After several hours of testing I think I was seeing the occasional ipv6 packet loss described in https://bugs.openwrt.org/index.php?do=details&task_id=3373 so I turned flow offloading back off. I don't think I really need the extra speed right now and I'd rather have ipv6 working reliably. Will wait for further bug fixes / releases from the openwrt developers.
Note - my PPPoE interface is eth0, and I've since limited the hardware queues on eth0 to 1. Prior to this change (which had zero affect on performance), I'd see something like -
..note *38 is fudged as an example, but basically zero traffic would hit the second hardware queue.
Ideally you want each receive queue bound to a different CPU to make use of a multi-core CPU....
I believe the RSS hashing algorithm was always placing PPPoE traffic in the same hardware queue - irq bound to the same CPU. Which means that PPPoE based traffic can not make use of multi-core CPUs.... would be good to see the output from your APU2 to confirm if you're getting 850Mb down in a similar scenerio....
This might interest you - https://blog.packagecloud.io/eng/2016/06/22/monitoring-tuning-linux-networking-stack-receiving-data/#hardware-interrupt-requests
Are you subscribed to our RSS feed? You can download the latest headlines and summaries from our stories directly
to your computer or smartphone by using a feed reader.
