[c-nsp] EC flow hash computation w/MPLS

[Oliver Boehmer (oboehmer)]

Peter Rathlev <> wrote on Tuesday, January 15, 2008 10:18 PM:

> On Tue, 2008-01-15 at 19:29 +0100, Oliver Boehmer (oboehmer) wrote:
>> assuming we're talking about PFC3*, we hash on the underlying IP
>> header for L3VPN and "regular" LDP LSPs and on bottom label
>> otherwise (e.g. L2VPN/EoMPLS). PFC3 can look at most 3 labels deep
>> for an IP header, which should be enough for most situations.
> 
> It's a 7600/Sup720-3B acting as MPLS PE in a data center, so we only
> have the two labels for regular L3VPN, nothing fancy. So the "MPLS:
> Label or IP" line from "show etherchannel load-balance" means IP if
> present, else label. And when IP is present, the EC uses the regular
> configured IP load-balance (from "port-channel load-balance <X>").
> Hope I got that right. :-)

Yes.

> On Tue, 2008-01-15 at 10:30 -0800, Tim Stevenson wrote:
>>> The context is that I'm trying to find out how the WS-X6704-10GE
>>> compares to using Etherchannels on e.g. a WS-X6748-SFP.
>> 
>> The EC decision criteria is identical on these two linecards.
> 
> Thanks, but I was more thinking about the difference between using
> 10Gig uplinks compared to using N-port 1Gig ECs. We're currently using
> two 4-port 1Gig ECs from a WS-X6748-GE-TX as uplinks and have to
> "expand" this. I'd prefer the 10Gig solution, but I think the bean
> counters would prefer we just add more ports to the EC, buying a new
> 6748 module if necessary. We have eight of these aggregators, each
> needing a 6704-10GE.
> 
> So I guess I'm looking for arguments for (and against) using 10Gig
> instead of EC, apart from the naturally non-perfect load sharing
> inherent in EC.

right, indeed..

	oli