The PC Engine APU2 is an SBC with an AMD SoC running at a max of 1 Ghz - 12 watts. They run Intel NICs but it wasn’t a great surprise the thing needed some tuning to get full 1gb/s speed out of it.
https://www.amd.com/en/system/files?file=2017-06/g-series-soc-product-brief.pdf
Super nice boards;
Oh Ok that’s why! 
I don’t know how much is an APU2 but a Mini PC is about 200 USD
It’s about the same 
I am looking at getting a Protectli box, but believe this is “just” a rebadged qotom machine.
Feel free to ping me if you have any recommendations for a good low power, silent box. Ideally I would want something that can take a 10gb/s sfp module as that is what comes into my apartment…
I also got an APU2 board, not sure if you got following settings allready (do this for all NICs) and reach somewhere around 920Mb/s.
ethtool -G ethX tx 4096 rx 4096
ethtool -K ethX tx on sg on tso on gso on gro on lro off
I put the commands in /config/scripts/vyos-postconfig-bootup.script which is run at boot.
And don’t have mitigation=off set.
Enabled mitigations again and implemented the changes @dannyw shared. getting over 900mb/s.
I will stick with that configuration.
@Dmitry, I can update the wiki with such information if you think this would be generally useful.
Intel NIC offloading may not be a good idea on a router, or so I have read somewhere.
Ring buffers - yes, rx 256 is too small resulting in lots of FIFO errors and dropped packets, rx 4096 works better.
Looking at this thread as I’m trying to get close to 1 Gb/s throughput on APU4 as BGP routers for small local ISP I’m running here. Unfortunately the APU4 has i211 NICs with only 2 queues. I’ve recently had a ~200k pps incoming DDoS which maxed out 2 of 4 CPU cores, looking to improve this - or do I really need to get faster (but more power-hungry) server boards to work well for this use?
I just need pure routing, no NAT etc. I wish there was an easy to setup router distro with something faster like DPDK but it seems it’s not there yet.
hey there, are you able to add more information about why offloading isn’t a good idea or have a link that explains more?
looks like your topic might deserve its own thread to get a better and more specific response.
Only LRO offloading shouldn’t be used on a router according to this:
Due to a known general compatibility issue with LRO and routing, do not use LRO
when routing packets.
On my apu4 board I was able to reduce CPU usage during download from 100% to <50% just by using
sudo ethtool -K ethX tx on sg on tso on gso on gro on lro off
LRO is not used as you can see.
I still don’t get full 1Gbps, but getting closer
Download: 863.05 Mbps (data used: 1.0 GB)
Upload: 938.31 Mbps (data used: 644.7 MB)
The problem with apu4 is interrupts don’t seem to distribute evenly over the 2 available CPUs but only for RX. TX looks fine:
CPU0 CPU1 CPU2 CPU3
56: 0 0 0 883331 PCI-MSI 2097153-edge eth3-rx-0
57: 1407 0 0 0 PCI-MSI 2097154-edge eth3-rx-1
58: 0 959698 0 0 PCI-MSI 2097155-edge eth3-tx-0
59: 0 0 946593 0 PCI-MSI 2097156-edge eth3-tx-1
Hey @phillipmcmahon,
I’ve been struggling with a big performance regression when updating from 1.2 to 1.3.
With the mitigations off, could you please share your output for:
grep . /sys/devices/system/cpu/vulnerabilities/*
I would like to compare it with my device to understand if the mitigations are actually off.
Thank you
Hi there, I no longer have vyos running on my apu, I moved to a protectli device running esxi so I could consolidate a bunch of stuff.
I will resurrect the unit this week, validate it’s doing 1gb and share my config. I do think getting it to 1gbs is absolutely at it’s peak and it’s not a great surprise to see it fluctuate due to this.
Thanks.
Well, in my case with Vyos 1.2 I could do 1Gbit/s easily, with having 2 cores at around 90% usage and the other 2 at around 50%.
Now with Vyos 1.3, it’s maxing out at 700Mbit/s and all cores maxed out.
It’s almost a 50% regression, which doesn’t make sense from a version upgrade.
Fast forward a year or two later, and it was time to go for 10 Gbps interfaces. Also, the APU has some stability issues under high load - I’ve seen unexplained random reboots. So it will have to find some other light use (one with a small hardware hack works as a GPS synced NTP server, with very good accuracy of hardware timestamping by the Intel NIC), and my new router platform of choice are Supermicro X9 series boards with Xeon E3-1220v2 CPUs and Chelsio T440-CR (quad SFP+) PCIe cards. It seems a reasonable compromise between electricity costs, hardware costs and performance. The hardware is about 10 years old (needs BIOS upgrade to fix a year 2021 bug), but routing about 1 Gbps of traffic over two 10 Gb interfaces is hardly visible as any CPU load (about 2-3%).
On my APU2E4 (with 3xi210AT NIC) running 1.3.1-S1 I easily get 1Gbps. Make sure you have an APU2E4 with i210AT NIC, not i211AT (like an APU2E5 or similar, see https://www.pcengines.ch/apu2e4.htm). The i211AT only has 2 queues per port, the i210AT has 4 queues per port. 4 transmit/receive queues per port will distribute interrupts better over all 4 CPU cores.
https://www.speedtest.net/result/c/7c8abcde-a80c-457f-af44-2295e147fb7c