1 // SPDX-License-Identifier: GPL-2.0 1 // SPDX-License-Identifier: GPL-2.0
2 /* 2 /*
3 * numa.c 3 * numa.c
4 * 4 *
5 * numa: Simulate NUMA-sensitive workload and 5 * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
6 */ 6 */
7 7
8 #include <inttypes.h> 8 #include <inttypes.h>
>> 9 /* For the CLR_() macros */
>> 10 #include <pthread.h>
9 11
10 #include <subcmd/parse-options.h> 12 #include <subcmd/parse-options.h>
11 #include "../util/cloexec.h" 13 #include "../util/cloexec.h"
12 14
13 #include "bench.h" 15 #include "bench.h"
14 16
15 #include <errno.h> 17 #include <errno.h>
16 #include <sched.h> 18 #include <sched.h>
17 #include <stdio.h> 19 #include <stdio.h>
18 #include <assert.h> 20 #include <assert.h>
19 #include <debug.h> <<
20 #include <malloc.h> 21 #include <malloc.h>
21 #include <signal.h> 22 #include <signal.h>
22 #include <stdlib.h> 23 #include <stdlib.h>
23 #include <string.h> 24 #include <string.h>
24 #include <unistd.h> 25 #include <unistd.h>
25 #include <sys/mman.h> 26 #include <sys/mman.h>
26 #include <sys/time.h> 27 #include <sys/time.h>
27 #include <sys/resource.h> 28 #include <sys/resource.h>
28 #include <sys/wait.h> 29 #include <sys/wait.h>
29 #include <sys/prctl.h> 30 #include <sys/prctl.h>
30 #include <sys/types.h> 31 #include <sys/types.h>
31 #include <linux/kernel.h> 32 #include <linux/kernel.h>
32 #include <linux/time64.h> 33 #include <linux/time64.h>
33 #include <linux/numa.h> 34 #include <linux/numa.h>
34 #include <linux/zalloc.h> 35 #include <linux/zalloc.h>
35 36
36 #include "../util/header.h" <<
37 #include "../util/mutex.h" <<
38 #include <numa.h> 37 #include <numa.h>
39 #include <numaif.h> 38 #include <numaif.h>
40 39
41 #ifndef RUSAGE_THREAD 40 #ifndef RUSAGE_THREAD
42 # define RUSAGE_THREAD 1 41 # define RUSAGE_THREAD 1
43 #endif 42 #endif
44 43
45 /* 44 /*
46 * Regular printout to the terminal, suppresse 45 * Regular printout to the terminal, suppressed if -q is specified:
47 */ 46 */
48 #define tprintf(x...) do { if (g && g->p.show_ 47 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
49 48
50 /* 49 /*
51 * Debug printf: 50 * Debug printf:
52 */ 51 */
53 #undef dprintf 52 #undef dprintf
54 #define dprintf(x...) do { if (g && g->p.show_ 53 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
55 54
56 struct thread_data { 55 struct thread_data {
57 int curr_cpu; 56 int curr_cpu;
58 cpu_set_t *bind_cpumask; !! 57 cpu_set_t bind_cpumask;
59 int bind_node; 58 int bind_node;
60 u8 *process_data; 59 u8 *process_data;
61 int process_nr; 60 int process_nr;
62 int thread_nr; 61 int thread_nr;
63 int task_nr; 62 int task_nr;
64 unsigned int loops_done; 63 unsigned int loops_done;
65 u64 val; 64 u64 val;
66 u64 runtime_ns; 65 u64 runtime_ns;
67 u64 system_time_ns 66 u64 system_time_ns;
68 u64 user_time_ns; 67 u64 user_time_ns;
69 double speed_gbs; 68 double speed_gbs;
70 struct mutex *process_lock; !! 69 pthread_mutex_t *process_lock;
71 }; 70 };
72 71
73 /* Parameters set by options: */ 72 /* Parameters set by options: */
74 73
75 struct params { 74 struct params {
76 /* Startup synchronization: */ 75 /* Startup synchronization: */
77 bool serialize_star 76 bool serialize_startup;
78 77
79 /* Task hierarchy: */ 78 /* Task hierarchy: */
80 int nr_proc; 79 int nr_proc;
81 int nr_threads; 80 int nr_threads;
82 81
83 /* Working set sizes: */ 82 /* Working set sizes: */
84 const char *mb_global_str 83 const char *mb_global_str;
85 const char *mb_proc_str; 84 const char *mb_proc_str;
86 const char *mb_proc_locke 85 const char *mb_proc_locked_str;
87 const char *mb_thread_str 86 const char *mb_thread_str;
88 87
89 double mb_global; 88 double mb_global;
90 double mb_proc; 89 double mb_proc;
91 double mb_proc_locked 90 double mb_proc_locked;
92 double mb_thread; 91 double mb_thread;
93 92
94 /* Access patterns to the working set: 93 /* Access patterns to the working set: */
95 bool data_reads; 94 bool data_reads;
96 bool data_writes; 95 bool data_writes;
97 bool data_backwards 96 bool data_backwards;
98 bool data_zero_mems 97 bool data_zero_memset;
99 bool data_rand_walk 98 bool data_rand_walk;
100 u32 nr_loops; 99 u32 nr_loops;
101 u32 nr_secs; 100 u32 nr_secs;
102 u32 sleep_usecs; 101 u32 sleep_usecs;
103 102
104 /* Working set initialization: */ 103 /* Working set initialization: */
105 bool init_zero; 104 bool init_zero;
106 bool init_random; 105 bool init_random;
107 bool init_cpu0; 106 bool init_cpu0;
108 107
109 /* Misc options: */ 108 /* Misc options: */
110 int show_details; 109 int show_details;
111 int run_all; 110 int run_all;
112 int thp; 111 int thp;
113 112
114 long bytes_global; 113 long bytes_global;
115 long bytes_process; 114 long bytes_process;
116 long bytes_process_ 115 long bytes_process_locked;
117 long bytes_thread; 116 long bytes_thread;
118 117
119 int nr_tasks; 118 int nr_tasks;
>> 119 bool show_quiet;
120 120
121 bool show_convergen 121 bool show_convergence;
122 bool measure_conver 122 bool measure_convergence;
123 123
124 int perturb_secs; 124 int perturb_secs;
125 int nr_cpus; 125 int nr_cpus;
126 int nr_nodes; 126 int nr_nodes;
127 127
128 /* Affinity options -C and -N: */ 128 /* Affinity options -C and -N: */
129 char *cpu_list_str; 129 char *cpu_list_str;
130 char *node_list_str 130 char *node_list_str;
131 }; 131 };
132 132
133 133
134 /* Global, read-writable area, accessible to a 134 /* Global, read-writable area, accessible to all processes and threads: */
135 135
136 struct global_info { 136 struct global_info {
137 u8 *data; 137 u8 *data;
138 138
139 struct mutex startup_mutex; !! 139 pthread_mutex_t startup_mutex;
140 struct cond startup_cond; !! 140 pthread_cond_t startup_cond;
141 int nr_tasks_start 141 int nr_tasks_started;
142 142
143 struct mutex start_work_mut !! 143 pthread_mutex_t start_work_mutex;
144 struct cond start_work_con !! 144 pthread_cond_t start_work_cond;
145 int nr_tasks_worki 145 int nr_tasks_working;
146 bool start_work; 146 bool start_work;
147 147
148 struct mutex stop_work_mute !! 148 pthread_mutex_t stop_work_mutex;
149 u64 bytes_done; 149 u64 bytes_done;
150 150
151 struct thread_data *threads; 151 struct thread_data *threads;
152 152
153 /* Convergence latency measurement: */ 153 /* Convergence latency measurement: */
154 bool all_converged; 154 bool all_converged;
155 bool stop_work; 155 bool stop_work;
156 156
157 int print_once; 157 int print_once;
158 158
159 struct params p; 159 struct params p;
160 }; 160 };
161 161
162 static struct global_info *g = NULL; 162 static struct global_info *g = NULL;
163 163
164 static int parse_cpus_opt(const struct option 164 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
165 static int parse_nodes_opt(const struct option 165 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
166 166
167 struct params p0; 167 struct params p0;
168 168
169 static const struct option options[] = { 169 static const struct option options[] = {
170 OPT_INTEGER('p', "nr_proc" , &p0. 170 OPT_INTEGER('p', "nr_proc" , &p0.nr_proc, "number of processes"),
171 OPT_INTEGER('t', "nr_threads" , &p0. 171 OPT_INTEGER('t', "nr_threads" , &p0.nr_threads, "number of threads per process"),
172 172
173 OPT_STRING('G', "mb_global" , &p0. 173 OPT_STRING('G', "mb_global" , &p0.mb_global_str, "MB", "global memory (MBs)"),
174 OPT_STRING('P', "mb_proc" , &p0. 174 OPT_STRING('P', "mb_proc" , &p0.mb_proc_str, "MB", "process memory (MBs)"),
175 OPT_STRING('L', "mb_proc_locked", &p0. 175 OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
176 OPT_STRING('T', "mb_thread" , &p0. 176 OPT_STRING('T', "mb_thread" , &p0.mb_thread_str, "MB", "thread memory (MBs)"),
177 177
178 OPT_UINTEGER('l', "nr_loops" , &p0. 178 OPT_UINTEGER('l', "nr_loops" , &p0.nr_loops, "max number of loops to run (default: unlimited)"),
179 OPT_UINTEGER('s', "nr_secs" , &p0. 179 OPT_UINTEGER('s', "nr_secs" , &p0.nr_secs, "max number of seconds to run (default: 5 secs)"),
180 OPT_UINTEGER('u', "usleep" , &p0. 180 OPT_UINTEGER('u', "usleep" , &p0.sleep_usecs, "usecs to sleep per loop iteration"),
181 181
182 OPT_BOOLEAN('R', "data_reads" , &p0. 182 OPT_BOOLEAN('R', "data_reads" , &p0.data_reads, "access the data via reads (can be mixed with -W)"),
183 OPT_BOOLEAN('W', "data_writes" , &p0. 183 OPT_BOOLEAN('W', "data_writes" , &p0.data_writes, "access the data via writes (can be mixed with -R)"),
184 OPT_BOOLEAN('B', "data_backwards", &p0 184 OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards, "access the data backwards as well"),
185 OPT_BOOLEAN('Z', "data_zero_memset", & 185 OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
186 OPT_BOOLEAN('r', "data_rand_walk", &p0 186 OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk, "access the data with random (32bit LFSR) walk"),
187 187
188 188
189 OPT_BOOLEAN('z', "init_zero" , &p0. 189 OPT_BOOLEAN('z', "init_zero" , &p0.init_zero, "bzero the initial allocations"),
190 OPT_BOOLEAN('I', "init_random" , &p0. 190 OPT_BOOLEAN('I', "init_random" , &p0.init_random, "randomize the contents of the initial allocations"),
191 OPT_BOOLEAN('', "init_cpu0" , &p0.i 191 OPT_BOOLEAN('', "init_cpu0" , &p0.init_cpu0, "do the initial allocations on CPU#0"),
192 OPT_INTEGER('x', "perturb_secs", &p0.p 192 OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs, "perturb thread 0/0 every X secs, to test convergence stability"),
193 193
194 OPT_INCR ('d', "show_details" , &p0. 194 OPT_INCR ('d', "show_details" , &p0.show_details, "Show details"),
195 OPT_INCR ('a', "all" , &p0. 195 OPT_INCR ('a', "all" , &p0.run_all, "Run all tests in the suite"),
196 OPT_INTEGER('H', "thp" , &p0. 196 OPT_INTEGER('H', "thp" , &p0.thp, "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
197 OPT_BOOLEAN('c', "show_convergence", & 197 OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
198 "convergence is reached wh 198 "convergence is reached when each process (all its threads) is running on a single NUMA node."),
199 OPT_BOOLEAN('m', "measure_convergence" 199 OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
200 OPT_BOOLEAN('q', "quiet" , &qui !! 200 OPT_BOOLEAN('q', "quiet" , &p0.show_quiet, "quiet mode"),
201 "quiet mode (do not show a <<
202 OPT_BOOLEAN('S', "serialize-startup", 201 OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
203 202
204 /* Special option string parsing callb 203 /* Special option string parsing callbacks: */
205 OPT_CALLBACK('C', "cpus", NULL, "cpu[, 204 OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
206 "bind the first N task 205 "bind the first N tasks to these specific cpus (the rest is unbound)",
207 parse_cpus_opt), 206 parse_cpus_opt),
208 OPT_CALLBACK('M', "memnodes", NULL, "n 207 OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
209 "bind the first N task 208 "bind the first N tasks to these specific memory nodes (the rest is unbound)",
210 parse_nodes_opt), 209 parse_nodes_opt),
211 OPT_END() 210 OPT_END()
212 }; 211 };
213 212
214 static const char * const bench_numa_usage[] = 213 static const char * const bench_numa_usage[] = {
215 "perf bench numa <options>", 214 "perf bench numa <options>",
216 NULL 215 NULL
217 }; 216 };
218 217
219 static const char * const numa_usage[] = { 218 static const char * const numa_usage[] = {
220 "perf bench numa mem [<options>]", 219 "perf bench numa mem [<options>]",
221 NULL 220 NULL
222 }; 221 };
223 222
224 /* 223 /*
225 * To get number of numa nodes present. 224 * To get number of numa nodes present.
226 */ 225 */
227 static int nr_numa_nodes(void) 226 static int nr_numa_nodes(void)
228 { 227 {
229 int i, nr_nodes = 0; 228 int i, nr_nodes = 0;
230 229
231 for (i = 0; i < g->p.nr_nodes; i++) { 230 for (i = 0; i < g->p.nr_nodes; i++) {
232 if (numa_bitmask_isbitset(numa 231 if (numa_bitmask_isbitset(numa_nodes_ptr, i))
233 nr_nodes++; 232 nr_nodes++;
234 } 233 }
235 234
236 return nr_nodes; 235 return nr_nodes;
237 } 236 }
238 237
239 /* 238 /*
240 * To check if given numa node is present. 239 * To check if given numa node is present.
241 */ 240 */
242 static int is_node_present(int node) 241 static int is_node_present(int node)
243 { 242 {
244 return numa_bitmask_isbitset(numa_node 243 return numa_bitmask_isbitset(numa_nodes_ptr, node);
245 } 244 }
246 245
247 /* 246 /*
248 * To check given numa node has cpus. 247 * To check given numa node has cpus.
249 */ 248 */
250 static bool node_has_cpus(int node) 249 static bool node_has_cpus(int node)
251 { 250 {
252 struct bitmask *cpumask = numa_allocat 251 struct bitmask *cpumask = numa_allocate_cpumask();
253 bool ret = false; /* fall back to nocp 252 bool ret = false; /* fall back to nocpus */
254 int cpu; 253 int cpu;
255 254
256 BUG_ON(!cpumask); 255 BUG_ON(!cpumask);
257 if (!numa_node_to_cpus(node, cpumask)) 256 if (!numa_node_to_cpus(node, cpumask)) {
258 for (cpu = 0; cpu < (int)cpuma 257 for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
259 if (numa_bitmask_isbit 258 if (numa_bitmask_isbitset(cpumask, cpu)) {
260 ret = true; 259 ret = true;
261 break; 260 break;
262 } 261 }
263 } 262 }
264 } 263 }
265 numa_free_cpumask(cpumask); 264 numa_free_cpumask(cpumask);
266 265
267 return ret; 266 return ret;
268 } 267 }
269 268
270 static cpu_set_t *bind_to_cpu(int target_cpu) !! 269 static cpu_set_t bind_to_cpu(int target_cpu)
271 { 270 {
272 int nrcpus = numa_num_possible_cpus(); !! 271 cpu_set_t orig_mask, mask;
273 cpu_set_t *orig_mask, *mask; !! 272 int ret;
274 size_t size; <<
275 <<
276 orig_mask = CPU_ALLOC(nrcpus); <<
277 BUG_ON(!orig_mask); <<
278 size = CPU_ALLOC_SIZE(nrcpus); <<
279 CPU_ZERO_S(size, orig_mask); <<
280 <<
281 if (sched_getaffinity(0, size, orig_ma <<
282 goto err_out; <<
283 273
284 mask = CPU_ALLOC(nrcpus); !! 274 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
285 if (!mask) !! 275 BUG_ON(ret);
286 goto err_out; <<
287 276
288 CPU_ZERO_S(size, mask); !! 277 CPU_ZERO(&mask);
289 278
290 if (target_cpu == -1) { 279 if (target_cpu == -1) {
291 int cpu; 280 int cpu;
292 281
293 for (cpu = 0; cpu < g->p.nr_cp 282 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
294 CPU_SET_S(cpu, size, m !! 283 CPU_SET(cpu, &mask);
295 } else { 284 } else {
296 if (target_cpu < 0 || target_c !! 285 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
297 goto err; !! 286 CPU_SET(target_cpu, &mask);
298 <<
299 CPU_SET_S(target_cpu, size, ma <<
300 } 287 }
301 288
302 if (sched_setaffinity(0, size, mask)) !! 289 ret = sched_setaffinity(0, sizeof(mask), &mask);
303 goto err; !! 290 BUG_ON(ret);
304 291
305 return orig_mask; 292 return orig_mask;
306 <<
307 err: <<
308 CPU_FREE(mask); <<
309 err_out: <<
310 CPU_FREE(orig_mask); <<
311 <<
312 /* BUG_ON due to failure in allocation <<
313 BUG_ON(-1); <<
314 return NULL; <<
315 } 293 }
316 294
317 static cpu_set_t *bind_to_node(int target_node !! 295 static cpu_set_t bind_to_node(int target_node)
318 { 296 {
319 int nrcpus = numa_num_possible_cpus(); !! 297 cpu_set_t orig_mask, mask;
320 size_t size; <<
321 cpu_set_t *orig_mask, *mask; <<
322 int cpu; 298 int cpu;
>> 299 int ret;
323 300
324 orig_mask = CPU_ALLOC(nrcpus); !! 301 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
325 BUG_ON(!orig_mask); !! 302 BUG_ON(ret);
326 size = CPU_ALLOC_SIZE(nrcpus); <<
327 CPU_ZERO_S(size, orig_mask); <<
328 <<
329 if (sched_getaffinity(0, size, orig_ma <<
330 goto err_out; <<
331 <<
332 mask = CPU_ALLOC(nrcpus); <<
333 if (!mask) <<
334 goto err_out; <<
335 303
336 CPU_ZERO_S(size, mask); !! 304 CPU_ZERO(&mask);
337 305
338 if (target_node == NUMA_NO_NODE) { 306 if (target_node == NUMA_NO_NODE) {
339 for (cpu = 0; cpu < g->p.nr_cp 307 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
340 CPU_SET_S(cpu, size, m !! 308 CPU_SET(cpu, &mask);
341 } else { 309 } else {
342 struct bitmask *cpumask = numa 310 struct bitmask *cpumask = numa_allocate_cpumask();
343 311
344 if (!cpumask) !! 312 BUG_ON(!cpumask);
345 goto err; <<
346 <<
347 if (!numa_node_to_cpus(target_ 313 if (!numa_node_to_cpus(target_node, cpumask)) {
348 for (cpu = 0; cpu < (i 314 for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
349 if (numa_bitma 315 if (numa_bitmask_isbitset(cpumask, cpu))
350 CPU_SE !! 316 CPU_SET(cpu, &mask);
351 } 317 }
352 } 318 }
353 numa_free_cpumask(cpumask); 319 numa_free_cpumask(cpumask);
354 } 320 }
355 321
356 if (sched_setaffinity(0, size, mask)) !! 322 ret = sched_setaffinity(0, sizeof(mask), &mask);
357 goto err; !! 323 BUG_ON(ret);
358 324
359 return orig_mask; 325 return orig_mask;
360 <<
361 err: <<
362 CPU_FREE(mask); <<
363 err_out: <<
364 CPU_FREE(orig_mask); <<
365 <<
366 /* BUG_ON due to failure in allocation <<
367 BUG_ON(-1); <<
368 return NULL; <<
369 } 326 }
370 327
371 static void bind_to_cpumask(cpu_set_t *mask) !! 328 static void bind_to_cpumask(cpu_set_t mask)
372 { 329 {
373 int ret; 330 int ret;
374 size_t size = CPU_ALLOC_SIZE(numa_num_ <<
375 331
376 ret = sched_setaffinity(0, size, mask) !! 332 ret = sched_setaffinity(0, sizeof(mask), &mask);
377 if (ret) { !! 333 BUG_ON(ret);
378 CPU_FREE(mask); <<
379 BUG_ON(ret); <<
380 } <<
381 } 334 }
382 335
383 static void mempol_restore(void) 336 static void mempol_restore(void)
384 { 337 {
385 int ret; 338 int ret;
386 339
387 ret = set_mempolicy(MPOL_DEFAULT, NULL 340 ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
388 341
389 BUG_ON(ret); 342 BUG_ON(ret);
390 } 343 }
391 344
392 static void bind_to_memnode(int node) 345 static void bind_to_memnode(int node)
393 { 346 {
394 struct bitmask *node_mask; 347 struct bitmask *node_mask;
395 int ret; 348 int ret;
396 349
397 if (node == NUMA_NO_NODE) 350 if (node == NUMA_NO_NODE)
398 return; 351 return;
399 352
400 node_mask = numa_allocate_nodemask(); 353 node_mask = numa_allocate_nodemask();
401 BUG_ON(!node_mask); 354 BUG_ON(!node_mask);
402 355
403 numa_bitmask_clearall(node_mask); 356 numa_bitmask_clearall(node_mask);
404 numa_bitmask_setbit(node_mask, node); 357 numa_bitmask_setbit(node_mask, node);
405 358
406 ret = set_mempolicy(MPOL_BIND, node_ma 359 ret = set_mempolicy(MPOL_BIND, node_mask->maskp, node_mask->size + 1);
407 dprintf("binding to node %d, mask: %01 360 dprintf("binding to node %d, mask: %016lx => %d\n", node, *node_mask->maskp, ret);
408 361
409 numa_bitmask_free(node_mask); 362 numa_bitmask_free(node_mask);
410 BUG_ON(ret); 363 BUG_ON(ret);
411 } 364 }
412 365
413 #define HPSIZE (2*1024*1024) 366 #define HPSIZE (2*1024*1024)
414 367
415 #define set_taskname(fmt...) 368 #define set_taskname(fmt...) \
416 do { 369 do { \
417 char name[20]; 370 char name[20]; \
418 371 \
419 snprintf(name, 20, fmt); 372 snprintf(name, 20, fmt); \
420 prctl(PR_SET_NAME, name); 373 prctl(PR_SET_NAME, name); \
421 } while (0) 374 } while (0)
422 375
423 static u8 *alloc_data(ssize_t bytes0, int map_ 376 static u8 *alloc_data(ssize_t bytes0, int map_flags,
424 int init_zero, int init_ 377 int init_zero, int init_cpu0, int thp, int init_random)
425 { 378 {
426 cpu_set_t *orig_mask = NULL; !! 379 cpu_set_t orig_mask;
427 ssize_t bytes; 380 ssize_t bytes;
428 u8 *buf; 381 u8 *buf;
429 int ret; 382 int ret;
430 383
431 if (!bytes0) 384 if (!bytes0)
432 return NULL; 385 return NULL;
433 386
434 /* Allocate and initialize all memory 387 /* Allocate and initialize all memory on CPU#0: */
435 if (init_cpu0) { 388 if (init_cpu0) {
436 int node = numa_node_of_cpu(0) 389 int node = numa_node_of_cpu(0);
437 390
438 orig_mask = bind_to_node(node) 391 orig_mask = bind_to_node(node);
439 bind_to_memnode(node); 392 bind_to_memnode(node);
440 } 393 }
441 394
442 bytes = bytes0 + HPSIZE; 395 bytes = bytes0 + HPSIZE;
443 396
444 buf = (void *)mmap(0, bytes, PROT_READ 397 buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
445 BUG_ON(buf == (void *)-1); 398 BUG_ON(buf == (void *)-1);
446 399
447 if (map_flags == MAP_PRIVATE) { 400 if (map_flags == MAP_PRIVATE) {
448 if (thp > 0) { 401 if (thp > 0) {
449 ret = madvise(buf, byt 402 ret = madvise(buf, bytes, MADV_HUGEPAGE);
450 if (ret && !g->print_o 403 if (ret && !g->print_once) {
451 g->print_once 404 g->print_once = 1;
452 printf("WARNIN 405 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
453 } 406 }
454 } 407 }
455 if (thp < 0) { 408 if (thp < 0) {
456 ret = madvise(buf, byt 409 ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
457 if (ret && !g->print_o 410 if (ret && !g->print_once) {
458 g->print_once 411 g->print_once = 1;
459 printf("WARNIN 412 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
460 } 413 }
461 } 414 }
462 } 415 }
463 416
464 if (init_zero) { 417 if (init_zero) {
465 bzero(buf, bytes); 418 bzero(buf, bytes);
466 } else { 419 } else {
467 /* Initialize random contents, 420 /* Initialize random contents, different in each word: */
468 if (init_random) { 421 if (init_random) {
469 u64 *wbuf = (void *)bu 422 u64 *wbuf = (void *)buf;
470 long off = rand(); 423 long off = rand();
471 long i; 424 long i;
472 425
473 for (i = 0; i < bytes/ 426 for (i = 0; i < bytes/8; i++)
474 wbuf[i] = i + 427 wbuf[i] = i + off;
475 } 428 }
476 } 429 }
477 430
478 /* Align to 2MB boundary: */ 431 /* Align to 2MB boundary: */
479 buf = (void *)(((unsigned long)buf + H 432 buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
480 433
481 /* Restore affinity: */ 434 /* Restore affinity: */
482 if (init_cpu0) { 435 if (init_cpu0) {
483 bind_to_cpumask(orig_mask); 436 bind_to_cpumask(orig_mask);
484 CPU_FREE(orig_mask); <<
485 mempol_restore(); 437 mempol_restore();
486 } 438 }
487 439
488 return buf; 440 return buf;
489 } 441 }
490 442
491 static void free_data(void *data, ssize_t byte 443 static void free_data(void *data, ssize_t bytes)
492 { 444 {
493 int ret; 445 int ret;
494 446
495 if (!data) 447 if (!data)
496 return; 448 return;
497 449
498 ret = munmap(data, bytes); 450 ret = munmap(data, bytes);
499 BUG_ON(ret); 451 BUG_ON(ret);
500 } 452 }
501 453
502 /* 454 /*
503 * Create a shared memory buffer that can be s 455 * Create a shared memory buffer that can be shared between processes, zeroed:
504 */ 456 */
505 static void * zalloc_shared_data(ssize_t bytes 457 static void * zalloc_shared_data(ssize_t bytes)
506 { 458 {
507 return alloc_data(bytes, MAP_SHARED, 1 459 return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0, g->p.thp, g->p.init_random);
508 } 460 }
509 461
510 /* 462 /*
511 * Create a shared memory buffer that can be s 463 * Create a shared memory buffer that can be shared between processes:
512 */ 464 */
513 static void * setup_shared_data(ssize_t bytes) 465 static void * setup_shared_data(ssize_t bytes)
514 { 466 {
515 return alloc_data(bytes, MAP_SHARED, 0 467 return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
516 } 468 }
517 469
518 /* 470 /*
519 * Allocate process-local memory - this will e 471 * Allocate process-local memory - this will either be shared between
520 * threads of this process, or only be accesse 472 * threads of this process, or only be accessed by this thread:
521 */ 473 */
522 static void * setup_private_data(ssize_t bytes 474 static void * setup_private_data(ssize_t bytes)
523 { 475 {
524 return alloc_data(bytes, MAP_PRIVATE, 476 return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
525 } 477 }
526 478
>> 479 /*
>> 480 * Return a process-shared (global) mutex:
>> 481 */
>> 482 static void init_global_mutex(pthread_mutex_t *mutex)
>> 483 {
>> 484 pthread_mutexattr_t attr;
>> 485
>> 486 pthread_mutexattr_init(&attr);
>> 487 pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
>> 488 pthread_mutex_init(mutex, &attr);
>> 489 }
>> 490
>> 491 /*
>> 492 * Return a process-shared (global) condition variable:
>> 493 */
>> 494 static void init_global_cond(pthread_cond_t *cond)
>> 495 {
>> 496 pthread_condattr_t attr;
>> 497
>> 498 pthread_condattr_init(&attr);
>> 499 pthread_condattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
>> 500 pthread_cond_init(cond, &attr);
>> 501 }
>> 502
527 static int parse_cpu_list(const char *arg) 503 static int parse_cpu_list(const char *arg)
528 { 504 {
529 p0.cpu_list_str = strdup(arg); 505 p0.cpu_list_str = strdup(arg);
530 506
531 dprintf("got CPU list: {%s}\n", p0.cpu 507 dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
532 508
533 return 0; 509 return 0;
534 } 510 }
535 511
536 static int parse_setup_cpu_list(void) 512 static int parse_setup_cpu_list(void)
537 { 513 {
538 struct thread_data *td; 514 struct thread_data *td;
539 char *str0, *str; 515 char *str0, *str;
540 int t; 516 int t;
541 517
542 if (!g->p.cpu_list_str) 518 if (!g->p.cpu_list_str)
543 return 0; 519 return 0;
544 520
545 dprintf("g->p.nr_tasks: %d\n", g->p.nr 521 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
546 522
547 str0 = str = strdup(g->p.cpu_list_str) 523 str0 = str = strdup(g->p.cpu_list_str);
548 t = 0; 524 t = 0;
549 525
550 BUG_ON(!str); 526 BUG_ON(!str);
551 527
552 tprintf("# binding tasks to CPUs:\n"); 528 tprintf("# binding tasks to CPUs:\n");
553 tprintf("# "); 529 tprintf("# ");
554 530
555 while (true) { 531 while (true) {
556 int bind_cpu, bind_cpu_0, bind 532 int bind_cpu, bind_cpu_0, bind_cpu_1;
557 char *tok, *tok_end, *tok_step 533 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
558 int bind_len; 534 int bind_len;
559 int step; 535 int step;
560 int mul; 536 int mul;
561 537
562 tok = strsep(&str, ","); 538 tok = strsep(&str, ",");
563 if (!tok) 539 if (!tok)
564 break; 540 break;
565 541
566 tok_end = strstr(tok, "-"); 542 tok_end = strstr(tok, "-");
567 543
568 dprintf("\ntoken: {%s}, end: { 544 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
569 if (!tok_end) { 545 if (!tok_end) {
570 /* Single CPU specifie 546 /* Single CPU specified: */
571 bind_cpu_0 = bind_cpu_ 547 bind_cpu_0 = bind_cpu_1 = atol(tok);
572 } else { 548 } else {
573 /* CPU range specified 549 /* CPU range specified (for example: "5-11"): */
574 bind_cpu_0 = atol(tok) 550 bind_cpu_0 = atol(tok);
575 bind_cpu_1 = atol(tok_ 551 bind_cpu_1 = atol(tok_end + 1);
576 } 552 }
577 553
578 step = 1; 554 step = 1;
579 tok_step = strstr(tok, "#"); 555 tok_step = strstr(tok, "#");
580 if (tok_step) { 556 if (tok_step) {
581 step = atol(tok_step + 557 step = atol(tok_step + 1);
582 BUG_ON(step <= 0 || st 558 BUG_ON(step <= 0 || step >= g->p.nr_cpus);
583 } 559 }
584 560
585 /* 561 /*
586 * Mask length. 562 * Mask length.
587 * Eg: "--cpus 8_4-16#4" means 563 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
588 * where the _4 means the next 564 * where the _4 means the next 4 CPUs are allowed.
589 */ 565 */
590 bind_len = 1; 566 bind_len = 1;
591 tok_len = strstr(tok, "_"); 567 tok_len = strstr(tok, "_");
592 if (tok_len) { 568 if (tok_len) {
593 bind_len = atol(tok_le 569 bind_len = atol(tok_len + 1);
594 BUG_ON(bind_len <= 0 | 570 BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
595 } 571 }
596 572
597 /* Multiplicator shortcut, "0x 573 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
598 mul = 1; 574 mul = 1;
599 tok_mul = strstr(tok, "x"); 575 tok_mul = strstr(tok, "x");
600 if (tok_mul) { 576 if (tok_mul) {
601 mul = atol(tok_mul + 1 577 mul = atol(tok_mul + 1);
602 BUG_ON(mul <= 0); 578 BUG_ON(mul <= 0);
603 } 579 }
604 580
605 dprintf("CPUs: %d_%d-%d#%dx%d\ 581 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
606 582
607 if (bind_cpu_0 >= g->p.nr_cpus 583 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
608 printf("\nTest not app 584 printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
609 return -1; 585 return -1;
610 } 586 }
611 587
612 if (is_cpu_online(bind_cpu_0) <<
613 printf("\nTest not app <<
614 return -1; <<
615 } <<
616 <<
617 BUG_ON(bind_cpu_0 < 0 || bind_ 588 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
618 BUG_ON(bind_cpu_0 > bind_cpu_1 589 BUG_ON(bind_cpu_0 > bind_cpu_1);
619 590
620 for (bind_cpu = bind_cpu_0; bi 591 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
621 size_t size = CPU_ALLO <<
622 int i; 592 int i;
623 593
624 for (i = 0; i < mul; i 594 for (i = 0; i < mul; i++) {
625 int cpu; 595 int cpu;
626 596
627 if (t >= g->p. 597 if (t >= g->p.nr_tasks) {
628 printf 598 printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
629 goto o 599 goto out;
630 } 600 }
631 td = g->thread 601 td = g->threads + t;
632 602
633 if (t) 603 if (t)
634 tprint 604 tprintf(",");
635 if (bind_len > 605 if (bind_len > 1) {
636 tprint 606 tprintf("%2d/%d", bind_cpu, bind_len);
637 } else { 607 } else {
638 tprint 608 tprintf("%2d", bind_cpu);
639 } 609 }
640 610
641 td->bind_cpuma !! 611 CPU_ZERO(&td->bind_cpumask);
642 BUG_ON(!td->bi <<
643 CPU_ZERO_S(siz <<
644 for (cpu = bin 612 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
645 if (cp !! 613 BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
646 !! 614 CPU_SET(cpu, &td->bind_cpumask);
647 <<
648 } <<
649 CPU_SE <<
650 } 615 }
651 t++; 616 t++;
652 } 617 }
653 } 618 }
654 } 619 }
655 out: 620 out:
656 621
657 tprintf("\n"); 622 tprintf("\n");
658 623
659 if (t < g->p.nr_tasks) 624 if (t < g->p.nr_tasks)
660 printf("# NOTE: %d tasks bound 625 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
661 626
662 free(str0); 627 free(str0);
663 return 0; 628 return 0;
664 } 629 }
665 630
666 static int parse_cpus_opt(const struct option 631 static int parse_cpus_opt(const struct option *opt __maybe_unused,
667 const char *arg, int 632 const char *arg, int unset __maybe_unused)
668 { 633 {
669 if (!arg) 634 if (!arg)
670 return -1; 635 return -1;
671 636
672 return parse_cpu_list(arg); 637 return parse_cpu_list(arg);
673 } 638 }
674 639
675 static int parse_node_list(const char *arg) 640 static int parse_node_list(const char *arg)
676 { 641 {
677 p0.node_list_str = strdup(arg); 642 p0.node_list_str = strdup(arg);
678 643
679 dprintf("got NODE list: {%s}\n", p0.no 644 dprintf("got NODE list: {%s}\n", p0.node_list_str);
680 645
681 return 0; 646 return 0;
682 } 647 }
683 648
684 static int parse_setup_node_list(void) 649 static int parse_setup_node_list(void)
685 { 650 {
686 struct thread_data *td; 651 struct thread_data *td;
687 char *str0, *str; 652 char *str0, *str;
688 int t; 653 int t;
689 654
690 if (!g->p.node_list_str) 655 if (!g->p.node_list_str)
691 return 0; 656 return 0;
692 657
693 dprintf("g->p.nr_tasks: %d\n", g->p.nr 658 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
694 659
695 str0 = str = strdup(g->p.node_list_str 660 str0 = str = strdup(g->p.node_list_str);
696 t = 0; 661 t = 0;
697 662
698 BUG_ON(!str); 663 BUG_ON(!str);
699 664
700 tprintf("# binding tasks to NODEs:\n") 665 tprintf("# binding tasks to NODEs:\n");
701 tprintf("# "); 666 tprintf("# ");
702 667
703 while (true) { 668 while (true) {
704 int bind_node, bind_node_0, bi 669 int bind_node, bind_node_0, bind_node_1;
705 char *tok, *tok_end, *tok_step 670 char *tok, *tok_end, *tok_step, *tok_mul;
706 int step; 671 int step;
707 int mul; 672 int mul;
708 673
709 tok = strsep(&str, ","); 674 tok = strsep(&str, ",");
710 if (!tok) 675 if (!tok)
711 break; 676 break;
712 677
713 tok_end = strstr(tok, "-"); 678 tok_end = strstr(tok, "-");
714 679
715 dprintf("\ntoken: {%s}, end: { 680 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
716 if (!tok_end) { 681 if (!tok_end) {
717 /* Single NODE specifi 682 /* Single NODE specified: */
718 bind_node_0 = bind_nod 683 bind_node_0 = bind_node_1 = atol(tok);
719 } else { 684 } else {
720 /* NODE range specifie 685 /* NODE range specified (for example: "5-11"): */
721 bind_node_0 = atol(tok 686 bind_node_0 = atol(tok);
722 bind_node_1 = atol(tok 687 bind_node_1 = atol(tok_end + 1);
723 } 688 }
724 689
725 step = 1; 690 step = 1;
726 tok_step = strstr(tok, "#"); 691 tok_step = strstr(tok, "#");
727 if (tok_step) { 692 if (tok_step) {
728 step = atol(tok_step + 693 step = atol(tok_step + 1);
729 BUG_ON(step <= 0 || st 694 BUG_ON(step <= 0 || step >= g->p.nr_nodes);
730 } 695 }
731 696
732 /* Multiplicator shortcut, "0x 697 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
733 mul = 1; 698 mul = 1;
734 tok_mul = strstr(tok, "x"); 699 tok_mul = strstr(tok, "x");
735 if (tok_mul) { 700 if (tok_mul) {
736 mul = atol(tok_mul + 1 701 mul = atol(tok_mul + 1);
737 BUG_ON(mul <= 0); 702 BUG_ON(mul <= 0);
738 } 703 }
739 704
740 dprintf("NODEs: %d-%d #%d\n", 705 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
741 706
742 if (bind_node_0 >= g->p.nr_nod 707 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
743 printf("\nTest not app 708 printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
744 return -1; 709 return -1;
745 } 710 }
746 711
747 BUG_ON(bind_node_0 < 0 || bind 712 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
748 BUG_ON(bind_node_0 > bind_node 713 BUG_ON(bind_node_0 > bind_node_1);
749 714
750 for (bind_node = bind_node_0; 715 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
751 int i; 716 int i;
752 717
753 for (i = 0; i < mul; i 718 for (i = 0; i < mul; i++) {
754 if (t >= g->p. 719 if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
755 printf 720 printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
756 goto o 721 goto out;
757 } 722 }
758 td = g->thread 723 td = g->threads + t;
759 724
760 if (!t) 725 if (!t)
761 tprint 726 tprintf(" %2d", bind_node);
762 else 727 else
763 tprint 728 tprintf(",%2d", bind_node);
764 729
765 td->bind_node 730 td->bind_node = bind_node;
766 t++; 731 t++;
767 } 732 }
768 } 733 }
769 } 734 }
770 out: 735 out:
771 736
772 tprintf("\n"); 737 tprintf("\n");
773 738
774 if (t < g->p.nr_tasks) 739 if (t < g->p.nr_tasks)
775 printf("# NOTE: %d tasks mem-b 740 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
776 741
777 free(str0); 742 free(str0);
778 return 0; 743 return 0;
779 } 744 }
780 745
781 static int parse_nodes_opt(const struct option 746 static int parse_nodes_opt(const struct option *opt __maybe_unused,
782 const char *arg, int 747 const char *arg, int unset __maybe_unused)
783 { 748 {
784 if (!arg) 749 if (!arg)
785 return -1; 750 return -1;
786 751
787 return parse_node_list(arg); 752 return parse_node_list(arg);
788 } 753 }
789 754
>> 755 #define BIT(x) (1ul << x)
>> 756
790 static inline uint32_t lfsr_32(uint32_t lfsr) 757 static inline uint32_t lfsr_32(uint32_t lfsr)
791 { 758 {
792 const uint32_t taps = BIT(1) | BIT(5) 759 const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
793 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x 760 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
794 } 761 }
795 762
796 /* 763 /*
797 * Make sure there's real data dependency to R 764 * Make sure there's real data dependency to RAM (when read
798 * accesses are enabled), so the compiler, the 765 * accesses are enabled), so the compiler, the CPU and the
799 * kernel (KSM, zero page, etc.) cannot optimi 766 * kernel (KSM, zero page, etc.) cannot optimize away RAM
800 * accesses: 767 * accesses:
801 */ 768 */
802 static inline u64 access_data(u64 *data, u64 v 769 static inline u64 access_data(u64 *data, u64 val)
803 { 770 {
804 if (g->p.data_reads) 771 if (g->p.data_reads)
805 val += *data; 772 val += *data;
806 if (g->p.data_writes) 773 if (g->p.data_writes)
807 *data = val + 1; 774 *data = val + 1;
808 return val; 775 return val;
809 } 776 }
810 777
811 /* 778 /*
812 * The worker process does two types of work, 779 * The worker process does two types of work, a forwards going
813 * loop and a backwards going loop. 780 * loop and a backwards going loop.
814 * 781 *
815 * We do this so that on multiprocessor system 782 * We do this so that on multiprocessor systems we do not create
816 * a 'train' of processing, with highly synchr 783 * a 'train' of processing, with highly synchronized processes,
817 * skewing the whole benchmark. 784 * skewing the whole benchmark.
818 */ 785 */
819 static u64 do_work(u8 *__data, long bytes, int 786 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
820 { 787 {
821 long words = bytes/sizeof(u64); 788 long words = bytes/sizeof(u64);
822 u64 *data = (void *)__data; 789 u64 *data = (void *)__data;
823 long chunk_0, chunk_1; 790 long chunk_0, chunk_1;
824 u64 *d0, *d, *d1; 791 u64 *d0, *d, *d1;
825 long off; 792 long off;
826 long i; 793 long i;
827 794
828 BUG_ON(!data && words); 795 BUG_ON(!data && words);
829 BUG_ON(data && !words); 796 BUG_ON(data && !words);
830 797
831 if (!data) 798 if (!data)
832 return val; 799 return val;
833 800
834 /* Very simple memset() work variant: 801 /* Very simple memset() work variant: */
835 if (g->p.data_zero_memset && !g->p.dat 802 if (g->p.data_zero_memset && !g->p.data_rand_walk) {
836 bzero(data, bytes); 803 bzero(data, bytes);
837 return val; 804 return val;
838 } 805 }
839 806
840 /* Spread out by PID/TID nr and by loo 807 /* Spread out by PID/TID nr and by loop nr: */
841 chunk_0 = words/nr_max; 808 chunk_0 = words/nr_max;
842 chunk_1 = words/g->p.nr_loops; 809 chunk_1 = words/g->p.nr_loops;
843 off = nr*chunk_0 + loop*chunk_1; 810 off = nr*chunk_0 + loop*chunk_1;
844 811
845 while (off >= words) 812 while (off >= words)
846 off -= words; 813 off -= words;
847 814
848 if (g->p.data_rand_walk) { 815 if (g->p.data_rand_walk) {
849 u32 lfsr = nr + loop + val; 816 u32 lfsr = nr + loop + val;
850 long j; !! 817 int j;
851 818
852 for (i = 0; i < words/1024; i+ 819 for (i = 0; i < words/1024; i++) {
853 long start, end; 820 long start, end;
854 821
855 lfsr = lfsr_32(lfsr); 822 lfsr = lfsr_32(lfsr);
856 823
857 start = lfsr % words; 824 start = lfsr % words;
858 end = min(start + 1024 825 end = min(start + 1024, words-1);
859 826
860 if (g->p.data_zero_mem 827 if (g->p.data_zero_memset) {
861 bzero(data + s 828 bzero(data + start, (end-start) * sizeof(u64));
862 } else { 829 } else {
863 for (j = start 830 for (j = start; j < end; j++)
864 val = 831 val = access_data(data + j, val);
865 } 832 }
866 } 833 }
867 } else if (!g->p.data_backwards || (nr 834 } else if (!g->p.data_backwards || (nr + loop) & 1) {
868 /* Process data forwards: */ 835 /* Process data forwards: */
869 836
870 d0 = data + off; 837 d0 = data + off;
871 d = data + off + 1; 838 d = data + off + 1;
872 d1 = data + words; 839 d1 = data + words;
873 840
874 for (;;) { 841 for (;;) {
875 if (unlikely(d >= d1)) 842 if (unlikely(d >= d1))
876 d = data; 843 d = data;
877 if (unlikely(d == d0)) 844 if (unlikely(d == d0))
878 break; 845 break;
879 846
880 val = access_data(d, v 847 val = access_data(d, val);
881 848
882 d++; 849 d++;
883 } 850 }
884 } else { 851 } else {
885 /* Process data backwards: */ 852 /* Process data backwards: */
886 853
887 d0 = data + off; 854 d0 = data + off;
888 d = data + off - 1; 855 d = data + off - 1;
889 d1 = data + words; 856 d1 = data + words;
890 857
891 for (;;) { 858 for (;;) {
892 if (unlikely(d < data) 859 if (unlikely(d < data))
893 d = data + wor 860 d = data + words-1;
894 if (unlikely(d == d0)) 861 if (unlikely(d == d0))
895 break; 862 break;
896 863
897 val = access_data(d, v 864 val = access_data(d, val);
898 865
899 d--; 866 d--;
900 } 867 }
901 } 868 }
902 869
903 return val; 870 return val;
904 } 871 }
905 872
906 static void update_curr_cpu(int task_nr, unsig 873 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
907 { 874 {
908 unsigned int cpu; 875 unsigned int cpu;
909 876
910 cpu = sched_getcpu(); 877 cpu = sched_getcpu();
911 878
912 g->threads[task_nr].curr_cpu = cpu; 879 g->threads[task_nr].curr_cpu = cpu;
913 prctl(0, bytes_worked); 880 prctl(0, bytes_worked);
914 } 881 }
915 882
916 /* 883 /*
917 * Count the number of nodes a process's threa 884 * Count the number of nodes a process's threads
918 * are spread out on. 885 * are spread out on.
919 * 886 *
920 * A count of 1 means that the process is comp 887 * A count of 1 means that the process is compressed
921 * to a single node. A count of g->p.nr_nodes 888 * to a single node. A count of g->p.nr_nodes means it's
922 * spread out on the whole system. 889 * spread out on the whole system.
923 */ 890 */
924 static int count_process_nodes(int process_nr) 891 static int count_process_nodes(int process_nr)
925 { 892 {
926 char *node_present; 893 char *node_present;
927 int nodes; 894 int nodes;
928 int n, t; 895 int n, t;
929 896
930 node_present = (char *)malloc(g->p.nr_ 897 node_present = (char *)malloc(g->p.nr_nodes * sizeof(char));
931 BUG_ON(!node_present); 898 BUG_ON(!node_present);
932 for (nodes = 0; nodes < g->p.nr_nodes; 899 for (nodes = 0; nodes < g->p.nr_nodes; nodes++)
933 node_present[nodes] = 0; 900 node_present[nodes] = 0;
934 901
935 for (t = 0; t < g->p.nr_threads; t++) 902 for (t = 0; t < g->p.nr_threads; t++) {
936 struct thread_data *td; 903 struct thread_data *td;
937 int task_nr; 904 int task_nr;
938 int node; 905 int node;
939 906
940 task_nr = process_nr*g->p.nr_t 907 task_nr = process_nr*g->p.nr_threads + t;
941 td = g->threads + task_nr; 908 td = g->threads + task_nr;
942 909
943 node = numa_node_of_cpu(td->cu 910 node = numa_node_of_cpu(td->curr_cpu);
944 if (node < 0) /* curr_cpu was 911 if (node < 0) /* curr_cpu was likely still -1 */ {
945 free(node_present); 912 free(node_present);
946 return 0; 913 return 0;
947 } 914 }
948 915
949 node_present[node] = 1; 916 node_present[node] = 1;
950 } 917 }
951 918
952 nodes = 0; 919 nodes = 0;
953 920
954 for (n = 0; n < g->p.nr_nodes; n++) 921 for (n = 0; n < g->p.nr_nodes; n++)
955 nodes += node_present[n]; 922 nodes += node_present[n];
956 923
957 free(node_present); 924 free(node_present);
958 return nodes; 925 return nodes;
959 } 926 }
960 927
961 /* 928 /*
962 * Count the number of distinct process-thread 929 * Count the number of distinct process-threads a node contains.
963 * 930 *
964 * A count of 1 means that the node contains o 931 * A count of 1 means that the node contains only a single
965 * process. If all nodes on the system contain 932 * process. If all nodes on the system contain at most one
966 * process then we are well-converged. 933 * process then we are well-converged.
967 */ 934 */
968 static int count_node_processes(int node) 935 static int count_node_processes(int node)
969 { 936 {
970 int processes = 0; 937 int processes = 0;
971 int t, p; 938 int t, p;
972 939
973 for (p = 0; p < g->p.nr_proc; p++) { 940 for (p = 0; p < g->p.nr_proc; p++) {
974 for (t = 0; t < g->p.nr_thread 941 for (t = 0; t < g->p.nr_threads; t++) {
975 struct thread_data *td 942 struct thread_data *td;
976 int task_nr; 943 int task_nr;
977 int n; 944 int n;
978 945
979 task_nr = p*g->p.nr_th 946 task_nr = p*g->p.nr_threads + t;
980 td = g->threads + task 947 td = g->threads + task_nr;
981 948
982 n = numa_node_of_cpu(t 949 n = numa_node_of_cpu(td->curr_cpu);
983 if (n == node) { 950 if (n == node) {
984 processes++; 951 processes++;
985 break; 952 break;
986 } 953 }
987 } 954 }
988 } 955 }
989 956
990 return processes; 957 return processes;
991 } 958 }
992 959
993 static void calc_convergence_compression(int * 960 static void calc_convergence_compression(int *strong)
994 { 961 {
995 unsigned int nodes_min, nodes_max; 962 unsigned int nodes_min, nodes_max;
996 int p; 963 int p;
997 964
998 nodes_min = -1; 965 nodes_min = -1;
999 nodes_max = 0; 966 nodes_max = 0;
1000 967
1001 for (p = 0; p < g->p.nr_proc; p++) { 968 for (p = 0; p < g->p.nr_proc; p++) {
1002 unsigned int nodes = count_pr 969 unsigned int nodes = count_process_nodes(p);
1003 970
1004 if (!nodes) { 971 if (!nodes) {
1005 *strong = 0; 972 *strong = 0;
1006 return; 973 return;
1007 } 974 }
1008 975
1009 nodes_min = min(nodes, nodes_ 976 nodes_min = min(nodes, nodes_min);
1010 nodes_max = max(nodes, nodes_ 977 nodes_max = max(nodes, nodes_max);
1011 } 978 }
1012 979
1013 /* Strong convergence: all threads co 980 /* Strong convergence: all threads compress on a single node: */
1014 if (nodes_min == 1 && nodes_max == 1) 981 if (nodes_min == 1 && nodes_max == 1) {
1015 *strong = 1; 982 *strong = 1;
1016 } else { 983 } else {
1017 *strong = 0; 984 *strong = 0;
1018 tprintf(" {%d-%d}", nodes_min 985 tprintf(" {%d-%d}", nodes_min, nodes_max);
1019 } 986 }
1020 } 987 }
1021 988
1022 static void calc_convergence(double runtime_n 989 static void calc_convergence(double runtime_ns_max, double *convergence)
1023 { 990 {
1024 unsigned int loops_done_min, loops_do 991 unsigned int loops_done_min, loops_done_max;
1025 int process_groups; 992 int process_groups;
1026 int *nodes; 993 int *nodes;
1027 int distance; 994 int distance;
1028 int nr_min; 995 int nr_min;
1029 int nr_max; 996 int nr_max;
1030 int strong; 997 int strong;
1031 int sum; 998 int sum;
1032 int nr; 999 int nr;
1033 int node; 1000 int node;
1034 int cpu; 1001 int cpu;
1035 int t; 1002 int t;
1036 1003
1037 if (!g->p.show_convergence && !g->p.m 1004 if (!g->p.show_convergence && !g->p.measure_convergence)
1038 return; 1005 return;
1039 1006
1040 nodes = (int *)malloc(g->p.nr_nodes * 1007 nodes = (int *)malloc(g->p.nr_nodes * sizeof(int));
1041 BUG_ON(!nodes); 1008 BUG_ON(!nodes);
1042 for (node = 0; node < g->p.nr_nodes; 1009 for (node = 0; node < g->p.nr_nodes; node++)
1043 nodes[node] = 0; 1010 nodes[node] = 0;
1044 1011
1045 loops_done_min = -1; 1012 loops_done_min = -1;
1046 loops_done_max = 0; 1013 loops_done_max = 0;
1047 1014
1048 for (t = 0; t < g->p.nr_tasks; t++) { 1015 for (t = 0; t < g->p.nr_tasks; t++) {
1049 struct thread_data *td = g->t 1016 struct thread_data *td = g->threads + t;
1050 unsigned int loops_done; 1017 unsigned int loops_done;
1051 1018
1052 cpu = td->curr_cpu; 1019 cpu = td->curr_cpu;
1053 1020
1054 /* Not all threads have writt 1021 /* Not all threads have written it yet: */
1055 if (cpu < 0) 1022 if (cpu < 0)
1056 continue; 1023 continue;
1057 1024
1058 node = numa_node_of_cpu(cpu); 1025 node = numa_node_of_cpu(cpu);
1059 1026
1060 nodes[node]++; 1027 nodes[node]++;
1061 1028
1062 loops_done = td->loops_done; 1029 loops_done = td->loops_done;
1063 loops_done_min = min(loops_do 1030 loops_done_min = min(loops_done, loops_done_min);
1064 loops_done_max = max(loops_do 1031 loops_done_max = max(loops_done, loops_done_max);
1065 } 1032 }
1066 1033
1067 nr_max = 0; 1034 nr_max = 0;
1068 nr_min = g->p.nr_tasks; 1035 nr_min = g->p.nr_tasks;
1069 sum = 0; 1036 sum = 0;
1070 1037
1071 for (node = 0; node < g->p.nr_nodes; 1038 for (node = 0; node < g->p.nr_nodes; node++) {
1072 if (!is_node_present(node)) 1039 if (!is_node_present(node))
1073 continue; 1040 continue;
1074 nr = nodes[node]; 1041 nr = nodes[node];
1075 nr_min = min(nr, nr_min); 1042 nr_min = min(nr, nr_min);
1076 nr_max = max(nr, nr_max); 1043 nr_max = max(nr, nr_max);
1077 sum += nr; 1044 sum += nr;
1078 } 1045 }
1079 BUG_ON(nr_min > nr_max); 1046 BUG_ON(nr_min > nr_max);
1080 1047
1081 BUG_ON(sum > g->p.nr_tasks); 1048 BUG_ON(sum > g->p.nr_tasks);
1082 1049
1083 if (0 && (sum < g->p.nr_tasks)) { 1050 if (0 && (sum < g->p.nr_tasks)) {
1084 free(nodes); 1051 free(nodes);
1085 return; 1052 return;
1086 } 1053 }
1087 1054
1088 /* 1055 /*
1089 * Count the number of distinct proce 1056 * Count the number of distinct process groups present
1090 * on nodes - when we are converged t 1057 * on nodes - when we are converged this will decrease
1091 * to g->p.nr_proc: 1058 * to g->p.nr_proc:
1092 */ 1059 */
1093 process_groups = 0; 1060 process_groups = 0;
1094 1061
1095 for (node = 0; node < g->p.nr_nodes; 1062 for (node = 0; node < g->p.nr_nodes; node++) {
1096 int processes; 1063 int processes;
1097 1064
1098 if (!is_node_present(node)) 1065 if (!is_node_present(node))
1099 continue; 1066 continue;
1100 processes = count_node_proces 1067 processes = count_node_processes(node);
1101 nr = nodes[node]; 1068 nr = nodes[node];
1102 tprintf(" %2d/%-2d", nr, proc 1069 tprintf(" %2d/%-2d", nr, processes);
1103 1070
1104 process_groups += processes; 1071 process_groups += processes;
1105 } 1072 }
1106 1073
1107 distance = nr_max - nr_min; 1074 distance = nr_max - nr_min;
1108 1075
1109 tprintf(" [%2d/%-2d]", distance, proc 1076 tprintf(" [%2d/%-2d]", distance, process_groups);
1110 1077
1111 tprintf(" l:%3d-%-3d (%3d)", 1078 tprintf(" l:%3d-%-3d (%3d)",
1112 loops_done_min, loops_done_ma 1079 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1113 1080
1114 if (loops_done_min && loops_done_max) 1081 if (loops_done_min && loops_done_max) {
1115 double skew = 1.0 - (double)l 1082 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1116 1083
1117 tprintf(" [%4.1f%%]", skew * 1084 tprintf(" [%4.1f%%]", skew * 100.0);
1118 } 1085 }
1119 1086
1120 calc_convergence_compression(&strong) 1087 calc_convergence_compression(&strong);
1121 1088
1122 if (strong && process_groups == g->p. 1089 if (strong && process_groups == g->p.nr_proc) {
1123 if (!*convergence) { 1090 if (!*convergence) {
1124 *convergence = runtim 1091 *convergence = runtime_ns_max;
1125 tprintf(" (%6.1fs con 1092 tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1126 if (g->p.measure_conv 1093 if (g->p.measure_convergence) {
1127 g->all_conver 1094 g->all_converged = true;
1128 g->stop_work 1095 g->stop_work = true;
1129 } 1096 }
1130 } 1097 }
1131 } else { 1098 } else {
1132 if (*convergence) { 1099 if (*convergence) {
1133 tprintf(" (%6.1fs de- 1100 tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1134 *convergence = 0; 1101 *convergence = 0;
1135 } 1102 }
1136 tprintf("\n"); 1103 tprintf("\n");
1137 } 1104 }
1138 1105
1139 free(nodes); 1106 free(nodes);
1140 } 1107 }
1141 1108
1142 static void show_summary(double runtime_ns_ma 1109 static void show_summary(double runtime_ns_max, int l, double *convergence)
1143 { 1110 {
1144 tprintf("\r # %5.1f%% [%.1f mins]", 1111 tprintf("\r # %5.1f%% [%.1f mins]",
1145 (double)(l+1)/g->p.nr_loops*1 1112 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1146 1113
1147 calc_convergence(runtime_ns_max, conv 1114 calc_convergence(runtime_ns_max, convergence);
1148 1115
1149 if (g->p.show_details >= 0) 1116 if (g->p.show_details >= 0)
1150 fflush(stdout); 1117 fflush(stdout);
1151 } 1118 }
1152 1119
1153 static void *worker_thread(void *__tdata) 1120 static void *worker_thread(void *__tdata)
1154 { 1121 {
1155 struct thread_data *td = __tdata; 1122 struct thread_data *td = __tdata;
1156 struct timeval start0, start, stop, d 1123 struct timeval start0, start, stop, diff;
1157 int process_nr = td->process_nr; 1124 int process_nr = td->process_nr;
1158 int thread_nr = td->thread_nr; 1125 int thread_nr = td->thread_nr;
1159 unsigned long last_perturbance; 1126 unsigned long last_perturbance;
1160 int task_nr = td->task_nr; 1127 int task_nr = td->task_nr;
1161 int details = g->p.show_details; 1128 int details = g->p.show_details;
1162 int first_task, last_task; 1129 int first_task, last_task;
1163 double convergence = 0; 1130 double convergence = 0;
1164 u64 val = td->val; 1131 u64 val = td->val;
1165 double runtime_ns_max; 1132 double runtime_ns_max;
1166 u8 *global_data; 1133 u8 *global_data;
1167 u8 *process_data; 1134 u8 *process_data;
1168 u8 *thread_data; 1135 u8 *thread_data;
1169 u64 bytes_done, secs; 1136 u64 bytes_done, secs;
1170 long work_done; 1137 long work_done;
1171 u32 l; 1138 u32 l;
1172 struct rusage rusage; 1139 struct rusage rusage;
1173 1140
1174 bind_to_cpumask(td->bind_cpumask); 1141 bind_to_cpumask(td->bind_cpumask);
1175 bind_to_memnode(td->bind_node); 1142 bind_to_memnode(td->bind_node);
1176 1143
1177 set_taskname("thread %d/%d", process_ 1144 set_taskname("thread %d/%d", process_nr, thread_nr);
1178 1145
1179 global_data = g->data; 1146 global_data = g->data;
1180 process_data = td->process_data; 1147 process_data = td->process_data;
1181 thread_data = setup_private_data(g->p 1148 thread_data = setup_private_data(g->p.bytes_thread);
1182 1149
1183 bytes_done = 0; 1150 bytes_done = 0;
1184 1151
1185 last_task = 0; 1152 last_task = 0;
1186 if (process_nr == g->p.nr_proc-1 && t 1153 if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1187 last_task = 1; 1154 last_task = 1;
1188 1155
1189 first_task = 0; 1156 first_task = 0;
1190 if (process_nr == 0 && thread_nr == 0 1157 if (process_nr == 0 && thread_nr == 0)
1191 first_task = 1; 1158 first_task = 1;
1192 1159
1193 if (details >= 2) { 1160 if (details >= 2) {
1194 printf("# thread %2d / %2d g 1161 printf("# thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1195 process_nr, thread_nr 1162 process_nr, thread_nr, global_data, process_data, thread_data);
1196 } 1163 }
1197 1164
1198 if (g->p.serialize_startup) { 1165 if (g->p.serialize_startup) {
1199 mutex_lock(&g->startup_mutex) !! 1166 pthread_mutex_lock(&g->startup_mutex);
1200 g->nr_tasks_started++; 1167 g->nr_tasks_started++;
1201 /* The last thread wakes the 1168 /* The last thread wakes the main process. */
1202 if (g->nr_tasks_started == g- 1169 if (g->nr_tasks_started == g->p.nr_tasks)
1203 cond_signal(&g->start !! 1170 pthread_cond_signal(&g->startup_cond);
1204 1171
1205 mutex_unlock(&g->startup_mute !! 1172 pthread_mutex_unlock(&g->startup_mutex);
1206 1173
1207 /* Here we will wait for the 1174 /* Here we will wait for the main process to start us all at once: */
1208 mutex_lock(&g->start_work_mut !! 1175 pthread_mutex_lock(&g->start_work_mutex);
1209 g->start_work = false; 1176 g->start_work = false;
1210 g->nr_tasks_working++; 1177 g->nr_tasks_working++;
1211 while (!g->start_work) 1178 while (!g->start_work)
1212 cond_wait(&g->start_w !! 1179 pthread_cond_wait(&g->start_work_cond, &g->start_work_mutex);
1213 1180
1214 mutex_unlock(&g->start_work_m !! 1181 pthread_mutex_unlock(&g->start_work_mutex);
1215 } 1182 }
1216 1183
1217 gettimeofday(&start0, NULL); 1184 gettimeofday(&start0, NULL);
1218 1185
1219 start = stop = start0; 1186 start = stop = start0;
1220 last_perturbance = start.tv_sec; 1187 last_perturbance = start.tv_sec;
1221 1188
1222 for (l = 0; l < g->p.nr_loops; l++) { 1189 for (l = 0; l < g->p.nr_loops; l++) {
1223 start = stop; 1190 start = stop;
1224 1191
1225 if (g->stop_work) 1192 if (g->stop_work)
1226 break; 1193 break;
1227 1194
1228 val += do_work(global_data, 1195 val += do_work(global_data, g->p.bytes_global, process_nr, g->p.nr_proc, l, val);
1229 val += do_work(process_data, 1196 val += do_work(process_data, g->p.bytes_process, thread_nr, g->p.nr_threads, l, val);
1230 val += do_work(thread_data, 1197 val += do_work(thread_data, g->p.bytes_thread, 0, 1, l, val);
1231 1198
1232 if (g->p.sleep_usecs) { 1199 if (g->p.sleep_usecs) {
1233 mutex_lock(td->proces !! 1200 pthread_mutex_lock(td->process_lock);
1234 usleep(g->p.sleep_use 1201 usleep(g->p.sleep_usecs);
1235 mutex_unlock(td->proc !! 1202 pthread_mutex_unlock(td->process_lock);
1236 } 1203 }
1237 /* 1204 /*
1238 * Amount of work to be done 1205 * Amount of work to be done under a process-global lock:
1239 */ 1206 */
1240 if (g->p.bytes_process_locked 1207 if (g->p.bytes_process_locked) {
1241 mutex_lock(td->proces !! 1208 pthread_mutex_lock(td->process_lock);
1242 val += do_work(proces 1209 val += do_work(process_data, g->p.bytes_process_locked, thread_nr, g->p.nr_threads, l, val);
1243 mutex_unlock(td->proc !! 1210 pthread_mutex_unlock(td->process_lock);
1244 } 1211 }
1245 1212
1246 work_done = g->p.bytes_global 1213 work_done = g->p.bytes_global + g->p.bytes_process +
1247 g->p.bytes_proces 1214 g->p.bytes_process_locked + g->p.bytes_thread;
1248 1215
1249 update_curr_cpu(task_nr, work 1216 update_curr_cpu(task_nr, work_done);
1250 bytes_done += work_done; 1217 bytes_done += work_done;
1251 1218
1252 if (details < 0 && !g->p.pert 1219 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1253 continue; 1220 continue;
1254 1221
1255 td->loops_done = l; 1222 td->loops_done = l;
1256 1223
1257 gettimeofday(&stop, NULL); 1224 gettimeofday(&stop, NULL);
1258 1225
1259 /* Check whether our max runt 1226 /* Check whether our max runtime timed out: */
1260 if (g->p.nr_secs) { 1227 if (g->p.nr_secs) {
1261 timersub(&stop, &star 1228 timersub(&stop, &start0, &diff);
1262 if ((u32)diff.tv_sec 1229 if ((u32)diff.tv_sec >= g->p.nr_secs) {
1263 g->stop_work 1230 g->stop_work = true;
1264 break; 1231 break;
1265 } 1232 }
1266 } 1233 }
1267 1234
1268 /* Update the summary at most 1235 /* Update the summary at most once per second: */
1269 if (start.tv_sec == stop.tv_s 1236 if (start.tv_sec == stop.tv_sec)
1270 continue; 1237 continue;
1271 1238
1272 /* 1239 /*
1273 * Perturb the first task's e 1240 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1274 * by migrating to CPU#0: 1241 * by migrating to CPU#0:
1275 */ 1242 */
1276 if (first_task && g->p.pertur 1243 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1277 cpu_set_t *orig_mask; !! 1244 cpu_set_t orig_mask;
1278 int target_cpu; 1245 int target_cpu;
1279 int this_cpu; 1246 int this_cpu;
1280 1247
1281 last_perturbance = st 1248 last_perturbance = stop.tv_sec;
1282 1249
1283 /* 1250 /*
1284 * Depending on where 1251 * Depending on where we are running, move into
1285 * the other half of 1252 * the other half of the system, to create some
1286 * real disturbance: 1253 * real disturbance:
1287 */ 1254 */
1288 this_cpu = g->threads 1255 this_cpu = g->threads[task_nr].curr_cpu;
1289 if (this_cpu < g->p.n 1256 if (this_cpu < g->p.nr_cpus/2)
1290 target_cpu = 1257 target_cpu = g->p.nr_cpus-1;
1291 else 1258 else
1292 target_cpu = 1259 target_cpu = 0;
1293 1260
1294 orig_mask = bind_to_c 1261 orig_mask = bind_to_cpu(target_cpu);
1295 1262
1296 /* Here we are runnin 1263 /* Here we are running on the target CPU already */
1297 if (details >= 1) 1264 if (details >= 1)
1298 printf(" (inj 1265 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1299 1266
1300 bind_to_cpumask(orig_ 1267 bind_to_cpumask(orig_mask);
1301 CPU_FREE(orig_mask); <<
1302 } 1268 }
1303 1269
1304 if (details >= 3) { 1270 if (details >= 3) {
1305 timersub(&stop, &star 1271 timersub(&stop, &start, &diff);
1306 runtime_ns_max = diff 1272 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1307 runtime_ns_max += dif 1273 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1308 1274
1309 if (details >= 0) { 1275 if (details >= 0) {
1310 printf(" #%2d 1276 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1311 proce 1277 process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1312 } 1278 }
1313 fflush(stdout); 1279 fflush(stdout);
1314 } 1280 }
1315 if (!last_task) 1281 if (!last_task)
1316 continue; 1282 continue;
1317 1283
1318 timersub(&stop, &start0, &dif 1284 timersub(&stop, &start0, &diff);
1319 runtime_ns_max = diff.tv_sec 1285 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1320 runtime_ns_max += diff.tv_use 1286 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1321 1287
1322 show_summary(runtime_ns_max, 1288 show_summary(runtime_ns_max, l, &convergence);
1323 } 1289 }
1324 1290
1325 gettimeofday(&stop, NULL); 1291 gettimeofday(&stop, NULL);
1326 timersub(&stop, &start0, &diff); 1292 timersub(&stop, &start0, &diff);
1327 td->runtime_ns = diff.tv_sec * NSEC_P 1293 td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1328 td->runtime_ns += diff.tv_usec * NSEC 1294 td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1329 secs = td->runtime_ns / NSEC_PER_SEC; 1295 secs = td->runtime_ns / NSEC_PER_SEC;
1330 td->speed_gbs = secs ? bytes_done / s 1296 td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1331 1297
1332 getrusage(RUSAGE_THREAD, &rusage); 1298 getrusage(RUSAGE_THREAD, &rusage);
1333 td->system_time_ns = rusage.ru_stime. 1299 td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1334 td->system_time_ns += rusage.ru_stime 1300 td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1335 td->user_time_ns = rusage.ru_utime.tv 1301 td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1336 td->user_time_ns += rusage.ru_utime.t 1302 td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1337 1303
1338 free_data(thread_data, g->p.bytes_thr 1304 free_data(thread_data, g->p.bytes_thread);
1339 1305
1340 mutex_lock(&g->stop_work_mutex); !! 1306 pthread_mutex_lock(&g->stop_work_mutex);
1341 g->bytes_done += bytes_done; 1307 g->bytes_done += bytes_done;
1342 mutex_unlock(&g->stop_work_mutex); !! 1308 pthread_mutex_unlock(&g->stop_work_mutex);
1343 1309
1344 return NULL; 1310 return NULL;
1345 } 1311 }
1346 1312
1347 /* 1313 /*
1348 * A worker process starts a couple of thread 1314 * A worker process starts a couple of threads:
1349 */ 1315 */
1350 static void worker_process(int process_nr) 1316 static void worker_process(int process_nr)
1351 { 1317 {
1352 struct mutex process_lock; !! 1318 pthread_mutex_t process_lock;
1353 struct thread_data *td; 1319 struct thread_data *td;
1354 pthread_t *pthreads; 1320 pthread_t *pthreads;
1355 u8 *process_data; 1321 u8 *process_data;
1356 int task_nr; 1322 int task_nr;
1357 int ret; 1323 int ret;
1358 int t; 1324 int t;
1359 1325
1360 mutex_init(&process_lock); !! 1326 pthread_mutex_init(&process_lock, NULL);
1361 set_taskname("process %d", process_nr 1327 set_taskname("process %d", process_nr);
1362 1328
1363 /* 1329 /*
1364 * Pick up the memory policy and the 1330 * Pick up the memory policy and the CPU binding of our first thread,
1365 * so that we initialize memory accor 1331 * so that we initialize memory accordingly:
1366 */ 1332 */
1367 task_nr = process_nr*g->p.nr_threads; 1333 task_nr = process_nr*g->p.nr_threads;
1368 td = g->threads + task_nr; 1334 td = g->threads + task_nr;
1369 1335
1370 bind_to_memnode(td->bind_node); 1336 bind_to_memnode(td->bind_node);
1371 bind_to_cpumask(td->bind_cpumask); 1337 bind_to_cpumask(td->bind_cpumask);
1372 1338
1373 pthreads = zalloc(g->p.nr_threads * s 1339 pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1374 process_data = setup_private_data(g-> 1340 process_data = setup_private_data(g->p.bytes_process);
1375 1341
1376 if (g->p.show_details >= 3) { 1342 if (g->p.show_details >= 3) {
1377 printf(" # process %2d global 1343 printf(" # process %2d global mem: %p, process mem: %p\n",
1378 process_nr, g->data, 1344 process_nr, g->data, process_data);
1379 } 1345 }
1380 1346
1381 for (t = 0; t < g->p.nr_threads; t++) 1347 for (t = 0; t < g->p.nr_threads; t++) {
1382 task_nr = process_nr*g->p.nr_ 1348 task_nr = process_nr*g->p.nr_threads + t;
1383 td = g->threads + task_nr; 1349 td = g->threads + task_nr;
1384 1350
1385 td->process_data = process_da 1351 td->process_data = process_data;
1386 td->process_nr = process_nr 1352 td->process_nr = process_nr;
1387 td->thread_nr = t; 1353 td->thread_nr = t;
1388 td->task_nr = task_nr; 1354 td->task_nr = task_nr;
1389 td->val = rand(); 1355 td->val = rand();
1390 td->curr_cpu = -1; 1356 td->curr_cpu = -1;
1391 td->process_lock = &process_l 1357 td->process_lock = &process_lock;
1392 1358
1393 ret = pthread_create(pthreads 1359 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1394 BUG_ON(ret); 1360 BUG_ON(ret);
1395 } 1361 }
1396 1362
1397 for (t = 0; t < g->p.nr_threads; t++) 1363 for (t = 0; t < g->p.nr_threads; t++) {
1398 ret = pthread_join(pthreads[t 1364 ret = pthread_join(pthreads[t], NULL);
1399 BUG_ON(ret); 1365 BUG_ON(ret);
1400 } 1366 }
1401 1367
1402 free_data(process_data, g->p.bytes_pr 1368 free_data(process_data, g->p.bytes_process);
1403 free(pthreads); 1369 free(pthreads);
1404 } 1370 }
1405 1371
1406 static void print_summary(void) 1372 static void print_summary(void)
1407 { 1373 {
1408 if (g->p.show_details < 0) 1374 if (g->p.show_details < 0)
1409 return; 1375 return;
1410 1376
1411 printf("\n ###\n"); 1377 printf("\n ###\n");
1412 printf(" # %d %s will execute (on %d 1378 printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1413 g->p.nr_tasks, g->p.nr_tasks 1379 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1414 printf(" # %5dx %5ldMB global s 1380 printf(" # %5dx %5ldMB global shared mem operations\n",
1415 g->p.nr_loops, g->p.b 1381 g->p.nr_loops, g->p.bytes_global/1024/1024);
1416 printf(" # %5dx %5ldMB process s 1382 printf(" # %5dx %5ldMB process shared mem operations\n",
1417 g->p.nr_loops, g->p.b 1383 g->p.nr_loops, g->p.bytes_process/1024/1024);
1418 printf(" # %5dx %5ldMB thread l 1384 printf(" # %5dx %5ldMB thread local mem operations\n",
1419 g->p.nr_loops, g->p.b 1385 g->p.nr_loops, g->p.bytes_thread/1024/1024);
1420 1386
1421 printf(" ###\n"); 1387 printf(" ###\n");
1422 1388
1423 printf("\n ###\n"); fflush(stdout); 1389 printf("\n ###\n"); fflush(stdout);
1424 } 1390 }
1425 1391
1426 static void init_thread_data(void) 1392 static void init_thread_data(void)
1427 { 1393 {
1428 ssize_t size = sizeof(*g->threads)*g- 1394 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1429 int t; 1395 int t;
1430 1396
1431 g->threads = zalloc_shared_data(size) 1397 g->threads = zalloc_shared_data(size);
1432 1398
1433 for (t = 0; t < g->p.nr_tasks; t++) { 1399 for (t = 0; t < g->p.nr_tasks; t++) {
1434 struct thread_data *td = g->t 1400 struct thread_data *td = g->threads + t;
1435 size_t cpuset_size = CPU_ALLO <<
1436 int cpu; 1401 int cpu;
1437 1402
1438 /* Allow all nodes by default 1403 /* Allow all nodes by default: */
1439 td->bind_node = NUMA_NO_NODE; 1404 td->bind_node = NUMA_NO_NODE;
1440 1405
1441 /* Allow all CPUs by default: 1406 /* Allow all CPUs by default: */
1442 td->bind_cpumask = CPU_ALLOC( !! 1407 CPU_ZERO(&td->bind_cpumask);
1443 BUG_ON(!td->bind_cpumask); <<
1444 CPU_ZERO_S(cpuset_size, td->b <<
1445 for (cpu = 0; cpu < g->p.nr_c 1408 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1446 CPU_SET_S(cpu, cpuset !! 1409 CPU_SET(cpu, &td->bind_cpumask);
1447 } 1410 }
1448 } 1411 }
1449 1412
1450 static void deinit_thread_data(void) 1413 static void deinit_thread_data(void)
1451 { 1414 {
1452 ssize_t size = sizeof(*g->threads)*g- 1415 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1453 int t; <<
1454 <<
1455 /* Free the bind_cpumask allocated fo <<
1456 for (t = 0; t < g->p.nr_tasks; t++) { <<
1457 struct thread_data *td = g->t <<
1458 CPU_FREE(td->bind_cpumask); <<
1459 } <<
1460 1416
1461 free_data(g->threads, size); 1417 free_data(g->threads, size);
1462 } 1418 }
1463 1419
1464 static int init(void) 1420 static int init(void)
1465 { 1421 {
1466 g = (void *)alloc_data(sizeof(*g), MA 1422 g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1467 1423
1468 /* Copy over options: */ 1424 /* Copy over options: */
1469 g->p = p0; 1425 g->p = p0;
1470 1426
1471 g->p.nr_cpus = numa_num_configured_cp 1427 g->p.nr_cpus = numa_num_configured_cpus();
1472 1428
1473 g->p.nr_nodes = numa_max_node() + 1; 1429 g->p.nr_nodes = numa_max_node() + 1;
1474 1430
1475 /* char array in count_process_nodes( 1431 /* char array in count_process_nodes(): */
1476 BUG_ON(g->p.nr_nodes < 0); 1432 BUG_ON(g->p.nr_nodes < 0);
1477 1433
1478 if (quiet && !g->p.show_details) !! 1434 if (g->p.show_quiet && !g->p.show_details)
1479 g->p.show_details = -1; 1435 g->p.show_details = -1;
1480 1436
1481 /* Some memory should be specified: * 1437 /* Some memory should be specified: */
1482 if (!g->p.mb_global_str && !g->p.mb_p 1438 if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1483 return -1; 1439 return -1;
1484 1440
1485 if (g->p.mb_global_str) { 1441 if (g->p.mb_global_str) {
1486 g->p.mb_global = atof(g->p.mb 1442 g->p.mb_global = atof(g->p.mb_global_str);
1487 BUG_ON(g->p.mb_global < 0); 1443 BUG_ON(g->p.mb_global < 0);
1488 } 1444 }
1489 1445
1490 if (g->p.mb_proc_str) { 1446 if (g->p.mb_proc_str) {
1491 g->p.mb_proc = atof(g->p.mb_p 1447 g->p.mb_proc = atof(g->p.mb_proc_str);
1492 BUG_ON(g->p.mb_proc < 0); 1448 BUG_ON(g->p.mb_proc < 0);
1493 } 1449 }
1494 1450
1495 if (g->p.mb_proc_locked_str) { 1451 if (g->p.mb_proc_locked_str) {
1496 g->p.mb_proc_locked = atof(g- 1452 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1497 BUG_ON(g->p.mb_proc_locked < 1453 BUG_ON(g->p.mb_proc_locked < 0);
1498 BUG_ON(g->p.mb_proc_locked > 1454 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1499 } 1455 }
1500 1456
1501 if (g->p.mb_thread_str) { 1457 if (g->p.mb_thread_str) {
1502 g->p.mb_thread = atof(g->p.mb 1458 g->p.mb_thread = atof(g->p.mb_thread_str);
1503 BUG_ON(g->p.mb_thread < 0); 1459 BUG_ON(g->p.mb_thread < 0);
1504 } 1460 }
1505 1461
1506 BUG_ON(g->p.nr_threads <= 0); 1462 BUG_ON(g->p.nr_threads <= 0);
1507 BUG_ON(g->p.nr_proc <= 0); 1463 BUG_ON(g->p.nr_proc <= 0);
1508 1464
1509 g->p.nr_tasks = g->p.nr_proc*g->p.nr_ 1465 g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1510 1466
1511 g->p.bytes_global = g-> 1467 g->p.bytes_global = g->p.mb_global *1024L*1024L;
1512 g->p.bytes_process = g-> 1468 g->p.bytes_process = g->p.mb_proc *1024L*1024L;
1513 g->p.bytes_process_locked = g-> 1469 g->p.bytes_process_locked = g->p.mb_proc_locked *1024L*1024L;
1514 g->p.bytes_thread = g-> 1470 g->p.bytes_thread = g->p.mb_thread *1024L*1024L;
1515 1471
1516 g->data = setup_shared_data(g->p.byte 1472 g->data = setup_shared_data(g->p.bytes_global);
1517 1473
1518 /* Startup serialization: */ 1474 /* Startup serialization: */
1519 mutex_init_pshared(&g->start_work_mut !! 1475 init_global_mutex(&g->start_work_mutex);
1520 cond_init_pshared(&g->start_work_cond !! 1476 init_global_cond(&g->start_work_cond);
1521 mutex_init_pshared(&g->startup_mutex) !! 1477 init_global_mutex(&g->startup_mutex);
1522 cond_init_pshared(&g->startup_cond); !! 1478 init_global_cond(&g->startup_cond);
1523 mutex_init_pshared(&g->stop_work_mute !! 1479 init_global_mutex(&g->stop_work_mutex);
1524 1480
1525 init_thread_data(); 1481 init_thread_data();
1526 1482
1527 tprintf("#\n"); 1483 tprintf("#\n");
1528 if (parse_setup_cpu_list() || parse_s 1484 if (parse_setup_cpu_list() || parse_setup_node_list())
1529 return -1; 1485 return -1;
1530 tprintf("#\n"); 1486 tprintf("#\n");
1531 1487
1532 print_summary(); 1488 print_summary();
1533 1489
1534 return 0; 1490 return 0;
1535 } 1491 }
1536 1492
1537 static void deinit(void) 1493 static void deinit(void)
1538 { 1494 {
1539 free_data(g->data, g->p.bytes_global) 1495 free_data(g->data, g->p.bytes_global);
1540 g->data = NULL; 1496 g->data = NULL;
1541 1497
1542 deinit_thread_data(); 1498 deinit_thread_data();
1543 1499
1544 free_data(g, sizeof(*g)); 1500 free_data(g, sizeof(*g));
1545 g = NULL; 1501 g = NULL;
1546 } 1502 }
1547 1503
1548 /* 1504 /*
1549 * Print a short or long result, depending on 1505 * Print a short or long result, depending on the verbosity setting:
1550 */ 1506 */
1551 static void print_res(const char *name, doubl 1507 static void print_res(const char *name, double val,
1552 const char *txt_unit, c 1508 const char *txt_unit, const char *txt_short, const char *txt_long)
1553 { 1509 {
1554 if (!name) 1510 if (!name)
1555 name = "main,"; 1511 name = "main,";
1556 1512
1557 if (!quiet) !! 1513 if (!g->p.show_quiet)
1558 printf(" %-30s %15.3f, %-15s 1514 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1559 else 1515 else
1560 printf(" %14.3f %s\n", val, t 1516 printf(" %14.3f %s\n", val, txt_long);
1561 } 1517 }
1562 1518
1563 static int __bench_numa(const char *name) 1519 static int __bench_numa(const char *name)
1564 { 1520 {
1565 struct timeval start, stop, diff; 1521 struct timeval start, stop, diff;
1566 u64 runtime_ns_min, runtime_ns_sum; 1522 u64 runtime_ns_min, runtime_ns_sum;
1567 pid_t *pids, pid, wpid; 1523 pid_t *pids, pid, wpid;
1568 double delta_runtime; 1524 double delta_runtime;
1569 double runtime_avg; 1525 double runtime_avg;
1570 double runtime_sec_max; 1526 double runtime_sec_max;
1571 double runtime_sec_min; 1527 double runtime_sec_min;
1572 int wait_stat; 1528 int wait_stat;
1573 double bytes; 1529 double bytes;
1574 int i, t, p; 1530 int i, t, p;
1575 1531
1576 if (init()) 1532 if (init())
1577 return -1; 1533 return -1;
1578 1534
1579 pids = zalloc(g->p.nr_proc * sizeof(* 1535 pids = zalloc(g->p.nr_proc * sizeof(*pids));
1580 pid = -1; 1536 pid = -1;
1581 1537
1582 if (g->p.serialize_startup) { 1538 if (g->p.serialize_startup) {
1583 tprintf(" #\n"); 1539 tprintf(" #\n");
1584 tprintf(" # Startup synchroni 1540 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1585 } 1541 }
1586 1542
1587 gettimeofday(&start, NULL); 1543 gettimeofday(&start, NULL);
1588 1544
1589 for (i = 0; i < g->p.nr_proc; i++) { 1545 for (i = 0; i < g->p.nr_proc; i++) {
1590 pid = fork(); 1546 pid = fork();
1591 dprintf(" # process %2d: PID 1547 dprintf(" # process %2d: PID %d\n", i, pid);
1592 1548
1593 BUG_ON(pid < 0); 1549 BUG_ON(pid < 0);
1594 if (!pid) { 1550 if (!pid) {
1595 /* Child process: */ 1551 /* Child process: */
1596 worker_process(i); 1552 worker_process(i);
1597 1553
1598 exit(0); 1554 exit(0);
1599 } 1555 }
1600 pids[i] = pid; 1556 pids[i] = pid;
1601 1557
1602 } 1558 }
1603 1559
1604 if (g->p.serialize_startup) { 1560 if (g->p.serialize_startup) {
1605 bool threads_ready = false; 1561 bool threads_ready = false;
1606 double startup_sec; 1562 double startup_sec;
1607 1563
1608 /* 1564 /*
1609 * Wait for all the threads t 1565 * Wait for all the threads to start up. The last thread will
1610 * signal this process. 1566 * signal this process.
1611 */ 1567 */
1612 mutex_lock(&g->startup_mutex) !! 1568 pthread_mutex_lock(&g->startup_mutex);
1613 while (g->nr_tasks_started != 1569 while (g->nr_tasks_started != g->p.nr_tasks)
1614 cond_wait(&g->startup !! 1570 pthread_cond_wait(&g->startup_cond, &g->startup_mutex);
1615 1571
1616 mutex_unlock(&g->startup_mute !! 1572 pthread_mutex_unlock(&g->startup_mutex);
1617 1573
1618 /* Wait for all threads to be 1574 /* Wait for all threads to be at the start_work_cond. */
1619 while (!threads_ready) { 1575 while (!threads_ready) {
1620 mutex_lock(&g->start_ !! 1576 pthread_mutex_lock(&g->start_work_mutex);
1621 threads_ready = (g->n 1577 threads_ready = (g->nr_tasks_working == g->p.nr_tasks);
1622 mutex_unlock(&g->star !! 1578 pthread_mutex_unlock(&g->start_work_mutex);
1623 if (!threads_ready) 1579 if (!threads_ready)
1624 usleep(1); 1580 usleep(1);
1625 } 1581 }
1626 1582
1627 gettimeofday(&stop, NULL); 1583 gettimeofday(&stop, NULL);
1628 1584
1629 timersub(&stop, &start, &diff 1585 timersub(&stop, &start, &diff);
1630 1586
1631 startup_sec = diff.tv_sec * N 1587 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1632 startup_sec += diff.tv_usec * 1588 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1633 startup_sec /= NSEC_PER_SEC; 1589 startup_sec /= NSEC_PER_SEC;
1634 1590
1635 tprintf(" threads initialized 1591 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1636 tprintf(" #\n"); 1592 tprintf(" #\n");
1637 1593
1638 start = stop; 1594 start = stop;
1639 /* Start all threads running. 1595 /* Start all threads running. */
1640 mutex_lock(&g->start_work_mut !! 1596 pthread_mutex_lock(&g->start_work_mutex);
1641 g->start_work = true; 1597 g->start_work = true;
1642 mutex_unlock(&g->start_work_m !! 1598 pthread_mutex_unlock(&g->start_work_mutex);
1643 cond_broadcast(&g->start_work !! 1599 pthread_cond_broadcast(&g->start_work_cond);
1644 } else { 1600 } else {
1645 gettimeofday(&start, NULL); 1601 gettimeofday(&start, NULL);
1646 } 1602 }
1647 1603
1648 /* Parent process: */ 1604 /* Parent process: */
1649 1605
1650 1606
1651 for (i = 0; i < g->p.nr_proc; i++) { 1607 for (i = 0; i < g->p.nr_proc; i++) {
1652 wpid = waitpid(pids[i], &wait 1608 wpid = waitpid(pids[i], &wait_stat, 0);
1653 BUG_ON(wpid < 0); 1609 BUG_ON(wpid < 0);
1654 BUG_ON(!WIFEXITED(wait_stat)) 1610 BUG_ON(!WIFEXITED(wait_stat));
1655 1611
1656 } 1612 }
1657 1613
1658 runtime_ns_sum = 0; 1614 runtime_ns_sum = 0;
1659 runtime_ns_min = -1LL; 1615 runtime_ns_min = -1LL;
1660 1616
1661 for (t = 0; t < g->p.nr_tasks; t++) { 1617 for (t = 0; t < g->p.nr_tasks; t++) {
1662 u64 thread_runtime_ns = g->th 1618 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1663 1619
1664 runtime_ns_sum += thread_runt 1620 runtime_ns_sum += thread_runtime_ns;
1665 runtime_ns_min = min(thread_r 1621 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1666 } 1622 }
1667 1623
1668 gettimeofday(&stop, NULL); 1624 gettimeofday(&stop, NULL);
1669 timersub(&stop, &start, &diff); 1625 timersub(&stop, &start, &diff);
1670 1626
1671 BUG_ON(bench_format != BENCH_FORMAT_D 1627 BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1672 1628
1673 tprintf("\n ###\n"); 1629 tprintf("\n ###\n");
1674 tprintf("\n"); 1630 tprintf("\n");
1675 1631
1676 runtime_sec_max = diff.tv_sec * NSEC_ 1632 runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1677 runtime_sec_max += diff.tv_usec * NSE 1633 runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1678 runtime_sec_max /= NSEC_PER_SEC; 1634 runtime_sec_max /= NSEC_PER_SEC;
1679 1635
1680 runtime_sec_min = runtime_ns_min / NS 1636 runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1681 1637
1682 bytes = g->bytes_done; 1638 bytes = g->bytes_done;
1683 runtime_avg = (double)runtime_ns_sum 1639 runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1684 1640
1685 if (g->p.measure_convergence) { 1641 if (g->p.measure_convergence) {
1686 print_res(name, runtime_sec_m 1642 print_res(name, runtime_sec_max,
1687 "secs,", "NUMA-conver 1643 "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1688 } 1644 }
1689 1645
1690 print_res(name, runtime_sec_max, 1646 print_res(name, runtime_sec_max,
1691 "secs,", "runtime-max/thread" 1647 "secs,", "runtime-max/thread", "secs slowest (max) thread-runtime");
1692 1648
1693 print_res(name, runtime_sec_min, 1649 print_res(name, runtime_sec_min,
1694 "secs,", "runtime-min/thread" 1650 "secs,", "runtime-min/thread", "secs fastest (min) thread-runtime");
1695 1651
1696 print_res(name, runtime_avg, 1652 print_res(name, runtime_avg,
1697 "secs,", "runtime-avg/thread" 1653 "secs,", "runtime-avg/thread", "secs average thread-runtime");
1698 1654
1699 delta_runtime = (runtime_sec_max - ru 1655 delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1700 print_res(name, delta_runtime / runti 1656 print_res(name, delta_runtime / runtime_sec_max * 100.0,
1701 "%,", "spread-runtime/thread" 1657 "%,", "spread-runtime/thread", "% difference between max/avg runtime");
1702 1658
1703 print_res(name, bytes / g->p.nr_tasks 1659 print_res(name, bytes / g->p.nr_tasks / 1e9,
1704 "GB,", "data/thread", 1660 "GB,", "data/thread", "GB data processed, per thread");
1705 1661
1706 print_res(name, bytes / 1e9, 1662 print_res(name, bytes / 1e9,
1707 "GB,", "data-total", 1663 "GB,", "data-total", "GB data processed, total");
1708 1664
1709 print_res(name, runtime_sec_max * NSE 1665 print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1710 "nsecs,", "runtime/byte/threa 1666 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1711 1667
1712 print_res(name, bytes / g->p.nr_tasks 1668 print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1713 "GB/sec,", "thread-speed", 1669 "GB/sec,", "thread-speed", "GB/sec/thread speed");
1714 1670
1715 print_res(name, bytes / runtime_sec_m 1671 print_res(name, bytes / runtime_sec_max / 1e9,
1716 "GB/sec,", "total-speed", 1672 "GB/sec,", "total-speed", "GB/sec total speed");
1717 1673
1718 if (g->p.show_details >= 2) { 1674 if (g->p.show_details >= 2) {
1719 char tname[14 + 2 * 11 + 1]; !! 1675 char tname[14 + 2 * 10 + 1];
1720 struct thread_data *td; 1676 struct thread_data *td;
1721 for (p = 0; p < g->p.nr_proc; 1677 for (p = 0; p < g->p.nr_proc; p++) {
1722 for (t = 0; t < g->p. 1678 for (t = 0; t < g->p.nr_threads; t++) {
1723 memset(tname, 1679 memset(tname, 0, sizeof(tname));
1724 td = g->threa 1680 td = g->threads + p*g->p.nr_threads + t;
1725 snprintf(tnam 1681 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1726 print_res(tna 1682 print_res(tname, td->speed_gbs,
1727 "GB/s 1683 "GB/sec", "thread-speed", "GB/sec/thread speed");
1728 print_res(tna 1684 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1729 "secs 1685 "secs", "thread-system-time", "system CPU time/thread");
1730 print_res(tna 1686 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1731 "secs 1687 "secs", "thread-user-time", "user CPU time/thread");
1732 } 1688 }
1733 } 1689 }
1734 } 1690 }
1735 1691
1736 free(pids); 1692 free(pids);
1737 1693
1738 deinit(); 1694 deinit();
1739 1695
1740 return 0; 1696 return 0;
1741 } 1697 }
1742 1698
1743 #define MAX_ARGS 50 1699 #define MAX_ARGS 50
1744 1700
1745 static int command_size(const char **argv) 1701 static int command_size(const char **argv)
1746 { 1702 {
1747 int size = 0; 1703 int size = 0;
1748 1704
1749 while (*argv) { 1705 while (*argv) {
1750 size++; 1706 size++;
1751 argv++; 1707 argv++;
1752 } 1708 }
1753 1709
1754 BUG_ON(size >= MAX_ARGS); 1710 BUG_ON(size >= MAX_ARGS);
1755 1711
1756 return size; 1712 return size;
1757 } 1713 }
1758 1714
1759 static void init_params(struct params *p, con 1715 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1760 { 1716 {
1761 int i; 1717 int i;
1762 1718
1763 printf("\n # Running %s \"perf bench 1719 printf("\n # Running %s \"perf bench numa", name);
1764 1720
1765 for (i = 0; i < argc; i++) 1721 for (i = 0; i < argc; i++)
1766 printf(" %s", argv[i]); 1722 printf(" %s", argv[i]);
1767 1723
1768 printf("\"\n"); 1724 printf("\"\n");
1769 1725
1770 memset(p, 0, sizeof(*p)); 1726 memset(p, 0, sizeof(*p));
1771 1727
1772 /* Initialize nonzero defaults: */ 1728 /* Initialize nonzero defaults: */
1773 1729
1774 p->serialize_startup = 1; 1730 p->serialize_startup = 1;
1775 p->data_reads = tru 1731 p->data_reads = true;
1776 p->data_writes = tru 1732 p->data_writes = true;
1777 p->data_backwards = tru 1733 p->data_backwards = true;
1778 p->data_rand_walk = tru 1734 p->data_rand_walk = true;
1779 p->nr_loops = -1; 1735 p->nr_loops = -1;
1780 p->init_random = tru 1736 p->init_random = true;
1781 p->mb_global_str = "1" 1737 p->mb_global_str = "1";
1782 p->nr_proc = 1; 1738 p->nr_proc = 1;
1783 p->nr_threads = 1; 1739 p->nr_threads = 1;
1784 p->nr_secs = 5; 1740 p->nr_secs = 5;
1785 p->run_all = arg 1741 p->run_all = argc == 1;
1786 } 1742 }
1787 1743
1788 static int run_bench_numa(const char *name, c 1744 static int run_bench_numa(const char *name, const char **argv)
1789 { 1745 {
1790 int argc = command_size(argv); 1746 int argc = command_size(argv);
1791 1747
1792 init_params(&p0, name, argc, argv); 1748 init_params(&p0, name, argc, argv);
1793 argc = parse_options(argc, argv, opti 1749 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1794 if (argc) 1750 if (argc)
1795 goto err; 1751 goto err;
1796 1752
1797 if (__bench_numa(name)) 1753 if (__bench_numa(name))
1798 goto err; 1754 goto err;
1799 1755
1800 return 0; 1756 return 0;
1801 1757
1802 err: 1758 err:
1803 return -1; 1759 return -1;
1804 } 1760 }
1805 1761
1806 #define OPT_BW_RAM "-s", "20", 1762 #define OPT_BW_RAM "-s", "20", "-zZq", "--thp", " 1", "--no-data_rand_walk"
1807 #define OPT_BW_RAM_NOTHP OPT_BW_RAM, 1763 #define OPT_BW_RAM_NOTHP OPT_BW_RAM, "--thp", "-1"
1808 1764
1809 #define OPT_CONV "-s", "100", 1765 #define OPT_CONV "-s", "100", "-zZ0qcm", "--thp", " 1"
1810 #define OPT_CONV_NOTHP OPT_CONV, 1766 #define OPT_CONV_NOTHP OPT_CONV, "--thp", "-1"
1811 1767
1812 #define OPT_BW "-s", "20", 1768 #define OPT_BW "-s", "20", "-zZ0q", "--thp", " 1"
1813 #define OPT_BW_NOTHP OPT_BW, 1769 #define OPT_BW_NOTHP OPT_BW, "--thp", "-1"
1814 1770
1815 /* 1771 /*
1816 * The built-in test-suite executed by "perf 1772 * The built-in test-suite executed by "perf bench numa -a".
1817 * 1773 *
1818 * (A minimum of 4 nodes and 16 GB of RAM is 1774 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1819 */ 1775 */
1820 static const char *tests[][MAX_ARGS] = { 1776 static const char *tests[][MAX_ARGS] = {
1821 /* Basic single-stream NUMA bandwidth meas 1777 /* Basic single-stream NUMA bandwidth measurements: */
1822 { "RAM-bw-local,", "mem", "-p", "1", 1778 { "RAM-bw-local,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1823 "-C" , "", "-M", 1779 "-C" , "", "-M", "", OPT_BW_RAM },
1824 { "RAM-bw-local-NOTHP,", 1780 { "RAM-bw-local-NOTHP,",
1825 "mem", "-p", "1", 1781 "mem", "-p", "1", "-t", "1", "-P", "1024",
1826 "-C" , "", "-M", 1782 "-C" , "", "-M", "", OPT_BW_RAM_NOTHP },
1827 { "RAM-bw-remote,", "mem", "-p", "1", 1783 { "RAM-bw-remote,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1828 "-C" , "", "-M", 1784 "-C" , "", "-M", "1", OPT_BW_RAM },
1829 1785
1830 /* 2-stream NUMA bandwidth measurements: * 1786 /* 2-stream NUMA bandwidth measurements: */
1831 { "RAM-bw-local-2x,", "mem", "-p", "2", 1787 { "RAM-bw-local-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1832 "-C", "0,2", "-M", 1788 "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1833 { "RAM-bw-remote-2x,", "mem", "-p", "2", 1789 { "RAM-bw-remote-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1834 "-C", "0,2", "-M", 1790 "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1835 1791
1836 /* Cross-stream NUMA bandwidth measurement 1792 /* Cross-stream NUMA bandwidth measurement: */
1837 { "RAM-bw-cross,", "mem", "-p", "2", 1793 { "RAM-bw-cross,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1838 "-C", "0,8", "-M", 1794 "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1839 1795
1840 /* Convergence latency measurements: */ 1796 /* Convergence latency measurements: */
1841 { " 1x3-convergence,", "mem", "-p", "1", 1797 { " 1x3-convergence,", "mem", "-p", "1", "-t", "3", "-P", "512", OPT_CONV },
1842 { " 1x4-convergence,", "mem", "-p", "1", 1798 { " 1x4-convergence,", "mem", "-p", "1", "-t", "4", "-P", "512", OPT_CONV },
1843 { " 1x6-convergence,", "mem", "-p", "1", 1799 { " 1x6-convergence,", "mem", "-p", "1", "-t", "6", "-P", "1020", OPT_CONV },
1844 { " 2x3-convergence,", "mem", "-p", "2", 1800 { " 2x3-convergence,", "mem", "-p", "2", "-t", "3", "-P", "1020", OPT_CONV },
1845 { " 3x3-convergence,", "mem", "-p", "3", 1801 { " 3x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1846 { " 4x4-convergence,", "mem", "-p", "4", 1802 { " 4x4-convergence,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV },
1847 { " 4x4-convergence-NOTHP,", 1803 { " 4x4-convergence-NOTHP,",
1848 "mem", "-p", "4", 1804 "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1849 { " 4x6-convergence,", "mem", "-p", "4", 1805 { " 4x6-convergence,", "mem", "-p", "4", "-t", "6", "-P", "1020", OPT_CONV },
1850 { " 4x8-convergence,", "mem", "-p", "4", 1806 { " 4x8-convergence,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_CONV },
1851 { " 8x4-convergence,", "mem", "-p", "8", 1807 { " 8x4-convergence,", "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV },
1852 { " 8x4-convergence-NOTHP,", 1808 { " 8x4-convergence-NOTHP,",
1853 "mem", "-p", "8", 1809 "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1854 { " 3x1-convergence,", "mem", "-p", "3", 1810 { " 3x1-convergence,", "mem", "-p", "3", "-t", "1", "-P", "512", OPT_CONV },
1855 { " 4x1-convergence,", "mem", "-p", "4", 1811 { " 4x1-convergence,", "mem", "-p", "4", "-t", "1", "-P", "512", OPT_CONV },
1856 { " 8x1-convergence,", "mem", "-p", "8", 1812 { " 8x1-convergence,", "mem", "-p", "8", "-t", "1", "-P", "512", OPT_CONV },
1857 { "16x1-convergence,", "mem", "-p", "16", 1813 { "16x1-convergence,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_CONV },
1858 { "32x1-convergence,", "mem", "-p", "32", 1814 { "32x1-convergence,", "mem", "-p", "32", "-t", "1", "-P", "128", OPT_CONV },
1859 1815
1860 /* Various NUMA process/thread layout band 1816 /* Various NUMA process/thread layout bandwidth measurements: */
1861 { " 2x1-bw-process,", "mem", "-p", "2", 1817 { " 2x1-bw-process,", "mem", "-p", "2", "-t", "1", "-P", "1024", OPT_BW },
1862 { " 3x1-bw-process,", "mem", "-p", "3", 1818 { " 3x1-bw-process,", "mem", "-p", "3", "-t", "1", "-P", "1024", OPT_BW },
1863 { " 4x1-bw-process,", "mem", "-p", "4", 1819 { " 4x1-bw-process,", "mem", "-p", "4", "-t", "1", "-P", "1024", OPT_BW },
1864 { " 8x1-bw-process,", "mem", "-p", "8", 1820 { " 8x1-bw-process,", "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW },
1865 { " 8x1-bw-process-NOTHP,", 1821 { " 8x1-bw-process-NOTHP,",
1866 "mem", "-p", "8", 1822 "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW_NOTHP },
1867 { "16x1-bw-process,", "mem", "-p", "16", 1823 { "16x1-bw-process,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_BW },
1868 1824
1869 { " 1x4-bw-thread,", "mem", "-p", "1", 1825 { " 1x4-bw-thread,", "mem", "-p", "1", "-t", "4", "-T", "256", OPT_BW },
1870 { " 1x8-bw-thread,", "mem", "-p", "1", 1826 { " 1x8-bw-thread,", "mem", "-p", "1", "-t", "8", "-T", "256", OPT_BW },
1871 { "1x16-bw-thread,", "mem", "-p", "1", 1827 { "1x16-bw-thread,", "mem", "-p", "1", "-t", "16", "-T", "128", OPT_BW },
1872 { "1x32-bw-thread,", "mem", "-p", "1", 1828 { "1x32-bw-thread,", "mem", "-p", "1", "-t", "32", "-T", "64", OPT_BW },
1873 1829
1874 { " 2x3-bw-process,", "mem", "-p", "2", 1830 { " 2x3-bw-process,", "mem", "-p", "2", "-t", "3", "-P", "512", OPT_BW },
1875 { " 4x4-bw-process,", "mem", "-p", "4", 1831 { " 4x4-bw-process,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_BW },
1876 { " 4x6-bw-process,", "mem", "-p", "4", 1832 { " 4x6-bw-process,", "mem", "-p", "4", "-t", "6", "-P", "512", OPT_BW },
1877 { " 4x8-bw-process,", "mem", "-p", "4", 1833 { " 4x8-bw-process,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW },
1878 { " 4x8-bw-process-NOTHP,", 1834 { " 4x8-bw-process-NOTHP,",
1879 "mem", "-p", "4", 1835 "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW_NOTHP },
1880 { " 3x3-bw-process,", "mem", "-p", "3", 1836 { " 3x3-bw-process,", "mem", "-p", "3", "-t", "3", "-P", "512", OPT_BW },
1881 { " 5x5-bw-process,", "mem", "-p", "5", 1837 { " 5x5-bw-process,", "mem", "-p", "5", "-t", "5", "-P", "512", OPT_BW },
1882 1838
1883 { "2x16-bw-process,", "mem", "-p", "2", 1839 { "2x16-bw-process,", "mem", "-p", "2", "-t", "16", "-P", "512", OPT_BW },
1884 { "1x32-bw-process,", "mem", "-p", "1", 1840 { "1x32-bw-process,", "mem", "-p", "1", "-t", "32", "-P", "2048", OPT_BW },
1885 1841
1886 { "numa02-bw,", "mem", "-p", "1", 1842 { "numa02-bw,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW },
1887 { "numa02-bw-NOTHP,", "mem", "-p", "1", 1843 { "numa02-bw-NOTHP,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW_NOTHP },
1888 { "numa01-bw-thread,", "mem", "-p", "2", 1844 { "numa01-bw-thread,", "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW },
1889 { "numa01-bw-thread-NOTHP,", 1845 { "numa01-bw-thread-NOTHP,",
1890 "mem", "-p", "2", 1846 "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW_NOTHP },
1891 }; 1847 };
1892 1848
1893 static int bench_all(void) 1849 static int bench_all(void)
1894 { 1850 {
1895 int nr = ARRAY_SIZE(tests); 1851 int nr = ARRAY_SIZE(tests);
1896 int ret; 1852 int ret;
1897 int i; 1853 int i;
1898 1854
1899 ret = system("echo ' #'; echo ' # Run 1855 ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1900 BUG_ON(ret < 0); 1856 BUG_ON(ret < 0);
1901 1857
1902 for (i = 0; i < nr; i++) { 1858 for (i = 0; i < nr; i++) {
1903 run_bench_numa(tests[i][0], t 1859 run_bench_numa(tests[i][0], tests[i] + 1);
1904 } 1860 }
1905 1861
1906 printf("\n"); 1862 printf("\n");
1907 1863
1908 return 0; 1864 return 0;
1909 } 1865 }
1910 1866
1911 int bench_numa(int argc, const char **argv) 1867 int bench_numa(int argc, const char **argv)
1912 { 1868 {
1913 init_params(&p0, "main,", argc, argv) 1869 init_params(&p0, "main,", argc, argv);
1914 argc = parse_options(argc, argv, opti 1870 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1915 if (argc) 1871 if (argc)
1916 goto err; 1872 goto err;
1917 1873
1918 if (p0.run_all) 1874 if (p0.run_all)
1919 return bench_all(); 1875 return bench_all();
1920 1876
1921 if (__bench_numa(NULL)) 1877 if (__bench_numa(NULL))
1922 goto err; 1878 goto err;
1923 1879
1924 return 0; 1880 return 0;
1925 1881
1926 err: 1882 err:
1927 usage_with_options(numa_usage, option 1883 usage_with_options(numa_usage, options);
1928 return -1; 1884 return -1;
1929 } 1885 }
1930 1886
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.