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