[ovs-dev] [PATCH] ofproto: Remove per-flow miss hash table from upcall handler.

Ethan Jackson ethan at nicira.com
Fri May 9 22:52:54 UTC 2014


I agree completely.  I think since this optimization was introduced,
enough as changed that I suspect we can remove it without a negative
impact.  Furthermore, we're so sensitive to memory access patterns
now, that I further suspect that removing it may actually help the
average case.  That said, we need to actually measure the change
before we consider merging it.

Ethan

On Fri, May 9, 2014 at 2:45 PM, Ben Pfaff <blp at nicira.com> wrote:
> On Fri, May 09, 2014 at 02:38:33PM -0700, Ryan Wilson wrote:
>> Sorry for the delayed response.
>
> No problem.
>
>> Unfortunately, I've had some issues with setting up an NVP perf
>> environment lately so I haven't gotten any numbers. When I do, I'll
>> publish them on this thread.
>>
>> Ethan's rationale is that it simplifies the code and we don't incur
>> the memory overhead of allocating a hash map and this would outweigh
>> the negative effect of possibly re-processing a packet from the same
>> flow. Since userspace forwarding has gotten so much faster, this
>> wouldn't be so bad.
>>
>> Obviously, I need perf numbers to justify these claims; I will post
>> them when I do.
>
> Thanks.
>
> Let me clarify that I would like to drop the complexity here too.  I
> just want to be fairly confident that we won't have to reintroduce it.



More information about the dev mailing list