[j-nsp] Best practice for igp/bgp metrics

[Mark Tinka]

On 25/Oct/17 20:47, Saku Ytti wrote:

> There are to principal IGP designs. One would be role based, where
> metric is static depending on roles of A and B end of connection. Such
> as P-P, P-PE, PE-PE. This strategy is well suited for networks where
> hop count is similar to geographic distance or where network spans so
> short geographic distance that it is not relevant.
> This is very simple, requires almost 0 maintenance. If this works for
> your topology, I high recommend it.
>
> Another strategy would be to encode latency cost to link, and always
> choose lowest latency path. Hybrid could penalise for example P-PE
> links, but otherwise use mostly latency on P-P, and document
> exceptions. Particularly if you rely on strategic TE as opposed to
> tactical TE,  latency cost makes lot of sense, as then TE can find
> next-lowest-latency path when you are out of capacity, and if you keep
> monitoring paths which are not on SPT, you will know where you need to
> buy more capacity.
> If your network spans large geographic areas with lot of redundancy,
> you will need this or hybrid, role based alone will cause too many
> exceptions to document.
>
> Luckily SPF is super simple to implement, and I think there is even
> free webtool for playing with costs.

I've implemented both strategies at all ISP's I've built, i.e., device
roles, where P-P links > P-PE links > PE-Metro links, e.t.c., as well as
encoding latency into the IGP metric. Works beautifully, and as with
Saku, I highly recommend it, as it's simple.

Mark.