ECMP vs. UCMP Load Balancing
|
Feature Name |
Release Information |
Feature Description |
|
ECMP vs. UCMP Load Balancing |
Release 24.4.1 |
Introduced in this release on: Fixed Systems (8700 [ASIC:K100])(select variants only*). Enhanced ECMP improves routing scalability by supporting a higher number of equal-cost paths, facilitating efficient load distribution across the network. This feature allows you to maximize throughput by evenly distributing traffic across available paths, which minimizes potential bottlenecks. Enhanced ECMP is designed to work seamlessly with various routing protocols, ensuring consistent performance and reliability in large-scale network environments. *Previously this feature was supported on Q200 and Q100. It is now extended to Cisco 8712-MOD-M routers. |
Load balancing is a forwarding mechanism that distributes traffic over multiple links based on certain parameters. Equal Cost Multi Path (ECMP) is a forwarding mechanism for routing packets along multiple paths of equal cost with the goal to achieve almost equally distributed link load sharing. This significantly impacts a router's next-hop (path) decision.
In ECMP, it is assumed that all links available are of similar speed which inherently means that the hash values that are computed are equally shared over the multiple paths available.
For instance, if we have two paths available, the buckets (which in the end identify the links to be chosen) will be assigned in a 50% / 50% loadsharing. This can be problematic when one path is say a 10G link and the other link is a 1G link. In this case, you probably want to assign a (near) 90/10 type deviation, but considering that BGP is not bandwidth aware, the 10G path is still chosen 50% of the time as much as the 1G link. In this scenario, not all paths are of equal cost path.
What UCMP does in this case is apply a weight to a path which means that we are giving more hash buckets to one path that has a higher weight. The weight applied is static in the sense that it is derived by the DMZ bandwidth extended community either assigned to a peer or as configured via the Route Policy Language (RPL) route manipulation functionality.
In general, a routing protocol decides a best path to a destination based on a metric. This metric is generally driven by the bandwidth of the circuit. When we have 3 paths available, say 1G/10G/100G, routing protocols generally discard the 1G/10G paths available. In defined cases, one may want to spread the load over the circuits based on the load they can carry. In this example, one may want to distribute traffic in a 1%/10%/89% fashion over the 1G/10G/100G paths available.
Feedback