]> git.proxmox.com Git - ceph.git/blob - ceph/src/spdk/dpdk/app/test/test_stack_perf.c
import 15.2.0 Octopus source
[ceph.git] / ceph / src / spdk / dpdk / app / test / test_stack_perf.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2019 Intel Corporation
3 */
4
5
6 #include <stdio.h>
7 #include <inttypes.h>
8
9 #include <rte_atomic.h>
10 #include <rte_cycles.h>
11 #include <rte_launch.h>
12 #include <rte_pause.h>
13 #include <rte_stack.h>
14
15 #include "test.h"
16
17 #define STACK_NAME "STACK_PERF"
18 #define MAX_BURST 32
19 #define STACK_SIZE (RTE_MAX_LCORE * MAX_BURST)
20
21 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
22
23 /*
24 * Push/pop bulk sizes, marked volatile so they aren't treated as compile-time
25 * constants.
26 */
27 static volatile unsigned int bulk_sizes[] = {8, MAX_BURST};
28
29 static rte_atomic32_t lcore_barrier;
30
31 struct lcore_pair {
32 unsigned int c1;
33 unsigned int c2;
34 };
35
36 static int
37 get_two_hyperthreads(struct lcore_pair *lcp)
38 {
39 unsigned int socket[2];
40 unsigned int core[2];
41 unsigned int id[2];
42
43 RTE_LCORE_FOREACH(id[0]) {
44 RTE_LCORE_FOREACH(id[1]) {
45 if (id[0] == id[1])
46 continue;
47 core[0] = lcore_config[id[0]].core_id;
48 core[1] = lcore_config[id[1]].core_id;
49 socket[0] = lcore_config[id[0]].socket_id;
50 socket[1] = lcore_config[id[1]].socket_id;
51 if ((core[0] == core[1]) && (socket[0] == socket[1])) {
52 lcp->c1 = id[0];
53 lcp->c2 = id[1];
54 return 0;
55 }
56 }
57 }
58
59 return 1;
60 }
61
62 static int
63 get_two_cores(struct lcore_pair *lcp)
64 {
65 unsigned int socket[2];
66 unsigned int core[2];
67 unsigned int id[2];
68
69 RTE_LCORE_FOREACH(id[0]) {
70 RTE_LCORE_FOREACH(id[1]) {
71 if (id[0] == id[1])
72 continue;
73 core[0] = lcore_config[id[0]].core_id;
74 core[1] = lcore_config[id[1]].core_id;
75 socket[0] = lcore_config[id[0]].socket_id;
76 socket[1] = lcore_config[id[1]].socket_id;
77 if ((core[0] != core[1]) && (socket[0] == socket[1])) {
78 lcp->c1 = id[0];
79 lcp->c2 = id[1];
80 return 0;
81 }
82 }
83 }
84
85 return 1;
86 }
87
88 static int
89 get_two_sockets(struct lcore_pair *lcp)
90 {
91 unsigned int socket[2];
92 unsigned int id[2];
93
94 RTE_LCORE_FOREACH(id[0]) {
95 RTE_LCORE_FOREACH(id[1]) {
96 if (id[0] == id[1])
97 continue;
98 socket[0] = lcore_config[id[0]].socket_id;
99 socket[1] = lcore_config[id[1]].socket_id;
100 if (socket[0] != socket[1]) {
101 lcp->c1 = id[0];
102 lcp->c2 = id[1];
103 return 0;
104 }
105 }
106 }
107
108 return 1;
109 }
110
111 /* Measure the cycle cost of popping an empty stack. */
112 static void
113 test_empty_pop(struct rte_stack *s)
114 {
115 unsigned int iterations = 100000000;
116 void *objs[MAX_BURST];
117 unsigned int i;
118
119 uint64_t start = rte_rdtsc();
120
121 for (i = 0; i < iterations; i++)
122 rte_stack_pop(s, objs, bulk_sizes[0]);
123
124 uint64_t end = rte_rdtsc();
125
126 printf("Stack empty pop: %.2F\n",
127 (double)(end - start) / iterations);
128 }
129
130 struct thread_args {
131 struct rte_stack *s;
132 unsigned int sz;
133 double avg;
134 };
135
136 /* Measure the average per-pointer cycle cost of stack push and pop */
137 static int
138 bulk_push_pop(void *p)
139 {
140 unsigned int iterations = 1000000;
141 struct thread_args *args = p;
142 void *objs[MAX_BURST] = {0};
143 unsigned int size, i;
144 struct rte_stack *s;
145
146 s = args->s;
147 size = args->sz;
148
149 rte_atomic32_sub(&lcore_barrier, 1);
150 while (rte_atomic32_read(&lcore_barrier) != 0)
151 rte_pause();
152
153 uint64_t start = rte_rdtsc();
154
155 for (i = 0; i < iterations; i++) {
156 rte_stack_push(s, objs, size);
157 rte_stack_pop(s, objs, size);
158 }
159
160 uint64_t end = rte_rdtsc();
161
162 args->avg = ((double)(end - start))/(iterations * size);
163
164 return 0;
165 }
166
167 /*
168 * Run bulk_push_pop() simultaneously on pairs of cores, to measure stack
169 * perf when between hyperthread siblings, cores on the same socket, and cores
170 * on different sockets.
171 */
172 static void
173 run_on_core_pair(struct lcore_pair *cores, struct rte_stack *s,
174 lcore_function_t fn)
175 {
176 struct thread_args args[2];
177 unsigned int i;
178
179 for (i = 0; i < ARRAY_SIZE(bulk_sizes); i++) {
180 rte_atomic32_set(&lcore_barrier, 2);
181
182 args[0].sz = args[1].sz = bulk_sizes[i];
183 args[0].s = args[1].s = s;
184
185 if (cores->c1 == rte_get_master_lcore()) {
186 rte_eal_remote_launch(fn, &args[1], cores->c2);
187 fn(&args[0]);
188 rte_eal_wait_lcore(cores->c2);
189 } else {
190 rte_eal_remote_launch(fn, &args[0], cores->c1);
191 rte_eal_remote_launch(fn, &args[1], cores->c2);
192 rte_eal_wait_lcore(cores->c1);
193 rte_eal_wait_lcore(cores->c2);
194 }
195
196 printf("Average cycles per object push/pop (bulk size: %u): %.2F\n",
197 bulk_sizes[i], (args[0].avg + args[1].avg) / 2);
198 }
199 }
200
201 /* Run bulk_push_pop() simultaneously on 1+ cores. */
202 static void
203 run_on_n_cores(struct rte_stack *s, lcore_function_t fn, int n)
204 {
205 struct thread_args args[RTE_MAX_LCORE];
206 unsigned int i;
207
208 for (i = 0; i < ARRAY_SIZE(bulk_sizes); i++) {
209 unsigned int lcore_id;
210 int cnt = 0;
211 double avg;
212
213 rte_atomic32_set(&lcore_barrier, n);
214
215 RTE_LCORE_FOREACH_SLAVE(lcore_id) {
216 if (++cnt >= n)
217 break;
218
219 args[lcore_id].s = s;
220 args[lcore_id].sz = bulk_sizes[i];
221
222 if (rte_eal_remote_launch(fn, &args[lcore_id],
223 lcore_id))
224 rte_panic("Failed to launch lcore %d\n",
225 lcore_id);
226 }
227
228 lcore_id = rte_lcore_id();
229
230 args[lcore_id].s = s;
231 args[lcore_id].sz = bulk_sizes[i];
232
233 fn(&args[lcore_id]);
234
235 rte_eal_mp_wait_lcore();
236
237 avg = args[rte_lcore_id()].avg;
238
239 cnt = 0;
240 RTE_LCORE_FOREACH_SLAVE(lcore_id) {
241 if (++cnt >= n)
242 break;
243 avg += args[lcore_id].avg;
244 }
245
246 printf("Average cycles per object push/pop (bulk size: %u): %.2F\n",
247 bulk_sizes[i], avg / n);
248 }
249 }
250
251 /*
252 * Measure the cycle cost of pushing and popping a single pointer on a single
253 * lcore.
254 */
255 static void
256 test_single_push_pop(struct rte_stack *s)
257 {
258 unsigned int iterations = 16000000;
259 void *obj = NULL;
260 unsigned int i;
261
262 uint64_t start = rte_rdtsc();
263
264 for (i = 0; i < iterations; i++) {
265 rte_stack_push(s, &obj, 1);
266 rte_stack_pop(s, &obj, 1);
267 }
268
269 uint64_t end = rte_rdtsc();
270
271 printf("Average cycles per single object push/pop: %.2F\n",
272 ((double)(end - start)) / iterations);
273 }
274
275 /* Measure the cycle cost of bulk pushing and popping on a single lcore. */
276 static void
277 test_bulk_push_pop(struct rte_stack *s)
278 {
279 unsigned int iterations = 8000000;
280 void *objs[MAX_BURST];
281 unsigned int sz, i;
282
283 for (sz = 0; sz < ARRAY_SIZE(bulk_sizes); sz++) {
284 uint64_t start = rte_rdtsc();
285
286 for (i = 0; i < iterations; i++) {
287 rte_stack_push(s, objs, bulk_sizes[sz]);
288 rte_stack_pop(s, objs, bulk_sizes[sz]);
289 }
290
291 uint64_t end = rte_rdtsc();
292
293 double avg = ((double)(end - start) /
294 (iterations * bulk_sizes[sz]));
295
296 printf("Average cycles per object push/pop (bulk size: %u): %.2F\n",
297 bulk_sizes[sz], avg);
298 }
299 }
300
301 static int
302 __test_stack_perf(uint32_t flags)
303 {
304 struct lcore_pair cores;
305 struct rte_stack *s;
306
307 rte_atomic32_init(&lcore_barrier);
308
309 s = rte_stack_create(STACK_NAME, STACK_SIZE, rte_socket_id(), flags);
310 if (s == NULL) {
311 printf("[%s():%u] failed to create a stack\n",
312 __func__, __LINE__);
313 return -1;
314 }
315
316 printf("### Testing single element push/pop ###\n");
317 test_single_push_pop(s);
318
319 printf("\n### Testing empty pop ###\n");
320 test_empty_pop(s);
321
322 printf("\n### Testing using a single lcore ###\n");
323 test_bulk_push_pop(s);
324
325 if (get_two_hyperthreads(&cores) == 0) {
326 printf("\n### Testing using two hyperthreads ###\n");
327 run_on_core_pair(&cores, s, bulk_push_pop);
328 }
329 if (get_two_cores(&cores) == 0) {
330 printf("\n### Testing using two physical cores ###\n");
331 run_on_core_pair(&cores, s, bulk_push_pop);
332 }
333 if (get_two_sockets(&cores) == 0) {
334 printf("\n### Testing using two NUMA nodes ###\n");
335 run_on_core_pair(&cores, s, bulk_push_pop);
336 }
337
338 printf("\n### Testing on all %u lcores ###\n", rte_lcore_count());
339 run_on_n_cores(s, bulk_push_pop, rte_lcore_count());
340
341 rte_stack_free(s);
342 return 0;
343 }
344
345 static int
346 test_stack_perf(void)
347 {
348 return __test_stack_perf(0);
349 }
350
351 static int
352 test_lf_stack_perf(void)
353 {
354 return __test_stack_perf(RTE_STACK_F_LF);
355 }
356
357 REGISTER_TEST_COMMAND(stack_perf_autotest, test_stack_perf);
358 REGISTER_TEST_COMMAND(stack_lf_perf_autotest, test_lf_stack_perf);