Re: Backporting DFly patches to FreeBSD?

[EM1897@xxxxxxx]

In a message dated 2/28/2005 2:15:39 PM Eastern Standard Time, Freddie Cash <fcash-ml@xxxxxxxxxx> writes:

>On February 28, 2005 11:03 am, EM1897@xxxxxxx wrote:
>> >I don't have the reference handy at the moment, but there was an
>> > article about a network test where FreeBSD was able to process and
>> > route 1 million IP packets per second.  Linux on the same hardware
>> > was only able to process approx. one hundred thousand IP packets
>> > per second.
>
>> That doesn't sound even remotely credible. Maybe FreeBSD
>> was "routing" to the loopback device?
>
>The following message was posted to the FreeBSD net mailing list:
>http://lists.freebsd.org/pipermail/freebsd-net/2004-September/004840.html
>
>With a link to the following paper on the subject:
>http://people.freebsd.org/~andre/FreeBSD-5.3-Networking.pdf
>
>Page 11 mentions two different projects that both reached 1 million pps
>throughput using IP FastForwarding.  These were mentiond in a mailing
>list thread on the FreeBSD lists, although I have not yet bothered to
>track down the exact threads.

As I said, its not credible. The fact that it says it is 
meaningless. Linux easily routes more than 100K pps,
so whoever wrote this is just a fool, because anyone
in the linux camp that he told that to would laugh in 
his face.

FreeBSD 4.11 can't do a million with a 3.4Ghz processor
using forced-forward bridging, which completely bypasses
the IP stack  and has actual zero lookup time. The guy who
wrote this (Andre I assume?) can't do better than zero time
no matter how wonderful his code is. You're not going to get
a 30% gain no matter what, because the IP stack is 
probably only 10% of the processing. 5.3 is 15-20% slower
 than FreeBSD 4.x, a notion that Robert Watson doesn't disagree with. What you are reading there is all a lot
of garbage. Maybe, if you get some code that fits in its
entirety in the CPU cache and turn off a lot of things
that usually need to run you might be able to do it, but
thats not a credible test either.

Its also not credible to make such a claim without a
complete hardware description, since the ethernet card 
and driver is more of a factor than the networking stack
for such a test. You can't compare linux and freebsd
without such a qualification, because different drivers
are "good" or "not good" in different OS/s. If you use
a card that has a poorly performing driver in Linux but a
 good one in FreeBSD, then your results are tainted, 
because you're not testing what you claim to be testing.

The FreeBSD "team" has been making a lot of claims that
are easily countered on the bench. I don't think that they
really know how to test real-world networking.