xref: /dpdk/app/test/test_ring_perf.c (revision 089e5ed727a15da2729cfee9b63533dd120bd04c)
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 
56 			c1 = rte_lcore_to_cpu_id(id1);
57 			c2 = rte_lcore_to_cpu_id(id2);
58 			s1 = rte_lcore_to_socket_id(id1);
59 			s2 = rte_lcore_to_socket_id(id2);
60 			if ((c1 == c2) && (s1 == s2)){
61 				lcp->c1 = id1;
62 				lcp->c2 = id2;
63 				return 0;
64 			}
65 		}
66 	}
67 	return 1;
68 }
69 
70 static int
71 get_two_cores(struct lcore_pair *lcp)
72 {
73 	unsigned id1, id2;
74 	unsigned c1, c2, s1, s2;
75 	RTE_LCORE_FOREACH(id1) {
76 		RTE_LCORE_FOREACH(id2) {
77 			if (id1 == id2)
78 				continue;
79 
80 			c1 = rte_lcore_to_cpu_id(id1);
81 			c2 = rte_lcore_to_cpu_id(id2);
82 			s1 = rte_lcore_to_socket_id(id1);
83 			s2 = rte_lcore_to_socket_id(id2);
84 			if ((c1 != c2) && (s1 == s2)){
85 				lcp->c1 = id1;
86 				lcp->c2 = id2;
87 				return 0;
88 			}
89 		}
90 	}
91 	return 1;
92 }
93 
94 static int
95 get_two_sockets(struct lcore_pair *lcp)
96 {
97 	unsigned id1, id2;
98 	unsigned s1, s2;
99 	RTE_LCORE_FOREACH(id1) {
100 		RTE_LCORE_FOREACH(id2) {
101 			if (id1 == id2)
102 				continue;
103 			s1 = rte_lcore_to_socket_id(id1);
104 			s2 = rte_lcore_to_socket_id(id2);
105 			if (s1 != s2){
106 				lcp->c1 = id1;
107 				lcp->c2 = id2;
108 				return 0;
109 			}
110 		}
111 	}
112 	return 1;
113 }
114 
115 /* Get cycle counts for dequeuing from an empty ring. Should be 2 or 3 cycles */
116 static void
117 test_empty_dequeue(struct rte_ring *r)
118 {
119 	const unsigned iter_shift = 26;
120 	const unsigned iterations = 1<<iter_shift;
121 	unsigned i = 0;
122 	void *burst[MAX_BURST];
123 
124 	const uint64_t sc_start = rte_rdtsc();
125 	for (i = 0; i < iterations; i++)
126 		rte_ring_sc_dequeue_bulk(r, burst, bulk_sizes[0], NULL);
127 	const uint64_t sc_end = rte_rdtsc();
128 
129 	const uint64_t mc_start = rte_rdtsc();
130 	for (i = 0; i < iterations; i++)
131 		rte_ring_mc_dequeue_bulk(r, burst, bulk_sizes[0], NULL);
132 	const uint64_t mc_end = rte_rdtsc();
133 
134 	printf("SC empty dequeue: %.2F\n",
135 			(double)(sc_end-sc_start) / iterations);
136 	printf("MC empty dequeue: %.2F\n",
137 			(double)(mc_end-mc_start) / iterations);
138 }
139 
140 /*
141  * for the separate enqueue and dequeue threads they take in one param
142  * and return two. Input = burst size, output = cycle average for sp/sc & mp/mc
143  */
144 struct thread_params {
145 	struct rte_ring *r;
146 	unsigned size;        /* input value, the burst size */
147 	double spsc, mpmc;    /* output value, the single or multi timings */
148 };
149 
150 /*
151  * Function that uses rdtsc to measure timing for ring enqueue. Needs pair
152  * thread running dequeue_bulk function
153  */
154 static int
155 enqueue_bulk(void *p)
156 {
157 	const unsigned iter_shift = 23;
158 	const unsigned iterations = 1<<iter_shift;
159 	struct thread_params *params = p;
160 	struct rte_ring *r = params->r;
161 	const unsigned size = params->size;
162 	unsigned i;
163 	void *burst[MAX_BURST] = {0};
164 
165 #ifdef RTE_USE_C11_MEM_MODEL
166 	if (__atomic_add_fetch(&lcore_count, 1, __ATOMIC_RELAXED) != 2)
167 #else
168 	if (__sync_add_and_fetch(&lcore_count, 1) != 2)
169 #endif
170 		while(lcore_count != 2)
171 			rte_pause();
172 
173 	const uint64_t sp_start = rte_rdtsc();
174 	for (i = 0; i < iterations; i++)
175 		while (rte_ring_sp_enqueue_bulk(r, burst, size, NULL) == 0)
176 			rte_pause();
177 	const uint64_t sp_end = rte_rdtsc();
178 
179 	const uint64_t mp_start = rte_rdtsc();
180 	for (i = 0; i < iterations; i++)
181 		while (rte_ring_mp_enqueue_bulk(r, burst, size, NULL) == 0)
182 			rte_pause();
183 	const uint64_t mp_end = rte_rdtsc();
184 
185 	params->spsc = ((double)(sp_end - sp_start))/(iterations*size);
186 	params->mpmc = ((double)(mp_end - mp_start))/(iterations*size);
187 	return 0;
188 }
189 
190 /*
191  * Function that uses rdtsc to measure timing for ring dequeue. Needs pair
192  * thread running enqueue_bulk function
193  */
194 static int
195 dequeue_bulk(void *p)
196 {
197 	const unsigned iter_shift = 23;
198 	const unsigned iterations = 1<<iter_shift;
199 	struct thread_params *params = p;
200 	struct rte_ring *r = params->r;
201 	const unsigned size = params->size;
202 	unsigned i;
203 	void *burst[MAX_BURST] = {0};
204 
205 #ifdef RTE_USE_C11_MEM_MODEL
206 	if (__atomic_add_fetch(&lcore_count, 1, __ATOMIC_RELAXED) != 2)
207 #else
208 	if (__sync_add_and_fetch(&lcore_count, 1) != 2)
209 #endif
210 		while(lcore_count != 2)
211 			rte_pause();
212 
213 	const uint64_t sc_start = rte_rdtsc();
214 	for (i = 0; i < iterations; i++)
215 		while (rte_ring_sc_dequeue_bulk(r, burst, size, NULL) == 0)
216 			rte_pause();
217 	const uint64_t sc_end = rte_rdtsc();
218 
219 	const uint64_t mc_start = rte_rdtsc();
220 	for (i = 0; i < iterations; i++)
221 		while (rte_ring_mc_dequeue_bulk(r, burst, size, NULL) == 0)
222 			rte_pause();
223 	const uint64_t mc_end = rte_rdtsc();
224 
225 	params->spsc = ((double)(sc_end - sc_start))/(iterations*size);
226 	params->mpmc = ((double)(mc_end - mc_start))/(iterations*size);
227 	return 0;
228 }
229 
230 /*
231  * Function that calls the enqueue and dequeue bulk functions on pairs of cores.
232  * used to measure ring perf between hyperthreads, cores and sockets.
233  */
234 static void
235 run_on_core_pair(struct lcore_pair *cores, struct rte_ring *r,
236 		lcore_function_t f1, lcore_function_t f2)
237 {
238 	struct thread_params param1 = {0}, param2 = {0};
239 	unsigned i;
240 	for (i = 0; i < sizeof(bulk_sizes)/sizeof(bulk_sizes[0]); i++) {
241 		lcore_count = 0;
242 		param1.size = param2.size = bulk_sizes[i];
243 		param1.r = param2.r = r;
244 		if (cores->c1 == rte_get_master_lcore()) {
245 			rte_eal_remote_launch(f2, &param2, cores->c2);
246 			f1(&param1);
247 			rte_eal_wait_lcore(cores->c2);
248 		} else {
249 			rte_eal_remote_launch(f1, &param1, cores->c1);
250 			rte_eal_remote_launch(f2, &param2, cores->c2);
251 			rte_eal_wait_lcore(cores->c1);
252 			rte_eal_wait_lcore(cores->c2);
253 		}
254 		printf("SP/SC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[i],
255 				param1.spsc + param2.spsc);
256 		printf("MP/MC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[i],
257 				param1.mpmc + param2.mpmc);
258 	}
259 }
260 
261 /*
262  * Test function that determines how long an enqueue + dequeue of a single item
263  * takes on a single lcore. Result is for comparison with the bulk enq+deq.
264  */
265 static void
266 test_single_enqueue_dequeue(struct rte_ring *r)
267 {
268 	const unsigned iter_shift = 24;
269 	const unsigned iterations = 1<<iter_shift;
270 	unsigned i = 0;
271 	void *burst = NULL;
272 
273 	const uint64_t sc_start = rte_rdtsc();
274 	for (i = 0; i < iterations; i++) {
275 		rte_ring_sp_enqueue(r, burst);
276 		rte_ring_sc_dequeue(r, &burst);
277 	}
278 	const uint64_t sc_end = rte_rdtsc();
279 
280 	const uint64_t mc_start = rte_rdtsc();
281 	for (i = 0; i < iterations; i++) {
282 		rte_ring_mp_enqueue(r, burst);
283 		rte_ring_mc_dequeue(r, &burst);
284 	}
285 	const uint64_t mc_end = rte_rdtsc();
286 
287 	printf("SP/SC single enq/dequeue: %"PRIu64"\n",
288 			(sc_end-sc_start) >> iter_shift);
289 	printf("MP/MC single enq/dequeue: %"PRIu64"\n",
290 			(mc_end-mc_start) >> iter_shift);
291 }
292 
293 /*
294  * Test that does both enqueue and dequeue on a core using the burst() API calls
295  * instead of the bulk() calls used in other tests. Results should be the same
296  * as for the bulk function called on a single lcore.
297  */
298 static void
299 test_burst_enqueue_dequeue(struct rte_ring *r)
300 {
301 	const unsigned iter_shift = 23;
302 	const unsigned iterations = 1<<iter_shift;
303 	unsigned sz, i = 0;
304 	void *burst[MAX_BURST] = {0};
305 
306 	for (sz = 0; sz < sizeof(bulk_sizes)/sizeof(bulk_sizes[0]); sz++) {
307 		const uint64_t sc_start = rte_rdtsc();
308 		for (i = 0; i < iterations; i++) {
309 			rte_ring_sp_enqueue_burst(r, burst,
310 					bulk_sizes[sz], NULL);
311 			rte_ring_sc_dequeue_burst(r, burst,
312 					bulk_sizes[sz], NULL);
313 		}
314 		const uint64_t sc_end = rte_rdtsc();
315 
316 		const uint64_t mc_start = rte_rdtsc();
317 		for (i = 0; i < iterations; i++) {
318 			rte_ring_mp_enqueue_burst(r, burst,
319 					bulk_sizes[sz], NULL);
320 			rte_ring_mc_dequeue_burst(r, burst,
321 					bulk_sizes[sz], NULL);
322 		}
323 		const uint64_t mc_end = rte_rdtsc();
324 
325 		uint64_t mc_avg = ((mc_end-mc_start) >> iter_shift) / bulk_sizes[sz];
326 		uint64_t sc_avg = ((sc_end-sc_start) >> iter_shift) / bulk_sizes[sz];
327 
328 		printf("SP/SC burst enq/dequeue (size: %u): %"PRIu64"\n", bulk_sizes[sz],
329 				sc_avg);
330 		printf("MP/MC burst enq/dequeue (size: %u): %"PRIu64"\n", bulk_sizes[sz],
331 				mc_avg);
332 	}
333 }
334 
335 /* Times enqueue and dequeue on a single lcore */
336 static void
337 test_bulk_enqueue_dequeue(struct rte_ring *r)
338 {
339 	const unsigned iter_shift = 23;
340 	const unsigned iterations = 1<<iter_shift;
341 	unsigned sz, i = 0;
342 	void *burst[MAX_BURST] = {0};
343 
344 	for (sz = 0; sz < sizeof(bulk_sizes)/sizeof(bulk_sizes[0]); sz++) {
345 		const uint64_t sc_start = rte_rdtsc();
346 		for (i = 0; i < iterations; i++) {
347 			rte_ring_sp_enqueue_bulk(r, burst,
348 					bulk_sizes[sz], NULL);
349 			rte_ring_sc_dequeue_bulk(r, burst,
350 					bulk_sizes[sz], NULL);
351 		}
352 		const uint64_t sc_end = rte_rdtsc();
353 
354 		const uint64_t mc_start = rte_rdtsc();
355 		for (i = 0; i < iterations; i++) {
356 			rte_ring_mp_enqueue_bulk(r, burst,
357 					bulk_sizes[sz], NULL);
358 			rte_ring_mc_dequeue_bulk(r, burst,
359 					bulk_sizes[sz], NULL);
360 		}
361 		const uint64_t mc_end = rte_rdtsc();
362 
363 		double sc_avg = ((double)(sc_end-sc_start) /
364 				(iterations * bulk_sizes[sz]));
365 		double mc_avg = ((double)(mc_end-mc_start) /
366 				(iterations * bulk_sizes[sz]));
367 
368 		printf("SP/SC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[sz],
369 				sc_avg);
370 		printf("MP/MC bulk enq/dequeue (size: %u): %.2F\n", bulk_sizes[sz],
371 				mc_avg);
372 	}
373 }
374 
375 static int
376 test_ring_perf(void)
377 {
378 	struct lcore_pair cores;
379 	struct rte_ring *r = NULL;
380 
381 	r = rte_ring_create(RING_NAME, RING_SIZE, rte_socket_id(), 0);
382 	if (r == NULL)
383 		return -1;
384 
385 	printf("### Testing single element and burst enq/deq ###\n");
386 	test_single_enqueue_dequeue(r);
387 	test_burst_enqueue_dequeue(r);
388 
389 	printf("\n### Testing empty dequeue ###\n");
390 	test_empty_dequeue(r);
391 
392 	printf("\n### Testing using a single lcore ###\n");
393 	test_bulk_enqueue_dequeue(r);
394 
395 	if (get_two_hyperthreads(&cores) == 0) {
396 		printf("\n### Testing using two hyperthreads ###\n");
397 		run_on_core_pair(&cores, r, enqueue_bulk, dequeue_bulk);
398 	}
399 	if (get_two_cores(&cores) == 0) {
400 		printf("\n### Testing using two physical cores ###\n");
401 		run_on_core_pair(&cores, r, enqueue_bulk, dequeue_bulk);
402 	}
403 	if (get_two_sockets(&cores) == 0) {
404 		printf("\n### Testing using two NUMA nodes ###\n");
405 		run_on_core_pair(&cores, r, enqueue_bulk, dequeue_bulk);
406 	}
407 	rte_ring_free(r);
408 	return 0;
409 }
410 
411 REGISTER_TEST_COMMAND(ring_perf_autotest, test_ring_perf);
412