10gbps peering subscriber switch recommendation

Good morning,
We're in the market to move our IX peering off of our core (too much
BGP/CPU :-/ ) and onto a dedicated switch.

Anybody have a recommendation on a switch that can do the following
without costing a fortune? I have scoured Cisco, and bang for the buck
is ... ASR9k (way over powered for handling zero-feature IX traffic),

3-8x 10gbps ports
64k routes minimum, preferably 128k
Must be able to speak BGP
Native/functional IPv6 would be sharp!
Basic QoS to police our ports

The prefix count seems to be the killer, as our exchange table is
getting pretty big (42k+ currently). I'm really tempted to build a
vyatta box or similar, but would rather do something off the shelf --
especially if it can be 1-2 gens old and cost effective.

I'm certain that this same situation is scratching many other folks as
exchanges become more important.

Thanks for your input in advance -- stay warm!
Randal

Good morning,
We're in the market to move our IX peering off of our core (too much
BGP/CPU :-/ ) and onto a dedicated switch.

Brocade ICX 7750 Switch seems to satisfy all the requirements.

If you don't need to carry a full Internet table, the Cisco 4500-X has
plenty of features and the 32 port model can accommodate 256k IPv4 routes.
It also does IPv6 in hardware (128k routes)

Aled

except qos (which needs switch port buffer space). There are no cheap 10G
boxes on the market at the moment which have reasonable numbers of 10G
ports and reasonable sized. Plenty which have 2-4 10G ports with
reasonable buffers and lots more which have plenty of 10G ports with hardly
any buffer space.

Nick

FIB space requirements in a switch are also going to limit
your options.

Also, many "non-service provider" switches don't do egress
policing (they might do shaping, but then if the buffers are
small...).

Mark.

A little bit overkill in term of number of ports but you can consider the
new Trident 2 switches Juniper EX-5100, Cisco Nexus 3100 .....
They have unified TCAM that can store 128K v4 routes

Nitzan

A little bit overkill in term of number of ports but you can consider
the new Trident 2 switches Juniper EX-5100, Cisco Nexus 3100 .....
They have unified TCAM that can store 128K v4 routes

the nice thing about buying bgp devices that can not hold a full table
is that you can expense them in the year of purchase as opposed to
amortizing them over 5 years or so.

randy

it's the merchant silicon boxes which are driving high density 10g prices
down, but most of these boxes tends to come with small fibs and tiny
buffers which limits their deployment usefulness. Still, if they work for
your requirements, they are completely awesome.

Nick

it's the merchant silicon boxes which are driving high
density 10g prices down,...

As they should, and good news for us all, but...

but most of these boxes tends
to come with small fibs and tiny buffers which limits
their deployment usefulness. Still, if they work for
your requirements, they are completely awesome.

My thinking is that provided they don't limit themselves in
the QoS side of things (particularly, how different services
going into the CPE can be policed/SLA'd), then they'd make
good FTTH access nodes that can compete with GPON.

But yes, as an IP route, pretty useless.

Mark.

If only the bean counter saw things our way :-).

Mark.

That�s actually a topic, I was thinking ago some time ago. Why not take
a current TOR switch with 1. BGP support and 2. high buffer. Like
mentioned above we have Trident 2 bases switches. HP (no recommendation)
has its HP 5930 series but tells "Routing table size 16000 entries
(IPv4), 8000 entries (IPv6)", but this one has 4GB RAM, so plenty of
space for full tables. I haven�t tried it out myself, perhaps someone
tried on any other device: What will happen, if I give the switch a full
table? Is there a software limit by the vendor, which will simply cut
everything above? Or would it simply work?

Michael

The 4GB RAM is control plane memory.

The problem is FIB memory, since switches generally forward
Layer 2 and Layer 3 traffic in hardware, and this relies on
forwarding entries being recorded into the FIB.

The 16,000 IPv4 entries or 8,000 IPv6 entries is because of
limited FIB memory.

It's, typically, a switch limitation.

Mark.

That´s actually a topic, I was thinking ago some time ago. Why not take
a current TOR switch with 1. BGP support and 2. high buffer. Like
mentioned above we have Trident 2 bases switches. HP (no recommendation)
has its HP 5930 series but tells "Routing table size 16000 entries
(IPv4), 8000 entries (IPv6)", but this one has 4GB RAM, so plenty of
space for full tables. I haven´t tried it out myself, perhaps someone
tried on any other device: What will happen, if I give the switch a full
table? Is there a software limit by the vendor, which will simply cut
everything above? Or would it simply work?

There are various reasons why one might take a full table on a switch
with not not enough FIB, the important part of course being the part
where you don't install them all.

I have taken a full bgp feed on an broadcom based Arista.

with respect to what happens if you don't filter them.

Either you get continuous fib churn and you only get to forward to the
routes you currently have installed at that time (this is if you're
lucky) or it explodes and you get to keep the pieces.

In Metro-E deployments, this is a good use-case when the box
is providing both IP and Ethernet services to the same or
different customers out of the same chassis.

It avoids having to run 2x eBGP sessions for the IP services
(the first being point-to-point eBGP between the switch and
the customer to get their routes into the network, and the
second being an eBGP Multi-Hop between the customer and a
"bigger" box in your core to send them the full BGP table).

If a switch allows you to keep the routes in control plane
RAM without downloading them into the FIB, you can maintain
a single point-to-point eBGP session to the customer,
including sending them the full table, provided you have a
default route in the switch's FIB to handle actual data
plane traffic flow from the customer upstream.

Mark.

Xtreme x480 can do this and has upto 6 * 10G ports.
It can actually hold a full bgp table also and is preatty cheap.
// Andreas

Med vänlig hälsning
Andreas Larsen

IP-Only Telecommunication AB| Postadress: 753 81 UPPSALA | Besöksadress: S:t Persgatan 6, Uppsala |
Telefon: +46 (0)18 843 10 00 | Direkt: +46 (0)18 843 10 56
www.ip-only.se

-----Ursprungligt meddelande-----