BGP Router - Page 3

alexcf · 7th September 2008

Quote:

Originally Posted by bluewave

Just to chime in if your using a Cisco 7 series router your going to need a minimum of 512mb of RAM as there are over 250k routes in a typical full transit routing table. Best bet as a medium cost solution is to build a linux box with BGPd as mentioned earlier.

I'll also add that there's no point using anything below an NPE-400, and even that would be pushing it these days, well, not unless you want to take a default+partial perhaps?

Cranky · 7th September 2008

Quote:

Originally Posted by CrystalVPS

Obviously, the J-Series cost a bit more than the equivilant Quagga/OpenBGPD stuff, but it's nice to know that if it goes wrong we phone Juniper and say "fix please". We've only ever had to use JTAC twice (that's accross a range of SSG and J-Series devices) and both times they've been knowledgeable and helpful. Both times were for obscure bugs. Our current Juniper core routers then Cisco Layer 3 switches at the edge is working absolutely brilliantly. I'd recommend them, but I know a lot of people are of the opinion "Quagga works, so why invest in something with commercial support", the only thing I can say against Juniper (or their resellers) are the lead times. We've ended up paying more on quite a few occassions as we didn't want to wait 6 weeks.

How good are the J-series? Is VRRP reliable in the J's? The specs look good for the price, I wrongly assumed Juniper were very expensive. What Cisco layer3's are you using?

CrystalVPS · 7th September 2008

Just a 3750 per rack, each 3750 peered with the Junipers, then a partial mesh with the 3750s, with some metrics/route-maps to prefer the ciscos for internal routes over the Junipers (to preserve their routing capacity to the Internet). I've not used VRRP on the J-Series, but it's been rock solid with this set up, and I've seen session uptimes of almost a year, and it only reset as I killed it to move a cable. the 6350s that we run are basically just a 3.6ghz P4 with some fancy bits and pieces, like a hardware crypto card. The *only* thing I can say is they're a bit stingy with the CF card supplied, when we upgraded ours to JunOS 9.x we had to upgrade the CF cards first, but that was very easy to do. It's fair to say, compared to an equivilant spec Quagga box, you're paying a huge margin for Juniper, but JunOS makes up for that in my opinion - it's so nice, so flexible, so logical. It took me a little while to get used to it after the ciscos, but it is logical - easy roll backs, the fact you commit changes as opposed to on Ciscos where they apply instantly. It's just a shame you have to pay for a J-flow license.

markcastle · 7th September 2008

Quote:

Originally Posted by Cranky

How good are the J-series? Is VRRP reliable in the J's?

VRRP works just as it should on all Juniper Kit in my experience.

I agree about the CF card size, Juniper really need to get their act together... when supplying a brand new router the CF card really should be able to take at least a years worth of Junos upgrades before having to upgrade it IMHO.

goscombtech · 7th September 2008

the 2 x J series I had at home for a while (hey, we had spares so i figured i'd try some DSL PIMs) did VRRP just fine. We just use them for terminating leased lines now

Cranky · 7th September 2008

I'm certainly not a networking expert, but I'm curious as to what people would use instead of VRRP. For example, no-wires, you say you have (I assume) a pair of Cisco switches with your transit providers terminating to both switches, and both switches connected to both routers? What happens if a switch or router dies?

Please excuse my blatant lack of experience in this area.

goscombtech · 7th September 2008

It depends on the kit really... Cisco would mainly be HSRP, CARP on BSD, VRRP Juniper, etc.

There are also other options, such as running LACP over a pair of 3750s in a stack

Cranky · 7th September 2008

Quote:

Originally Posted by goscombtech

It depends on the kit really... Cisco would mainly be HSRP, CARP on BSD, VRRP Juniper, etc.

There are also other options, such as running LACP over a pair of 3750s in a stack

Sorry, I must not have made myself clear. No-wires above said they don't use VRRP on their Juniper's - does that mean if a router dies so does their connectivity, or are there alternative options? (for Juniper in particular)

goscombtech · 7th September 2008

ah, gotcha

as they are running an IGP/iBGP there will (or should) be multiple paths off each cisco switch to each router... if a router fails the traffic just flows over another path... this is not interface failover though... its pure routing

Cranky · 7th September 2008

Quote:

Originally Posted by goscombtech

ah, gotcha

as they are running an IGP/iBGP there will (or should) be multiple paths off each cisco switch to each router... if a router fails the traffic just flows over another path... this is not interface failover though... its pure routing

Ah, I understand.

What about if a switch fails? I assume they'll need something there for the upstream routers to know to send traffic to the working switch... would this require something like HSRP on the Cisco switches or is there a less complex failover?

goscombtech · 7th September 2008

its a little more complicated... best would be a stack with LACP to the server over the pair... that way there is no single point of failure... most providers though only connect each server to a single switch, so there is no redundancy here and HSRP/VRRP in this situation are pointless if the switch to which you connect is your gateway. If that fails there is no connectivity to a backup gateway anyhow!

bluewave · 7th September 2008

Quote:

Originally Posted by administrator

Quagga is the fish`n`chips of the routing world

Id say low end cisco are sort of pizza express

hehe

Loving the fast food analogy here!

CrystalVPS · 7th September 2008

gateway is on the access switch, we've not had a switch fail (we've had a couple crash, but a quick reboot has fixed that - and we've never had an issue on the 3750s), then we send defaults + more specifics of our AS to each switch, this set up isn't complex, requires little maintenence and has proven itself to be reliable. We figure adding vrrp/lacp/whatever just adds complexity - We've got staff within 5 minutes from the datacentre at all times, so if a switch does die I can be onsite within 5 minutes and swap out the failed switch. We have a layer 2 switch in all racks, which connects to the secondary NIC, worse case scenario we use the layer 2 switches if we have a total failure of our 3750s. We don't even need STP, as we do it all at layer 3.

I know possibly not the most effective network topology, but it works, it's reliable, we only use BGP and I'm working on the keep it simple principle. All our network issues in the last 6 months have been upstream related, and we're actively looking at alternative transits. For someone who doesn't need to do much, I think our setup is brilliant. Maybe if you're providing HA transit feeds, or 100% SLA services then you would want a more complex set up, but we don't do that. If we did that we'd be using 3750 stacks with LACP to the servers, UPS and so on. Our intention is to eventually run 2 3750 stacks spread between racks, with 2960G access switches, as this is more cost effective and I think it'll be equally reliable, when we get to this stage we'll probably look in to VRRP/HRSP. However, anyone who wants a good network, I would happily recommend a couple of 3750s and a couple of J4350s and I think they'd have a very good network that will last them for a few years.

samb · 8th September 2008

Quote:

Originally Posted by goscombtech

not to mention the hardware replacement warranties and 24/7 JTAC access which is far from a "please help me" mailing list!

There are plenty of companies giving 24/7 hardware replacement warranties on commodity x86/SPARC kit (of the sort Quagga runs on), and as we've already established, if you want better support than the mailing list you can come to us.

Quagga certainly has its downsides, but they aren't the ones you've mentioned.
As a rule of thumb, I'd say if you need to route more than 3Gbps through a single router you'd probably be better off with Cisco/Juniper kit. It's worth bearing in mind that because the cost of each router is so small for Quagga, it's perfectly feasible to have, for example, one router per upstream, where a more traditional network layout might have a single edge router (or active/passive pair).

CrystalVPS · 8th September 2008

I know a lot of people are of the opinion that a load of quaggas then some switches that support multipathing makes more business sense than having fewer Juniper/Cisco routers. I can see where they're coming from, but I still prefer my nice shiny Junipers.