(no title)
hapless | 3 months ago
This was doable back in 2008 with about $15k of x86 gear and a Linux kernel and a little trickery with pf_ring. The minute AMD K10 and Intel Nehalem dropped, high routing performance was mostly a software problem... Which is cool as hell, compared to the era when it required elaborate dedicated hardware, but it does not make it cheap or easy. Just, commodity. Expensive commodity.
Now you can buy a device off the shelf for $800 that will do it on the CPU, to avoid the cost of Cisco or Juniper, and it has a super simple configuration interface for all the software-based features. Everything you could do in L3/L4 on a Linux platform in 2008, for like, 1/16th the price, with vastly less engineering effort. It is just like, a thing you buy, and it all kinda works outta the box.
No pf_ring trickery, no deep in-house experience, just a box you buy on a web site and it moves 10 gbps with filtering for $800
There's no real magic here: they use absolutely shockingly enormous ARM chips from Amazon/Annapurna. You can build an $800 commodity platform that rivals a $15k commodity platform in 2008, and both of them replace what used to cost $500k.
Is it as good as Cisco or Juniper? oh, certainly not. Will it route and filter traffic at much greater rates, for $800, than anything they have ever been bothered to offer? ABSOLUTELY
versteegen|3 months ago
hapless|3 months ago
Although, their original paper says they used a 2-socket prototype and got some very impressive numbers: https://www.sigops.org/s/conferences/sosp/2009/papers/dobres...
So maybe you could skate by with a slightly cheaper machine ;)