Re: [PATCH 2/6] mempool: add stack (lifo) based external mempool handler

[Olivier MATZ <olivier.matz@6wind.com>]

Hi,

>> Hi David,
>>
>> On 02/16/2016 03:48 PM, David Hunt wrote:
>>> adds a simple stack based mempool handler
>>>
>>> Signed-off-by: David Hunt <david.hunt@intel.com>
>>> ---
>>>  lib/librte_mempool/Makefile            |   2 +-
>>>  lib/librte_mempool/rte_mempool.c       |   4 +-
>>>  lib/librte_mempool/rte_mempool.h       |   1 +
>>>  lib/librte_mempool/rte_mempool_stack.c | 164
>>> +++++++++++++++++++++++++++++++++
>>>  4 files changed, 169 insertions(+), 2 deletions(-)  create mode
>>> 100644 lib/librte_mempool/rte_mempool_stack.c
>>>
>>
>> I don't get what is the purpose of this handler. Is it an example or is it
>> something that could be useful for dpdk applications?
>>
> This is actually something that is useful for pipelining apps,
> where the mempool cache doesn't really work - example, where we
> have one core doing rx (and alloc), and another core doing
> Tx (and return). In such a case, the mempool ring simply cycles
> through all the mbufs, resulting in a LLC miss on every mbuf
> allocated when the number of mbufs is large. A stack recycles
> buffers more effectively in this case.
> 

While I agree on the principle, if this is the case the commit should
come with an explanation about when this handler should be used, a
small test report showing the performance numbers and probably an
example app.

Also, I think there is a some room for optimizations, especially I
don't think that the spinlock will scale with many cores.

Regards,
Olivier