app/test-crypto-perf: improve dequeue logic

Message ID 20190327113823.13481-1-akhil.goyal@nxp.com (mailing list archive)
State Superseded, archived
Delegated to: akhil goyal
Headers
Series app/test-crypto-perf: improve dequeue logic |

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/mellanox-Performance-Testing fail Performance Testing issues
ci/intel-Performance-Testing success Performance Testing PASS
ci/Intel-compilation success Compilation OK

Commit Message

Akhil Goyal March 27, 2019, 11:47 a.m. UTC
  In case of hardware PMDs (which may take a longer duration
for processing the packets), there may be a case when the
number of enqueued packets are more than the dequeued.

So if the difference is more than 8 times the burst size,
more dequeue operations should be performed otherwise all
the buffers will be blocked in hardware.

Signed-off-by: Akhil Goyal <akhil.goyal@nxp.com>
---
 app/test-crypto-perf/cperf_test_throughput.c | 40 +++++++++++---------
 1 file changed, 23 insertions(+), 17 deletions(-)
  

Comments

Fan Zhang March 27, 2019, 3:20 p.m. UTC | #1
Hi Akhil,

This patch seems introduced performance degradation on my server with
QAT DH895XCC by more than 10% for big packets (>=256).

Fiona/Pablo, could you have a try with it?

Regards,
Fan


> -----Original Message-----
> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Akhil Goyal
> Sent: Wednesday, March 27, 2019 11:48 AM
> To: dev@dpdk.org
> Cc: Doherty, Declan <declan.doherty@intel.com>; Akhil Goyal
> <akhil.goyal@nxp.com>
> Subject: [dpdk-dev] [PATCH] app/test-crypto-perf: improve dequeue logic
> 
> In case of hardware PMDs (which may take a longer duration for processing
> the packets), there may be a case when the number of enqueued packets
> are more than the dequeued.
> 
> So if the difference is more than 8 times the burst size, more dequeue
> operations should be performed otherwise all the buffers will be blocked in
> hardware.
> 
> Signed-off-by: Akhil Goyal <akhil.goyal@nxp.com>
> ---
  
Kevin Traynor March 27, 2019, 5:31 p.m. UTC | #2
On 27/03/2019 11:47, Akhil Goyal wrote:
> In case of hardware PMDs (which may take a longer duration
> for processing the packets), there may be a case when the
> number of enqueued packets are more than the dequeued.
> 
> So if the difference is more than 8 times the burst size,
> more dequeue operations should be performed otherwise all
> the buffers will be blocked in hardware.
> 

Please add Fixes: tag and stable tag if appropriate.

> Signed-off-by: Akhil Goyal <akhil.goyal@nxp.com>
  

Patch

diff --git a/app/test-crypto-perf/cperf_test_throughput.c b/app/test-crypto-perf/cperf_test_throughput.c
index 2767f4ea8..d24a6dda0 100644
--- a/app/test-crypto-perf/cperf_test_throughput.c
+++ b/app/test-crypto-perf/cperf_test_throughput.c
@@ -208,23 +208,29 @@  cperf_throughput_test_runner(void *test_ctx)
 
 
 			/* Dequeue processed burst of ops from crypto device */
-			ops_deqd = rte_cryptodev_dequeue_burst(ctx->dev_id, ctx->qp_id,
-					ops_processed, test_burst_size);
-
-			if (likely(ops_deqd))  {
-				/* Free crypto ops so they can be reused. */
-				rte_mempool_put_bulk(ctx->pool,
-						(void **)ops_processed, ops_deqd);
-
-				ops_deqd_total += ops_deqd;
-			} else {
-				/**
-				 * Count dequeue polls which didn't return any
-				 * processed operations. This statistic is mainly
-				 * relevant to hw accelerators.
-				 */
-				ops_deqd_failed++;
-			}
+			do {
+				ops_deqd = rte_cryptodev_dequeue_burst(
+						ctx->dev_id, ctx->qp_id,
+						ops_processed, test_burst_size);
+
+				if (likely(ops_deqd))  {
+					/* Free crypto ops for reuse */
+					rte_mempool_put_bulk(ctx->pool,
+							(void **)ops_processed,
+							ops_deqd);
+
+					ops_deqd_total += ops_deqd;
+				} else {
+					/**
+					 * Count dequeue polls which didn't
+					 * return any processed operations.
+					 * This statistic is mainly relevant
+					 * to hw accelerators.
+					 */
+					ops_deqd_failed++;
+				}
+			} while (ops_enqd_total - ops_deqd_total >
+					test_burst_size * 8);
 
 		}