I’m running into performance issues on the macchiatobin single shot where the pppoe session is local to the macchiatobin, running on a bridge vlan interface.
My current setup is with eth3(2500BASE-X) and eth1(10GBASE-X) making up br0 (1550 mtu), eth3 contains a GPON SFP, eth1 is directly connected to a server. eth0 (10GBASE-X, 9000 mtu) connects to the network, and is not part of any bridge.
If I’m running a PPPOE session on my server, going through the bridge to reach the GPON SFP, speed tests give me my line speed of 1500Mbps down, 1000Mbps up, with 100% softirq usage on eth3_rx_shared for download.
If the PPPOE session(1500 mtu) is on the macchiatobin (via br0.35, 1520mtu), I max out around 1170Mbps download, and 700-800 upload. This was an improvement from ~900 download before using irqbalance to optimize the irqs (At a very least I needed the smp_affinity for eth3_tx_shared to be not on CPU0, as when it was all performance was even worse).
I’m also very unclear on where the advertised “packet processor” comes in to play, and if it is being used or not. Most offloading options in ethtool -k are [fixed], but the ones I can turn on have been turned on by default.