On Fri, 2012-06-01 at 23:34 +0200, Hans Schillström wrote:

It think we are on the right way now,

Some results from one of our testers:
before applying "reflect SYN queue_mapping into SYNACK"

"(The latest one from Eric is not included. I am building with
that one right now.)
Results were that with the same number of SYN/s, load went down
30% on each of the three Cpus that were handling the SYNs.
Great !!!"

I am not sure reflecting queue_mapping will help your workload, since
you specifically asked to your NIC to queue all SYN packets on one
single queue.

Eventually not relying on skb->queue_mapping but skb->rxhash to chose an
outgoing queue for the SYNACKS to not harm a single tx queue ?

Then it might be not needed, if the queue is dedicated to SYN and SYNACK
packets, since net_rx_action/net_tx_action should both dequeue 64
packets each round, in a round robin fashion.

(I had problems in a standard setup, where you can have a single cpu
(CPU0 in my case) servicing all NAPI interrupts, so with 16 queues, the
rx_action/tx_action ratio is 16/1 if all synack go to a single queue,
while SYN are distributed to all 16 rx queues)

I'm looking forward to see the results of the latests patch.

Then I think conntrack need a little shape up, like a "mini-conntrack"
it is way to expensive to alloc a full "coontack for every SYN.

I have a bunch of patches and ideas for that...

Cool ! the conntrack issue is a real one for sure.


Given the conntrack current requirement (being protected by a central
lock), I guess your best bet would be following setup :

One single CPU to handle all SYN packets.

Eventually not relying on skb->queue_mapping but skb->rxhash to chose an
outgoing queue for the SYNACKS to not harm a single tx queue.

Thanks Eric for a great job

Thanks for giving testing results and ideas !