Hello!
> ie it depends on the size of frame, and for eth sized chunks overhead is ~83%.
>
> > Ah. That's pretty convincing. But it would be a huge win for zero-copy
>
> hmm, the quick check i just did seems to suggest the difference might not
> disappear with a cold cache either. Did anyone try replacing the
> checksum-copy with a plain copy, made the kernel ignore checksum errors
> and benchmarked against a normal one? That would give an idea of the
> true sw checksum impact (ignoring the extra hw support overhead).
I've said, I tried. On loopback improvement is invisible inside
statistical bounds. On 100Mbit ethernet the difference does not affect
throughput (it is saturated in any case), so that I looked at CPU consumption,
which is very random. Again, the difference is inside statistical errors.
Things can change drastically on Giga ethernet, but it is almost
impossible to measure because we are bottlenecked at receiver side.
Voila.
Alexey
|