With regards to Juniper Networks , how does 'targeted-distribution' knob works?

Question

Can someone please explain how does the knob 'targeted-distribution' works in Juniper MX routers.

We created aggregated Ethernet (AE) links on our Juniper MX router acting as BNG. However , after creating bundle links and terminate subscribers on it , it was revealed that subscribers are getting more bandwidth than what was being policed. JTAC suggested to apply knob 'targeted-distribution' under dynamic profile to address the issue. After applying the knob , the problem got solved.

I have tried to learn from Juniper documentation but here is what i understand.

As per my understanding , this knob distributes the dynamic interfaces (created for BNG subscribers) among the member links in a bundle. Some subscribers may be on one member link while other subscribers on other member link. The policier which is being pushed by AAA/SRC against any subscriber will then rate limit traffic on that member link at which subscriber resides because without this knob , subscriber traffic will be hashed on all member links in a bundle and rate limit will be active for each individual member link in a bundle , hence all subscribers getting more speed than their profile on BNG.

Please correct me if I am missing something or I am wrong altogether. And How does it differs from 'shared-bandwidth-policer' knob

Answer 1

It helps to explain the problems that targeted distribution is solving, so let's take a look at that. Basically, without targeted distribution, you have 2 ways to allocate shaping resources to a subscriber. For the sake of example, let's say you have a 4x1G LAG with a 100M subscriber connected.

1. Split the 100M shaper over the 4 LAG members (25M each). This works fine, but large flows will max out at 25M. Not a great experience for your customers. This is effectively what a shared-bandwidth-policer does, it will statically distribute the configured shaping rate across n number of interfaces that a subscriber can use for access (LAG members). You can check out this KB to get a bit more context with regard to subscriber management.

2. Replicate the 100M shaper to all 4 LAG members (100M each). This does fix the first problem, but it also presents a new one in that even though the customer pays for 100M, they could get up to 400M depending on what their traffic patterns. This is what you're seeing, I think. This over subscription may also cause congestion for other subscribers.

Finally, both options consume more hardware scheduling resources than is really required.

In comes targeted-distribution. Essentially, instead of having to choose per-flow OR per-subscriber shaping, you get both. It does this by assigning the subscriber (and corollary resources) to a single LAG member. In addition there are redundancy options where resources are replicated to a backup link within the LAG, either on the same FPC (link level redundancy) or another FPC (FPC redundancy). This makes link failures more transparent to the end customer.

As subscribers connect, they will be assigned to the least loaded LAG member based on the number of current subscriber sessions on that LAG member (not bandwidth). Furthermore, you can assign a distribution-list to ensure that certain LAG members only get subscribers associated with a specific list.

NOTE: In your particular case, you have 2 LAG members per AE. These could be primary and backup if you decide to deploy it this way. If your network eventually demanded increasing the size of those LAGs, then the other links would not become primary or backup until failure conditions arose, this is how the resource allocations are more efficient with targeted distribution.