cancel
Showing results for 
Search instead for 
Did you mean: 

VSP (ISIS/SPB) Load balancing across numerous links

VSP (ISIS/SPB) Load balancing across numerous links

Andre_Brits_Kan
Contributor II
Hi Community

I am looking for some more information on how ECMP would work in an Extreme Automated Campus deployment.

If we take a simple Spine leaf design:

13d3cf4109f24a6c906fb0ad3cf06f2f_2427276b-468b-4da9-8913-67159be6fc7e.jpg


We configure the above SPB Fabric (With the two Backbone Vlans)
If we have a single service (ISID) that is configured across all three Leaf VSP's.
A single layer 2 environment for all servers connected to the leafs.

If traffic had to move from one VSP 7200 to another we have two links (Via Spine) to the remote VSP 7200.
Will the underlying ISIS forward the traffic using ECMP and load balance across both both the links?

From reading documentation I understand that in SPB different I-Sid's (Odd's and Even's) is load balanced across the two backbone vlans, but in the event of having a single I-Sid how does load balancing work?

Thx
1 REPLY 1

Ludovico_Steven
Extreme Employee
For a given, single I-SID, all traffic belonging to that service will only go via the same single spine (no ECMP).
If you compare this with TRILL (and TRILL-derived proprietary fabrics) it is a trade-off in the fabric protocol architecture. Being able to spray the same I-SID (VLAN) traffic over multiple paths sounds attractive, but what if you wanted to verify the network dataplane for that traffic? In TRILL you are clueless as to which path your traffic is taking. With Fabric Connect (SPB) you know upfront exactly what the path is, and you can test that path from the ingress VSP7200 BEB using L2pings or L2traceroutes or even L2tracemroutes; these capabilities are based on 802.1ag CFM which can be leveraged with SPB (but not with TRILL based fabrics).
That said, it would be technically possible to hash L2VSN traffic across both BVLANs, but it kind of defeats the CFM benefit, so we have not productized it (yet!).

How VSN traffic is assigned to the two BVLANs varies with BEB type and serviced type.
The even/odd assignment of I-SIDs mentioned, is only applicable to standalone BEBs (which are not part of an MLAG/SMLT construct) and for L2 VSNs.
If the VSP7200s BEBs in your diagram are paired up in MLAG/SMLT Clusters, then the even/odd distribution no longer applies and one of the BEBs will forward all L2VSN traffic to one BVLAN while the other BEB will use the other BVLAN. In this scenario you would typically have x2 7200s in the same server Rack, with all servers dual homed with NIC teaming into both.
If you have L3 flows, then the distribution is different again (and unrelated to MLAG/SMLT mode), depending on whether or not IP ECMP is enabled in the VRF for the destination IP subnets (the node advertising the destination IP route can be reached via both BVLANs, so the IP route can be applied twice in the routing table in ECMP mode). DVR further enhances this as it becomes host based routing and can spread across both spines regardless of what the ingress or egress VLANs are.
GTM-P2G8KFN