Router Board Opinions?

[LRL]

I'm wondering what the general rule of thumb is some of you follow in regards to the load for different router boards.

Specifically, how many subs would you recommend as the limit for a RB493G as a tower router?
It would be running OSPF, no NAT, no MPLS/VPLS (at least not initially), possibly some Queues, that's it. I know I've seen or heard that >120 is recommended but I believe that was if the RB was preforming NAT for the customers. With the number of DMCA letters we get I can't see why anyone wouldn't hand out public IPs to CYA.

RB1100AH, thoughts?

RB450G, thoughts?

[adairw]

I easily run 100-200Mb through 2011's running OSPF routing, mpls/vpls tunnels, vlan's and bridges. all of my shaping is at the core.

[LRL]

I curious, what do you run at the core for shaping?

We currently run a pair of CCR1036's at our core and we currently do all our shaping there. I'm starting to see some performance issues, I believe due to the number of queues and firewall rules.

I'm working on transitioning most of our shaping to closer to the network edge. Ideally I'd like to have the client CPE preform all of our shaping but bursting isn't configurable in a manor consistent with our plan descriptions.

I consider doing away with the bursting, but that would instantly raise our traffic 30%.

We currently see peaks of 190'sMbps from a couple of our towers. The backhauls are capable of 280Mbps. The OSPF overhead is the question mark for me.

[rebelwireless]

The 600Mhz CPU models of Mikrotik are good for up to about 200Mbps routed/bridged in the real world, in my experience.

If you shape on the device, more like 50Mbps.

Sub count doesn't matter that much because you're likely to run into throughput issues before you run out of RAM.

Basically, the model number isn't so important, it's the CPU speed and architecture. Multiple models share the same CPU.

A 493G and an 2011 are pretty similar.

I run a x86 RouterOS box with a Core2Duo, Sandisk U110 SSD, and 2GB RAM. I can push line speed through it with queues using iPerf. The gigabit port is my limit as the CPU has plenty of headroom.

[LRL]

I'm debating going with CCR1009 for our AC powered sites and RB1100 for our alternative sites.

Mikrotik's site leaves you with a incomplete picture. According to their performance tests the RG493G can route 499Mbps of 512B packets while employing 25 simple queues, and 678Mbps of 1500B packets with 25 queues. I guess the question is what % of CPU does OSPF consume, but there are to many variables to OSPF to determine that. We redistribute some of our BGP learned roues via OSPF because we have two gateways on different providers and on different sides of the network. Mostly just the upstream local routes.

[rebelwireless]

OSPF isn't very heavy at all. I've run largish OSPF areas on Mikrotiks with <5% CPU (rb951).

The crappy thing about the CCRs is that they aren't the powerhouses they are claimed to be for queues. Not really much better than the PPC models. I prefer the RB1100ahx2 for queues as they are more tried and true. It's really a shame that the x86 platform (and PPC) only gets 2GB of ram with routeros and no apparent plans to update it to 64bit. An i5 based RouterOS with 16GB would make for a handy and familiar shaping box :/

With that in mind I've been playing with creating a script to create a tcnq based queue tree on linux from a customer database. Ideally something that would run on Edgerouters and x86_64 linux. Currently trying to get nDPI and nDPI-netfilter to play nice for identifying connections. The devs over at nTop are even interested in helping.