RE: [PATCH] mempool perf test: test random bulk sizes

["Morten Brørup" <mb@smartsharesystems.com>]

Second PING for review.

Med venlig hilsen / Kind regards,
-Morten Brørup


> From: Morten Brørup [mailto:mb@smartsharesystems.com]
> Sent: Thursday, 13 March 2025 09.23
> 
> PING for review.
> 
> This could still make it into 25.03-rc3 (deadline: 14 March 2025).
> 
> Med venlig hilsen / Kind regards,
> -Morten Brørup
> 
> 
> > From: Morten Brørup [mailto:mb@smartsharesystems.com]
> > Sent: Friday, 28 February 2025 17.49
> >
> > Bulk requests to get or put objects in a mempool often vary in size.
> > A series of tests with pseudo random request sizes, to mitigate the
> > benefits of the CPU's dynamic branch predictor, was added.
> >
> > Also, various other minor changes:
> > - Improved the output formatting for readability.
> > - Added test for the "default" mempool with cache.
> > - Skip the tests for the "default" mempool, if it happens to use the
> > same
> >   driver (i.e. operations) as already tested.
> > - Replaced bare use of "unsigned" with "unsigned int",
> >   to make checkpatches happy.
> >
> > Signed-off-by: Morten Brørup <mb@smartsharesystems.com>
> > ---
> >  app/test/test_mempool_perf.c | 219 +++++++++++++++++++++++++++------
> --
> >  1 file changed, 172 insertions(+), 47 deletions(-)
> >
> > diff --git a/app/test/test_mempool_perf.c
> > b/app/test/test_mempool_perf.c
> > index 4dd74ef75a..5e29797f02 100644
> > --- a/app/test/test_mempool_perf.c
> > +++ b/app/test/test_mempool_perf.c
> > @@ -33,6 +33,13 @@
> >   * Mempool performance
> >   * =======
> >   *
> > + *    Each core get *n_keep* objects per bulk of a pseudorandom
> number
> > + *    between 1 and *n_max_bulk*.
> > + *    Objects are put back in the pool per bulk of a similar
> > pseudorandom number.
> > + *    Note: The very low entropy of the randomization algorithm is
> > harmless, because
> > + *          the sole purpose of randomization is to prevent the
> CPU's
> > dynamic branch
> > + *          predictor from enhancing the test results.
> > + *
> >   *    Each core get *n_keep* objects per bulk of *n_get_bulk*. Then,
> >   *    objects are put back in the pool per bulk of *n_put_bulk*.
> >   *
> > @@ -52,7 +59,12 @@
> >   *      - Two cores with user-owned cache
> >   *      - Max. cores with user-owned cache
> >   *
> > - *    - Bulk size (*n_get_bulk*, *n_put_bulk*)
> > + *    - Pseudorandom max bulk size (*n_max_bulk*)
> > + *
> > + *      - Max bulk from CACHE_LINE_BURST to 256, and
> > RTE_MEMPOOL_CACHE_MAX_SIZE,
> > + *        where CACHE_LINE_BURST is the number of pointers fitting
> > into one CPU cache line.
> > + *
> > + *    - Fixed bulk size (*n_get_bulk*, *n_put_bulk*)
> >   *
> >   *      - Bulk get from 1 to 256, and RTE_MEMPOOL_CACHE_MAX_SIZE
> >   *      - Bulk put from 1 to 256, and RTE_MEMPOOL_CACHE_MAX_SIZE
> > @@ -89,16 +101,19 @@
> >  	} while (0)
> >
> >  static int use_external_cache;
> > -static unsigned external_cache_size = RTE_MEMPOOL_CACHE_MAX_SIZE;
> > +static unsigned int external_cache_size =
> RTE_MEMPOOL_CACHE_MAX_SIZE;
> >
> >  static RTE_ATOMIC(uint32_t) synchro;
> >
> > +/* max random number of objects in one bulk operation (get and put)
> */
> > +static unsigned int n_max_bulk;
> > +
> >  /* number of objects in one bulk operation (get or put) */
> > -static unsigned n_get_bulk;
> > -static unsigned n_put_bulk;
> > +static unsigned int n_get_bulk;
> > +static unsigned int n_put_bulk;
> >
> >  /* number of objects retrieved from mempool before putting them back
> > */
> > -static unsigned n_keep;
> > +static unsigned int n_keep;
> >
> >  /* true if we want to test with constant n_get_bulk and n_put_bulk
> */
> >  static int use_constant_values;
> > @@ -118,7 +133,7 @@ static struct mempool_test_stats
> > stats[RTE_MAX_LCORE];
> >   */
> >  static void
> >  my_obj_init(struct rte_mempool *mp, __rte_unused void *arg,
> > -	    void *obj, unsigned i)
> > +	    void *obj, unsigned int i)
> >  {
> >  	uint32_t *objnum = obj;
> >  	memset(obj, 0, mp->elt_size);
> > @@ -159,11 +174,55 @@ test_loop(struct rte_mempool *mp, struct
> > rte_mempool_cache *cache,
> >  	return 0;
> >  }
> >
> > +static __rte_always_inline int
> > +test_loop_random(struct rte_mempool *mp, struct rte_mempool_cache
> > *cache,
> > +	  unsigned int x_keep, unsigned int x_max_bulk)
> > +{
> > +	alignas(RTE_CACHE_LINE_SIZE) void *obj_table[MAX_KEEP];
> > +	unsigned int idx;
> > +	unsigned int i;
> > +	unsigned int r = 0;
> > +	unsigned int x_bulk;
> > +	int ret;
> > +
> > +	for (i = 0; likely(i < (N / x_keep)); i++) {
> > +		/* get x_keep objects by bulk of random [1 .. x_max_bulk]
> > */
> > +		for (idx = 0; idx < x_keep; idx += x_bulk, r++) {
> > +			/* Generate a pseudorandom number [1 .. x_max_bulk].
> > */
> > +			x_bulk = ((r ^ (r >> 2) ^ (r << 3)) & (x_max_bulk -
> > 1)) + 1;
> > +			if (unlikely(idx + x_bulk > x_keep))
> > +				x_bulk = x_keep - idx;
> > +			ret = rte_mempool_generic_get(mp,
> > +						      &obj_table[idx],
> > +						      x_bulk,
> > +						      cache);
> > +			if (unlikely(ret < 0)) {
> > +				rte_mempool_dump(stdout, mp);
> > +				return ret;
> > +			}
> > +		}
> > +
> > +		/* put the objects back by bulk of random [1 .. x_max_bulk]
> > */
> > +		for (idx = 0; idx < x_keep; idx += x_bulk, r++) {
> > +			/* Generate a pseudorandom number [1 .. x_max_bulk].
> > */
> > +			x_bulk = ((r ^ (r >> 2) ^ (r << 3)) & (x_max_bulk -
> > 1)) + 1;
> > +			if (unlikely(idx + x_bulk > x_keep))
> > +				x_bulk = x_keep - idx;
> > +			rte_mempool_generic_put(mp,
> > +						&obj_table[idx],
> > +						x_bulk,
> > +						cache);
> > +		}
> > +	}
> > +
> > +	return 0;
> > +}
> > +
> >  static int
> >  per_lcore_mempool_test(void *arg)
> >  {
> >  	struct rte_mempool *mp = arg;
> > -	unsigned lcore_id = rte_lcore_id();
> > +	unsigned int lcore_id = rte_lcore_id();
> >  	int ret = 0;
> >  	uint64_t start_cycles, end_cycles;
> >  	uint64_t time_diff = 0, hz = rte_get_timer_hz();
> > @@ -181,9 +240,9 @@ per_lcore_mempool_test(void *arg)
> >  	}
> >
> >  	/* n_get_bulk and n_put_bulk must be divisors of n_keep */
> > -	if (((n_keep / n_get_bulk) * n_get_bulk) != n_keep)
> > +	if (!n_max_bulk && (((n_keep / n_get_bulk) * n_get_bulk) !=
> > n_keep))
> >  		GOTO_ERR(ret, out);
> > -	if (((n_keep / n_put_bulk) * n_put_bulk) != n_keep)
> > +	if (!n_max_bulk && (((n_keep / n_put_bulk) * n_put_bulk) !=
> > n_keep))
> >  		GOTO_ERR(ret, out);
> >  	/* for constant n, n_get_bulk and n_put_bulk must be the same */
> >  	if (use_constant_values && n_put_bulk != n_get_bulk)
> > @@ -200,7 +259,9 @@ per_lcore_mempool_test(void *arg)
> >  	start_cycles = rte_get_timer_cycles();
> >
> >  	while (time_diff/hz < TIME_S) {
> > -		if (!use_constant_values)
> > +		if (n_max_bulk)
> > +			ret = test_loop_random(mp, cache, n_keep,
> > n_max_bulk);
> > +		else if (!use_constant_values)
> >  			ret = test_loop(mp, cache, n_keep, n_get_bulk,
> > n_put_bulk);
> >  		else if (n_get_bulk == 1)
> >  			ret = test_loop(mp, cache, n_keep, 1, 1);
> > @@ -246,10 +307,10 @@ per_lcore_mempool_test(void *arg)
> >  static int
> >  launch_cores(struct rte_mempool *mp, unsigned int cores)
> >  {
> > -	unsigned lcore_id;
> > +	unsigned int lcore_id;
> >  	uint64_t rate;
> >  	int ret;
> > -	unsigned cores_save = cores;
> > +	unsigned int cores_save = cores;
> >  	double hz = rte_get_timer_hz();
> >
> >  	rte_atomic_store_explicit(&synchro, 0, rte_memory_order_relaxed);
> > @@ -257,11 +318,18 @@ launch_cores(struct rte_mempool *mp, unsigned
> int
> > cores)
> >  	/* reset stats */
> >  	memset(stats, 0, sizeof(stats));
> >
> > -	printf("mempool_autotest cache=%u cores=%u n_get_bulk=%u "
> > -	       "n_put_bulk=%u n_keep=%u constant_n=%u ",
> > +	printf("mempool_autotest cache=%u cores=%u n_keep=%5u ",
> >  	       use_external_cache ?
> >  		   external_cache_size : (unsigned) mp->cache_size,
> > -	       cores, n_get_bulk, n_put_bulk, n_keep,
> > use_constant_values);
> > +	       cores,
> > +	       n_keep);
> > +	if (n_max_bulk)
> > +		printf("n_max_bulk=%3u ",
> > +		       n_max_bulk);
> > +	else
> > +		printf("n_get_bulk=%3u n_put_bulk=%3u constant_n=%u ",
> > +		       n_get_bulk, n_put_bulk,
> > +		       use_constant_values);
> >
> >  	if (rte_mempool_avail_count(mp) != MEMPOOL_SIZE) {
> >  		printf("mempool is not full\n");
> > @@ -301,7 +369,7 @@ launch_cores(struct rte_mempool *mp, unsigned int
> > cores)
> >  			rate += (double)stats[lcore_id].enq_count * hz /
> >  					(double)stats[lcore_id].duration_cycles;
> >
> > -	printf("rate_persec=%" PRIu64 "\n", rate);
> > +	printf("rate_persec=%10" PRIu64 "\n", rate);
> >
> >  	return 0;
> >  }
> > @@ -310,25 +378,47 @@ launch_cores(struct rte_mempool *mp, unsigned
> int
> > cores)
> >  static int
> >  do_one_mempool_test(struct rte_mempool *mp, unsigned int cores, int
> > external_cache)
> >  {
> > +	unsigned int bulk_tab_max[] = { CACHE_LINE_BURST, 32, 64, 128,
> > 256,
> > +			RTE_MEMPOOL_CACHE_MAX_SIZE, 0 };
> >  	unsigned int bulk_tab_get[] = { 1, 4, CACHE_LINE_BURST, 32, 64,
> > 128, 256,
> >  			RTE_MEMPOOL_CACHE_MAX_SIZE, 0 };
> >  	unsigned int bulk_tab_put[] = { 1, 4, CACHE_LINE_BURST, 32, 64,
> > 128, 256,
> >  			RTE_MEMPOOL_CACHE_MAX_SIZE, 0 };
> >  	unsigned int keep_tab[] = { 32, 128, 512, 2048, 8192, 32768, 0 };
> > -	unsigned *get_bulk_ptr;
> > -	unsigned *put_bulk_ptr;
> > -	unsigned *keep_ptr;
> > +	unsigned int *max_bulk_ptr;
> > +	unsigned int *get_bulk_ptr;
> > +	unsigned int *put_bulk_ptr;
> > +	unsigned int *keep_ptr;
> >  	int ret;
> >
> > -	for (get_bulk_ptr = bulk_tab_get; *get_bulk_ptr; get_bulk_ptr++)
> > {
> > -		for (put_bulk_ptr = bulk_tab_put; *put_bulk_ptr;
> > put_bulk_ptr++) {
> > -			for (keep_ptr = keep_tab; *keep_ptr; keep_ptr++) {
> > +	for (keep_ptr = keep_tab; *keep_ptr; keep_ptr++) {
> > +		for (max_bulk_ptr = bulk_tab_max; *max_bulk_ptr;
> > max_bulk_ptr++) {
> > +
> > +			if (*keep_ptr < *max_bulk_ptr)
> > +				continue;
> > +
> > +			use_external_cache = external_cache;
> > +			use_constant_values = 0;
> > +			n_max_bulk = *max_bulk_ptr;
> > +			n_get_bulk = 0;
> > +			n_put_bulk = 0;
> > +			n_keep = *keep_ptr;
> > +			ret = launch_cores(mp, cores);
> > +			if (ret < 0)
> > +				return -1;
> > +		}
> > +	}
> > +
> > +	for (keep_ptr = keep_tab; *keep_ptr; keep_ptr++) {
> > +		for (get_bulk_ptr = bulk_tab_get; *get_bulk_ptr;
> > get_bulk_ptr++) {
> > +			for (put_bulk_ptr = bulk_tab_put; *put_bulk_ptr;
> > put_bulk_ptr++) {
> >
> >  				if (*keep_ptr < *get_bulk_ptr || *keep_ptr <
> > *put_bulk_ptr)
> >  					continue;
> >
> >  				use_external_cache = external_cache;
> >  				use_constant_values = 0;
> > +				n_max_bulk = 0;
> >  				n_get_bulk = *get_bulk_ptr;
> >  				n_put_bulk = *put_bulk_ptr;
> >  				n_keep = *keep_ptr;
> > @@ -346,6 +436,7 @@ do_one_mempool_test(struct rte_mempool *mp,
> > unsigned int cores, int external_cac
> >  			}
> >  		}
> >  	}
> > +
> >  	return 0;
> >  }
> >
> > @@ -354,7 +445,10 @@ do_all_mempool_perf_tests(unsigned int cores)
> >  {
> >  	struct rte_mempool *mp_cache = NULL;
> >  	struct rte_mempool *mp_nocache = NULL;
> > -	struct rte_mempool *default_pool = NULL;
> > +	struct rte_mempool *default_pool_cache = NULL;
> > +	struct rte_mempool *default_pool_nocache = NULL;
> > +	const char *mp_cache_ops;
> > +	const char *mp_nocache_ops;
> >  	const char *default_pool_ops;
> >  	int ret = -1;
> >
> > @@ -368,6 +462,7 @@ do_all_mempool_perf_tests(unsigned int cores)
> >  		printf("cannot allocate mempool (without cache)\n");
> >  		goto err;
> >  	}
> > +	mp_nocache_ops = rte_mempool_get_ops(mp_nocache->ops_index)-
> > >name;
> >
> >  	/* create a mempool (with cache) */
> >  	mp_cache = rte_mempool_create("perf_test_cache", MEMPOOL_SIZE,
> > @@ -380,47 +475,76 @@ do_all_mempool_perf_tests(unsigned int cores)
> >  		printf("cannot allocate mempool (with cache)\n");
> >  		goto err;
> >  	}
> > +	mp_cache_ops = rte_mempool_get_ops(mp_cache->ops_index)->name;
> >
> >  	default_pool_ops = rte_mbuf_best_mempool_ops();
> > -	/* Create a mempool based on Default handler */
> > -	default_pool = rte_mempool_create_empty("default_pool",
> > -						MEMPOOL_SIZE,
> > -						MEMPOOL_ELT_SIZE,
> > -						0, 0,
> > -						SOCKET_ID_ANY, 0);
> > -
> > -	if (default_pool == NULL) {
> > -		printf("cannot allocate %s mempool\n", default_pool_ops);
> > +
> > +	/* Create a mempool (without cache) based on Default handler */
> > +	default_pool_nocache =
> > rte_mempool_create_empty("default_pool_nocache",
> > +			MEMPOOL_SIZE,
> > +			MEMPOOL_ELT_SIZE,
> > +			0, 0,
> > +			SOCKET_ID_ANY, 0);
> > +	if (default_pool_nocache == NULL) {
> > +		printf("cannot allocate %s mempool (without cache)\n",
> > default_pool_ops);
> >  		goto err;
> >  	}
> > -
> > -	if (rte_mempool_set_ops_byname(default_pool, default_pool_ops,
> > NULL)
> > -				       < 0) {
> > +	if (rte_mempool_set_ops_byname(default_pool_nocache,
> > default_pool_ops, NULL) < 0) {
> >  		printf("cannot set %s handler\n", default_pool_ops);
> >  		goto err;
> >  	}
> > -
> > -	if (rte_mempool_populate_default(default_pool) < 0) {
> > +	if (rte_mempool_populate_default(default_pool_nocache) < 0) {
> >  		printf("cannot populate %s mempool\n", default_pool_ops);
> >  		goto err;
> >  	}
> > +	rte_mempool_obj_iter(default_pool_nocache, my_obj_init, NULL);
> > +
> > +	/* Create a mempool (with cache) based on Default handler */
> > +	default_pool_cache =
> > rte_mempool_create_empty("default_pool_cache",
> > +			MEMPOOL_SIZE,
> > +			MEMPOOL_ELT_SIZE,
> > +			RTE_MEMPOOL_CACHE_MAX_SIZE, 0,
> > +			SOCKET_ID_ANY, 0);
> > +	if (default_pool_cache == NULL) {
> > +		printf("cannot allocate %s mempool (with cache)\n",
> > default_pool_ops);
> > +		goto err;
> > +	}
> > +	if (rte_mempool_set_ops_byname(default_pool_cache,
> > default_pool_ops, NULL) < 0) {
> > +		printf("cannot set %s handler\n", default_pool_ops);
> > +		goto err;
> > +	}
> > +	if (rte_mempool_populate_default(default_pool_cache) < 0) {
> > +		printf("cannot populate %s mempool\n", default_pool_ops);
> > +		goto err;
> > +	}
> > +	rte_mempool_obj_iter(default_pool_cache, my_obj_init, NULL);
> >
> > -	rte_mempool_obj_iter(default_pool, my_obj_init, NULL);
> > -
> > -	printf("start performance test (without cache)\n");
> > +	printf("start performance test (using %s, without cache)\n",
> > +	       mp_nocache_ops);
> >  	if (do_one_mempool_test(mp_nocache, cores, 0) < 0)
> >  		goto err;
> >
> > -	printf("start performance test for %s (without cache)\n",
> > -	       default_pool_ops);
> > -	if (do_one_mempool_test(default_pool, cores, 0) < 0)
> > -		goto err;
> > +	if (strcmp(default_pool_ops, mp_nocache_ops) != 0) {
> > +		printf("start performance test for %s (without cache)\n",
> > +		       default_pool_ops);
> > +		if (do_one_mempool_test(default_pool_nocache, cores, 0) <
> > 0)
> > +			goto err;
> > +	}
> >
> > -	printf("start performance test (with cache)\n");
> > +	printf("start performance test (using %s, with cache)\n",
> > +	       mp_cache_ops);
> >  	if (do_one_mempool_test(mp_cache, cores, 0) < 0)
> >  		goto err;
> >
> > -	printf("start performance test (with user-owned cache)\n");
> > +	if (strcmp(default_pool_ops, mp_cache_ops) != 0) {
> > +		printf("start performance test for %s (with cache)\n",
> > +		       default_pool_ops);
> > +		if (do_one_mempool_test(default_pool_cache, cores, 0) < 0)
> > +			goto err;
> > +	}
> > +
> > +	printf("start performance test (using %s, with user-owned
> > cache)\n",
> > +	       mp_nocache_ops);
> >  	if (do_one_mempool_test(mp_nocache, cores, 1) < 0)
> >  		goto err;
> >
> > @@ -431,7 +555,8 @@ do_all_mempool_perf_tests(unsigned int cores)
> >  err:
> >  	rte_mempool_free(mp_cache);
> >  	rte_mempool_free(mp_nocache);
> > -	rte_mempool_free(default_pool);
> > +	rte_mempool_free(default_pool_cache);
> > +	rte_mempool_free(default_pool_nocache);
> >  	return ret;
> >  }
> >
> > --
> > 2.43.0