On Jan 25, 2006, at 11:50 AM, Pavel Cahyna wrote:

> On Wed, Jan 25, 2006 at 08:33:57PM +0100, Manuel Bouyer wrote:
>> On Wed, Jan 25, 2006 at 08:01:47PM +0100, Pavel Cahyna wrote:
>>> On Wed, Jan 25, 2006 at 07:50:57PM +0100, Manuel Bouyer wrote:
>>>> On Wed, Jan 25, 2006 at 12:22:06PM +0100, Pavel Cahyna wrote:
>>>>> the itended effect...
>>>>>
>>>>> What is the purpose of hardware VLAN tagging, anyway? The  
>>>>> performance
>>>>> impact will be rather negative, because searching the VLAN tag  
>>>>> in the
>>>>> linear list of mbuf tags will be probably slower than simply  
>>>>> prepending it
>>>>> to the mbuf. It also seriously limits usability of tcpdump  
>>>>> (tcpdump sees
>>>>
>>>> I can tell you that it has a positive inpact on a Alpha DS20 with a
>>>> gigabit interface; I get a 10% increase with tcpdump.
>>>
>>> tcpdump?
>>
>> Sorry, I meant ttcp
>
> Do you have any explanation why does it help?

No need to copy data coming in from the chip?

> Could be HW tagging made conditional on some setting, like hardware
> checksums (if the hardware allows turning them off)?

Why would you want to turn them off if the hardware supports it?  How  
many VLANs are you working with, anyway?  Looking them up is almost  
certainly cheaper than copying data that is guaranteed not to be in  
cache.

-- thorpej