demo stuff

[Dave Taht]

On Sat, Mar 17, 2012 at 2:32 PM, Francis Dupont <fdupont at isc.org> wrote:
>> An example of a problem that we ran into was that packets can come in
>> misaligned, and the default code generated spent much time in
>> unaligned traps.
>
> => I don't code things as *(u_short *)foo when foo is a random pointer.

well good, but you probably assume that things like ipv6 addresses
are aligned, and if they are not, merely comparing two addresses
can cause 8 times the unaligned instruction traps than ipv4 does.

I note that the fix (and why I mentioned it in regards to your
reference to checksums) for ipv4 was in the csum_partial
routine for ipv4, where only 8 traps cut forwarding performance nearly
in half....

I'll stop whining now and get to work...

oh, I only pointed to the cerowrt bug list. this is the overall
openwrt one. just the open bugs

https://dev.openwrt.org/query?status=assigned&status=new&status=reopened&order=id&col=id&col=summary&col=milestone&col=status&col=type&col=priority&col=component&milestone=!Kamikaze+Bugs+Paradise&milestone=!Kamikaze+Features+Paradise&report=12&desc=1

>> http://www.bufferbloat.net/issues/195
>
> => I agree it is a sure way to get poor performances.
>
>> For further elucidation on the problems working on embedded hardware
>> can induce vs working on easy stuff like the mainline architectures,
>> please view
>>
>> http://www.bufferbloat.net/projects/cerowrt/issues
>
> => for me the mainline architecture is not the x86, in particular
> in its 32 bit modes.

x86_64?
>
> Regards
>
> Francis Dupont <fdupont at isc.org>



-- 
Dave Täht
SKYPE: davetaht
US Tel: 1-239-829-5608
http://www.bufferbloat.net