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