[c-nsp] Assigning VLANs on a per-subnet basis

[jamie baddeley]

On Fri, 2006-04-21 at 08:54 -0500, Frank Bulk wrote:

> We're leaning toward a 3750 because there are models with 4 SFP's, while the
> 4948 only has up to 2, I believe.  There's also consideration for a Extreme
> switch.  Does anyone have an opinion on the 3750's horsepower to route up to
> a 1 Gbps in what are likely mostly small-sized packets?  

We have customers doing just this and it's all good. We pummelled the
thing with iPerf too to confirm. Watch out for tcp windowing effects on
the hosts. (http://www.psc.edu/networking/projects/tcptune/)
we found that 1ms extra delay in a network 2-4 ms long *does* make a
difference.

The disadvantage of the 4948's is they can't be stacked. A stacked
system gives you redundant PSU which is nice.

jamie

> We're planning on
> using the Fujitsu to perform QoS (each company, for example, will say they
> need x Mbps with 20% burst), but perhaps we should put QoS on the VLAN
> egress rather than then Fuji ingress.
> 
> Regards,
> 
> Frank
> 
> -----Original Message-----
> From: Robert Blayzor [mailto:rblayzor at inoc.net] 
> Sent: Sunday, April 16, 2006 8:46 PM
> To: frnkblk at iname.com
> Cc: bep at whack.org; cisco-nsp at puck.nether.net
> Subject: Re: [c-nsp] Assigning VLANs on a per-subnet basis
> 
> Frank Bulk wrote:
> > At this point I have to admit my ignorance.  I'll have to talk to the P.E.
> > at our consultancy to find out if an RPR without VLAN tags (or with just
> one
> > VLAN tag) acts like a hub or if it does have the intelligence that you're
> > describing.  
> 
> 
> I think the big misconception is that RPR is just an ethernet handoff.
> The only part of that is true is the port you're connecting to the
> network with, in your case 100BaseTX/FX.
> 
> Again, I really only have experience with RPR on the Cisco platforms,
> mostly being the ONS15454 via the ML card.  If you look at the ML card
> it's basically a layer3 switch on a blade that has two virtual POS ports
> that tie the card into the SONET network.  I guess that's where the ML
> shines because you then have the flexibility to route or bridge over the
> SPR interface (RPR interface to the ML card).
> 
> I guess where I'm getting confused is that you're the customer, I'm
> looking at it from the transport side.  If four ISP's came to me and
> wanted to use a packet ring and have an aggregation point where they can
> share the entire alloted bandwidth they can be VLAN tagged (internal to
> the transport provider) and then each customers interface would exist in
> that VLAN.  Customers could then use a common subnet and route to each
> other and the upstream.  It's a really easy application.
> 
> The ML will also do QinQ if you needed to do your own tagging around the
> ring.  Another plus on the ML is that you can use service policies to
> rate-limit traffic at each customers port, should you need to do that.
> 
-- 
jamie baddeley <jamie.baddeley at fx.net.nz>
FX Networks