xref: /dpdk/app/test/test_mempool_perf.c (revision 2f45703c17acb943aaded9f79676fd56a72542b2)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
5  *   All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include <string.h>
35 #include <stdio.h>
36 #include <stdlib.h>
37 #include <stdint.h>
38 #include <inttypes.h>
39 #include <stdarg.h>
40 #include <errno.h>
41 #include <sys/queue.h>
42 
43 #include <rte_common.h>
44 #include <rte_log.h>
45 #include <rte_debug.h>
46 #include <rte_memory.h>
47 #include <rte_memzone.h>
48 #include <rte_launch.h>
49 #include <rte_cycles.h>
50 #include <rte_eal.h>
51 #include <rte_per_lcore.h>
52 #include <rte_lcore.h>
53 #include <rte_atomic.h>
54 #include <rte_branch_prediction.h>
55 #include <rte_mempool.h>
56 #include <rte_spinlock.h>
57 #include <rte_malloc.h>
58 
59 #include "test.h"
60 
61 /*
62  * Mempool performance
63  * =======
64  *
65  *    Each core get *n_keep* objects per bulk of *n_get_bulk*. Then,
66  *    objects are put back in the pool per bulk of *n_put_bulk*.
67  *
68  *    This sequence is done during TIME_S seconds.
69  *
70  *    This test is done on the following configurations:
71  *
72  *    - Cores configuration (*cores*)
73  *
74  *      - One core with cache
75  *      - Two cores with cache
76  *      - Max. cores with cache
77  *      - One core without cache
78  *      - Two cores without cache
79  *      - Max. cores without cache
80  *      - One core with user-owned cache
81  *      - Two cores with user-owned cache
82  *      - Max. cores with user-owned cache
83  *
84  *    - Bulk size (*n_get_bulk*, *n_put_bulk*)
85  *
86  *      - Bulk get from 1 to 32
87  *      - Bulk put from 1 to 32
88  *
89  *    - Number of kept objects (*n_keep*)
90  *
91  *      - 32
92  *      - 128
93  */
94 
95 #define N 65536
96 #define TIME_S 5
97 #define MEMPOOL_ELT_SIZE 2048
98 #define MAX_KEEP 128
99 #define MEMPOOL_SIZE ((rte_lcore_count()*(MAX_KEEP+RTE_MEMPOOL_CACHE_MAX_SIZE))-1)
100 
101 #define LOG_ERR() printf("test failed at %s():%d\n", __func__, __LINE__)
102 #define RET_ERR() do {							\
103 		LOG_ERR();						\
104 		return -1;						\
105 	} while (0)
106 #define GOTO_ERR(var, label) do {					\
107 		LOG_ERR();						\
108 		var = -1;						\
109 		goto label;						\
110 	} while (0)
111 
112 static struct rte_mempool *mp;
113 static struct rte_mempool *mp_cache, *mp_nocache;
114 static int use_external_cache;
115 static unsigned external_cache_size = RTE_MEMPOOL_CACHE_MAX_SIZE;
116 
117 static rte_atomic32_t synchro;
118 
119 /* number of objects in one bulk operation (get or put) */
120 static unsigned n_get_bulk;
121 static unsigned n_put_bulk;
122 
123 /* number of objects retrived from mempool before putting them back */
124 static unsigned n_keep;
125 
126 /* number of enqueues / dequeues */
127 struct mempool_test_stats {
128 	uint64_t enq_count;
129 } __rte_cache_aligned;
130 
131 static struct mempool_test_stats stats[RTE_MAX_LCORE];
132 
133 /*
134  * save the object number in the first 4 bytes of object data. All
135  * other bytes are set to 0.
136  */
137 static void
138 my_obj_init(struct rte_mempool *mp, __attribute__((unused)) void *arg,
139 	    void *obj, unsigned i)
140 {
141 	uint32_t *objnum = obj;
142 	memset(obj, 0, mp->elt_size);
143 	*objnum = i;
144 }
145 
146 static int
147 per_lcore_mempool_test(__attribute__((unused)) void *arg)
148 {
149 	void *obj_table[MAX_KEEP];
150 	unsigned i, idx;
151 	unsigned lcore_id = rte_lcore_id();
152 	int ret = 0;
153 	uint64_t start_cycles, end_cycles;
154 	uint64_t time_diff = 0, hz = rte_get_timer_hz();
155 	struct rte_mempool_cache *cache;
156 
157 	if (use_external_cache) {
158 		/* Create a user-owned mempool cache. */
159 		cache = rte_mempool_cache_create(external_cache_size,
160 						 SOCKET_ID_ANY);
161 		if (cache == NULL)
162 			RET_ERR();
163 	} else {
164 		/* May be NULL if cache is disabled. */
165 		cache = rte_mempool_default_cache(mp, lcore_id);
166 	}
167 
168 	/* n_get_bulk and n_put_bulk must be divisors of n_keep */
169 	if (((n_keep / n_get_bulk) * n_get_bulk) != n_keep)
170 		GOTO_ERR(ret, out);
171 	if (((n_keep / n_put_bulk) * n_put_bulk) != n_keep)
172 		GOTO_ERR(ret, out);
173 
174 	stats[lcore_id].enq_count = 0;
175 
176 	/* wait synchro for slaves */
177 	if (lcore_id != rte_get_master_lcore())
178 		while (rte_atomic32_read(&synchro) == 0);
179 
180 	start_cycles = rte_get_timer_cycles();
181 
182 	while (time_diff/hz < TIME_S) {
183 		for (i = 0; likely(i < (N/n_keep)); i++) {
184 			/* get n_keep objects by bulk of n_bulk */
185 			idx = 0;
186 			while (idx < n_keep) {
187 				ret = rte_mempool_generic_get(mp,
188 							      &obj_table[idx],
189 							      n_get_bulk,
190 							      cache, 0);
191 				if (unlikely(ret < 0)) {
192 					rte_mempool_dump(stdout, mp);
193 					/* in this case, objects are lost... */
194 					GOTO_ERR(ret, out);
195 				}
196 				idx += n_get_bulk;
197 			}
198 
199 			/* put the objects back */
200 			idx = 0;
201 			while (idx < n_keep) {
202 				rte_mempool_generic_put(mp, &obj_table[idx],
203 							n_put_bulk,
204 							cache, 0);
205 				idx += n_put_bulk;
206 			}
207 		}
208 		end_cycles = rte_get_timer_cycles();
209 		time_diff = end_cycles - start_cycles;
210 		stats[lcore_id].enq_count += N;
211 	}
212 
213 out:
214 	if (use_external_cache) {
215 		rte_mempool_cache_flush(cache, mp);
216 		rte_mempool_cache_free(cache);
217 	}
218 
219 	return ret;
220 }
221 
222 /* launch all the per-lcore test, and display the result */
223 static int
224 launch_cores(unsigned cores)
225 {
226 	unsigned lcore_id;
227 	uint64_t rate;
228 	int ret;
229 	unsigned cores_save = cores;
230 
231 	rte_atomic32_set(&synchro, 0);
232 
233 	/* reset stats */
234 	memset(stats, 0, sizeof(stats));
235 
236 	printf("mempool_autotest cache=%u cores=%u n_get_bulk=%u "
237 	       "n_put_bulk=%u n_keep=%u ",
238 	       use_external_cache ?
239 		   external_cache_size : (unsigned) mp->cache_size,
240 	       cores, n_get_bulk, n_put_bulk, n_keep);
241 
242 	if (rte_mempool_avail_count(mp) != MEMPOOL_SIZE) {
243 		printf("mempool is not full\n");
244 		return -1;
245 	}
246 
247 	RTE_LCORE_FOREACH_SLAVE(lcore_id) {
248 		if (cores == 1)
249 			break;
250 		cores--;
251 		rte_eal_remote_launch(per_lcore_mempool_test,
252 				      NULL, lcore_id);
253 	}
254 
255 	/* start synchro and launch test on master */
256 	rte_atomic32_set(&synchro, 1);
257 
258 	ret = per_lcore_mempool_test(NULL);
259 
260 	cores = cores_save;
261 	RTE_LCORE_FOREACH_SLAVE(lcore_id) {
262 		if (cores == 1)
263 			break;
264 		cores--;
265 		if (rte_eal_wait_lcore(lcore_id) < 0)
266 			ret = -1;
267 	}
268 
269 	if (ret < 0) {
270 		printf("per-lcore test returned -1\n");
271 		return -1;
272 	}
273 
274 	rate = 0;
275 	for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++)
276 		rate += (stats[lcore_id].enq_count / TIME_S);
277 
278 	printf("rate_persec=%" PRIu64 "\n", rate);
279 
280 	return 0;
281 }
282 
283 /* for a given number of core, launch all test cases */
284 static int
285 do_one_mempool_test(unsigned cores)
286 {
287 	unsigned bulk_tab_get[] = { 1, 4, 32, 0 };
288 	unsigned bulk_tab_put[] = { 1, 4, 32, 0 };
289 	unsigned keep_tab[] = { 32, 128, 0 };
290 	unsigned *get_bulk_ptr;
291 	unsigned *put_bulk_ptr;
292 	unsigned *keep_ptr;
293 	int ret;
294 
295 	for (get_bulk_ptr = bulk_tab_get; *get_bulk_ptr; get_bulk_ptr++) {
296 		for (put_bulk_ptr = bulk_tab_put; *put_bulk_ptr; put_bulk_ptr++) {
297 			for (keep_ptr = keep_tab; *keep_ptr; keep_ptr++) {
298 
299 				n_get_bulk = *get_bulk_ptr;
300 				n_put_bulk = *put_bulk_ptr;
301 				n_keep = *keep_ptr;
302 				ret = launch_cores(cores);
303 
304 				if (ret < 0)
305 					return -1;
306 			}
307 		}
308 	}
309 	return 0;
310 }
311 
312 static int
313 test_mempool_perf(void)
314 {
315 	rte_atomic32_init(&synchro);
316 
317 	/* create a mempool (without cache) */
318 	if (mp_nocache == NULL)
319 		mp_nocache = rte_mempool_create("perf_test_nocache", MEMPOOL_SIZE,
320 						MEMPOOL_ELT_SIZE, 0, 0,
321 						NULL, NULL,
322 						my_obj_init, NULL,
323 						SOCKET_ID_ANY, 0);
324 	if (mp_nocache == NULL)
325 		return -1;
326 
327 	/* create a mempool (with cache) */
328 	if (mp_cache == NULL)
329 		mp_cache = rte_mempool_create("perf_test_cache", MEMPOOL_SIZE,
330 					      MEMPOOL_ELT_SIZE,
331 					      RTE_MEMPOOL_CACHE_MAX_SIZE, 0,
332 					      NULL, NULL,
333 					      my_obj_init, NULL,
334 					      SOCKET_ID_ANY, 0);
335 	if (mp_cache == NULL)
336 		return -1;
337 
338 	/* performance test with 1, 2 and max cores */
339 	printf("start performance test (without cache)\n");
340 	mp = mp_nocache;
341 
342 	if (do_one_mempool_test(1) < 0)
343 		return -1;
344 
345 	if (do_one_mempool_test(2) < 0)
346 		return -1;
347 
348 	if (do_one_mempool_test(rte_lcore_count()) < 0)
349 		return -1;
350 
351 	/* performance test with 1, 2 and max cores */
352 	printf("start performance test (with cache)\n");
353 	mp = mp_cache;
354 
355 	if (do_one_mempool_test(1) < 0)
356 		return -1;
357 
358 	if (do_one_mempool_test(2) < 0)
359 		return -1;
360 
361 	if (do_one_mempool_test(rte_lcore_count()) < 0)
362 		return -1;
363 
364 	/* performance test with 1, 2 and max cores */
365 	printf("start performance test (with user-owned cache)\n");
366 	mp = mp_nocache;
367 	use_external_cache = 1;
368 
369 	if (do_one_mempool_test(1) < 0)
370 		return -1;
371 
372 	if (do_one_mempool_test(2) < 0)
373 		return -1;
374 
375 	if (do_one_mempool_test(rte_lcore_count()) < 0)
376 		return -1;
377 
378 	rte_mempool_list_dump(stdout);
379 
380 	return 0;
381 }
382 
383 REGISTER_TEST_COMMAND(mempool_perf_autotest, test_mempool_perf);
384