[c-nsp] Cisco Nexus as MetroE switch?

[Nathan Ward]

> On 19/10/2015, at 14:46, James Jun <james at towardex.com> wrote:
> 
> On Sun, Oct 18, 2015 at 07:42:27PM +1300, Nathan Ward wrote:
>> Sorry, I should look better.
> 
> I would say Juniper ACX is more comparable to ASR 901 Series meant for cell sites with simple l3vpn instance or small-scale l2circuit'ing.  From my past experience, it wasn't really suited for MetroE use cases.
> 
> For the similar price footprint, ASR 920 is a clear winning choice over ACX for 1GE MetroE access.  I agree with Mark, Juniper has been lagging in MetroE space.

We put in ASR920s actually, but only for 10G - they stopped forwarding traffic under high load, so got whipped out. We plugged our 10G in to ASR9000s which is obviously a higher cost per port so won’t scale well.

>> 
>> I???ve heard this about the label depth limitation. It???s not clear to me what this means in practice - the max depth is 3 labels, but, does that mean it can???t forward packets with more than 3 labels, or does this mean that only certain features (perhaps hashing for load balancing across multiple physicals in a bundle) can happen with 3 or less labels? The latter makes sense, the former seems kind of odd.
>> I know that there is a limitation where flow labels cannot be used, and where it won???t load-balance across multiple paths. We do TE, and if we need to we???ll load balance across multiple tunnels at the head end, so I can???t see that being relevant either - it???s just another TE path from the ACX point of view.
> 
> FWIW, the 3 label stack limitation on QFX 5100 has to do with label imposition & disposition operations -- in other words, if you have a QFX5100/ACX5k (same thing) acting as  PE.  If you are using it as P LSR (core router just forwarding outer transport labels), then label stack doesn't matter as it just swaps outer label.  ECMP/Lagg hashing isn't too shabby on QFX either, it works fairly well.

What’s what I figured. We were told that MPLS ECMP doesn’t work on ACX though - which I assume means if a labelled packet comes in, it can only go out a single interface - whether that means physical or logical (i.e. LAG) is probably interesting but I don’t have an answer on that yet. Obviously the only case where that matters is for LDP where the ACX advertises a next hop label for a destination that is then balanced across multiple physical (or logical) links to a destination by the ACX. That won’t matter for TE where we’re always going out a certain link.

> That's really what ACX/QFX5k is awesome for -- it gives you a cheap 48x10GE P router at teh core, perfect for small networks needing LSRs.  It also has nice control plane so you can use it as route-reflector too if you want.  But for edge duties (i.e MetroE access, l2/l3vpn terminations, hQOS, etc), I really would not recommend it.

Our use case is pulling in several 10G interfaces from a 3rd party backhaul network provider with a handful of VLANs each, each VLAN corresponding to a remote regional customer aggregation point. We want to map each VLAN in to a ethernet pseudowire - there’s maybe 5-10 per 10G, for a total of maybe 40-50 per ACX. We have no need for HQoS or anything - just per VLAN QoS. These are aggregation for us, and don’t see individual customers.

> That being said, we've been extremely pleased with ASR 920 and 902/903 Series for MetroE access duties at very reasonable price points.  With RSP3 and ASR 907, it's an exciting time for this product; now... if they could just bring IOS XR into 9xx series, oh I could only dream.. :>

Yeah, the ASR920 seemed like a good box, and we’ve got a couple doing some really simple stuff, but found the MPLS stuff broke under load. Unfortunately we didn’t get much time for TAC to look in to the faults as stopping forwarding traffic at peak time for many thousands of customers is a pretty impacting fault, and it was only fixable by a reload.

Sorry to OP for side tracking this :-)

--
Nathan Ward