Robert Olsson wrote:
Ben Greear writes:
I still doubt ;-)
With e1000 I played with various settings for RX-buffers rather recently
when the 82544 increased the number of available buffers from 256 to 4096.
And I guess my test looks a bit like yours... Injecting an "overload" of
packets. I found was 256 buffers was the optimum. Approximative of course.
It's possible that it is a particular issue with my NICs (Old Phobos
4-port).
Phobos folks said the bridge chipset has errata that make it un-suitable for
high speeds. And something about a memory divide-by-four error. It may be
that the extra buffers help hide the hardware defects in some manner.
I was, for instance, seeing cases where packets just dissappeared...and no
error counters were being bumped.
With the latest kernel drivers, the 570tx NIC seems to have
trouble autonegotiating full-duplex again, so I have not been testing
with it lately (I think it uses the same bridge chipset anyway.)
I will try changing around those numbers again now that I have a baseline
to work from.
And as you saw for SMP with recycle it is easy to feed the recycled skb
back to CPU were it was created/processed.
Yes, I can see how that would be useful on SMP, so there may be less gain
for single-proc systems. I actually am pretty fuzzy on cache-line optimizations
and the like...
Thanks,
Ben
--
Ben Greear <greearb@xxxxxxxxxxxxxxx> <Ben_Greear AT excite.com>
President of Candela Technologies Inc http://www.candelatech.com
ScryMUD: http://scry.wanfear.com http://scry.wanfear.com/~greear
|