Re: [PATCH net-next 00/13] Control action percpu counters allocation by netlink flag

[Vlad Buslov <vladbu@mellanox.com>]

On Wed 23 Oct 2019 at 17:21, Jamal Hadi Salim <jhs@mojatatu.com> wrote:
> On 2019-10-23 9:04 a.m., Vlad Buslov wrote:
>>
>> On Wed 23 Oct 2019 at 15:49, Jamal Hadi Salim <jhs@mojatatu.com> wrote:
>>> Hi Vlad,
>>>
>
>>> I understand your use case being different since it is for h/w
>>> offload. If you have time can you test with batching many actions
>>> and seeing the before/after improvement?
>>
>> Will do.
>
> Thanks.
>
> I think you may have published number before, but would be interesting
> to see the before and after of adding the action first and measuring the
> filter improvement without caring about the allocator.

For filter with single gact drop action (first line in insertion rate
table in the cover letter) I get insertion rate of 412k rules/sec with
all of the actions preallocated in advance, which is 2x improvement.

>
>>
>>>
>>> Note: even for h/w offload it makes sense to first create the actions
>>> then bind to filters (in my world thats what we end up doing).
>>> If we can improve the first phase it is a win for both s/w and hw use
>>> cases.
>>>
>>> Question:
>>> Given TCA_ACT_FLAGS_FAST_INIT is common to all actions would it make
>>> sense to use Could you have used a TLV in the namespace of TCA_ACT_MAX
>>> (outer TLV)? You will have to pass a param to ->init().
>>
>> It is not common for all actions. I omitted modifying actions that are
>> not offloaded and some actions don't user percpu allocator at all
>> (pedit, for example) and have no use for this flag at the moment.
>
> pedit just never got updated (its simple to update). There is
> value in the software to have _all_ the actions use per cpu stats.
> It improves fast path performance.
>
> Jiri complains constantly about all these new per-action TLVs
> which are generic. He promised to "fix it all" someday. Jiri i notice
> your ack here, what happened? ;->
>
> cheers,
> jamal