ksoftirqd behaviour on high network traffic

Question

I am running Windriver Linux on a MIPS (octeon) based hardware. Linux runs on 16 cores and we have koftirqd/0 to ksoftirq/15 running. I observe the following behavior of load balancing on high incoming traffic ( like ping flood):

First, kostfirqd/0 takes all the load until it reaches some where around 96-97% of cpu. Once cpu0 reaches 96-97% of usage, koftirqd/1 starts taking load and % of CPU for cpu1 starts increasing. On more traffic being pumped in, cpu 1 reaches 96 -97% and cpu2 starts taking load. And it goes on till ksoftirqd/15 takes 96-97% as the incoming traffic increases.

Is this an expected behaviour? Could you please let me know whether it is the default linux behavior or a possible improvement done by Windriver.

Thanks a lot,

Vasudev

Answer 1

Cavium Mips ethernet driver has the logic to send inter processor interrupt to other cores to take the load given the conditions. When ever backlog crosses certain limit, then IPI is sent to other cores. And the handler for the IPI in turn is nothing but the NAPI poll logic.

Hence the behavior.