Won De Erick wrote: > Another thing, I observed that in the above test, the net.isr is enabled by > default. When I tried disabling this, > > # sysctl net.isr.direct=0 > net.isr.direct: 1 -> 0 > > the result: > > 52 root 1 -68 - 0K 16K WAIT b 64:00 42.97% irq32: bce1 > 51 root 1 -68 - 0K 16K WAIT a 38:22 12.26% irq31: bce0 > > The CPU utilizations considerably dropped!
You will probably find a "swi" process that has picked up the difference (when isr.direct is disabled, some of network protocol processing is offloaded to a swi thread). This might help spread the load across CPU but in my testing it didn't help real-world throughput.
signature.asc
Description: OpenPGP digital signature
