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

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

[Mallesh Koujalagi]

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);
+	rte_atomic64_add(&h->tx_pkts, nb_bufs);
 
-	rte_atomic64_add(&(h->tx_pkts), i);
-
-	return i;
+	return nb_bufs;
 }
 
 static uint16_t
-- 
2.7.4

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

[Ferruh Yigit]

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?

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

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

[Ananyev, Konstantin]

> -----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);
Konstantin

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

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

[Ferruh Yigit]

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
>>>
> 

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

[Koujalagi, MalleshX]

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
>>>
> 

[PATCH v2] net/null: support bulk allocation

[Mallesh Koujalagi]

Bulk allocation of multiple mbufs increased more than ~2%  and less
than 8% throughput on single core (1.8 GHz), based on usage for example
1: Testpmd case: Two null devices with copy 8% improvement.
    testpmd -c 0x3 -n 4 --socket-mem 1024,1024
	--vdev 'eth_null0,size=64,copy=1' --vdev 'eth_null1,size=64,copy=1'
	-- -i -a --coremask=0x2 --txrst=64 --txfreet=64 --txd=256
	--rxd=256 --rxfreet=64 --burst=64 --txpt=64 --txq=1 --rxq=1 --numa
2. Ovs switch case: 2% improvement.
$VSCTL add-port ovs-br dpdk1 -- set Interface dpdk1 type=dpdk \
        options:dpdk-devargs=eth_null0,size=64,copy=1
$VSCTL add-port ovs-br dpdk2 -- set Interface dpdk2 type=dpdk \
        options:dpdk-devargs=eth_null1,size=64,copy=1

Signed-off-by: Mallesh Koujalagi <malleshx.koujalagi@intel.com>
---
 drivers/net/null/rte_eth_null.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/drivers/net/null/rte_eth_null.c b/drivers/net/null/rte_eth_null.c
index 9385ffd..c019d2d 100644
--- a/drivers/net/null/rte_eth_null.c
+++ b/drivers/net/null/rte_eth_null.c
@@ -105,10 +105,10 @@ eth_null_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;
 		bufs[i]->data_len = (uint16_t)packet_size;
 		bufs[i]->pkt_len = packet_size;
 		bufs[i]->port = h->internals->port_id;
@@ -130,10 +130,10 @@ 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;
-- 
2.7.4

Re: [PATCH v2] net/null: support bulk allocation

[Ferruh Yigit]

On 3/8/2018 11:40 PM, Mallesh Koujalagi wrote:
> Bulk allocation of multiple mbufs increased more than ~2%  and less
> than 8% throughput on single core (1.8 GHz), based on usage for example
> 1: Testpmd case: Two null devices with copy 8% improvement.
>     testpmd -c 0x3 -n 4 --socket-mem 1024,1024
> 	--vdev 'eth_null0,size=64,copy=1' --vdev 'eth_null1,size=64,copy=1'
> 	-- -i -a --coremask=0x2 --txrst=64 --txfreet=64 --txd=256
> 	--rxd=256 --rxfreet=64 --burst=64 --txpt=64 --txq=1 --rxq=1 --numa
> 2. Ovs switch case: 2% improvement.
> $VSCTL add-port ovs-br dpdk1 -- set Interface dpdk1 type=dpdk \
>         options:dpdk-devargs=eth_null0,size=64,copy=1
> $VSCTL add-port ovs-br dpdk2 -- set Interface dpdk2 type=dpdk \
>         options:dpdk-devargs=eth_null1,size=64,copy=1
> 
> Signed-off-by: Mallesh Koujalagi <malleshx.koujalagi@intel.com>

Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>

Re: [PATCH v2] net/null: support bulk allocation

[Ferruh Yigit]

On 3/9/2018 11:09 AM, Ferruh Yigit wrote:
> On 3/8/2018 11:40 PM, Mallesh Koujalagi wrote:
>> Bulk allocation of multiple mbufs increased more than ~2%  and less
>> than 8% throughput on single core (1.8 GHz), based on usage for example
>> 1: Testpmd case: Two null devices with copy 8% improvement.
>>     testpmd -c 0x3 -n 4 --socket-mem 1024,1024
>> 	--vdev 'eth_null0,size=64,copy=1' --vdev 'eth_null1,size=64,copy=1'
>> 	-- -i -a --coremask=0x2 --txrst=64 --txfreet=64 --txd=256
>> 	--rxd=256 --rxfreet=64 --burst=64 --txpt=64 --txq=1 --rxq=1 --numa
>> 2. Ovs switch case: 2% improvement.
>> $VSCTL add-port ovs-br dpdk1 -- set Interface dpdk1 type=dpdk \
>>         options:dpdk-devargs=eth_null0,size=64,copy=1
>> $VSCTL add-port ovs-br dpdk2 -- set Interface dpdk2 type=dpdk \
>>         options:dpdk-devargs=eth_null1,size=64,copy=1
>>
>> Signed-off-by: Mallesh Koujalagi <malleshx.koujalagi@intel.com>
> 
> Reviewed-by: Ferruh Yigit <ferruh.yigit@intel.com>

Applied to dpdk-next-net/master, thanks.