From: Herbert Xu <herbert@gondor.apana.org.au>
Date: Thu, 29 Jan 2009 15:41:48 +1100

But yeah if a single TX queue isn't big enough to take care of
the output from a single core then you will need multiple TX
queues.  Although from an system-wide point of view it would
seem to be more optimal to make that TX queue bigger rather than
forcing the software to use multiple queues (which may come with
a bigger overhead) to make up for it.

Of course for us software developers we don't really have a choice :)

Yes, this is one of the things I told Robert, make the
TX queues larger :-)

Besides, a plain iterating RX --> TX function can't work because
that doesn't take flows into consideration and thus can introduce
reordering.