Hi Adam, My understanding was that you might be able to oversubscribe only using PIR > (need to test). > And in that case all the queues are in the excess region. > So only the excess priorities are honoured (HI and LO in strict priority > fashion) and queues with the same priority are serviced round robin. >
You can oversubscribe both PIR and CIR (G-rate) so you have to be careful how much G-rate you allocate (if you are selling a PIR/CIR service that is). In H-CoS mode with only PIR set all queues are in excess even if the aggregate of all the shapers does not oversubscribe the interface bandwidth (or aggregate shaper). In per-unit mode with only PIR set, you have to oversubscribe the shapers to end up with all queues in excess. > I also thought that weight with which the queues in excess region are > served is proportional to the transmit-rate (as percentage of VLAN PIR). Though looking at the show outputs and reading your test results it looks > like it's not the case. > Well, the weights are determined from the transmit-rates but the queues aren't proportioned in the way you'd expect them to be relative to the transmit-rates. For example, you'll find that you can get packet loss in a queue that is sending less than contracted rate if you oversubscribe another queue at the same priority level. The weightings don't translate to an exact percentage of bandwidth in reality. > But have you tried setting excess-rate for the queues - that should be > honoured while a given queue is in excess region right? > Yep, but if you set excess-rate = transmit-rate then the weights are just the same as if you hadn't set them and it doesn't affect the behaviour of the queues. > I just can't believe that once in excess region the queues (using %) are > using main interface PIR -is it a bug please? > It's more a case that per-queue guaranteed rates are set to zero when you are in H-CoS or oversubscribed per-unit mode which means that you have to factor in G-rate for each node when it would be nice not to have to bother (and you can not bother, but things don't work quite the way you expect them to). > With regards to buffers > May I oversubscribe buffers by configuring say "delay-buffer-rate percent > 10" for 11 VLANs on a physical interface please? Not sure on that one, never tried to be honest. Dan _______________________________________________ juniper-nsp mailing list juniper-nsp@puck.nether.net https://puck.nether.net/mailman/listinfo/juniper-nsp