[PATCH] event/eth_tx: prefetch mbuf headers

[PATCH] event/eth_tx: prefetch mbuf headers

[Mattias Rönnblom]

Prefetch mbuf headers, resulting in ~10% throughput improvement when
the Ethernet RX and TX Adapters are hosted on the same core (likely
~2x in case a dedicated TX core is used).

Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
Tested-by: Peter Nilsson <peter.j.nilsson@ericsson.com>
---
 lib/eventdev/rte_event_eth_tx_adapter.c | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

diff --git a/lib/eventdev/rte_event_eth_tx_adapter.c b/lib/eventdev/rte_event_eth_tx_adapter.c
index 67fff8b7d6..d740ae00f9 100644
--- a/lib/eventdev/rte_event_eth_tx_adapter.c
+++ b/lib/eventdev/rte_event_eth_tx_adapter.c
@@ -598,6 +598,12 @@ txa_process_event_vector(struct txa_service_data *txa,
 	return nb_tx;
 }
 
+static inline void
+txa_prefetch_mbuf(struct rte_mbuf *mbuf)
+{
+	rte_mbuf_prefetch_part1(mbuf);
+}
+
 static void
 txa_service_tx(struct txa_service_data *txa, struct rte_event *ev,
 	uint32_t n)
@@ -608,6 +614,20 @@ txa_service_tx(struct txa_service_data *txa, struct rte_event *ev,
 
 	stats = &txa->stats;
 
+	for (i = 0; i < n; i++) {
+		struct rte_event *event = &ev[i];
+
+		if (unlikely(event->event_type & RTE_EVENT_TYPE_VECTOR)) {
+			struct rte_event_vector *vec = event->vec;
+			struct rte_mbuf **mbufs = vec->mbufs;
+			uint32_t k;
+
+			for (k = 0; k < vec->nb_elem; k++)
+				txa_prefetch_mbuf(mbufs[k]);
+		} else
+			txa_prefetch_mbuf(event->mbuf);
+	}
+
 	nb_tx = 0;
 	for (i = 0; i < n; i++) {
 		uint16_t port;
-- 
2.43.0

Re: [PATCH] event/eth_tx: prefetch mbuf headers

[Mattias Rönnblom]

On 2025-03-28 06:43, Mattias Rönnblom wrote:
> Prefetch mbuf headers, resulting in ~10% throughput improvement when
> the Ethernet RX and TX Adapters are hosted on the same core (likely
> ~2x in case a dedicated TX core is used).
> 
> Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
> Tested-by: Peter Nilsson <peter.j.nilsson@ericsson.com>

What should be added is that what's been tested is the 
non-RTE_EVENT_TYPE_VECTOR case.

> ---
>   lib/eventdev/rte_event_eth_tx_adapter.c | 20 ++++++++++++++++++++
>   1 file changed, 20 insertions(+)
> 
> diff --git a/lib/eventdev/rte_event_eth_tx_adapter.c b/lib/eventdev/rte_event_eth_tx_adapter.c
> index 67fff8b7d6..d740ae00f9 100644
> --- a/lib/eventdev/rte_event_eth_tx_adapter.c
> +++ b/lib/eventdev/rte_event_eth_tx_adapter.c
> @@ -598,6 +598,12 @@ txa_process_event_vector(struct txa_service_data *txa,
>   	return nb_tx;
>   }
>   
> +static inline void
> +txa_prefetch_mbuf(struct rte_mbuf *mbuf)
> +{
> +	rte_mbuf_prefetch_part1(mbuf);
> +}
> +
>   static void
>   txa_service_tx(struct txa_service_data *txa, struct rte_event *ev,
>   	uint32_t n)
> @@ -608,6 +614,20 @@ txa_service_tx(struct txa_service_data *txa, struct rte_event *ev,
>   
>   	stats = &txa->stats;
>   
> +	for (i = 0; i < n; i++) {
> +		struct rte_event *event = &ev[i];
> +
> +		if (unlikely(event->event_type & RTE_EVENT_TYPE_VECTOR)) {
> +			struct rte_event_vector *vec = event->vec;
> +			struct rte_mbuf **mbufs = vec->mbufs;
> +			uint32_t k;
> +
> +			for (k = 0; k < vec->nb_elem; k++)
> +				txa_prefetch_mbuf(mbufs[k]);
> +		} else
> +			txa_prefetch_mbuf(event->mbuf);
> +	}
> +
>   	nb_tx = 0;
>   	for (i = 0; i < n; i++) {
>   		uint16_t port;

Re: [PATCH] event/eth_tx: prefetch mbuf headers

[Mattias Rönnblom]

On 2025-03-28 07:07, Mattias Rönnblom wrote:
> On 2025-03-28 06:43, Mattias Rönnblom wrote:
>> Prefetch mbuf headers, resulting in ~10% throughput improvement when
>> the Ethernet RX and TX Adapters are hosted on the same core (likely
>> ~2x in case a dedicated TX core is used).
>>
>> Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
>> Tested-by: Peter Nilsson <peter.j.nilsson@ericsson.com>
> 

<snip>

Naga, could you comment on this patch?

RE: [EXTERNAL] [PATCH] event/eth_tx: prefetch mbuf headers

[Jerin Jacob]

[-- Attachment #1: Type: text/html, Size: 8086 bytes --]

RE: [PATCH] event/eth_tx: prefetch mbuf headers

[Naga Harish K, S V]

> -----Original Message-----
> From: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
> Sent: Friday, March 28, 2025 11:14 AM
> To: dev@dpdk.org
> Cc: Mattias Rönnblom <hofors@lysator.liu.se>; Naga Harish K, S V
> <s.v.naga.harish.k@intel.com>; Jerin Jacob <jerinj@marvell.com>; Mattias
> Rönnblom <mattias.ronnblom@ericsson.com>; Peter Nilsson
> <peter.j.nilsson@ericsson.com>
> Subject: [PATCH] event/eth_tx: prefetch mbuf headers
> 
> Prefetch mbuf headers, resulting in ~10% throughput improvement when the
> Ethernet RX and TX Adapters are hosted on the same core (likely ~2x in case a
> dedicated TX core is used).
> 
> Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
> Tested-by: Peter Nilsson <peter.j.nilsson@ericsson.com>
> ---
>  lib/eventdev/rte_event_eth_tx_adapter.c | 20 ++++++++++++++++++++
>  1 file changed, 20 insertions(+)
> 
> diff --git a/lib/eventdev/rte_event_eth_tx_adapter.c
> b/lib/eventdev/rte_event_eth_tx_adapter.c
> index 67fff8b7d6..d740ae00f9 100644
> --- a/lib/eventdev/rte_event_eth_tx_adapter.c
> +++ b/lib/eventdev/rte_event_eth_tx_adapter.c
> @@ -598,6 +598,12 @@ txa_process_event_vector(struct txa_service_data
> *txa,
>  	return nb_tx;
>  }
> 
> +static inline void
> +txa_prefetch_mbuf(struct rte_mbuf *mbuf) {
> +	rte_mbuf_prefetch_part1(mbuf);
> +}
> +
>  static void
>  txa_service_tx(struct txa_service_data *txa, struct rte_event *ev,
>  	uint32_t n)
> @@ -608,6 +614,20 @@ txa_service_tx(struct txa_service_data *txa, struct
> rte_event *ev,
> 
>  	stats = &txa->stats;
> 
> +	for (i = 0; i < n; i++) {
> +		struct rte_event *event = &ev[i];
> +
> +		if (unlikely(event->event_type & RTE_EVENT_TYPE_VECTOR))


This gives a branch prediction advantage to non-vector events. Is that the intention?

> {
> +			struct rte_event_vector *vec = event->vec;
> +			struct rte_mbuf **mbufs = vec->mbufs;
> +			uint32_t k;
> +
> +			for (k = 0; k < vec->nb_elem; k++)
> +				txa_prefetch_mbuf(mbufs[k]);
> +		} else
> +			txa_prefetch_mbuf(event->mbuf);
> +	}
> +
>  	nb_tx = 0;
>  	for (i = 0; i < n; i++) {
>  		uint16_t port;
> --
> 2.43.0

Re: [PATCH] event/eth_tx: prefetch mbuf headers

[Mattias Rönnblom]

On 2025-05-27 12:55, Naga Harish K, S V wrote:
> 
> 
>> -----Original Message-----
>> From: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
>> Sent: Friday, March 28, 2025 11:14 AM
>> To: dev@dpdk.org
>> Cc: Mattias Rönnblom <hofors@lysator.liu.se>; Naga Harish K, S V
>> <s.v.naga.harish.k@intel.com>; Jerin Jacob <jerinj@marvell.com>; Mattias
>> Rönnblom <mattias.ronnblom@ericsson.com>; Peter Nilsson
>> <peter.j.nilsson@ericsson.com>
>> Subject: [PATCH] event/eth_tx: prefetch mbuf headers
>>
>> Prefetch mbuf headers, resulting in ~10% throughput improvement when the
>> Ethernet RX and TX Adapters are hosted on the same core (likely ~2x in case a
>> dedicated TX core is used).
>>
>> Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
>> Tested-by: Peter Nilsson <peter.j.nilsson@ericsson.com>
>> ---
>>   lib/eventdev/rte_event_eth_tx_adapter.c | 20 ++++++++++++++++++++
>>   1 file changed, 20 insertions(+)
>>
>> diff --git a/lib/eventdev/rte_event_eth_tx_adapter.c
>> b/lib/eventdev/rte_event_eth_tx_adapter.c
>> index 67fff8b7d6..d740ae00f9 100644
>> --- a/lib/eventdev/rte_event_eth_tx_adapter.c
>> +++ b/lib/eventdev/rte_event_eth_tx_adapter.c
>> @@ -598,6 +598,12 @@ txa_process_event_vector(struct txa_service_data
>> *txa,
>>   	return nb_tx;
>>   }
>>
>> +static inline void
>> +txa_prefetch_mbuf(struct rte_mbuf *mbuf) {
>> +	rte_mbuf_prefetch_part1(mbuf);
>> +}
>> +
>>   static void
>>   txa_service_tx(struct txa_service_data *txa, struct rte_event *ev,
>>   	uint32_t n)
>> @@ -608,6 +614,20 @@ txa_service_tx(struct txa_service_data *txa, struct
>> rte_event *ev,
>>
>>   	stats = &txa->stats;
>>
>> +	for (i = 0; i < n; i++) {
>> +		struct rte_event *event = &ev[i];
>> +
>> +		if (unlikely(event->event_type & RTE_EVENT_TYPE_VECTOR))
> 
> 
> This gives a branch prediction advantage to non-vector events. Is that the intention?
> 

Yes.

>> {
>> +			struct rte_event_vector *vec = event->vec;
>> +			struct rte_mbuf **mbufs = vec->mbufs;
>> +			uint32_t k;
>> +
>> +			for (k = 0; k < vec->nb_elem; k++)
>> +				txa_prefetch_mbuf(mbufs[k]);
>> +		} else
>> +			txa_prefetch_mbuf(event->mbuf);
>> +	}
>> +
>>   	nb_tx = 0;
>>   	for (i = 0; i < n; i++) {
>>   		uint16_t port;
>> --
>> 2.43.0
> 

RE: [PATCH] event/eth_tx: prefetch mbuf headers

[Naga Harish K, S V]

> -----Original Message-----
> From: Mattias Rönnblom <hofors@lysator.liu.se>
> Sent: Thursday, July 3, 2025 1:50 AM
> To: Naga Harish K, S V <s.v.naga.harish.k@intel.com>; Mattias Rönnblom
> <mattias.ronnblom@ericsson.com>; dev@dpdk.org
> Cc: Jerin Jacob <jerinj@marvell.com>; Peter Nilsson
> <peter.j.nilsson@ericsson.com>
> Subject: Re: [PATCH] event/eth_tx: prefetch mbuf headers
> 
> On 2025-05-27 12:55, Naga Harish K, S V wrote:
> >
> >
> >> -----Original Message-----
> >> From: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
> >> Sent: Friday, March 28, 2025 11:14 AM
> >> To: dev@dpdk.org
> >> Cc: Mattias Rönnblom <hofors@lysator.liu.se>; Naga Harish K, S V
> >> <s.v.naga.harish.k@intel.com>; Jerin Jacob <jerinj@marvell.com>;
> >> Mattias Rönnblom <mattias.ronnblom@ericsson.com>; Peter Nilsson
> >> <peter.j.nilsson@ericsson.com>
> >> Subject: [PATCH] event/eth_tx: prefetch mbuf headers
> >>
> >> Prefetch mbuf headers, resulting in ~10% throughput improvement when
> >> the Ethernet RX and TX Adapters are hosted on the same core (likely
> >> ~2x in case a dedicated TX core is used).
> >>
> >> Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
> >> Tested-by: Peter Nilsson <peter.j.nilsson@ericsson.com>
> >> ---
> >>   lib/eventdev/rte_event_eth_tx_adapter.c | 20 ++++++++++++++++++++
> >>   1 file changed, 20 insertions(+)
> >>
> >> diff --git a/lib/eventdev/rte_event_eth_tx_adapter.c
> >> b/lib/eventdev/rte_event_eth_tx_adapter.c
> >> index 67fff8b7d6..d740ae00f9 100644
> >> --- a/lib/eventdev/rte_event_eth_tx_adapter.c
> >> +++ b/lib/eventdev/rte_event_eth_tx_adapter.c
> >> @@ -598,6 +598,12 @@ txa_process_event_vector(struct
> txa_service_data
> >> *txa,
> >>   	return nb_tx;
> >>   }
> >>
> >> +static inline void
> >> +txa_prefetch_mbuf(struct rte_mbuf *mbuf) {
> >> +	rte_mbuf_prefetch_part1(mbuf);
> >> +}
> >> +
> >>   static void
> >>   txa_service_tx(struct txa_service_data *txa, struct rte_event *ev,
> >>   	uint32_t n)
> >> @@ -608,6 +614,20 @@ txa_service_tx(struct txa_service_data *txa,
> >> struct rte_event *ev,
> >>
> >>   	stats = &txa->stats;
> >>
> >> +	for (i = 0; i < n; i++) {
> >> +		struct rte_event *event = &ev[i];
> >> +
> >> +		if (unlikely(event->event_type & RTE_EVENT_TYPE_VECTOR))
> >
> >
> > This gives a branch prediction advantage to non-vector events. Is that the
> intention?
> >
> 
> Yes.

I think all event-types need to be equally weighted. My ask was to remove the "unlikely" for vector events.

> 
> >> {
> >> +			struct rte_event_vector *vec = event->vec;
> >> +			struct rte_mbuf **mbufs = vec->mbufs;
> >> +			uint32_t k;
> >> +
> >> +			for (k = 0; k < vec->nb_elem; k++)
> >> +				txa_prefetch_mbuf(mbufs[k]);
> >> +		} else
> >> +			txa_prefetch_mbuf(event->mbuf);
> >> +	}
> >> +
> >>   	nb_tx = 0;
> >>   	for (i = 0; i < n; i++) {
> >>   		uint16_t port;
> >> --
> >> 2.43.0
> >

Re: [PATCH] event/eth_tx: prefetch mbuf headers

[Mattias Rönnblom]

On 2025-07-07 11:00, Naga Harish K, S V wrote:
> 
> 
>> -----Original Message-----
>> From: Mattias Rönnblom <hofors@lysator.liu.se>
>> Sent: Thursday, July 3, 2025 1:50 AM
>> To: Naga Harish K, S V <s.v.naga.harish.k@intel.com>; Mattias Rönnblom
>> <mattias.ronnblom@ericsson.com>; dev@dpdk.org
>> Cc: Jerin Jacob <jerinj@marvell.com>; Peter Nilsson
>> <peter.j.nilsson@ericsson.com>
>> Subject: Re: [PATCH] event/eth_tx: prefetch mbuf headers
>>
>> On 2025-05-27 12:55, Naga Harish K, S V wrote:
>>>
>>>
>>>> -----Original Message-----
>>>> From: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
>>>> Sent: Friday, March 28, 2025 11:14 AM
>>>> To: dev@dpdk.org
>>>> Cc: Mattias Rönnblom <hofors@lysator.liu.se>; Naga Harish K, S V
>>>> <s.v.naga.harish.k@intel.com>; Jerin Jacob <jerinj@marvell.com>;
>>>> Mattias Rönnblom <mattias.ronnblom@ericsson.com>; Peter Nilsson
>>>> <peter.j.nilsson@ericsson.com>
>>>> Subject: [PATCH] event/eth_tx: prefetch mbuf headers
>>>>
>>>> Prefetch mbuf headers, resulting in ~10% throughput improvement when
>>>> the Ethernet RX and TX Adapters are hosted on the same core (likely
>>>> ~2x in case a dedicated TX core is used).
>>>>
>>>> Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
>>>> Tested-by: Peter Nilsson <peter.j.nilsson@ericsson.com>
>>>> ---
>>>>    lib/eventdev/rte_event_eth_tx_adapter.c | 20 ++++++++++++++++++++
>>>>    1 file changed, 20 insertions(+)
>>>>
>>>> diff --git a/lib/eventdev/rte_event_eth_tx_adapter.c
>>>> b/lib/eventdev/rte_event_eth_tx_adapter.c
>>>> index 67fff8b7d6..d740ae00f9 100644
>>>> --- a/lib/eventdev/rte_event_eth_tx_adapter.c
>>>> +++ b/lib/eventdev/rte_event_eth_tx_adapter.c
>>>> @@ -598,6 +598,12 @@ txa_process_event_vector(struct
>> txa_service_data
>>>> *txa,
>>>>    	return nb_tx;
>>>>    }
>>>>
>>>> +static inline void
>>>> +txa_prefetch_mbuf(struct rte_mbuf *mbuf) {
>>>> +	rte_mbuf_prefetch_part1(mbuf);
>>>> +}
>>>> +
>>>>    static void
>>>>    txa_service_tx(struct txa_service_data *txa, struct rte_event *ev,
>>>>    	uint32_t n)
>>>> @@ -608,6 +614,20 @@ txa_service_tx(struct txa_service_data *txa,
>>>> struct rte_event *ev,
>>>>
>>>>    	stats = &txa->stats;
>>>>
>>>> +	for (i = 0; i < n; i++) {
>>>> +		struct rte_event *event = &ev[i];
>>>> +
>>>> +		if (unlikely(event->event_type & RTE_EVENT_TYPE_VECTOR))
>>>
>>>
>>> This gives a branch prediction advantage to non-vector events. Is that the
>> intention?
>>>
>>
>> Yes.
> 
> I think all event-types need to be equally weighted. My ask was to remove the "unlikely" for vector events.
> 

This is not possible. One branch will always be cheaper. If you leave 
out unlikely()/likely(), you leave all control to compiler heuristics. 
In this case, I think the resulting object code will be identical (on GCC).

RTE_EVENT_TYPE_VECTOR will result in fewer events, and thus the 
per-event overhead is less of an issue. So if you weigh the importance 
of vector and non-vector use cases equally, you should optimize for the 
non-vector case.

>>
>>>> {
>>>> +			struct rte_event_vector *vec = event->vec;
>>>> +			struct rte_mbuf **mbufs = vec->mbufs;
>>>> +			uint32_t k;
>>>> +
>>>> +			for (k = 0; k < vec->nb_elem; k++)
>>>> +				txa_prefetch_mbuf(mbufs[k]);
>>>> +		} else
>>>> +			txa_prefetch_mbuf(event->mbuf);
>>>> +	}
>>>> +
>>>>    	nb_tx = 0;
>>>>    	for (i = 0; i < n; i++) {
>>>>    		uint16_t port;
>>>> --
>>>> 2.43.0
>>>
> 

RE: [PATCH] event/eth_tx: prefetch mbuf headers

[Naga Harish K, S V]

> -----Original Message-----
> From: Mattias Rönnblom <hofors@lysator.liu.se>
> Sent: Monday, July 7, 2025 5:27 PM
> To: Naga Harish K, S V <s.v.naga.harish.k@intel.com>; Mattias Rönnblom
> <mattias.ronnblom@ericsson.com>; dev@dpdk.org
> Cc: Jerin Jacob <jerinj@marvell.com>; Peter Nilsson
> <peter.j.nilsson@ericsson.com>
> Subject: Re: [PATCH] event/eth_tx: prefetch mbuf headers
> 
> On 2025-07-07 11:00, Naga Harish K, S V wrote:
> >
> >
> >> -----Original Message-----
> >> From: Mattias Rönnblom <hofors@lysator.liu.se>
> >> Sent: Thursday, July 3, 2025 1:50 AM
> >> To: Naga Harish K, S V <s.v.naga.harish.k@intel.com>; Mattias
> >> Rönnblom <mattias.ronnblom@ericsson.com>; dev@dpdk.org
> >> Cc: Jerin Jacob <jerinj@marvell.com>; Peter Nilsson
> >> <peter.j.nilsson@ericsson.com>
> >> Subject: Re: [PATCH] event/eth_tx: prefetch mbuf headers
> >>
> >> On 2025-05-27 12:55, Naga Harish K, S V wrote:
> >>>
> >>>
> >>>> -----Original Message-----
> >>>> From: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
> >>>> Sent: Friday, March 28, 2025 11:14 AM
> >>>> To: dev@dpdk.org
> >>>> Cc: Mattias Rönnblom <hofors@lysator.liu.se>; Naga Harish K, S V
> >>>> <s.v.naga.harish.k@intel.com>; Jerin Jacob <jerinj@marvell.com>;
> >>>> Mattias Rönnblom <mattias.ronnblom@ericsson.com>; Peter Nilsson
> >>>> <peter.j.nilsson@ericsson.com>
> >>>> Subject: [PATCH] event/eth_tx: prefetch mbuf headers
> >>>>
> >>>> Prefetch mbuf headers, resulting in ~10% throughput improvement
> >>>> when the Ethernet RX and TX Adapters are hosted on the same core
> >>>> (likely ~2x in case a dedicated TX core is used).
> >>>>
> >>>> Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
> >>>> Tested-by: Peter Nilsson <peter.j.nilsson@ericsson.com>

Acked-by: Naga Harish K S V <s.v.naga.harish.k@intel.com>

> >>>> ---
> >>>>    lib/eventdev/rte_event_eth_tx_adapter.c | 20
> ++++++++++++++++++++
> >>>>    1 file changed, 20 insertions(+)
> >>>>
> >>>> diff --git a/lib/eventdev/rte_event_eth_tx_adapter.c
> >>>> b/lib/eventdev/rte_event_eth_tx_adapter.c
> >>>> index 67fff8b7d6..d740ae00f9 100644
> >>>> --- a/lib/eventdev/rte_event_eth_tx_adapter.c
> >>>> +++ b/lib/eventdev/rte_event_eth_tx_adapter.c
> >>>> @@ -598,6 +598,12 @@ txa_process_event_vector(struct
> >> txa_service_data
> >>>> *txa,
> >>>>    	return nb_tx;
> >>>>    }
> >>>>
> >>>> +static inline void
> >>>> +txa_prefetch_mbuf(struct rte_mbuf *mbuf) {
> >>>> +	rte_mbuf_prefetch_part1(mbuf);
> >>>> +}
> >>>> +
> >>>>    static void
> >>>>    txa_service_tx(struct txa_service_data *txa, struct rte_event *ev,
> >>>>    	uint32_t n)
> >>>> @@ -608,6 +614,20 @@ txa_service_tx(struct txa_service_data *txa,
> >>>> struct rte_event *ev,
> >>>>
> >>>>    	stats = &txa->stats;
> >>>>
> >>>> +	for (i = 0; i < n; i++) {
> >>>> +		struct rte_event *event = &ev[i];
> >>>> +
> >>>> +		if (unlikely(event->event_type & RTE_EVENT_TYPE_VECTOR))
> >>>
> >>>
> >>> This gives a branch prediction advantage to non-vector events. Is
> >>> that the
> >> intention?
> >>>
> >>
> >> Yes.
> >
> > I think all event-types need to be equally weighted. My ask was to remove
> the "unlikely" for vector events.
> >
> 
> This is not possible. One branch will always be cheaper. If you leave out
> unlikely()/likely(), you leave all control to compiler heuristics.
> In this case, I think the resulting object code will be identical (on GCC).
> 
> RTE_EVENT_TYPE_VECTOR will result in fewer events, and thus the per-event
> overhead is less of an issue. So if you weigh the importance of vector and non-
> vector use cases equally, you should optimize for the non-vector case.
> 

Fine, agreed.

> >>
> >>>> {
> >>>> +			struct rte_event_vector *vec = event->vec;
> >>>> +			struct rte_mbuf **mbufs = vec->mbufs;
> >>>> +			uint32_t k;
> >>>> +
> >>>> +			for (k = 0; k < vec->nb_elem; k++)
> >>>> +				txa_prefetch_mbuf(mbufs[k]);
> >>>> +		} else
> >>>> +			txa_prefetch_mbuf(event->mbuf);
> >>>> +	}
> >>>> +
> >>>>    	nb_tx = 0;
> >>>>    	for (i = 0; i < n; i++) {
> >>>>    		uint16_t port;
> >>>> --
> >>>> 2.43.0
> >>>
> >

Re: [PATCH] event/eth_tx: prefetch mbuf headers

[Stephen Hemminger]

On Fri, 28 Mar 2025 06:43:39 +0100
Mattias Rönnblom <mattias.ronnblom@ericsson.com> wrote:

> Prefetch mbuf headers, resulting in ~10% throughput improvement when
> the Ethernet RX and TX Adapters are hosted on the same core (likely
> ~2x in case a dedicated TX core is used).
> 
> Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
> Tested-by: Peter Nilsson <peter.j.nilsson@ericsson.com>

Prefetching all the mbufs can be counter productive on a big burst.

VPP does something similar but more unrolled.
See https://fd.io/docs/vpp/v2101/gettingstarted/developers/vnet.html#single-dual-loops

Re: [PATCH] event/eth_tx: prefetch mbuf headers

[Mattias Rönnblom]

On 2025-07-10 17:37, Stephen Hemminger wrote:
> On Fri, 28 Mar 2025 06:43:39 +0100
> Mattias Rönnblom <mattias.ronnblom@ericsson.com> wrote:
> 
>> Prefetch mbuf headers, resulting in ~10% throughput improvement when
>> the Ethernet RX and TX Adapters are hosted on the same core (likely
>> ~2x in case a dedicated TX core is used).
>>
>> Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
>> Tested-by: Peter Nilsson <peter.j.nilsson@ericsson.com>
> 
> Prefetching all the mbufs can be counter productive on a big burst.
> 

For the non-vector case, the burst is no larger than 32. From what's 
available in terms of public information, the number of load queue 
entries is 72 on Skylake. What it is on newer micro architecture 
generations, I don't know. So 32 is a lot of prefetches, but at least 
likely smaller than the load queue.

> VPP does something similar but more unrolled.
> See https://fd.io/docs/vpp/v2101/gettingstarted/developers/vnet.html#single-dual-loops

This pattern makes sense, if the do_something_to() function has 
non-trivial latency.

If it doesn't, which I suspect is the case in the TX adapter case, you 
will issue 4 prefetches, of which some or even all aren't resolved 
before the core need to data. Repeat.

Also - and I'm guessing now - the do_something_to() equivalent in the TX 
adapter case is likely not allocating a lot of load buffer entries, so 
little risk of the prefetches being discarded.

That said, I'm sure you can tweak non-vector TXA prefetching to further 
improve performance. For example, it may be little point in prefetching 
the first few mbuf headers, since you will need that data very soon indeed.

I no longer have the setup to further refine this patch. I suggest we 
live with only ~20% performance gain at this point.

For the vector case, I agree this loop may result in too many prefetches.

I can remove prefetching from the vector case, to maintain legacy 
performance. I could also cap the number of prefetches (e.g., to 32).