WAN to LAN Throughput

Hi!

what are the results of WAN to LAN Throughput benchmarks?
Martin Strbačka mentioned "Over the Christmas holidays we made some tests and the results are: Turris Omnia is able to route 975 Mbps (even with NAT)"

Is it WAN to LAN routing results? If not, what are the results of such benchmark? What about hardware NAT support? maybe it is not needed with that capable CPU?

Since they mention routing (as opposed to switching) I’m quite sure they mesured the speed between WAN and LAN ports, i.e. two ports where at least one had a dedicated connection to the SoC where the second might have been connected via the network chip.

I’m not aware of any hardware NAT component. If you want such an ASIC you’d have to add two to three zeros to the price tag. :wink: I think given the HW specs, the Linux kernel’s networking components’ maturity and the expected use cases (SOHO, not CGN) there’s little reason to worry about NAT performance.

Also this “hardware NAT”-thingie is quite a myth: Of cause some additional speed can be achieved by offloading tasks to hardware, but most SOC-vendors that talk about “HW-NAT” also provide a proprietary (precompiled, non-opens-ource - which makes this a no-go for me) kernel module to bybass the linux network stack (which is quite inefficient when it comes to forwarding packets), it’s also called “fastpath” or sth. similiar - and it seems most of the speed is achieved just by this kernel-module.

I really think this router is powerful enough to saturate a GBit link… all the guys that bought one of the new Linksys-WRT’s (the newer versions use the same SOC) are quite positive about it.

1 Like

For pure routing, yes, but the Omnia is supposed to do more than that (virtual server, VPN, NAS, firewall, etc.), so hardware offloading can be a good thing. The documentation mentions some form of acceleration at the controller level, so hopefully that will keep the CPU free for other tasks.

Yes of cause, i didn’t say it’s a bad thing, did i? :slightly_smiling:

Well, you did mention that proprietary blobs were a no-go ;), so let’s hope that it’s open source. It seems simple enough to turn on

Yes, this is absolutely true. Once i had an idea to implement HWNAT for the C7 router.

  1. Documentation is very limited.
  2. There are some leaked sources and its a full mess - many kernel modules are patched, some changes looks weird, etc.
  3. In this sources some chipset limitations are mentioned, like number of such hw fastpath channels. Moreover - it is very clear that it wont work correctly with nf_ infrastructure, so potentially is a huge security risk.

So after all it was decided no to do anything like that in OpenWRT and just to suggest some modern CPU boards for anyone who needs recent speeds.