Brian E Carpenter <brian.e.carpen...@gmail.com> writes: > I'm confused. We've been talking for months about recommending > pseudo-random flow label values as inputs to hash functions, > precisely to allow scaleable and stateless load balancing and ECMP.
I have a general issue with the above, but I'm not sure whether its a real issue or just terminology. The crux of the issue is the following: > 1. It is RECOMMENDED that source hosts support the flow label by > setting the flow label field for all packets of a flow to the > same pseudo-random value. I do not see a reason to require this. The goal of using the Flow Label with load balancing is to include it as input to the hash in order to produce a more uniform spread of the output of the hash, and specifically, so different flows from the same machine can be sent across different links. You do NOT need uniform spread on the input to the hash to get such an output. A decent hash algorithm is what you need. You also don't need Flow Labels selected in a psuedo random fashion. RFC 3697 says specifically you can assign Flow Label values sequentially. And that works just fine with the modulo(N) algorithm specified in the ECMP document. Note that I am in favor of suggesting that the Flow Label be included in the hash when doing load balancing. Its a no brainer. At worst, its a now op. But it certainly is never better to exclude it. What I don't think is needed, is *any* requirement for pseudo-randomness on the actual Flow Label values. Thomas -------------------------------------------------------------------- IETF IPv6 working group mailing list ipv6@ietf.org Administrative Requests: https://www.ietf.org/mailman/listinfo/ipv6 --------------------------------------------------------------------
