[c-nsp] ASR9k Bundle QoS in 6.0.1

[Saku Ytti]

On 12 May 2016 at 16:35, Mark Tinka <mark.tinka at seacom.mu> wrote:
> Remember that load sharing comes before policing. So if your load
> sharing is unbalanced, you will drop traffic before the actual policer
> limit has been hit. I've found that the only way to fix this is to
> perform per-packet load balancing, as per-flow load balancing may not
> always be equally distributed across member links of the LAG.

Juniper today also has adaptive load-balancing, which essentially adds
one layer of abstraction between hash-result and egress port. This
way, on this interim-table, they can change how many hash-results go
to interface A how many to B.
Compliment this with control-plane monitoring of traffic rates in A
and B, and if some elephant flow is staying in B, adaptive
load-balancing can reduce how many hash-results hit B, diverting
larger share of traffic to A.


-- 
  ++ytti