unsure whether section 5.3 is correct or the old code (in draft-ietf-aqm-pie-05.txt) with queue_.byte_length() == 0 was correct. In my PIE implementation, I found that without using queue_.byte_length() == 0, PIE turns off after a very short period and causes queue delay to become uncontrolled until queue length exceeds 1/3 drop-tail size. I put queue_.byte_length() == 0 condition and I got better results. Regards, Rasool _______________________________________________ aqm mailing list aqm@ietf.org https://www.ietf.org/mailman/listinfo/aqm
