On Thu, May 08, 2025 at 07:58:10AM +0200, Janne Johansson wrote: > Yeah, some part of the offloadings started at gig speeds, since when > those first arrived, the CPU could not be expected to do all the work > to fill a gig link like they could for 10 and 100Mbit/s, as cards now > come in 40,50,100,400Gbit/s speeds, you are forced to have offloading > or it would never get near linkspeed ever.
The ice(4) driver currently runs without any offloading enabled and manages about 3, or 4, or maybe 5 Gbit/s line-speed in practice, for both Rx and Tx. The hardware supports 1Gb/10Gb/25Gb/50Gb/100Gb. Users who buy such hardware will reasonably expect that effective throughput will be somewhere close to line-speed. Especially now that the network stack is increasingly getting unlocked to process packets in parallel. With wifi, which is still entirely giant-locked, we've seen that not offloading AES crypto to hardware means that 11n/11ac at 100 Mbit/s or higher keeps the CPU so busy during interrupts that mouse movement in X11 is stuttering. Without offloading we'd be stuck with either 11a/b/g speeds or unusable X11 on laptops. So far, we are only offloading crypto, and currently manage about 300 Mbit/s on 11ac wifi under ideal conditions. Rather than 1 Gbit/s which would be possible with more parallel processing and offloading of things like IP checkums and TCP LRO etc.
