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