Hi, Will -
> > OK, that confirms the suspicion that a sampled-metric type of pmda
> > approach suits this better than timestamped-line-of-trace-data one.
> Note that the number above was with really light traffic. It is
> quite possible that number of packets would be hundreds of thousands
> per second. The netdev-times perf script records a huge amount of
> trace data and does post processing on it.
Right, it's likely that in-situ statistics aggregation a la your stap
script is a winning approach here.
> The syscall can return before the packet is sent (or freed) so it is
> not clear what skb-free-to-sys_exit would show.
Good point.
> [...] My understanding is that some of the networking hardware is
> using DMA so the kernel is just giving it some pointers on a list to
> the networking hardware, so the kernel doesn't know exactly when a
> particular packet has been send. The kernel knows the data has been
> sent because the skb can be freed. [...]
Right, if we have zero-copy hardware, that makes sense.
- FChE
|