The Ars guide to building a Linux router from scratch

[System Error Message]

For routers - believe it or not, most of the ARMv7's and Intel x86-64's actually perform quite well - at least the recent ones do...

The Tilera, while interesting - under Linux, most of the cores are wasted - mostly due to poor thread scheduling in Linux, and also how to massively take serial tasks parallel - many cores are nice to process certain things - like OpenGL and graphics primitives - look at CPGPU/OpenCL/Cuda - fantastic numbers until one tries to do general work - and AMD has tried to do some interesting things with their APU's and HSA...

But Tilera and MicroTik ain't there - and likely not going to be...

The CCR1072 is actually limited by memory bandwidth. Using faster RAM yields better results but thats because they have a lot of CPU power just like a GPU which again has many times more memory bandwidth. The 72 core uses 125W whereas a GPU uses between 150W to 250W depending on how big. The problem however is that there arent many others selling Tile and mikrotik locks down their firmware so you cant install openWRT on it (which some have been trying) or install other programs on their firmware. So while mikrotik makes some really good routers and firmware you just cant use them for anything other than just being a router.

Linux does have issues when it comes to really powerful hardware but in the course of using a GPU for networking the code would be handled by the GPU or openCL or CUDA so all the CPU has to do is pass data and start/stop tasks but it will not have to do the massive threading as that is automatically done by the GPU. The only problem with packetshader was that the test was done with multiple 10G ports which linux doesnt handle that well with many PCIe bandwidth needed but in a home/business environment or even in a few years with at most 10G internet i think linux should be able to handle it.

 

[sfx2000]

The CCR1072 is actually limited by memory bandwidth. Using faster RAM yields better results but thats because they have a lot of CPU power just like a GPU which again has many times more memory bandwidth.

Let's do some simple figuring...

(note, read below and know - The Tilera is starved for memory - and below pretty much proves my point - the Tilera is an interesting chip, but it's a rest-stop on the CPU Interstate/Autobahn - folks have moved on...)

AMD Durango - this is the APU in the PS4 - the CPU's (x86-64 small cores * 8) run at 1.6 GHz - and the GPU which has 18 CU's which are available at half that clock speed) - depending on the workload - the OpenCL/Direct Compute speed - the CU's can parallel task things pretty fast using OpenCL/Direct Compute, but single thread performance, the Jaguar cores, even one, will outpace it...

And this requires GDDR-5 (55oo Mhz) , on the PS4, which gives about 176 GB/sec to the APU (CPU and GPU)

DDr3-1600 dual channel on an x86-64 gives you around 25 GB/sec...

Most Home Router/AP's - much less than this, and looking forward into 2016/2017, since packet mangling on the WAN/LAN interface is not compute bound, it's memory bound, is that devices will have to evolve as we get into the Gigabit WAN space...