[Soekris] net4801 throughput with no IPSEC
komarek at cmu.edu
Thu Apr 21 14:36:30 UTC 2005
Just a crazy question: is the Linux performance similar to the FreeBSD
performance on these chips? I have a net4501 running Linux that I could
use for testing, if the answer to this question is not already known and
a net4501 has enough horsepower to make the test worthwhile.
Jason Dixon wrote:
> On Apr 21, 2005, at 2:39 AM, Poul-Henning Kamp wrote:
>> I've spent considrable amounts of time on the if_sis driver, I don't
>> really have anything to show for it.
>> The one experiment that did turn out some improvement was to reuse
>> transmit buffers (when certain conditions were fulfilled) to replenish
>> the receive ring. That way the driver avoids freeing buffers and
>> immediately allocating them again.
>> There were a number of downsides though: it does badly in many
>> benchmarks (because only a single stream in a single direction is
>> tested) and it handles massive bursts of trafic worse than the
>> regular version.
>> For at typical routing or server application however it improved
>> things somewhat.
>> I've still not given up though.
> No discounting your work, since all of the BSDs lean on it, but perhaps
> Soekris might do well to reconsider another chipset in future revisions
> of the 4x01 systems. Seeing as the BSDs account for a large portion of
> usage on these machines, it's something to think about. I've personally
> run into a number of scenarios (particularly with IPSec, CARP, pfsync,
> etc) where the interfaces perform badly or fail altogether.
> Jason Dixon
> DixonGroup Consulting
> Soekris-tech mailing list
> Soekris-tech at lists.soekris.com
More information about the Soekris-tech