Re: [PATCH] net/null: Support bulk alloc and free.

["Koujalagi, MalleshX" <malleshx.koujalagi@intel.com>]

Hi Ferruh,

Bulk allocation gives benefit but how much, will check and provide patch.

Best regards
-/Mallesh

-----Original Message-----
From: Yigit, Ferruh 
Sent: Wednesday, March 7, 2018 2:57 AM
To: Ananyev, Konstantin <konstantin.ananyev@intel.com>; Koujalagi, MalleshX <malleshx.koujalagi@intel.com>; dev@dpdk.org
Cc: mtetsuyah@gmail.com
Subject: Re: [dpdk-dev] [PATCH] net/null: Support bulk alloc and free.

On 3/5/2018 3:36 PM, Ananyev, Konstantin wrote:
> 
> 
>> -----Original Message-----
>> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Ferruh Yigit
>> Sent: Monday, March 5, 2018 3:25 PM
>> To: Koujalagi, MalleshX <malleshx.koujalagi@intel.com>; dev@dpdk.org
>> Cc: mtetsuyah@gmail.com
>> Subject: Re: [dpdk-dev] [PATCH] net/null: Support bulk alloc and free.
>>
>> On 2/3/2018 3:11 AM, Mallesh Koujalagi wrote:
>>> After bulk allocation and freeing of multiple mbufs increase more 
>>> than ~2% throughput on single core.
>>>
>>> Signed-off-by: Mallesh Koujalagi <malleshx.koujalagi@intel.com>
>>> ---
>>>  drivers/net/null/rte_eth_null.c | 16 +++++++---------
>>>  1 file changed, 7 insertions(+), 9 deletions(-)
>>>
>>> diff --git a/drivers/net/null/rte_eth_null.c 
>>> b/drivers/net/null/rte_eth_null.c index 9385ffd..247ede0 100644
>>> --- a/drivers/net/null/rte_eth_null.c
>>> +++ b/drivers/net/null/rte_eth_null.c
>>> @@ -130,10 +130,11 @@ eth_null_copy_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
>>>  		return 0;
>>>
>>>  	packet_size = h->internals->packet_size;
>>> +
>>> +	if (rte_pktmbuf_alloc_bulk(h->mb_pool, bufs, nb_bufs) != 0)
>>> +		return 0;
>>> +
>>>  	for (i = 0; i < nb_bufs; i++) {
>>> -		bufs[i] = rte_pktmbuf_alloc(h->mb_pool);
>>> -		if (!bufs[i])
>>> -			break;
>>>  		rte_memcpy(rte_pktmbuf_mtod(bufs[i], void *), h->dummy_packet,
>>>  					packet_size);
>>>  		bufs[i]->data_len = (uint16_t)packet_size; @@ -149,18 +150,15 @@ 
>>> eth_null_copy_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)  
>>> static uint16_t  eth_null_tx(void *q, struct rte_mbuf **bufs, 
>>> uint16_t nb_bufs)  {
>>> -	int i;
>>>  	struct null_queue *h = q;
>>>
>>>  	if ((q == NULL) || (bufs == NULL))
>>>  		return 0;
>>>
>>> -	for (i = 0; i < nb_bufs; i++)
>>> -		rte_pktmbuf_free(bufs[i]);
>>> +	rte_mempool_put_bulk(bufs[0]->pool, (void **)bufs, nb_bufs);
>>
>> Is it guarantied that all mbufs will be from same mempool?
> 
> I don't think it does, plus
> rte_pktmbuf_free(mb) != rte_mempool_put_bulk(mb->pool, &mb, 1);

Perhaps we can just benefit from bulk alloc.

Hi Mallesh,

Does it give any performance improvement if we switch "rte_pktmbuf_alloc()" to "rte_pktmbuf_alloc_bulk()" but keep free functions untouched?

Thanks,
ferruh


> Konstantin
> 
>>
>>> +	rte_atomic64_add(&h->tx_pkts, nb_bufs);
>>>
>>> -	rte_atomic64_add(&(h->tx_pkts), i);
>>> -
>>> -	return i;
>>> +	return nb_bufs;
>>>  }
>>>
>>>  static uint16_t
>>>
>