xref: /dpdk/app/test/test_ring_perf.c (revision 742bde12)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2014 Intel Corporation
3  */
4 
5 
6 #include <stdio.h>
7 #include <inttypes.h>
8 #include <rte_ring.h>
9 #include <rte_cycles.h>
10 #include <rte_launch.h>
11 #include <rte_pause.h>
12 
13 #include "test.h"
14 
15 /*
16  * Ring
17  * ====
18  *
19  * Measures performance of various operations using rdtsc
20  *  * Empty ring dequeue
21  *  * Enqueue/dequeue of bursts in 1 threads
22  *  * Enqueue/dequeue of bursts in 2 threads
23  */
24 
25 #define RING_NAME "RING_PERF"
26 #define RING_SIZE 4096
27 #define MAX_BURST 32
28 
29 /*
30  * the sizes to enqueue and dequeue in testing
31  * (marked volatile so they won't be seen as compile-time constants)
32  */
33 static const volatile unsigned bulk_sizes[] = { 8, 32 };
34 
35 struct lcore_pair {
36 	unsigned c1, c2;
37 };
38 
39 static volatile unsigned lcore_count = 0;
40 
41 /**** Functions to analyse our core mask to get cores for different tests ***/
42 
43 static int
44 get_two_hyperthreads(struct lcore_pair *lcp)
45 {
46 	unsigned id1, id2;
47 	unsigned c1, c2, s1, s2;
48 	RTE_LCORE_FOREACH(id1) {
49 		/* inner loop just re-reads all id's. We could skip the first few
50 		 * elements, but since number of cores is small there is little point
51 		 */
52 		RTE_LCORE_FOREACH(id2) {
53 			if (id1 == id2)
54 				continue;
55 			c1 = lcore_config[id1].core_id;
56 			c2 = lcore_config[id2].core_id;
57 			s1 = lcore_config[id1].socket_id;
58 			s2 = lcore_config[id2].socket_id;
59 			if ((c1 == c2) && (s1 == s2)){
60 				lcp->c1 = id1;
61 				lcp->c2 = id2;
62 				return 0;
63 			}
64 		}
65 	}
66 	return 1;
67 }
68 
69 static int
70 get_two_cores(struct lcore_pair *lcp)
71 {
72 	unsigned id1, id2;
73 	unsigned c1, c2, s1, s2;
74 	RTE_LCORE_FOREACH(id1) {
75 		RTE_LCORE_FOREACH(id2) {
76 			if (id1 == id2)
77 				continue;
78 			c1 = lcore_config[id1].core_id;
79 			c2 = lcore_config[id2].core_id;
80 			s1 = lcore_config[id1].socket_id;
81 			s2 = lcore_config[id2].socket_id;
82 			if ((c1 != c2) && (s1 == s2)){
83 				lcp->c1 = id1;
84 				lcp->c2 = id2;
85 				return 0;
86 			}
87 		}
88 	}
89 	return 1;
90 }
91 
92 static int
93 get_two_sockets(struct lcore_pair *lcp)
94 {
95 	unsigned id1, id2;
96 	unsigned s1, s2;
97 	RTE_LCORE_FOREACH(id1) {
98 		RTE_LCORE_FOREACH(id2) {
99 			if (id1 == id2)
100 				continue;
101 			s1 = lcore_config[id1].socket_id;
102 			s2 = lcore_config[id2].socket_id;
103 			if (s1 != s2){
104 				lcp->c1 = id1;
105 				lcp->c2 = id2;
106 				return 0;
107 			}
108 		}
109 	}
110 	return 1;
111 }
112 
113 /* Get cycle counts for dequeuing from an empty ring. Should be 2 or 3 cycles */
114 static void
115 test_empty_dequeue(struct rte_ring *r)
116 {
117 	const unsigned iter_shift = 26;
118 	const unsigned iterations = 1<<iter_shift;
119 	unsigned i = 0;
120 	void *burst[MAX_BURST];
121 
122 	const uint64_t sc_start = rte_rdtsc();
123 	for (i = 0; i < iterations; i++)
124 		rte_ring_sc_dequeue_bulk(r, burst, bulk_sizes[0], NULL);
125 	const uint64_t sc_end = rte_rdtsc();
126 
127 	const uint64_t mc_start = rte_rdtsc();
128 	for (i = 0; i < iterations; i++)
129 		rte_ring_mc_dequeue_bulk(r, burst, bulk_sizes[0], NULL);
130 	const uint64_t mc_end = rte_rdtsc();
131 
132 	printf("SC empty dequeue: %.2F\n",
133 			(double)(sc_end-sc_start) / iterations);
134 	printf("MC empty dequeue: %.2F\n",
135 			(double)(mc_end-mc_start) / iterations);
136 }
137 
138 /*
139  * for the separate enqueue and dequeue threads they take in one param
140  * and return two. Input = burst size, output = cycle average for sp/sc & mp/mc
141  */
142 struct thread_params {
143 	struct rte_ring *r;
144 	unsigned size;        /* input value, the burst size */
145 	double spsc, mpmc;    /* output value, the single or multi timings */
146 };
147 
148 /*
149  * Function that uses rdtsc to measure timing for ring enqueue. Needs pair
150  * thread running dequeue_bulk function
151  */
152 static int
153 enqueue_bulk(void *p)
154 {
155 	const unsigned iter_shift = 23;
156 	const unsigned iterations = 1<<iter_shift;
157 	struct thread_params *params = p;
158 	struct rte_ring *r = params->r;
159 	const unsigned size = params->size;
160 	unsigned i;
161 	void *burst[MAX_BURST] = {0};
162 
163 	if ( __sync_add_and_fetch(&lcore_count, 1) != 2 )
164 		while(lcore_count != 2)
165 			rte_pause();
166 
167 	const uint64_t sp_start = rte_rdtsc();
168 	for (i = 0; i < iterations; i++)
169 		while (rte_ring_sp_enqueue_bulk(r, burst, size, NULL) == 0)
170 			rte_pause();
171 	const uint64_t sp_end = rte_rdtsc();
172 
173 	const uint64_t mp_start = rte_rdtsc();
174 	for (i = 0; i < iterations; i++)
175 		while (rte_ring_mp_enqueue_bulk(r, burst, size, NULL) == 0)
176 			rte_pause();
177 	const uint64_t mp_end = rte_rdtsc();
178 
179 	params->spsc = ((double)(sp_end - sp_start))/(iterations*size);
180 	params->mpmc = ((double)(mp_end - mp_start))/(iterations*size);
181 	return 0;
182 }
183 
184 /*
185  * Function that uses rdtsc to measure timing for ring dequeue. Needs pair
186  * thread running enqueue_bulk function
187  */
188 static int
189 dequeue_bulk(void *p)
190 {
191 	const unsigned iter_shift = 23;
192 	const unsigned iterations = 1<<iter_shift;
193 	struct thread_params *params = p;
194 	struct rte_ring *r = params->r;
195 	const unsigned size = params->size;
196 	unsigned i;
197 	void *burst[MAX_BURST] = {0};
198 
199 	if ( __sync_add_and_fetch(&lcore_count, 1) != 2 )
200 		while(lcore_count != 2)
201 			rte_pause();
202 
203 	const uint64_t sc_start = rte_rdtsc();
204 	for (i = 0; i < iterations; i++)
205 		while (rte_ring_sc_dequeue_bulk(r, burst, size, NULL) == 0)
206 			rte_pause();
207 	const uint64_t sc_end = rte_rdtsc();
208 
209 	const uint64_t mc_start = rte_rdtsc();
210 	for (i = 0; i < iterations; i++)
211 		while (rte_ring_mc_dequeue_bulk(r, burst, size, NULL) == 0)
212 			rte_pause();
213 	const uint64_t mc_end = rte_rdtsc();
214 
215 	params->spsc = ((double)(sc_end - sc_start))/(iterations*size);
216 	params->mpmc = ((double)(mc_end - mc_start))/(iterations*size);
217 	return 0;
218 }
219 
220 /*
221  * Function that calls the enqueue and dequeue bulk functions on pairs of cores.
222  * used to measure ring perf between hyperthreads, cores and sockets.
223  */
224 static void
225 run_on_core_pair(struct lcore_pair *cores, struct rte_ring *r,
226 		lcore_function_t f1, lcore_function_t f2)
227 {
228 	struct thread_params param1 = {0}, param2 = {0};
229 	unsigned i;
230 	for (i = 0; i < sizeof(bulk_sizes)/sizeof(bulk_sizes[0]); i++) {
231 		lcore_count = 0;
232 		param1.size = param2.size = bulk_sizes[i];
233 		param1.r = param2.r = r;
234 		if (cores->c1 == rte_get_master_lcore()) {
235 			rte_eal_remote_launch(f2, &param2, cores->c2);
236 			f1(&param1);
237 			rte_eal_wait_lcore(cores->c2);
238 		} else {
239 			rte_eal_remote_launch(f1, &param1, cores->c1);
240 			rte_eal_remote_launch(f2, &param2, cores->c2);
241 			rte_eal_wait_lcore(cores->c1);
242 			rte_eal_wait_lcore(cores->c2);
243 		}
244 		printf("SP/SC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[i],
245 				param1.spsc + param2.spsc);
246 		printf("MP/MC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[i],
247 				param1.mpmc + param2.mpmc);
248 	}
249 }
250 
251 /*
252  * Test function that determines how long an enqueue + dequeue of a single item
253  * takes on a single lcore. Result is for comparison with the bulk enq+deq.
254  */
255 static void
256 test_single_enqueue_dequeue(struct rte_ring *r)
257 {
258 	const unsigned iter_shift = 24;
259 	const unsigned iterations = 1<<iter_shift;
260 	unsigned i = 0;
261 	void *burst = NULL;
262 
263 	const uint64_t sc_start = rte_rdtsc();
264 	for (i = 0; i < iterations; i++) {
265 		rte_ring_sp_enqueue(r, burst);
266 		rte_ring_sc_dequeue(r, &burst);
267 	}
268 	const uint64_t sc_end = rte_rdtsc();
269 
270 	const uint64_t mc_start = rte_rdtsc();
271 	for (i = 0; i < iterations; i++) {
272 		rte_ring_mp_enqueue(r, burst);
273 		rte_ring_mc_dequeue(r, &burst);
274 	}
275 	const uint64_t mc_end = rte_rdtsc();
276 
277 	printf("SP/SC single enq/dequeue: %"PRIu64"\n",
278 			(sc_end-sc_start) >> iter_shift);
279 	printf("MP/MC single enq/dequeue: %"PRIu64"\n",
280 			(mc_end-mc_start) >> iter_shift);
281 }
282 
283 /*
284  * Test that does both enqueue and dequeue on a core using the burst() API calls
285  * instead of the bulk() calls used in other tests. Results should be the same
286  * as for the bulk function called on a single lcore.
287  */
288 static void
289 test_burst_enqueue_dequeue(struct rte_ring *r)
290 {
291 	const unsigned iter_shift = 23;
292 	const unsigned iterations = 1<<iter_shift;
293 	unsigned sz, i = 0;
294 	void *burst[MAX_BURST] = {0};
295 
296 	for (sz = 0; sz < sizeof(bulk_sizes)/sizeof(bulk_sizes[0]); sz++) {
297 		const uint64_t sc_start = rte_rdtsc();
298 		for (i = 0; i < iterations; i++) {
299 			rte_ring_sp_enqueue_burst(r, burst,
300 					bulk_sizes[sz], NULL);
301 			rte_ring_sc_dequeue_burst(r, burst,
302 					bulk_sizes[sz], NULL);
303 		}
304 		const uint64_t sc_end = rte_rdtsc();
305 
306 		const uint64_t mc_start = rte_rdtsc();
307 		for (i = 0; i < iterations; i++) {
308 			rte_ring_mp_enqueue_burst(r, burst,
309 					bulk_sizes[sz], NULL);
310 			rte_ring_mc_dequeue_burst(r, burst,
311 					bulk_sizes[sz], NULL);
312 		}
313 		const uint64_t mc_end = rte_rdtsc();
314 
315 		uint64_t mc_avg = ((mc_end-mc_start) >> iter_shift) / bulk_sizes[sz];
316 		uint64_t sc_avg = ((sc_end-sc_start) >> iter_shift) / bulk_sizes[sz];
317 
318 		printf("SP/SC burst enq/dequeue (size: %u): %"PRIu64"\n", bulk_sizes[sz],
319 				sc_avg);
320 		printf("MP/MC burst enq/dequeue (size: %u): %"PRIu64"\n", bulk_sizes[sz],
321 				mc_avg);
322 	}
323 }
324 
325 /* Times enqueue and dequeue on a single lcore */
326 static void
327 test_bulk_enqueue_dequeue(struct rte_ring *r)
328 {
329 	const unsigned iter_shift = 23;
330 	const unsigned iterations = 1<<iter_shift;
331 	unsigned sz, i = 0;
332 	void *burst[MAX_BURST] = {0};
333 
334 	for (sz = 0; sz < sizeof(bulk_sizes)/sizeof(bulk_sizes[0]); sz++) {
335 		const uint64_t sc_start = rte_rdtsc();
336 		for (i = 0; i < iterations; i++) {
337 			rte_ring_sp_enqueue_bulk(r, burst,
338 					bulk_sizes[sz], NULL);
339 			rte_ring_sc_dequeue_bulk(r, burst,
340 					bulk_sizes[sz], NULL);
341 		}
342 		const uint64_t sc_end = rte_rdtsc();
343 
344 		const uint64_t mc_start = rte_rdtsc();
345 		for (i = 0; i < iterations; i++) {
346 			rte_ring_mp_enqueue_bulk(r, burst,
347 					bulk_sizes[sz], NULL);
348 			rte_ring_mc_dequeue_bulk(r, burst,
349 					bulk_sizes[sz], NULL);
350 		}
351 		const uint64_t mc_end = rte_rdtsc();
352 
353 		double sc_avg = ((double)(sc_end-sc_start) /
354 				(iterations * bulk_sizes[sz]));
355 		double mc_avg = ((double)(mc_end-mc_start) /
356 				(iterations * bulk_sizes[sz]));
357 
358 		printf("SP/SC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[sz],
359 				sc_avg);
360 		printf("MP/MC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[sz],
361 				mc_avg);
362 	}
363 }
364 
365 static int
366 test_ring_perf(void)
367 {
368 	struct lcore_pair cores;
369 	struct rte_ring *r = NULL;
370 
371 	r = rte_ring_create(RING_NAME, RING_SIZE, rte_socket_id(), 0);
372 	if (r == NULL)
373 		return -1;
374 
375 	printf("### Testing single element and burst enq/deq ###\n");
376 	test_single_enqueue_dequeue(r);
377 	test_burst_enqueue_dequeue(r);
378 
379 	printf("\n### Testing empty dequeue ###\n");
380 	test_empty_dequeue(r);
381 
382 	printf("\n### Testing using a single lcore ###\n");
383 	test_bulk_enqueue_dequeue(r);
384 
385 	if (get_two_hyperthreads(&cores) == 0) {
386 		printf("\n### Testing using two hyperthreads ###\n");
387 		run_on_core_pair(&cores, r, enqueue_bulk, dequeue_bulk);
388 	}
389 	if (get_two_cores(&cores) == 0) {
390 		printf("\n### Testing using two physical cores ###\n");
391 		run_on_core_pair(&cores, r, enqueue_bulk, dequeue_bulk);
392 	}
393 	if (get_two_sockets(&cores) == 0) {
394 		printf("\n### Testing using two NUMA nodes ###\n");
395 		run_on_core_pair(&cores, r, enqueue_bulk, dequeue_bulk);
396 	}
397 	rte_ring_free(r);
398 	return 0;
399 }
400 
401 REGISTER_TEST_COMMAND(ring_perf_autotest, test_ring_perf);
402