Price?

Posted Nov 13, 2024 9:56 UTC (Wed) by danieldk (guest, #27876)
In reply to: Price? by Sesse
Parent article: The OpenWrt One system

I agree that Cortex-A53 are somewhat anemic. *However*, the performance of these routers is mostly determined by the capability of their network processing units (confusingly also called NPUs). Good NPUs can offload PPPoE, NAT, etc., so in practice these routers will be near-idle when it comes to regular cores during high network load because NPUs are doing the work.

The OpenWrt One uses a Mediatek Filogic 820 (I think), which generally use NPUs that are really well-supported by the kernel.

To give an example, I have a router that uses Filogic 830 and it has no issue at all doing PPPoE, NAT, and firewalling at line speed (2.5Gbit), while the CPU cores are idle. It even seems to do fq-codel in hardware and gives me a buffer bloat rating of A.

tl;dr: the CPU cores are not for processing packets, but for running the web interface.

to post comments

Price?

Posted Nov 13, 2024 10:59 UTC (Wed) by paulj (subscriber, #341) [Link] (1 responses)

NPUs are generally a different thing to the common forwarding logic in switch ASICs / SoCs. NPUs are programmable logic that are aimed at processing data as it streams through, applying transforms to the data (possibly with the aid of other fixed blocks of logic, e.g. cryptographic functions). The switching ASICs and the switching blocks in SoCs in these kinds of AP/routers are generally fixed-function pipelines, configured via tables (mostly). I.e. you setup lookup tables (e.g. enable or disable whatever features, and then create entries in the tables that control those features) - and the fixed-function pipeline applies its logic to packets, guided by lookups in tables. The logic is not generally programmable - it's not an NPU.

I can't find programming docs for the PSE in the MediaTek engine, but OpenWRT patches I can find for stuff like "Wireless Ethernet Dispatch (offload of wifi packets to PSE) look to be the typical programming of tables for fixed-function pipeline.

Generally, where you see NPUs, you'll still see a fixed-function packet processing pipeline. Cause the latter handles the common case much faster and with lower latency than NPUs. The NPUs will be there to accelerate the "slow path" - allowing uncommon case packets to be processed faster than entirely on CPU.

Price?

Posted Nov 13, 2024 19:51 UTC (Wed) by danieldk (guest, #27876) [Link]

I just wanted to say 'thank you' for this reply. For someone who just dabbled a bit in OpenWrt and OPNsense, it's very insightful. I guess I was slightly misled because some vendors call both types of offloading together 'NPU' in their marketing. At any rate, I hope the general point was clear though, that even with old Cortex cores, a device could still be pretty good as a router.