[ovs-dev] [PATCH net-next 2/2] openvswitch: Use zerocopy if applicable when performing the upcall

Jesse Gross jesse at nicira.com
Fri May 24 22:18:09 UTC 2013


On Fri, May 24, 2013 at 2:57 PM, Eric Dumazet <eric.dumazet at gmail.com> wrote:
> On Fri, 2013-05-24 at 14:23 -0700, Jesse Gross wrote:
>> On Fri, May 24, 2013 at 11:58 AM, Eric Dumazet <eric.dumazet at gmail.com> wrote:
>> > On Fri, 2013-05-24 at 10:24 -0700, Jesse Gross wrote:
>> >
>> >> Does this have any impact on small packets? Those are usually the
>> >> common case (i.e. TCP SYN) and I think this is slightly less optimal
>> >> for those.
>> >
>> > No difference at all, small packets are copied anyway in skb->head
>>
>> Yes, but it makes the Open vSwitch code slightly worse - for example,
>> currently checksumming and copying are done in a single step but this
>> prevents that. Actually, I'm also curious about the test case that was
>> used for large packets and the full profile output since checksumming
>> and GSO aren't listed in the one that Thomas gave.
>>
>
> GSO is fully supported in nfnetlink, I see no reason why Open vSwitch
> would not allow that.

Offloads are supported. What I want to know is how they affect
performance with this change.

>> My guess is that there isn't a real different for small packets since
>> everything will be in the cache but it seems worth checking given that
>> this is optimizing a rare case at the expense of the common one.
>
> I really doubt checksumming a SYN/ACK packet is that a performance
> issue. Do you have performance numbers ?
>
> You could always provide a patch to restore this copy/checksum if it
> really gives a benefit, and if people still use NIC not doing this
> checksum.

If it makes a difference then it needs to be addressed before this
patch goes in since it's the common case. I don't think it will but
that's why I'm asking for numbers.



More information about the dev mailing list