n4rfblatt

a bit of a random question, has there been any thought of an api (e.g. maybe setsockopt() parameter) to hint to the kernel to use a particular tx ring?

or maybe to put it a bit more generic so you can just make sure a certain set of sockets are fanned out as much as you can

(IIRC, right now it all ends up being how mac hashes things to a ring)

so can't really control two sockets hashing to the same ring

at least i'm not aware of any way to do it

jbk: linux has something in that space -- "transmit packet steering" aka "XPS". Basically map CPU to TX ring.

I seem to remember someone then doing a benchmark special where they selected the local port in TCP connect() such that the response traffic came back to the same CPU (they could predict the flow-hashing done in the last-hop switch and NIC).

otherwise for N rings you need something like N^2 flows to let things level out..

Reminds me of when I launched iperf clients on sequential local-ports just to remove the entropy from everything. (Might've been at Nexenta...)

:D

sommerfeld: didn't oxide implment that for illumos? I remember a repo of similar name or am I mixing it up with another linux thing starting with X?