Skip to contentSkip to conversation

test/mcslock: move performance test to perf tests

PATCH Superseded Phil Yang 4 participants 10 messages Mar 23, 2020 – Jun 13, 2023 Arm/Amlogic Meson SoC support

lore
Share & tools
lore

Phil Yang

author
#0 lore raw

PATCH 1/2 test/mcslock: move performance test to perf tests

The MCS lock performance test takes more than 10 seconds and leads to meson test timeout on some platforms. Move the performance test into perf tests.

---
 MAINTAINERS                  |   1 +
 app/test/Makefile            |   1 +
 app/test/autotest_data.py    |   6 +++
 app/test/meson.build         |   2 +
 app/test/test_mcslock.c      |  88 -------------------------------
 app/test/test_mcslock_perf.c | 121 +++++++++++++++++++++++++++++++++++++++++++
 6 files changed, 131 insertions(+), 88 deletions(-)
 create mode 100644 app/test/test_mcslock_perf.c
MAINTAINERS +1 -0
index db235c2..411bdeb 100644
@@ -247,6 +247,7 @@ MCSlock - EXPERIMENTAL
 M: Phil Yang <[email protected]>
 F: lib/librte_eal/common/include/generic/rte_mcslock.h
 F: app/test/test_mcslock.c
+F: app/test/test_mcslock_perf.c
 
 Ticketlock
 M: Joyce Kong <[email protected]>
app/test/Makefile +1 -0
index 1f080d1..97de3ac 100644
@@ -65,6 +65,7 @@ SRCS-y += test_barrier.c
 SRCS-y += test_malloc.c
 SRCS-y += test_cycles.c
 SRCS-y += test_mcslock.c
+SRCS-y += test_mcslock_perf.c
 SRCS-y += test_spinlock.c
 SRCS-y += test_ticketlock.c
 SRCS-y += test_memory.c
app/test/autotest_data.py +6 -0
index 7b1d013..2a4619d 100644
@@ -784,6 +784,12 @@
         "Func":    default_autotest,
         "Report":  None,
     },
+    {
+        "Name":    "MCS Lock performance autotest",
+        "Command": "mcslock_perf_autotest",
+        "Func":    default_autotest,
+        "Report":  None,
+    },
     #
     # Please always make sure that ring_perf is the last test!
     #
app/test/meson.build +2 -0
index 0a2ce71..335a869 100644
@@ -82,6 +82,7 @@ test_sources = files('commands.c',
 	'test_meter.c',
 	'test_metrics.c',
 	'test_mcslock.c',
+	'test_mcslock_perf.c',
 	'test_mp_secondary.c',
 	'test_per_lcore.c',
 	'test_pmd_perf.c',
@@ -270,6 +271,7 @@ perf_test_names = [
         'rand_perf_autotest',
         'hash_readwrite_perf_autotest',
         'hash_readwrite_lf_perf_autotest',
+	'mcslock_perf_autotest',
 ]
 
 driver_test_names = [
app/test/test_mcslock.c +0 -88

118 lines not on this page · raw message

app/test/test_mcslock_perf.c new file +121 -0

121 lines not on this page · raw message

Signature · 1 line
2.7.4
author
#1 lore raw

PATCH 2/2 test/atomic: reduce the number of loops to avoid timeouts

Reduce the number of loops to 10K to avoid the meson test timeout warning.

---
 app/test/test_atomic.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
app/test/test_atomic.c +1 -1
index de3030d..661fecc 100644
@@ -99,7 +99,7 @@
 
 #define NUM_ATOMIC_TYPES 3
 
-#define N 1000000
+#define N 10000
 
 static rte_atomic16_t a16;
 static rte_atomic32_t a32;
Signature · 1 line
2.7.4
author
#2 lore raw

PATCH v2 1/1 test/atomic: reduce the number of loops to avoid timeouts

Reduce the number of loops to 10K to avoid the meson test timeout warning.

For example:

$ sudo meson test -C build --suite DPDK:fast-tests / atomic_autotest -t 50
[...]
1/1 DPDK:fast-tests / atomic_autotest OK             187.35s

---
v2:
Split from the original patchset.

 app/test/test_atomic.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
app/test/test_atomic.c +1 -1
index 214452e..9b70f44 100644
@@ -99,7 +99,7 @@
 
 #define NUM_ATOMIC_TYPES 3
 
-#define N 1000000
+#define N 10000
 
 static rte_atomic16_t a16;
 static rte_atomic32_t a32;
Signature · 1 line
2.7.4
Reviewed-by on v2 1/1
#3 lore raw

Re: [dpdk-dev] [PATCH v2] test/atomic: reduce the number of loops to avoid timeouts

Quoted 39 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Phil Yang
Sent: Tuesday, July 28, 2020 5:42 AM
To: [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]
Subject: [dpdk-dev] [PATCH v2] test/atomic: reduce the number of loops to avoid timeouts

Reduce the number of loops to 10K to avoid the meson test timeout warning.

For example:

$ sudo meson test -C build --suite DPDK:fast-tests / atomic_autotest -t 50 [...]
1/1 DPDK:fast-tests / atomic_autotest OK             187.35s

---
v2:
Split from the original patchset.

 app/test/test_atomic.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
app/test/test_atomic.c index renamed from app/test/test_atomic.c +0 -0

214452e..9b70f44 100644

app/test/test_atomic.c +1 -2
@@ -99,7 +99,7 @@
 
 #define NUM_ATOMIC_TYPES 3
 
-#define N 1000000
+#define N 10000
 
 static rte_atomic16_t a16;
 static rte_atomic32_t a32;
--
Quoted 2 lines

2.7.4

#4 lore raw

Re: [dpdk-dev] [PATCH v2] test/atomic: reduce the number of loops to avoid timeouts

In reply to Phil Yang
Quoted 9 lines · On Tue, 28 Jul 2020 11:41:37 +0800 Phil Yang <[email protected]> wrote:

Reduce the number of loops to 10K to avoid the meson test timeout warning.

For example:

$ sudo meson test -C build --suite DPDK:fast-tests / atomic_autotest -t 50
[...]
1/1 DPDK:fast-tests / atomic_autotest OK             187.35s

I don't have Arm server to test on but this looks like a serious issue with Arm performance. FYI the default 1000000 takes 10 seconds on slow NUC (AMD x86). Is this a problem with the lack of real compare-exchange and the default atomic retry model for that?

This patch didn't get applied in 3 years but still worth investigating if anyone has resources.

author
#5 lore raw
In reply to Phil Yang

Hi Aaron,

It seems Travis CI cannot capture this timeout issue.
But the local test records these two cases as TIMEOUT under the default timeout configuration.

1. The test results on x86 server (72 lcores online@ 2.60GHz) are:
a. Default setting:

$ sudo meson test -C build_test --suite DPDK:fast-tests
   3/92 DPDK:fast-tests / atomic_autotest                TIMEOUT        10.32s
39/92 DPDK:fast-tests / mcslock_autotest               TIMEOUT        10.32s

b. Extend timeout:

$ sudo meson test -C build_test --suite DPDK:fast-tests -t 30
   3/92 DPDK:fast-tests / atomic_autotest                OK             173.01s
39/92 DPDK:fast-tests / mcslock_autotest               OK             12.19s

2. The test results on aarch64 server (56 lcore online @ 2.0GHz) are:
a. Default setting:

$ sudo meson test -C build_test --suite DPDK:fast-tests
   3/92 DPDK:fast-tests / atomic_autotest       TIMEOUT 10.47 s
39/92 DPDK:fast-tests / mcslock_autotest      TIMEOUT 10.47 s

b. Extend timeout:

$ sudo meson test -C build_test --suite DPDK:fast-tests -t 60
   3/92 DPDK:fast-tests / atomic_autotest       OK      431.89 s
39/92 DPDK:fast-tests / mcslock_autotest      OK      21.25 s

So I think these two patches can resolve this issue. I'd love to hear your thoughts.

Thanks,
Phil

<snip>

Quoted 336 lines

Subject: [dpdk-dev] [PATCH 1/2] test/mcslock: move performance test to perf tests

The MCS lock performance test takes more than 10 seconds and leads to meson test timeout on some platforms. Move the performance test into perf tests.

---
 MAINTAINERS                  |   1 +
 app/test/Makefile            |   1 +
 app/test/autotest_data.py    |   6 +++
 app/test/meson.build         |   2 +
 app/test/test_mcslock.c      |  88 -------------------------------
 app/test/test_mcslock_perf.c | 121
+++++++++++++++++++++++++++++++++++++++++++
 6 files changed, 131 insertions(+), 88 deletions(-)
 create mode 100644 app/test/test_mcslock_perf.c
MAINTAINERS +1 -0
index db235c2..411bdeb 100644
@@ -247,6 +247,7 @@ MCSlock - EXPERIMENTAL
 M: Phil Yang <[email protected]>
 F: lib/librte_eal/common/include/generic/rte_mcslock.h
 F: app/test/test_mcslock.c
+F: app/test/test_mcslock_perf.c
 
 Ticketlock
 M: Joyce Kong <[email protected]>
app/test/Makefile +1 -0
index 1f080d1..97de3ac 100644
@@ -65,6 +65,7 @@ SRCS-y += test_barrier.c
 SRCS-y += test_malloc.c
 SRCS-y += test_cycles.c
 SRCS-y += test_mcslock.c
+SRCS-y += test_mcslock_perf.c
 SRCS-y += test_spinlock.c
 SRCS-y += test_ticketlock.c
 SRCS-y += test_memory.c
app/test/autotest_data.py +6 -0
index 7b1d013..2a4619d 100644
@@ -784,6 +784,12 @@
         "Func":    default_autotest,
         "Report":  None,
     },
+    {
+        "Name":    "MCS Lock performance autotest",
+        "Command": "mcslock_perf_autotest",
+        "Func":    default_autotest,
+        "Report":  None,
+    },
     #
     # Please always make sure that ring_perf is the last test!
     #
app/test/meson.build +2 -0
index 0a2ce71..335a869 100644
@@ -82,6 +82,7 @@ test_sources = files('commands.c',
 	'test_meter.c',
 	'test_metrics.c',
 	'test_mcslock.c',
+	'test_mcslock_perf.c',
 	'test_mp_secondary.c',
 	'test_per_lcore.c',
 	'test_pmd_perf.c',
@@ -270,6 +271,7 @@ perf_test_names = [
         'rand_perf_autotest',
         'hash_readwrite_perf_autotest',
         'hash_readwrite_lf_perf_autotest',
+	'mcslock_perf_autotest',
 ]
 
 driver_test_names = [
app/test/test_mcslock.c +0 -84
index e9359df..15f9751 100644
@@ -32,23 +32,16 @@
  *
  *   - The function takes the global lock, display something, then releases
  *     the global lock on each core.
- *
- * - A load test is carried out, with all cores attempting to lock a single
- *   lock multiple times.
  */
 
 RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_me);
 RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_try_me);
-RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_perf_me);
 
 rte_mcslock_t *p_ml;
 rte_mcslock_t *p_ml_try;
-rte_mcslock_t *p_ml_perf;
 
 static unsigned int count;
 
-static rte_atomic32_t synchro;
-
 static int
 test_mcslock_per_core(__attribute__((unused)) void *arg)
 {
@@ -63,85 +56,8 @@ test_mcslock_per_core(__attribute__((unused)) void
 *arg)
 	return 0;
 }
 
-static uint64_t time_count[RTE_MAX_LCORE] = {0};
-
 #define MAX_LOOP 1000000
 
-static int
-load_loop_fn(void *func_param)
-{
-	uint64_t time_diff = 0, begin;
-	uint64_t hz = rte_get_timer_hz();
-	volatile uint64_t lcount = 0;
-	const int use_lock = *(int *)func_param;
-	const unsigned int lcore = rte_lcore_id();
-
-	/**< Per core me node. */
-	rte_mcslock_t ml_perf_me = RTE_PER_LCORE(_ml_perf_me);
-
-	/* wait synchro */
-	while (rte_atomic32_read(&synchro) == 0)
-		;
-
-	begin = rte_get_timer_cycles();
-	while (lcount < MAX_LOOP) {
-		if (use_lock)
-			rte_mcslock_lock(&p_ml_perf, &ml_perf_me);
-
-		lcount++;
-		if (use_lock)
-			rte_mcslock_unlock(&p_ml_perf, &ml_perf_me);
-	}
-	time_diff = rte_get_timer_cycles() - begin;
-	time_count[lcore] = time_diff * 1000000 / hz;
-	return 0;
-}
-
-static int
-test_mcslock_perf(void)
-{
-	unsigned int i;
-	uint64_t total = 0;
-	int lock = 0;
-	const unsigned int lcore = rte_lcore_id();
-
-	printf("\nTest with no lock on single core...\n");
-	rte_atomic32_set(&synchro, 1);
-	load_loop_fn(&lock);
-	printf("Core [%u] Cost Time = %"PRIu64" us\n",
-			lcore, time_count[lcore]);
-	memset(time_count, 0, sizeof(time_count));
-
-	printf("\nTest with lock on single core...\n");
-	lock = 1;
-	rte_atomic32_set(&synchro, 1);
-	load_loop_fn(&lock);
-	printf("Core [%u] Cost Time = %"PRIu64" us\n",
-			lcore, time_count[lcore]);
-	memset(time_count, 0, sizeof(time_count));
-
-	printf("\nTest with lock on %u cores...\n", (rte_lcore_count()));
-
-	rte_atomic32_set(&synchro, 0);
-	rte_eal_mp_remote_launch(load_loop_fn, &lock, SKIP_MASTER);
-
-	/* start synchro and launch test on master */
-	rte_atomic32_set(&synchro, 1);
-	load_loop_fn(&lock);
-
-	rte_eal_mp_wait_lcore();
-
-	RTE_LCORE_FOREACH(i) {
-		printf("Core [%u] Cost Time = %"PRIu64" us\n",
-				i, time_count[i]);
-		total += time_count[i];
-	}
-
-	printf("Total Cost Time = %"PRIu64" us\n", total);
-
-	return 0;
-}
-
 /*
  * Use rte_mcslock_trylock() to trylock a mcs lock object,

* If it could not lock the object successfully, it would

app/test/test_mcslock.c +0 -4
@@ -240,10 +156,6 @@ test_mcslock(void)
 		ret = -1;
 	rte_mcslock_unlock(&p_ml, &ml_me);
 
-	/* mcs lock perf test */
-	if (test_mcslock_perf() < 0)
-		return -1;
-
 	return ret;
 }
 
app/test/test_mcslock_perf.c new file +121 -1
new file mode 100644
index 0000000..6948344
@@ -0,0 +1,121 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Arm Limited
+ */
+
+#include <stdio.h>
+#include <stdint.h>
+#include <inttypes.h>
+#include <string.h>
+#include <unistd.h>
+#include <sys/queue.h>
+
+#include <rte_common.h>
+#include <rte_memory.h>
+#include <rte_per_lcore.h>
+#include <rte_launch.h>
+#include <rte_eal.h>
+#include <rte_lcore.h>
+#include <rte_cycles.h>
+#include <rte_mcslock.h>
+#include <rte_atomic.h>
+
+#include "test.h"
+
+/*
+ * RTE MCS lock perf test
+ * ======================
+ *
+ * These tests are derived from spin lock perf test cases.
+ *
+ * - A load test is carried out, with all cores attempting to lock a single
+ *   lock multiple times.
+ */
+
+RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_perf_me);
+rte_mcslock_t *p_ml_perf;
+
+static rte_atomic32_t synchro;
+static uint64_t time_count[RTE_MAX_LCORE] = {0};
+
+#define MAX_LOOP 1000000
+
+static int
+load_loop_fn(void *func_param)
+{
+	uint64_t time_diff = 0, begin;
+	uint64_t hz = rte_get_timer_hz();
+	volatile uint64_t lcount = 0;
+	const int use_lock = *(int *)func_param;
+	const unsigned int lcore = rte_lcore_id();
+
+	/**< Per core me node. */
+	rte_mcslock_t ml_perf_me = RTE_PER_LCORE(_ml_perf_me);
+
+	/* wait synchro */
+	while (rte_atomic32_read(&synchro) == 0)
+		;
+
+	begin = rte_get_timer_cycles();
+	while (lcount < MAX_LOOP) {
+		if (use_lock)
+			rte_mcslock_lock(&p_ml_perf, &ml_perf_me);
+
+		lcount++;
+		if (use_lock)
+			rte_mcslock_unlock(&p_ml_perf, &ml_perf_me);
+	}
+	time_diff = rte_get_timer_cycles() - begin;
+	time_count[lcore] = time_diff * 1000000 / hz;
+	return 0;
+}
+
+/*
+ * Test rte_eal_get_lcore_state() in addition to mcs locks
+ * as we have "waiting" then "running" lcores.
+ */
+static int
+test_mcslock_perf(void)
+{
+	unsigned int i;
+	uint64_t total = 0;
+	int lock = 0;
+	const unsigned int lcore = rte_lcore_id();
+
+	printf("\nTest with no lock on single core...\n");
+	rte_atomic32_set(&synchro, 1);
+	load_loop_fn(&lock);
+	printf("Core [%u] Cost Time = %"PRIu64" us\n",
+			lcore, time_count[lcore]);
+	memset(time_count, 0, sizeof(time_count));
+
+	printf("\nTest with lock on single core...\n");
+	lock = 1;
+	rte_atomic32_set(&synchro, 1);
+	load_loop_fn(&lock);
+	printf("Core [%u] Cost Time = %"PRIu64" us\n",
+			lcore, time_count[lcore]);
+	memset(time_count, 0, sizeof(time_count));
+
+	printf("\nTest with lock on %u cores...\n", (rte_lcore_count()));
+
+	rte_atomic32_set(&synchro, 0);
+	rte_eal_mp_remote_launch(load_loop_fn, &lock, SKIP_MASTER);
+
+	/* start synchro and launch test on master */
+	rte_atomic32_set(&synchro, 1);
+	load_loop_fn(&lock);
+
+	rte_eal_mp_wait_lcore();
+
+	RTE_LCORE_FOREACH(i) {
+		printf("Core [%u] Cost Time = %"PRIu64" us\n",
+				i, time_count[i]);
+		total += time_count[i];
+	}
+
+	printf("Total Cost Time = %"PRIu64" us\n", total);
+
+	return 0;
+}
+
+REGISTER_TEST_COMMAND(mcslock_perf_autotest, test_mcslock_perf);
--

2.7.4

#6 lore raw
Quoted 3 lines · Phil Yang

Hi Aaron,

It seems Travis CI cannot capture this timeout issue.

That's probably because we use -t 3 to multiply the timeout. We needed to do that because a few of the tests were on the cusp of failure in the Travis environment since it's a bit lower power. In more brawny systems, probably they aren't an issue.

Quoted 15 lines

But the local test records these two cases as TIMEOUT under the default timeout configuration.

1. The test results on x86 server (72 lcores online@ 2.60GHz) are:
a. Default setting:

$ sudo meson test -C build_test --suite DPDK:fast-tests
   3/92 DPDK:fast-tests / atomic_autotest                TIMEOUT        10.32s
39/92 DPDK:fast-tests / mcslock_autotest               TIMEOUT        10.32s

b. Extend timeout:

$ sudo meson test -C build_test --suite DPDK:fast-tests -t 30
   3/92 DPDK:fast-tests / atomic_autotest                OK             173.01s
39/92 DPDK:fast-tests / mcslock_autotest               OK             12.19s

2. The test results on aarch64 server (56 lcore online @ 2.0GHz) are:
a. Default setting:

Quoted 8 lines
$ sudo meson test -C build_test --suite DPDK:fast-tests
   3/92 DPDK:fast-tests / atomic_autotest       TIMEOUT 10.47 s
39/92 DPDK:fast-tests / mcslock_autotest      TIMEOUT 10.47 s

b. Extend timeout:

$ sudo meson test -C build_test --suite DPDK:fast-tests -t 60
   3/92 DPDK:fast-tests / atomic_autotest       OK      431.89 s
39/92 DPDK:fast-tests / mcslock_autotest      OK      21.25 s

In all the cases it seems the mcslock_autotest is really not too far off, but the atomic_autotest looks really long no matter what.

Quoted 1 line

So I think these two patches can resolve this issue. I'd love to hear your thoughts.

In general, I'm always happy to see perf tests more appropriately put under perf area. I don't have much opinion on this specific patch, though.

Quoted 341 lines

Thanks,
Phil

<snip>

Quoted 336 lines

Subject: [dpdk-dev] [PATCH 1/2] test/mcslock: move performance test to perf tests

The MCS lock performance test takes more than 10 seconds and leads to meson test timeout on some platforms. Move the performance test into perf tests.

---
 MAINTAINERS                  |   1 +
 app/test/Makefile            |   1 +
 app/test/autotest_data.py    |   6 +++
 app/test/meson.build         |   2 +
 app/test/test_mcslock.c      |  88 -------------------------------
 app/test/test_mcslock_perf.c | 121
+++++++++++++++++++++++++++++++++++++++++++
 6 files changed, 131 insertions(+), 88 deletions(-)
 create mode 100644 app/test/test_mcslock_perf.c
MAINTAINERS +1 -0
index db235c2..411bdeb 100644
@@ -247,6 +247,7 @@ MCSlock - EXPERIMENTAL
 M: Phil Yang <[email protected]>
 F: lib/librte_eal/common/include/generic/rte_mcslock.h
 F: app/test/test_mcslock.c
+F: app/test/test_mcslock_perf.c
 
 Ticketlock
 M: Joyce Kong <[email protected]>
app/test/Makefile +1 -0
index 1f080d1..97de3ac 100644
@@ -65,6 +65,7 @@ SRCS-y += test_barrier.c
 SRCS-y += test_malloc.c
 SRCS-y += test_cycles.c
 SRCS-y += test_mcslock.c
+SRCS-y += test_mcslock_perf.c
 SRCS-y += test_spinlock.c
 SRCS-y += test_ticketlock.c
 SRCS-y += test_memory.c
app/test/autotest_data.py +6 -0
index 7b1d013..2a4619d 100644
@@ -784,6 +784,12 @@
         "Func":    default_autotest,
         "Report":  None,
     },
+    {
+        "Name":    "MCS Lock performance autotest",
+        "Command": "mcslock_perf_autotest",
+        "Func":    default_autotest,
+        "Report":  None,
+    },
     #
     # Please always make sure that ring_perf is the last test!
     #
app/test/meson.build +2 -0
index 0a2ce71..335a869 100644
@@ -82,6 +82,7 @@ test_sources = files('commands.c',
 	'test_meter.c',
 	'test_metrics.c',
 	'test_mcslock.c',
+	'test_mcslock_perf.c',
 	'test_mp_secondary.c',
 	'test_per_lcore.c',
 	'test_pmd_perf.c',
@@ -270,6 +271,7 @@ perf_test_names = [
         'rand_perf_autotest',
         'hash_readwrite_perf_autotest',
         'hash_readwrite_lf_perf_autotest',
+	'mcslock_perf_autotest',
 ]
 
 driver_test_names = [
app/test/test_mcslock.c +0 -84
index e9359df..15f9751 100644
@@ -32,23 +32,16 @@
  *
  *   - The function takes the global lock, display something, then releases
  *     the global lock on each core.
- *
- * - A load test is carried out, with all cores attempting to lock a single
- *   lock multiple times.
  */
 
 RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_me);
 RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_try_me);
-RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_perf_me);
 
 rte_mcslock_t *p_ml;
 rte_mcslock_t *p_ml_try;
-rte_mcslock_t *p_ml_perf;
 
 static unsigned int count;
 
-static rte_atomic32_t synchro;
-
 static int
 test_mcslock_per_core(__attribute__((unused)) void *arg)
 {
@@ -63,85 +56,8 @@ test_mcslock_per_core(__attribute__((unused)) void
 *arg)
 	return 0;
 }
 
-static uint64_t time_count[RTE_MAX_LCORE] = {0};
-
 #define MAX_LOOP 1000000
 
-static int
-load_loop_fn(void *func_param)
-{
-	uint64_t time_diff = 0, begin;
-	uint64_t hz = rte_get_timer_hz();
-	volatile uint64_t lcount = 0;
-	const int use_lock = *(int *)func_param;
-	const unsigned int lcore = rte_lcore_id();
-
-	/**< Per core me node. */
-	rte_mcslock_t ml_perf_me = RTE_PER_LCORE(_ml_perf_me);
-
-	/* wait synchro */
-	while (rte_atomic32_read(&synchro) == 0)
-		;
-
-	begin = rte_get_timer_cycles();
-	while (lcount < MAX_LOOP) {
-		if (use_lock)
-			rte_mcslock_lock(&p_ml_perf, &ml_perf_me);
-
-		lcount++;
-		if (use_lock)
-			rte_mcslock_unlock(&p_ml_perf, &ml_perf_me);
-	}
-	time_diff = rte_get_timer_cycles() - begin;
-	time_count[lcore] = time_diff * 1000000 / hz;
-	return 0;
-}
-
-static int
-test_mcslock_perf(void)
-{
-	unsigned int i;
-	uint64_t total = 0;
-	int lock = 0;
-	const unsigned int lcore = rte_lcore_id();
-
-	printf("\nTest with no lock on single core...\n");
-	rte_atomic32_set(&synchro, 1);
-	load_loop_fn(&lock);
-	printf("Core [%u] Cost Time = %"PRIu64" us\n",
-			lcore, time_count[lcore]);
-	memset(time_count, 0, sizeof(time_count));
-
-	printf("\nTest with lock on single core...\n");
-	lock = 1;
-	rte_atomic32_set(&synchro, 1);
-	load_loop_fn(&lock);
-	printf("Core [%u] Cost Time = %"PRIu64" us\n",
-			lcore, time_count[lcore]);
-	memset(time_count, 0, sizeof(time_count));
-
-	printf("\nTest with lock on %u cores...\n", (rte_lcore_count()));
-
-	rte_atomic32_set(&synchro, 0);
-	rte_eal_mp_remote_launch(load_loop_fn, &lock, SKIP_MASTER);
-
-	/* start synchro and launch test on master */
-	rte_atomic32_set(&synchro, 1);
-	load_loop_fn(&lock);
-
-	rte_eal_mp_wait_lcore();
-
-	RTE_LCORE_FOREACH(i) {
-		printf("Core [%u] Cost Time = %"PRIu64" us\n",
-				i, time_count[i]);
-		total += time_count[i];
-	}
-
-	printf("Total Cost Time = %"PRIu64" us\n", total);
-
-	return 0;
-}
-
 /*
  * Use rte_mcslock_trylock() to trylock a mcs lock object,

* If it could not lock the object successfully, it would

app/test/test_mcslock.c +0 -4
@@ -240,10 +156,6 @@ test_mcslock(void)
 		ret = -1;
 	rte_mcslock_unlock(&p_ml, &ml_me);
 
-	/* mcs lock perf test */
-	if (test_mcslock_perf() < 0)
-		return -1;
-
 	return ret;
 }
 
app/test/test_mcslock_perf.c new file +121 -1
new file mode 100644
index 0000000..6948344
@@ -0,0 +1,121 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Arm Limited
+ */
+
+#include <stdio.h>
+#include <stdint.h>
+#include <inttypes.h>
+#include <string.h>
+#include <unistd.h>
+#include <sys/queue.h>
+
+#include <rte_common.h>
+#include <rte_memory.h>
+#include <rte_per_lcore.h>
+#include <rte_launch.h>
+#include <rte_eal.h>
+#include <rte_lcore.h>
+#include <rte_cycles.h>
+#include <rte_mcslock.h>
+#include <rte_atomic.h>
+
+#include "test.h"
+
+/*
+ * RTE MCS lock perf test
+ * ======================
+ *
+ * These tests are derived from spin lock perf test cases.
+ *
+ * - A load test is carried out, with all cores attempting to lock a single
+ *   lock multiple times.
+ */
+
+RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_perf_me);
+rte_mcslock_t *p_ml_perf;
+
+static rte_atomic32_t synchro;
+static uint64_t time_count[RTE_MAX_LCORE] = {0};
+
+#define MAX_LOOP 1000000
+
+static int
+load_loop_fn(void *func_param)
+{
+	uint64_t time_diff = 0, begin;
+	uint64_t hz = rte_get_timer_hz();
+	volatile uint64_t lcount = 0;
+	const int use_lock = *(int *)func_param;
+	const unsigned int lcore = rte_lcore_id();
+
+	/**< Per core me node. */
+	rte_mcslock_t ml_perf_me = RTE_PER_LCORE(_ml_perf_me);
+
+	/* wait synchro */
+	while (rte_atomic32_read(&synchro) == 0)
+		;
+
+	begin = rte_get_timer_cycles();
+	while (lcount < MAX_LOOP) {
+		if (use_lock)
+			rte_mcslock_lock(&p_ml_perf, &ml_perf_me);
+
+		lcount++;
+		if (use_lock)
+			rte_mcslock_unlock(&p_ml_perf, &ml_perf_me);
+	}
+	time_diff = rte_get_timer_cycles() - begin;
+	time_count[lcore] = time_diff * 1000000 / hz;
+	return 0;
+}
+
+/*
+ * Test rte_eal_get_lcore_state() in addition to mcs locks
+ * as we have "waiting" then "running" lcores.
+ */
+static int
+test_mcslock_perf(void)
+{
+	unsigned int i;
+	uint64_t total = 0;
+	int lock = 0;
+	const unsigned int lcore = rte_lcore_id();
+
+	printf("\nTest with no lock on single core...\n");
+	rte_atomic32_set(&synchro, 1);
+	load_loop_fn(&lock);
+	printf("Core [%u] Cost Time = %"PRIu64" us\n",
+			lcore, time_count[lcore]);
+	memset(time_count, 0, sizeof(time_count));
+
+	printf("\nTest with lock on single core...\n");
+	lock = 1;
+	rte_atomic32_set(&synchro, 1);
+	load_loop_fn(&lock);
+	printf("Core [%u] Cost Time = %"PRIu64" us\n",
+			lcore, time_count[lcore]);
+	memset(time_count, 0, sizeof(time_count));
+
+	printf("\nTest with lock on %u cores...\n", (rte_lcore_count()));
+
+	rte_atomic32_set(&synchro, 0);
+	rte_eal_mp_remote_launch(load_loop_fn, &lock, SKIP_MASTER);
+
+	/* start synchro and launch test on master */
+	rte_atomic32_set(&synchro, 1);
+	load_loop_fn(&lock);
+
+	rte_eal_mp_wait_lcore();
+
+	RTE_LCORE_FOREACH(i) {
+		printf("Core [%u] Cost Time = %"PRIu64" us\n",
+				i, time_count[i]);
+		total += time_count[i];
+	}
+
+	printf("Total Cost Time = %"PRIu64" us\n", total);
+
+	return 0;
+}
+
+REGISTER_TEST_COMMAND(mcslock_perf_autotest, test_mcslock_perf);
--

2.7.4

author
#7 lore raw

PATCH v2 1/1 test/mcslock: move performance test to perf tests

In reply to Phil Yang

Move the MCS lock performance test into perf tests.

---
v2:
1. Rebase code.
2. Split from the original patchset.

 MAINTAINERS                  |   1 +
 app/test/Makefile            |   1 +
 app/test/autotest_data.py    |   6 +++
 app/test/meson.build         |   4 +-
 app/test/test_mcslock.c      |  88 -------------------------------
 app/test/test_mcslock_perf.c | 121 +++++++++++++++++++++++++++++++++++++++++++
 6 files changed, 132 insertions(+), 89 deletions(-)
 create mode 100644 app/test/test_mcslock_perf.c
MAINTAINERS +1 -0
index 171c4c7..eae6034 100644
@@ -265,6 +265,7 @@ MCSlock - EXPERIMENTAL
 M: Phil Yang <[email protected]>
 F: lib/librte_eal/include/generic/rte_mcslock.h
 F: app/test/test_mcslock.c
+F: app/test/test_mcslock_perf.c
 
 Ticketlock
 M: Joyce Kong <[email protected]>
app/test/Makefile +1 -0
index f406527..e4b3327 100644
@@ -65,6 +65,7 @@ SRCS-y += test_barrier.c
 SRCS-y += test_malloc.c
 SRCS-y += test_cycles.c
 SRCS-y += test_mcslock.c
+SRCS-y += test_mcslock_perf.c
 SRCS-y += test_spinlock.c
 SRCS-y += test_ticketlock.c
 SRCS-y += test_memory.c
app/test/autotest_data.py +6 -0
index 4b7da45..c813b93 100644
@@ -796,6 +796,12 @@
         "Func":    default_autotest,
         "Report":  None,
     },
+    {
+        "Name":    "MCS Lock performance autotest",
+        "Command": "mcslock_perf_autotest",
+        "Func":    default_autotest,
+        "Report":  None,
+    },
     #
     # Please always make sure that ring_perf is the last test!
     #
app/test/meson.build +3 -1
index 786a213..59df2fd 100644
@@ -86,6 +86,7 @@ test_sources = files('commands.c',
 	'test_meter.c',
 	'test_metrics.c',
 	'test_mcslock.c',
+	'test_mcslock_perf.c',
 	'test_mp_secondary.c',
 	'test_per_lcore.c',
 	'test_pmd_perf.c',
@@ -297,7 +298,8 @@ perf_test_names = [
         'hash_readwrite_perf_autotest',
         'hash_readwrite_lf_perf_autotest',
         'trace_perf_autotest',
-	'ipsec_perf_autotest',
+        'ipsec_perf_autotest',
+        'mcslock_perf_autotest',
 ]
 
 driver_test_names = [
app/test/test_mcslock.c +0 -88

118 lines not on this page · raw message

app/test/test_mcslock_perf.c new file +121 -0

121 lines not on this page · raw message

Signature · 1 line
2.7.4
Reviewed-by on v2 1/1
#8 lore raw
Quoted 343 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Phil Yang
Sent: Tuesday, July 28, 2020 5:24 AM
To: [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; Thomas Monjalon <[email protected]>
Subject: [dpdk-dev] [PATCH v2] test/mcslock: move performance test to perf tests

Move the MCS lock performance test into perf tests.

---
v2:
1. Rebase code.
2. Split from the original patchset.

 MAINTAINERS                  |   1 +
 app/test/Makefile            |   1 +
 app/test/autotest_data.py    |   6 +++
 app/test/meson.build         |   4 +-
 app/test/test_mcslock.c      |  88 -------------------------------
 app/test/test_mcslock_perf.c | 121
+++++++++++++++++++++++++++++++++++++++++++
 6 files changed, 132 insertions(+), 89 deletions(-)  create mode 100644

app/test/test_mcslock_perf.c

MAINTAINERS +1 -0
index 171c4c7..eae6034 100644
@@ -265,6 +265,7 @@ MCSlock - EXPERIMENTAL
 M: Phil Yang <[email protected]>
 F: lib/librte_eal/include/generic/rte_mcslock.h
 F: app/test/test_mcslock.c
+F: app/test/test_mcslock_perf.c
 
 Ticketlock
 M: Joyce Kong <[email protected]>
app/test/Makefile index f406527..e4b3327 renamed from app/test/Makefile +0 -0

100644

app/test/Makefile +2 -0
@@ -65,6 +65,7 @@ SRCS-y += test_barrier.c  SRCS-y += test_malloc.c  SRCS-y
+= test_cycles.c  SRCS-y += test_mcslock.c
+SRCS-y += test_mcslock_perf.c
 SRCS-y += test_spinlock.c
 SRCS-y += test_ticketlock.c
 SRCS-y += test_memory.c
app/test/autotest_data.py index renamed from app/test/autotest_data.py +0 -0

4b7da45..c813b93 100644

app/test/autotest_data.py +6 -0
@@ -796,6 +796,12 @@
         "Func":    default_autotest,
         "Report":  None,
     },
+    {
+        "Name":    "MCS Lock performance autotest",
+        "Command": "mcslock_perf_autotest",
+        "Func":    default_autotest,
+        "Report":  None,
+    },
     #
     # Please always make sure that ring_perf is the last test!
     #
app/test/meson.build index renamed from app/test/meson.build +0 -0

786a213..59df2fd 100644

app/test/meson.build +3 -1
@@ -86,6 +86,7 @@ test_sources = files('commands.c',
 	'test_meter.c',
 	'test_metrics.c',
 	'test_mcslock.c',
+	'test_mcslock_perf.c',
 	'test_mp_secondary.c',
 	'test_per_lcore.c',
 	'test_pmd_perf.c',
@@ -297,7 +298,8 @@ perf_test_names = [
         'hash_readwrite_perf_autotest',
         'hash_readwrite_lf_perf_autotest',
         'trace_perf_autotest',
-	'ipsec_perf_autotest',
+        'ipsec_perf_autotest',
+        'mcslock_perf_autotest',
 ]
 
 driver_test_names = [
app/test/test_mcslock.c index renamed from app/test/test_mcslock.c +0 -0

ddccaaf..a1dc610 100644

app/test/test_mcslock.c +0 -3
@@ -32,23 +32,16 @@
  *
  *   - The function takes the global lock, display something, then releases
  *     the global lock on each core.
- *
- * - A load test is carried out, with all cores attempting to lock a single
- *   lock multiple times.
  */
 
 RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_me);

RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_try_me); - RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_perf_me);

rte_mcslock_t *p_ml;
rte_mcslock_t *p_ml_try;

app/test/test_mcslock.c +0 -3
-rte_mcslock_t *p_ml_perf;
 
 static unsigned int count;
 
-static rte_atomic32_t synchro;
-
 static int
 test_mcslock_per_core(__rte_unused void *arg)  { @@ -63,85 +56,8 @@

test_mcslock_per_core(__rte_unused void *arg)

app/test/test_mcslock.c +0 -77
 	return 0;
 }
 
-static uint64_t time_count[RTE_MAX_LCORE] = {0};
-
 #define MAX_LOOP 1000000
 
-static int
-load_loop_fn(void *func_param)
-{
-	uint64_t time_diff = 0, begin;
-	uint64_t hz = rte_get_timer_hz();
-	volatile uint64_t lcount = 0;
-	const int use_lock = *(int *)func_param;
-	const unsigned int lcore = rte_lcore_id();
-
-	/**< Per core me node. */
-	rte_mcslock_t ml_perf_me = RTE_PER_LCORE(_ml_perf_me);
-
-	/* wait synchro */
-	while (rte_atomic32_read(&synchro) == 0)
-		;
-
-	begin = rte_get_timer_cycles();
-	while (lcount < MAX_LOOP) {
-		if (use_lock)
-			rte_mcslock_lock(&p_ml_perf, &ml_perf_me);
-
-		lcount++;
-		if (use_lock)
-			rte_mcslock_unlock(&p_ml_perf, &ml_perf_me);
-	}
-	time_diff = rte_get_timer_cycles() - begin;
-	time_count[lcore] = time_diff * 1000000 / hz;
-	return 0;
-}
-
-static int
-test_mcslock_perf(void)
-{
-	unsigned int i;
-	uint64_t total = 0;
-	int lock = 0;
-	const unsigned int lcore = rte_lcore_id();
-
-	printf("\nTest with no lock on single core...\n");
-	rte_atomic32_set(&synchro, 1);
-	load_loop_fn(&lock);
-	printf("Core [%u] Cost Time = %"PRIu64" us\n",
-			lcore, time_count[lcore]);
-	memset(time_count, 0, sizeof(time_count));
-
-	printf("\nTest with lock on single core...\n");
-	lock = 1;
-	rte_atomic32_set(&synchro, 1);
-	load_loop_fn(&lock);
-	printf("Core [%u] Cost Time = %"PRIu64" us\n",
-			lcore, time_count[lcore]);
-	memset(time_count, 0, sizeof(time_count));
-
-	printf("\nTest with lock on %u cores...\n", (rte_lcore_count()));
-
-	rte_atomic32_set(&synchro, 0);
-	rte_eal_mp_remote_launch(load_loop_fn, &lock, SKIP_MASTER);
-
-	/* start synchro and launch test on master */
-	rte_atomic32_set(&synchro, 1);
-	load_loop_fn(&lock);
-
-	rte_eal_mp_wait_lcore();
-
-	RTE_LCORE_FOREACH(i) {
-		printf("Core [%u] Cost Time = %"PRIu64" us\n",
-				i, time_count[i]);
-		total += time_count[i];
-	}
-
-	printf("Total Cost Time = %"PRIu64" us\n", total);
-
-	return 0;
-}
-
 /*
  * Use rte_mcslock_trylock() to trylock a mcs lock object,
  * If it could not lock the object successfully, it would @@ -240,10 +156,6 @@

test_mcslock(void)

app/test/test_mcslock.c +0 -4
 		ret = -1;
 	rte_mcslock_unlock(&p_ml, &ml_me);
 
-	/* mcs lock perf test */
-	if (test_mcslock_perf() < 0)
-		return -1;
-
 	return ret;
 }
app/test/test_mcslock_perf.c new renamed from app/test/test_mcslock_perf.c +0 -0

file mode 100644 index 0000000..6948344

app/test/test_mcslock_perf.c new file +121 -1
@@ -0,0 +1,121 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2019 Arm Limited
+ */
+
+#include <stdio.h>
+#include <stdint.h>
+#include <inttypes.h>
+#include <string.h>
+#include <unistd.h>
+#include <sys/queue.h>
+
+#include <rte_common.h>
+#include <rte_memory.h>
+#include <rte_per_lcore.h>
+#include <rte_launch.h>
+#include <rte_eal.h>
+#include <rte_lcore.h>
+#include <rte_cycles.h>
+#include <rte_mcslock.h>
+#include <rte_atomic.h>
+
+#include "test.h"
+
+/*
+ * RTE MCS lock perf test
+ * ======================
+ *
+ * These tests are derived from spin lock perf test cases.
+ *
+ * - A load test is carried out, with all cores attempting to lock a single
+ *   lock multiple times.
+ */
+
+RTE_DEFINE_PER_LCORE(rte_mcslock_t, _ml_perf_me); rte_mcslock_t
+*p_ml_perf;
+
+static rte_atomic32_t synchro;
+static uint64_t time_count[RTE_MAX_LCORE] = {0};
+
+#define MAX_LOOP 1000000
+
+static int
+load_loop_fn(void *func_param)
+{
+	uint64_t time_diff = 0, begin;
+	uint64_t hz = rte_get_timer_hz();
+	volatile uint64_t lcount = 0;
+	const int use_lock = *(int *)func_param;
+	const unsigned int lcore = rte_lcore_id();
+
+	/**< Per core me node. */
+	rte_mcslock_t ml_perf_me = RTE_PER_LCORE(_ml_perf_me);
+
+	/* wait synchro */
+	while (rte_atomic32_read(&synchro) == 0)
+		;
+
+	begin = rte_get_timer_cycles();
+	while (lcount < MAX_LOOP) {
+		if (use_lock)
+			rte_mcslock_lock(&p_ml_perf, &ml_perf_me);
+
+		lcount++;
+		if (use_lock)
+			rte_mcslock_unlock(&p_ml_perf, &ml_perf_me);
+	}
+	time_diff = rte_get_timer_cycles() - begin;
+	time_count[lcore] = time_diff * 1000000 / hz;
+	return 0;
+}
+
+/*
+ * Test rte_eal_get_lcore_state() in addition to mcs locks
+ * as we have "waiting" then "running" lcores.
+ */
+static int
+test_mcslock_perf(void)
+{
+	unsigned int i;
+	uint64_t total = 0;
+	int lock = 0;
+	const unsigned int lcore = rte_lcore_id();
+
+	printf("\nTest with no lock on single core...\n");
+	rte_atomic32_set(&synchro, 1);
+	load_loop_fn(&lock);
+	printf("Core [%u] Cost Time = %"PRIu64" us\n",
+			lcore, time_count[lcore]);
+	memset(time_count, 0, sizeof(time_count));
+
+	printf("\nTest with lock on single core...\n");
+	lock = 1;
+	rte_atomic32_set(&synchro, 1);
+	load_loop_fn(&lock);
+	printf("Core [%u] Cost Time = %"PRIu64" us\n",
+			lcore, time_count[lcore]);
+	memset(time_count, 0, sizeof(time_count));
+
+	printf("\nTest with lock on %u cores...\n", (rte_lcore_count()));
+
+	rte_atomic32_set(&synchro, 0);
+	rte_eal_mp_remote_launch(load_loop_fn, &lock, SKIP_MASTER);
+
+	/* start synchro and launch test on master */
+	rte_atomic32_set(&synchro, 1);
+	load_loop_fn(&lock);
+
+	rte_eal_mp_wait_lcore();
+
+	RTE_LCORE_FOREACH(i) {
+		printf("Core [%u] Cost Time = %"PRIu64" us\n",
+				i, time_count[i]);
+		total += time_count[i];
+	}
+
+	printf("Total Cost Time = %"PRIu64" us\n", total);
+
+	return 0;
+}
+
+REGISTER_TEST_COMMAND(mcslock_perf_autotest, test_mcslock_perf);
--
Quoted 2 lines

2.7.4

#9 lore raw
In reply to Phil Yang
Quoted 7 lines · On Mon, 23 Mar 2020 12:16:17 +0800 Phil Yang <[email protected]> wrote:

The MCS lock performance test takes more than 10 seconds and leads to meson test timeout on some platforms. Move the performance test into perf tests.

---

No, fix the test.
There is no reason that mcslock should be in perf section but not all the other lock types.