On Wed, Oct 14, 2020 at 10:54:49AM -0700, Ryan Hamel wrote: > > One would think that with 100GE interfaces, it would not be possible to > overrun the interface if we allowed full 10Gbps/flow, however most 100GE > interfaces, at the chip level are broken down into 10Gbps lanes and the > interfaces do not have a way to easily determine that a lane through the > interface is at capacity, so as new flows enter the interface, they could get > allocated to a lane that is already full and therefore experience packet loss.
This does raise an interesting point regarding ASR9K platform. IIRC, older Typhoon/NP4c 100GE cards (e.g. Juggernaut A9K-2X100GE-TR) had problems where on a 100GE port, you can't sustain more than 10-12 Gbps per individual flow. You had to hash to achieve aggregate bandwidth, which became a significant issue if you were trying to transport 10G L2 pseudowires with limited flow visibility. I'm not aware that it is an issue any longer on newer Tomahawk/NP5c cards? James
