[dpdk-dev,v2] event/sw: perform partial burst enqueues

Message ID 1520866522-13942-1-git-send-email-gage.eads@intel.com (mailing list archive)
State Accepted, archived
Delegated to: Jerin Jacob
Headers

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/Intel-compilation success Compilation OK

Commit Message

Eads, Gage March 12, 2018, 2:55 p.m. UTC
  Previously, the sw PMD would enqueue either all or no events, depending on
if enough inflight credits were available for the new events in the burst.
If a port is enqueueing a large burst (i.e. a multiple of the credit update
quanta), this can result in suboptimal performance, and requires an
understanding of the sw PMD implementation (in particular, its credit
scheme) to tune an application's burst size.

This affects software that enqueues large bursts of new events, such as the
ethernet event adapter which uses a 128-deep event buffer, when the input
packet rate is sufficiently high.

This change makes the sw PMD enqueue as many events as it has credits, if
there are any new events in the burst.

Signed-off-by: Gage Eads <gage.eads@intel.com>
---
v2: fix commit message bug ("sw event adapter" -> "sw PMD")

 drivers/event/sw/sw_evdev_worker.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
  

Comments

Van Haaren, Harry March 13, 2018, 2:08 p.m. UTC | #1
> From: Eads, Gage
> Sent: Monday, March 12, 2018 2:55 PM
> To: dev@dpdk.org
> Cc: jerin.jacob@caviumnetworks.com; Van Haaren, Harry
> <harry.van.haaren@intel.com>
> Subject: [PATCH v2] event/sw: perform partial burst enqueues
> 
> Previously, the sw PMD would enqueue either all or no events, depending on
> if enough inflight credits were available for the new events in the burst.
> If a port is enqueueing a large burst (i.e. a multiple of the credit update
> quanta), this can result in suboptimal performance, and requires an
> understanding of the sw PMD implementation (in particular, its credit
> scheme) to tune an application's burst size.
> 
> This affects software that enqueues large bursts of new events, such as the
> ethernet event adapter which uses a 128-deep event buffer, when the input
> packet rate is sufficiently high.
> 
> This change makes the sw PMD enqueue as many events as it has credits, if
> there are any new events in the burst.
> 
> Signed-off-by: Gage Eads <gage.eads@intel.com>


Agree that making forward progress by enqueuing what we have credits for
makes sense for the application.

No performance degradation here, thanks for the patch:

Acked-by: Harry van Haaren <harry.van.haaren@intel.com>
  
Jerin Jacob March 20, 2018, 7:01 a.m. UTC | #2
-----Original Message-----
> Date: Tue, 13 Mar 2018 14:08:20 +0000
> From: "Van Haaren, Harry" <harry.van.haaren@intel.com>
> To: "Eads, Gage" <gage.eads@intel.com>, "dev@dpdk.org" <dev@dpdk.org>
> CC: "jerin.jacob@caviumnetworks.com" <jerin.jacob@caviumnetworks.com>
> Subject: RE: [PATCH v2] event/sw: perform partial burst enqueues
> 
> > From: Eads, Gage
> > Sent: Monday, March 12, 2018 2:55 PM
> > To: dev@dpdk.org
> > Cc: jerin.jacob@caviumnetworks.com; Van Haaren, Harry
> > <harry.van.haaren@intel.com>
> > Subject: [PATCH v2] event/sw: perform partial burst enqueues
> > 
> > Previously, the sw PMD would enqueue either all or no events, depending on
> > if enough inflight credits were available for the new events in the burst.
> > If a port is enqueueing a large burst (i.e. a multiple of the credit update
> > quanta), this can result in suboptimal performance, and requires an
> > understanding of the sw PMD implementation (in particular, its credit
> > scheme) to tune an application's burst size.
> > 
> > This affects software that enqueues large bursts of new events, such as the
> > ethernet event adapter which uses a 128-deep event buffer, when the input
> > packet rate is sufficiently high.
> > 
> > This change makes the sw PMD enqueue as many events as it has credits, if
> > there are any new events in the burst.
> > 
> > Signed-off-by: Gage Eads <gage.eads@intel.com>
> 
> 
> Agree that making forward progress by enqueuing what we have credits for
> makes sense for the application.
> 
> No performance degradation here, thanks for the patch:
> 
> Acked-by: Harry van Haaren <harry.van.haaren@intel.com>

Applied to dpdk-next-eventdev/master. Thanks.
  

Patch

diff --git a/drivers/event/sw/sw_evdev_worker.c b/drivers/event/sw/sw_evdev_worker.c
index 67151f7..063b919 100644
--- a/drivers/event/sw/sw_evdev_worker.c
+++ b/drivers/event/sw/sw_evdev_worker.c
@@ -77,8 +77,10 @@  sw_event_enqueue_burst(void *port, const struct rte_event ev[], uint16_t num)
 		rte_atomic32_add(&sw->inflights, credit_update_quanta);
 		p->inflight_credits += (credit_update_quanta);
 
-		if (p->inflight_credits < new)
-			return 0;
+		/* If there are fewer inflight credits than new events, limit
+		 * the number of enqueued events.
+		 */
+		num = (p->inflight_credits < new) ? p->inflight_credits : new;
 	}
 
 	for (i = 0; i < num; i++) {