Loading...
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
55struct 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
74struct 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
136struct global_info {
137 u8 *data;
138
139 pthread_mutex_t startup_mutex;
140 int nr_tasks_started;
141
142 pthread_mutex_t startup_done_mutex;
143
144 pthread_mutex_t start_work_mutex;
145 int nr_tasks_working;
146
147 pthread_mutex_t stop_work_mutex;
148 u64 bytes_done;
149
150 struct thread_data *threads;
151
152 /* Convergence latency measurement: */
153 bool all_converged;
154 bool stop_work;
155
156 int print_once;
157
158 struct params p;
159};
160
161static struct global_info *g = NULL;
162
163static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
164static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
165
166struct params p0;
167
168static const struct option options[] = {
169 OPT_INTEGER('p', "nr_proc" , &p0.nr_proc, "number of processes"),
170 OPT_INTEGER('t', "nr_threads" , &p0.nr_threads, "number of threads per process"),
171
172 OPT_STRING('G', "mb_global" , &p0.mb_global_str, "MB", "global memory (MBs)"),
173 OPT_STRING('P', "mb_proc" , &p0.mb_proc_str, "MB", "process memory (MBs)"),
174 OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
175 OPT_STRING('T', "mb_thread" , &p0.mb_thread_str, "MB", "thread memory (MBs)"),
176
177 OPT_UINTEGER('l', "nr_loops" , &p0.nr_loops, "max number of loops to run (default: unlimited)"),
178 OPT_UINTEGER('s', "nr_secs" , &p0.nr_secs, "max number of seconds to run (default: 5 secs)"),
179 OPT_UINTEGER('u', "usleep" , &p0.sleep_usecs, "usecs to sleep per loop iteration"),
180
181 OPT_BOOLEAN('R', "data_reads" , &p0.data_reads, "access the data via reads (can be mixed with -W)"),
182 OPT_BOOLEAN('W', "data_writes" , &p0.data_writes, "access the data via writes (can be mixed with -R)"),
183 OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards, "access the data backwards as well"),
184 OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
185 OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk, "access the data with random (32bit LFSR) walk"),
186
187
188 OPT_BOOLEAN('z', "init_zero" , &p0.init_zero, "bzero the initial allocations"),
189 OPT_BOOLEAN('I', "init_random" , &p0.init_random, "randomize the contents of the initial allocations"),
190 OPT_BOOLEAN('0', "init_cpu0" , &p0.init_cpu0, "do the initial allocations on CPU#0"),
191 OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs, "perturb thread 0/0 every X secs, to test convergence stability"),
192
193 OPT_INCR ('d', "show_details" , &p0.show_details, "Show details"),
194 OPT_INCR ('a', "all" , &p0.run_all, "Run all tests in the suite"),
195 OPT_INTEGER('H', "thp" , &p0.thp, "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
196 OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
197 "convergence is reached when each process (all its threads) is running on a single NUMA node."),
198 OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
199 OPT_BOOLEAN('q', "quiet" , &p0.show_quiet, "quiet mode"),
200 OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
201
202 /* Special option string parsing callbacks: */
203 OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
204 "bind the first N tasks to these specific cpus (the rest is unbound)",
205 parse_cpus_opt),
206 OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
207 "bind the first N tasks to these specific memory nodes (the rest is unbound)",
208 parse_nodes_opt),
209 OPT_END()
210};
211
212static const char * const bench_numa_usage[] = {
213 "perf bench numa <options>",
214 NULL
215};
216
217static const char * const numa_usage[] = {
218 "perf bench numa mem [<options>]",
219 NULL
220};
221
222/*
223 * To get number of numa nodes present.
224 */
225static int nr_numa_nodes(void)
226{
227 int i, nr_nodes = 0;
228
229 for (i = 0; i < g->p.nr_nodes; i++) {
230 if (numa_bitmask_isbitset(numa_nodes_ptr, i))
231 nr_nodes++;
232 }
233
234 return nr_nodes;
235}
236
237/*
238 * To check if given numa node is present.
239 */
240static int is_node_present(int node)
241{
242 return numa_bitmask_isbitset(numa_nodes_ptr, node);
243}
244
245/*
246 * To check given numa node has cpus.
247 */
248static bool node_has_cpus(int node)
249{
250 struct bitmask *cpu = numa_allocate_cpumask();
251 unsigned int i;
252
253 if (cpu && !numa_node_to_cpus(node, cpu)) {
254 for (i = 0; i < cpu->size; i++) {
255 if (numa_bitmask_isbitset(cpu, i))
256 return true;
257 }
258 }
259
260 return false; /* lets fall back to nocpus safely */
261}
262
263static cpu_set_t bind_to_cpu(int target_cpu)
264{
265 cpu_set_t orig_mask, mask;
266 int ret;
267
268 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
269 BUG_ON(ret);
270
271 CPU_ZERO(&mask);
272
273 if (target_cpu == -1) {
274 int cpu;
275
276 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
277 CPU_SET(cpu, &mask);
278 } else {
279 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
280 CPU_SET(target_cpu, &mask);
281 }
282
283 ret = sched_setaffinity(0, sizeof(mask), &mask);
284 BUG_ON(ret);
285
286 return orig_mask;
287}
288
289static cpu_set_t bind_to_node(int target_node)
290{
291 int cpus_per_node = g->p.nr_cpus / nr_numa_nodes();
292 cpu_set_t orig_mask, mask;
293 int cpu;
294 int ret;
295
296 BUG_ON(cpus_per_node * nr_numa_nodes() != g->p.nr_cpus);
297 BUG_ON(!cpus_per_node);
298
299 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
300 BUG_ON(ret);
301
302 CPU_ZERO(&mask);
303
304 if (target_node == NUMA_NO_NODE) {
305 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
306 CPU_SET(cpu, &mask);
307 } else {
308 int cpu_start = (target_node + 0) * cpus_per_node;
309 int cpu_stop = (target_node + 1) * cpus_per_node;
310
311 BUG_ON(cpu_stop > g->p.nr_cpus);
312
313 for (cpu = cpu_start; cpu < cpu_stop; cpu++)
314 CPU_SET(cpu, &mask);
315 }
316
317 ret = sched_setaffinity(0, sizeof(mask), &mask);
318 BUG_ON(ret);
319
320 return orig_mask;
321}
322
323static void bind_to_cpumask(cpu_set_t mask)
324{
325 int ret;
326
327 ret = sched_setaffinity(0, sizeof(mask), &mask);
328 BUG_ON(ret);
329}
330
331static void mempol_restore(void)
332{
333 int ret;
334
335 ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
336
337 BUG_ON(ret);
338}
339
340static void bind_to_memnode(int node)
341{
342 unsigned long nodemask;
343 int ret;
344
345 if (node == NUMA_NO_NODE)
346 return;
347
348 BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
349 nodemask = 1L << node;
350
351 ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
352 dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
353
354 BUG_ON(ret);
355}
356
357#define HPSIZE (2*1024*1024)
358
359#define set_taskname(fmt...) \
360do { \
361 char name[20]; \
362 \
363 snprintf(name, 20, fmt); \
364 prctl(PR_SET_NAME, name); \
365} while (0)
366
367static u8 *alloc_data(ssize_t bytes0, int map_flags,
368 int init_zero, int init_cpu0, int thp, int init_random)
369{
370 cpu_set_t orig_mask;
371 ssize_t bytes;
372 u8 *buf;
373 int ret;
374
375 if (!bytes0)
376 return NULL;
377
378 /* Allocate and initialize all memory on CPU#0: */
379 if (init_cpu0) {
380 int node = numa_node_of_cpu(0);
381
382 orig_mask = bind_to_node(node);
383 bind_to_memnode(node);
384 }
385
386 bytes = bytes0 + HPSIZE;
387
388 buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
389 BUG_ON(buf == (void *)-1);
390
391 if (map_flags == MAP_PRIVATE) {
392 if (thp > 0) {
393 ret = madvise(buf, bytes, MADV_HUGEPAGE);
394 if (ret && !g->print_once) {
395 g->print_once = 1;
396 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
397 }
398 }
399 if (thp < 0) {
400 ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
401 if (ret && !g->print_once) {
402 g->print_once = 1;
403 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
404 }
405 }
406 }
407
408 if (init_zero) {
409 bzero(buf, bytes);
410 } else {
411 /* Initialize random contents, different in each word: */
412 if (init_random) {
413 u64 *wbuf = (void *)buf;
414 long off = rand();
415 long i;
416
417 for (i = 0; i < bytes/8; i++)
418 wbuf[i] = i + off;
419 }
420 }
421
422 /* Align to 2MB boundary: */
423 buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
424
425 /* Restore affinity: */
426 if (init_cpu0) {
427 bind_to_cpumask(orig_mask);
428 mempol_restore();
429 }
430
431 return buf;
432}
433
434static void free_data(void *data, ssize_t bytes)
435{
436 int ret;
437
438 if (!data)
439 return;
440
441 ret = munmap(data, bytes);
442 BUG_ON(ret);
443}
444
445/*
446 * Create a shared memory buffer that can be shared between processes, zeroed:
447 */
448static void * zalloc_shared_data(ssize_t bytes)
449{
450 return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0, g->p.thp, g->p.init_random);
451}
452
453/*
454 * Create a shared memory buffer that can be shared between processes:
455 */
456static void * setup_shared_data(ssize_t bytes)
457{
458 return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
459}
460
461/*
462 * Allocate process-local memory - this will either be shared between
463 * threads of this process, or only be accessed by this thread:
464 */
465static void * setup_private_data(ssize_t bytes)
466{
467 return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
468}
469
470/*
471 * Return a process-shared (global) mutex:
472 */
473static void init_global_mutex(pthread_mutex_t *mutex)
474{
475 pthread_mutexattr_t attr;
476
477 pthread_mutexattr_init(&attr);
478 pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
479 pthread_mutex_init(mutex, &attr);
480}
481
482static int parse_cpu_list(const char *arg)
483{
484 p0.cpu_list_str = strdup(arg);
485
486 dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
487
488 return 0;
489}
490
491static int parse_setup_cpu_list(void)
492{
493 struct thread_data *td;
494 char *str0, *str;
495 int t;
496
497 if (!g->p.cpu_list_str)
498 return 0;
499
500 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
501
502 str0 = str = strdup(g->p.cpu_list_str);
503 t = 0;
504
505 BUG_ON(!str);
506
507 tprintf("# binding tasks to CPUs:\n");
508 tprintf("# ");
509
510 while (true) {
511 int bind_cpu, bind_cpu_0, bind_cpu_1;
512 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
513 int bind_len;
514 int step;
515 int mul;
516
517 tok = strsep(&str, ",");
518 if (!tok)
519 break;
520
521 tok_end = strstr(tok, "-");
522
523 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
524 if (!tok_end) {
525 /* Single CPU specified: */
526 bind_cpu_0 = bind_cpu_1 = atol(tok);
527 } else {
528 /* CPU range specified (for example: "5-11"): */
529 bind_cpu_0 = atol(tok);
530 bind_cpu_1 = atol(tok_end + 1);
531 }
532
533 step = 1;
534 tok_step = strstr(tok, "#");
535 if (tok_step) {
536 step = atol(tok_step + 1);
537 BUG_ON(step <= 0 || step >= g->p.nr_cpus);
538 }
539
540 /*
541 * Mask length.
542 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
543 * where the _4 means the next 4 CPUs are allowed.
544 */
545 bind_len = 1;
546 tok_len = strstr(tok, "_");
547 if (tok_len) {
548 bind_len = atol(tok_len + 1);
549 BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
550 }
551
552 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
553 mul = 1;
554 tok_mul = strstr(tok, "x");
555 if (tok_mul) {
556 mul = atol(tok_mul + 1);
557 BUG_ON(mul <= 0);
558 }
559
560 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
561
562 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
563 printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
564 return -1;
565 }
566
567 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
568 BUG_ON(bind_cpu_0 > bind_cpu_1);
569
570 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
571 int i;
572
573 for (i = 0; i < mul; i++) {
574 int cpu;
575
576 if (t >= g->p.nr_tasks) {
577 printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
578 goto out;
579 }
580 td = g->threads + t;
581
582 if (t)
583 tprintf(",");
584 if (bind_len > 1) {
585 tprintf("%2d/%d", bind_cpu, bind_len);
586 } else {
587 tprintf("%2d", bind_cpu);
588 }
589
590 CPU_ZERO(&td->bind_cpumask);
591 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
592 BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
593 CPU_SET(cpu, &td->bind_cpumask);
594 }
595 t++;
596 }
597 }
598 }
599out:
600
601 tprintf("\n");
602
603 if (t < g->p.nr_tasks)
604 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
605
606 free(str0);
607 return 0;
608}
609
610static int parse_cpus_opt(const struct option *opt __maybe_unused,
611 const char *arg, int unset __maybe_unused)
612{
613 if (!arg)
614 return -1;
615
616 return parse_cpu_list(arg);
617}
618
619static int parse_node_list(const char *arg)
620{
621 p0.node_list_str = strdup(arg);
622
623 dprintf("got NODE list: {%s}\n", p0.node_list_str);
624
625 return 0;
626}
627
628static int parse_setup_node_list(void)
629{
630 struct thread_data *td;
631 char *str0, *str;
632 int t;
633
634 if (!g->p.node_list_str)
635 return 0;
636
637 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
638
639 str0 = str = strdup(g->p.node_list_str);
640 t = 0;
641
642 BUG_ON(!str);
643
644 tprintf("# binding tasks to NODEs:\n");
645 tprintf("# ");
646
647 while (true) {
648 int bind_node, bind_node_0, bind_node_1;
649 char *tok, *tok_end, *tok_step, *tok_mul;
650 int step;
651 int mul;
652
653 tok = strsep(&str, ",");
654 if (!tok)
655 break;
656
657 tok_end = strstr(tok, "-");
658
659 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
660 if (!tok_end) {
661 /* Single NODE specified: */
662 bind_node_0 = bind_node_1 = atol(tok);
663 } else {
664 /* NODE range specified (for example: "5-11"): */
665 bind_node_0 = atol(tok);
666 bind_node_1 = atol(tok_end + 1);
667 }
668
669 step = 1;
670 tok_step = strstr(tok, "#");
671 if (tok_step) {
672 step = atol(tok_step + 1);
673 BUG_ON(step <= 0 || step >= g->p.nr_nodes);
674 }
675
676 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
677 mul = 1;
678 tok_mul = strstr(tok, "x");
679 if (tok_mul) {
680 mul = atol(tok_mul + 1);
681 BUG_ON(mul <= 0);
682 }
683
684 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
685
686 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
687 printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
688 return -1;
689 }
690
691 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
692 BUG_ON(bind_node_0 > bind_node_1);
693
694 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
695 int i;
696
697 for (i = 0; i < mul; i++) {
698 if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
699 printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
700 goto out;
701 }
702 td = g->threads + t;
703
704 if (!t)
705 tprintf(" %2d", bind_node);
706 else
707 tprintf(",%2d", bind_node);
708
709 td->bind_node = bind_node;
710 t++;
711 }
712 }
713 }
714out:
715
716 tprintf("\n");
717
718 if (t < g->p.nr_tasks)
719 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
720
721 free(str0);
722 return 0;
723}
724
725static int parse_nodes_opt(const struct option *opt __maybe_unused,
726 const char *arg, int unset __maybe_unused)
727{
728 if (!arg)
729 return -1;
730
731 return parse_node_list(arg);
732
733 return 0;
734}
735
736#define BIT(x) (1ul << x)
737
738static inline uint32_t lfsr_32(uint32_t lfsr)
739{
740 const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
741 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
742}
743
744/*
745 * Make sure there's real data dependency to RAM (when read
746 * accesses are enabled), so the compiler, the CPU and the
747 * kernel (KSM, zero page, etc.) cannot optimize away RAM
748 * accesses:
749 */
750static inline u64 access_data(u64 *data, u64 val)
751{
752 if (g->p.data_reads)
753 val += *data;
754 if (g->p.data_writes)
755 *data = val + 1;
756 return val;
757}
758
759/*
760 * The worker process does two types of work, a forwards going
761 * loop and a backwards going loop.
762 *
763 * We do this so that on multiprocessor systems we do not create
764 * a 'train' of processing, with highly synchronized processes,
765 * skewing the whole benchmark.
766 */
767static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
768{
769 long words = bytes/sizeof(u64);
770 u64 *data = (void *)__data;
771 long chunk_0, chunk_1;
772 u64 *d0, *d, *d1;
773 long off;
774 long i;
775
776 BUG_ON(!data && words);
777 BUG_ON(data && !words);
778
779 if (!data)
780 return val;
781
782 /* Very simple memset() work variant: */
783 if (g->p.data_zero_memset && !g->p.data_rand_walk) {
784 bzero(data, bytes);
785 return val;
786 }
787
788 /* Spread out by PID/TID nr and by loop nr: */
789 chunk_0 = words/nr_max;
790 chunk_1 = words/g->p.nr_loops;
791 off = nr*chunk_0 + loop*chunk_1;
792
793 while (off >= words)
794 off -= words;
795
796 if (g->p.data_rand_walk) {
797 u32 lfsr = nr + loop + val;
798 int j;
799
800 for (i = 0; i < words/1024; i++) {
801 long start, end;
802
803 lfsr = lfsr_32(lfsr);
804
805 start = lfsr % words;
806 end = min(start + 1024, words-1);
807
808 if (g->p.data_zero_memset) {
809 bzero(data + start, (end-start) * sizeof(u64));
810 } else {
811 for (j = start; j < end; j++)
812 val = access_data(data + j, val);
813 }
814 }
815 } else if (!g->p.data_backwards || (nr + loop) & 1) {
816
817 d0 = data + off;
818 d = data + off + 1;
819 d1 = data + words;
820
821 /* Process data forwards: */
822 for (;;) {
823 if (unlikely(d >= d1))
824 d = data;
825 if (unlikely(d == d0))
826 break;
827
828 val = access_data(d, val);
829
830 d++;
831 }
832 } else {
833 /* Process data backwards: */
834
835 d0 = data + off;
836 d = data + off - 1;
837 d1 = data + words;
838
839 /* Process data forwards: */
840 for (;;) {
841 if (unlikely(d < data))
842 d = data + words-1;
843 if (unlikely(d == d0))
844 break;
845
846 val = access_data(d, val);
847
848 d--;
849 }
850 }
851
852 return val;
853}
854
855static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
856{
857 unsigned int cpu;
858
859 cpu = sched_getcpu();
860
861 g->threads[task_nr].curr_cpu = cpu;
862 prctl(0, bytes_worked);
863}
864
865#define MAX_NR_NODES 64
866
867/*
868 * Count the number of nodes a process's threads
869 * are spread out on.
870 *
871 * A count of 1 means that the process is compressed
872 * to a single node. A count of g->p.nr_nodes means it's
873 * spread out on the whole system.
874 */
875static int count_process_nodes(int process_nr)
876{
877 char node_present[MAX_NR_NODES] = { 0, };
878 int nodes;
879 int n, t;
880
881 for (t = 0; t < g->p.nr_threads; t++) {
882 struct thread_data *td;
883 int task_nr;
884 int node;
885
886 task_nr = process_nr*g->p.nr_threads + t;
887 td = g->threads + task_nr;
888
889 node = numa_node_of_cpu(td->curr_cpu);
890 if (node < 0) /* curr_cpu was likely still -1 */
891 return 0;
892
893 node_present[node] = 1;
894 }
895
896 nodes = 0;
897
898 for (n = 0; n < MAX_NR_NODES; n++)
899 nodes += node_present[n];
900
901 return nodes;
902}
903
904/*
905 * Count the number of distinct process-threads a node contains.
906 *
907 * A count of 1 means that the node contains only a single
908 * process. If all nodes on the system contain at most one
909 * process then we are well-converged.
910 */
911static int count_node_processes(int node)
912{
913 int processes = 0;
914 int t, p;
915
916 for (p = 0; p < g->p.nr_proc; p++) {
917 for (t = 0; t < g->p.nr_threads; t++) {
918 struct thread_data *td;
919 int task_nr;
920 int n;
921
922 task_nr = p*g->p.nr_threads + t;
923 td = g->threads + task_nr;
924
925 n = numa_node_of_cpu(td->curr_cpu);
926 if (n == node) {
927 processes++;
928 break;
929 }
930 }
931 }
932
933 return processes;
934}
935
936static void calc_convergence_compression(int *strong)
937{
938 unsigned int nodes_min, nodes_max;
939 int p;
940
941 nodes_min = -1;
942 nodes_max = 0;
943
944 for (p = 0; p < g->p.nr_proc; p++) {
945 unsigned int nodes = count_process_nodes(p);
946
947 if (!nodes) {
948 *strong = 0;
949 return;
950 }
951
952 nodes_min = min(nodes, nodes_min);
953 nodes_max = max(nodes, nodes_max);
954 }
955
956 /* Strong convergence: all threads compress on a single node: */
957 if (nodes_min == 1 && nodes_max == 1) {
958 *strong = 1;
959 } else {
960 *strong = 0;
961 tprintf(" {%d-%d}", nodes_min, nodes_max);
962 }
963}
964
965static void calc_convergence(double runtime_ns_max, double *convergence)
966{
967 unsigned int loops_done_min, loops_done_max;
968 int process_groups;
969 int nodes[MAX_NR_NODES];
970 int distance;
971 int nr_min;
972 int nr_max;
973 int strong;
974 int sum;
975 int nr;
976 int node;
977 int cpu;
978 int t;
979
980 if (!g->p.show_convergence && !g->p.measure_convergence)
981 return;
982
983 for (node = 0; node < g->p.nr_nodes; node++)
984 nodes[node] = 0;
985
986 loops_done_min = -1;
987 loops_done_max = 0;
988
989 for (t = 0; t < g->p.nr_tasks; t++) {
990 struct thread_data *td = g->threads + t;
991 unsigned int loops_done;
992
993 cpu = td->curr_cpu;
994
995 /* Not all threads have written it yet: */
996 if (cpu < 0)
997 continue;
998
999 node = numa_node_of_cpu(cpu);
1000
1001 nodes[node]++;
1002
1003 loops_done = td->loops_done;
1004 loops_done_min = min(loops_done, loops_done_min);
1005 loops_done_max = max(loops_done, loops_done_max);
1006 }
1007
1008 nr_max = 0;
1009 nr_min = g->p.nr_tasks;
1010 sum = 0;
1011
1012 for (node = 0; node < g->p.nr_nodes; node++) {
1013 if (!is_node_present(node))
1014 continue;
1015 nr = nodes[node];
1016 nr_min = min(nr, nr_min);
1017 nr_max = max(nr, nr_max);
1018 sum += nr;
1019 }
1020 BUG_ON(nr_min > nr_max);
1021
1022 BUG_ON(sum > g->p.nr_tasks);
1023
1024 if (0 && (sum < g->p.nr_tasks))
1025 return;
1026
1027 /*
1028 * Count the number of distinct process groups present
1029 * on nodes - when we are converged this will decrease
1030 * to g->p.nr_proc:
1031 */
1032 process_groups = 0;
1033
1034 for (node = 0; node < g->p.nr_nodes; node++) {
1035 int processes;
1036
1037 if (!is_node_present(node))
1038 continue;
1039 processes = count_node_processes(node);
1040 nr = nodes[node];
1041 tprintf(" %2d/%-2d", nr, processes);
1042
1043 process_groups += processes;
1044 }
1045
1046 distance = nr_max - nr_min;
1047
1048 tprintf(" [%2d/%-2d]", distance, process_groups);
1049
1050 tprintf(" l:%3d-%-3d (%3d)",
1051 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1052
1053 if (loops_done_min && loops_done_max) {
1054 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1055
1056 tprintf(" [%4.1f%%]", skew * 100.0);
1057 }
1058
1059 calc_convergence_compression(&strong);
1060
1061 if (strong && process_groups == g->p.nr_proc) {
1062 if (!*convergence) {
1063 *convergence = runtime_ns_max;
1064 tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1065 if (g->p.measure_convergence) {
1066 g->all_converged = true;
1067 g->stop_work = true;
1068 }
1069 }
1070 } else {
1071 if (*convergence) {
1072 tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1073 *convergence = 0;
1074 }
1075 tprintf("\n");
1076 }
1077}
1078
1079static void show_summary(double runtime_ns_max, int l, double *convergence)
1080{
1081 tprintf("\r # %5.1f%% [%.1f mins]",
1082 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1083
1084 calc_convergence(runtime_ns_max, convergence);
1085
1086 if (g->p.show_details >= 0)
1087 fflush(stdout);
1088}
1089
1090static void *worker_thread(void *__tdata)
1091{
1092 struct thread_data *td = __tdata;
1093 struct timeval start0, start, stop, diff;
1094 int process_nr = td->process_nr;
1095 int thread_nr = td->thread_nr;
1096 unsigned long last_perturbance;
1097 int task_nr = td->task_nr;
1098 int details = g->p.show_details;
1099 int first_task, last_task;
1100 double convergence = 0;
1101 u64 val = td->val;
1102 double runtime_ns_max;
1103 u8 *global_data;
1104 u8 *process_data;
1105 u8 *thread_data;
1106 u64 bytes_done, secs;
1107 long work_done;
1108 u32 l;
1109 struct rusage rusage;
1110
1111 bind_to_cpumask(td->bind_cpumask);
1112 bind_to_memnode(td->bind_node);
1113
1114 set_taskname("thread %d/%d", process_nr, thread_nr);
1115
1116 global_data = g->data;
1117 process_data = td->process_data;
1118 thread_data = setup_private_data(g->p.bytes_thread);
1119
1120 bytes_done = 0;
1121
1122 last_task = 0;
1123 if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1124 last_task = 1;
1125
1126 first_task = 0;
1127 if (process_nr == 0 && thread_nr == 0)
1128 first_task = 1;
1129
1130 if (details >= 2) {
1131 printf("# thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1132 process_nr, thread_nr, global_data, process_data, thread_data);
1133 }
1134
1135 if (g->p.serialize_startup) {
1136 pthread_mutex_lock(&g->startup_mutex);
1137 g->nr_tasks_started++;
1138 pthread_mutex_unlock(&g->startup_mutex);
1139
1140 /* Here we will wait for the main process to start us all at once: */
1141 pthread_mutex_lock(&g->start_work_mutex);
1142 g->nr_tasks_working++;
1143
1144 /* Last one wake the main process: */
1145 if (g->nr_tasks_working == g->p.nr_tasks)
1146 pthread_mutex_unlock(&g->startup_done_mutex);
1147
1148 pthread_mutex_unlock(&g->start_work_mutex);
1149 }
1150
1151 gettimeofday(&start0, NULL);
1152
1153 start = stop = start0;
1154 last_perturbance = start.tv_sec;
1155
1156 for (l = 0; l < g->p.nr_loops; l++) {
1157 start = stop;
1158
1159 if (g->stop_work)
1160 break;
1161
1162 val += do_work(global_data, g->p.bytes_global, process_nr, g->p.nr_proc, l, val);
1163 val += do_work(process_data, g->p.bytes_process, thread_nr, g->p.nr_threads, l, val);
1164 val += do_work(thread_data, g->p.bytes_thread, 0, 1, l, val);
1165
1166 if (g->p.sleep_usecs) {
1167 pthread_mutex_lock(td->process_lock);
1168 usleep(g->p.sleep_usecs);
1169 pthread_mutex_unlock(td->process_lock);
1170 }
1171 /*
1172 * Amount of work to be done under a process-global lock:
1173 */
1174 if (g->p.bytes_process_locked) {
1175 pthread_mutex_lock(td->process_lock);
1176 val += do_work(process_data, g->p.bytes_process_locked, thread_nr, g->p.nr_threads, l, val);
1177 pthread_mutex_unlock(td->process_lock);
1178 }
1179
1180 work_done = g->p.bytes_global + g->p.bytes_process +
1181 g->p.bytes_process_locked + g->p.bytes_thread;
1182
1183 update_curr_cpu(task_nr, work_done);
1184 bytes_done += work_done;
1185
1186 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1187 continue;
1188
1189 td->loops_done = l;
1190
1191 gettimeofday(&stop, NULL);
1192
1193 /* Check whether our max runtime timed out: */
1194 if (g->p.nr_secs) {
1195 timersub(&stop, &start0, &diff);
1196 if ((u32)diff.tv_sec >= g->p.nr_secs) {
1197 g->stop_work = true;
1198 break;
1199 }
1200 }
1201
1202 /* Update the summary at most once per second: */
1203 if (start.tv_sec == stop.tv_sec)
1204 continue;
1205
1206 /*
1207 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1208 * by migrating to CPU#0:
1209 */
1210 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1211 cpu_set_t orig_mask;
1212 int target_cpu;
1213 int this_cpu;
1214
1215 last_perturbance = stop.tv_sec;
1216
1217 /*
1218 * Depending on where we are running, move into
1219 * the other half of the system, to create some
1220 * real disturbance:
1221 */
1222 this_cpu = g->threads[task_nr].curr_cpu;
1223 if (this_cpu < g->p.nr_cpus/2)
1224 target_cpu = g->p.nr_cpus-1;
1225 else
1226 target_cpu = 0;
1227
1228 orig_mask = bind_to_cpu(target_cpu);
1229
1230 /* Here we are running on the target CPU already */
1231 if (details >= 1)
1232 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1233
1234 bind_to_cpumask(orig_mask);
1235 }
1236
1237 if (details >= 3) {
1238 timersub(&stop, &start, &diff);
1239 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1240 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1241
1242 if (details >= 0) {
1243 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1244 process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1245 }
1246 fflush(stdout);
1247 }
1248 if (!last_task)
1249 continue;
1250
1251 timersub(&stop, &start0, &diff);
1252 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1253 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1254
1255 show_summary(runtime_ns_max, l, &convergence);
1256 }
1257
1258 gettimeofday(&stop, NULL);
1259 timersub(&stop, &start0, &diff);
1260 td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1261 td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1262 secs = td->runtime_ns / NSEC_PER_SEC;
1263 td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1264
1265 getrusage(RUSAGE_THREAD, &rusage);
1266 td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1267 td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1268 td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1269 td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1270
1271 free_data(thread_data, g->p.bytes_thread);
1272
1273 pthread_mutex_lock(&g->stop_work_mutex);
1274 g->bytes_done += bytes_done;
1275 pthread_mutex_unlock(&g->stop_work_mutex);
1276
1277 return NULL;
1278}
1279
1280/*
1281 * A worker process starts a couple of threads:
1282 */
1283static void worker_process(int process_nr)
1284{
1285 pthread_mutex_t process_lock;
1286 struct thread_data *td;
1287 pthread_t *pthreads;
1288 u8 *process_data;
1289 int task_nr;
1290 int ret;
1291 int t;
1292
1293 pthread_mutex_init(&process_lock, NULL);
1294 set_taskname("process %d", process_nr);
1295
1296 /*
1297 * Pick up the memory policy and the CPU binding of our first thread,
1298 * so that we initialize memory accordingly:
1299 */
1300 task_nr = process_nr*g->p.nr_threads;
1301 td = g->threads + task_nr;
1302
1303 bind_to_memnode(td->bind_node);
1304 bind_to_cpumask(td->bind_cpumask);
1305
1306 pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1307 process_data = setup_private_data(g->p.bytes_process);
1308
1309 if (g->p.show_details >= 3) {
1310 printf(" # process %2d global mem: %p, process mem: %p\n",
1311 process_nr, g->data, process_data);
1312 }
1313
1314 for (t = 0; t < g->p.nr_threads; t++) {
1315 task_nr = process_nr*g->p.nr_threads + t;
1316 td = g->threads + task_nr;
1317
1318 td->process_data = process_data;
1319 td->process_nr = process_nr;
1320 td->thread_nr = t;
1321 td->task_nr = task_nr;
1322 td->val = rand();
1323 td->curr_cpu = -1;
1324 td->process_lock = &process_lock;
1325
1326 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1327 BUG_ON(ret);
1328 }
1329
1330 for (t = 0; t < g->p.nr_threads; t++) {
1331 ret = pthread_join(pthreads[t], NULL);
1332 BUG_ON(ret);
1333 }
1334
1335 free_data(process_data, g->p.bytes_process);
1336 free(pthreads);
1337}
1338
1339static void print_summary(void)
1340{
1341 if (g->p.show_details < 0)
1342 return;
1343
1344 printf("\n ###\n");
1345 printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1346 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1347 printf(" # %5dx %5ldMB global shared mem operations\n",
1348 g->p.nr_loops, g->p.bytes_global/1024/1024);
1349 printf(" # %5dx %5ldMB process shared mem operations\n",
1350 g->p.nr_loops, g->p.bytes_process/1024/1024);
1351 printf(" # %5dx %5ldMB thread local mem operations\n",
1352 g->p.nr_loops, g->p.bytes_thread/1024/1024);
1353
1354 printf(" ###\n");
1355
1356 printf("\n ###\n"); fflush(stdout);
1357}
1358
1359static void init_thread_data(void)
1360{
1361 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1362 int t;
1363
1364 g->threads = zalloc_shared_data(size);
1365
1366 for (t = 0; t < g->p.nr_tasks; t++) {
1367 struct thread_data *td = g->threads + t;
1368 int cpu;
1369
1370 /* Allow all nodes by default: */
1371 td->bind_node = NUMA_NO_NODE;
1372
1373 /* Allow all CPUs by default: */
1374 CPU_ZERO(&td->bind_cpumask);
1375 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1376 CPU_SET(cpu, &td->bind_cpumask);
1377 }
1378}
1379
1380static void deinit_thread_data(void)
1381{
1382 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1383
1384 free_data(g->threads, size);
1385}
1386
1387static int init(void)
1388{
1389 g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1390
1391 /* Copy over options: */
1392 g->p = p0;
1393
1394 g->p.nr_cpus = numa_num_configured_cpus();
1395
1396 g->p.nr_nodes = numa_max_node() + 1;
1397
1398 /* char array in count_process_nodes(): */
1399 BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1400
1401 if (g->p.show_quiet && !g->p.show_details)
1402 g->p.show_details = -1;
1403
1404 /* Some memory should be specified: */
1405 if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1406 return -1;
1407
1408 if (g->p.mb_global_str) {
1409 g->p.mb_global = atof(g->p.mb_global_str);
1410 BUG_ON(g->p.mb_global < 0);
1411 }
1412
1413 if (g->p.mb_proc_str) {
1414 g->p.mb_proc = atof(g->p.mb_proc_str);
1415 BUG_ON(g->p.mb_proc < 0);
1416 }
1417
1418 if (g->p.mb_proc_locked_str) {
1419 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1420 BUG_ON(g->p.mb_proc_locked < 0);
1421 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1422 }
1423
1424 if (g->p.mb_thread_str) {
1425 g->p.mb_thread = atof(g->p.mb_thread_str);
1426 BUG_ON(g->p.mb_thread < 0);
1427 }
1428
1429 BUG_ON(g->p.nr_threads <= 0);
1430 BUG_ON(g->p.nr_proc <= 0);
1431
1432 g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1433
1434 g->p.bytes_global = g->p.mb_global *1024L*1024L;
1435 g->p.bytes_process = g->p.mb_proc *1024L*1024L;
1436 g->p.bytes_process_locked = g->p.mb_proc_locked *1024L*1024L;
1437 g->p.bytes_thread = g->p.mb_thread *1024L*1024L;
1438
1439 g->data = setup_shared_data(g->p.bytes_global);
1440
1441 /* Startup serialization: */
1442 init_global_mutex(&g->start_work_mutex);
1443 init_global_mutex(&g->startup_mutex);
1444 init_global_mutex(&g->startup_done_mutex);
1445 init_global_mutex(&g->stop_work_mutex);
1446
1447 init_thread_data();
1448
1449 tprintf("#\n");
1450 if (parse_setup_cpu_list() || parse_setup_node_list())
1451 return -1;
1452 tprintf("#\n");
1453
1454 print_summary();
1455
1456 return 0;
1457}
1458
1459static void deinit(void)
1460{
1461 free_data(g->data, g->p.bytes_global);
1462 g->data = NULL;
1463
1464 deinit_thread_data();
1465
1466 free_data(g, sizeof(*g));
1467 g = NULL;
1468}
1469
1470/*
1471 * Print a short or long result, depending on the verbosity setting:
1472 */
1473static void print_res(const char *name, double val,
1474 const char *txt_unit, const char *txt_short, const char *txt_long)
1475{
1476 if (!name)
1477 name = "main,";
1478
1479 if (!g->p.show_quiet)
1480 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1481 else
1482 printf(" %14.3f %s\n", val, txt_long);
1483}
1484
1485static int __bench_numa(const char *name)
1486{
1487 struct timeval start, stop, diff;
1488 u64 runtime_ns_min, runtime_ns_sum;
1489 pid_t *pids, pid, wpid;
1490 double delta_runtime;
1491 double runtime_avg;
1492 double runtime_sec_max;
1493 double runtime_sec_min;
1494 int wait_stat;
1495 double bytes;
1496 int i, t, p;
1497
1498 if (init())
1499 return -1;
1500
1501 pids = zalloc(g->p.nr_proc * sizeof(*pids));
1502 pid = -1;
1503
1504 /* All threads try to acquire it, this way we can wait for them to start up: */
1505 pthread_mutex_lock(&g->start_work_mutex);
1506
1507 if (g->p.serialize_startup) {
1508 tprintf(" #\n");
1509 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1510 }
1511
1512 gettimeofday(&start, NULL);
1513
1514 for (i = 0; i < g->p.nr_proc; i++) {
1515 pid = fork();
1516 dprintf(" # process %2d: PID %d\n", i, pid);
1517
1518 BUG_ON(pid < 0);
1519 if (!pid) {
1520 /* Child process: */
1521 worker_process(i);
1522
1523 exit(0);
1524 }
1525 pids[i] = pid;
1526
1527 }
1528 /* Wait for all the threads to start up: */
1529 while (g->nr_tasks_started != g->p.nr_tasks)
1530 usleep(USEC_PER_MSEC);
1531
1532 BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1533
1534 if (g->p.serialize_startup) {
1535 double startup_sec;
1536
1537 pthread_mutex_lock(&g->startup_done_mutex);
1538
1539 /* This will start all threads: */
1540 pthread_mutex_unlock(&g->start_work_mutex);
1541
1542 /* This mutex is locked - the last started thread will wake us: */
1543 pthread_mutex_lock(&g->startup_done_mutex);
1544
1545 gettimeofday(&stop, NULL);
1546
1547 timersub(&stop, &start, &diff);
1548
1549 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1550 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1551 startup_sec /= NSEC_PER_SEC;
1552
1553 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1554 tprintf(" #\n");
1555
1556 start = stop;
1557 pthread_mutex_unlock(&g->startup_done_mutex);
1558 } else {
1559 gettimeofday(&start, NULL);
1560 }
1561
1562 /* Parent process: */
1563
1564
1565 for (i = 0; i < g->p.nr_proc; i++) {
1566 wpid = waitpid(pids[i], &wait_stat, 0);
1567 BUG_ON(wpid < 0);
1568 BUG_ON(!WIFEXITED(wait_stat));
1569
1570 }
1571
1572 runtime_ns_sum = 0;
1573 runtime_ns_min = -1LL;
1574
1575 for (t = 0; t < g->p.nr_tasks; t++) {
1576 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1577
1578 runtime_ns_sum += thread_runtime_ns;
1579 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1580 }
1581
1582 gettimeofday(&stop, NULL);
1583 timersub(&stop, &start, &diff);
1584
1585 BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1586
1587 tprintf("\n ###\n");
1588 tprintf("\n");
1589
1590 runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1591 runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1592 runtime_sec_max /= NSEC_PER_SEC;
1593
1594 runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1595
1596 bytes = g->bytes_done;
1597 runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1598
1599 if (g->p.measure_convergence) {
1600 print_res(name, runtime_sec_max,
1601 "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1602 }
1603
1604 print_res(name, runtime_sec_max,
1605 "secs,", "runtime-max/thread", "secs slowest (max) thread-runtime");
1606
1607 print_res(name, runtime_sec_min,
1608 "secs,", "runtime-min/thread", "secs fastest (min) thread-runtime");
1609
1610 print_res(name, runtime_avg,
1611 "secs,", "runtime-avg/thread", "secs average thread-runtime");
1612
1613 delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1614 print_res(name, delta_runtime / runtime_sec_max * 100.0,
1615 "%,", "spread-runtime/thread", "% difference between max/avg runtime");
1616
1617 print_res(name, bytes / g->p.nr_tasks / 1e9,
1618 "GB,", "data/thread", "GB data processed, per thread");
1619
1620 print_res(name, bytes / 1e9,
1621 "GB,", "data-total", "GB data processed, total");
1622
1623 print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1624 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1625
1626 print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1627 "GB/sec,", "thread-speed", "GB/sec/thread speed");
1628
1629 print_res(name, bytes / runtime_sec_max / 1e9,
1630 "GB/sec,", "total-speed", "GB/sec total speed");
1631
1632 if (g->p.show_details >= 2) {
1633 char tname[14 + 2 * 10 + 1];
1634 struct thread_data *td;
1635 for (p = 0; p < g->p.nr_proc; p++) {
1636 for (t = 0; t < g->p.nr_threads; t++) {
1637 memset(tname, 0, sizeof(tname));
1638 td = g->threads + p*g->p.nr_threads + t;
1639 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1640 print_res(tname, td->speed_gbs,
1641 "GB/sec", "thread-speed", "GB/sec/thread speed");
1642 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1643 "secs", "thread-system-time", "system CPU time/thread");
1644 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1645 "secs", "thread-user-time", "user CPU time/thread");
1646 }
1647 }
1648 }
1649
1650 free(pids);
1651
1652 deinit();
1653
1654 return 0;
1655}
1656
1657#define MAX_ARGS 50
1658
1659static int command_size(const char **argv)
1660{
1661 int size = 0;
1662
1663 while (*argv) {
1664 size++;
1665 argv++;
1666 }
1667
1668 BUG_ON(size >= MAX_ARGS);
1669
1670 return size;
1671}
1672
1673static void init_params(struct params *p, const char *name, int argc, const char **argv)
1674{
1675 int i;
1676
1677 printf("\n # Running %s \"perf bench numa", name);
1678
1679 for (i = 0; i < argc; i++)
1680 printf(" %s", argv[i]);
1681
1682 printf("\"\n");
1683
1684 memset(p, 0, sizeof(*p));
1685
1686 /* Initialize nonzero defaults: */
1687
1688 p->serialize_startup = 1;
1689 p->data_reads = true;
1690 p->data_writes = true;
1691 p->data_backwards = true;
1692 p->data_rand_walk = true;
1693 p->nr_loops = -1;
1694 p->init_random = true;
1695 p->mb_global_str = "1";
1696 p->nr_proc = 1;
1697 p->nr_threads = 1;
1698 p->nr_secs = 5;
1699 p->run_all = argc == 1;
1700}
1701
1702static int run_bench_numa(const char *name, const char **argv)
1703{
1704 int argc = command_size(argv);
1705
1706 init_params(&p0, name, argc, argv);
1707 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1708 if (argc)
1709 goto err;
1710
1711 if (__bench_numa(name))
1712 goto err;
1713
1714 return 0;
1715
1716err:
1717 return -1;
1718}
1719
1720#define OPT_BW_RAM "-s", "20", "-zZq", "--thp", " 1", "--no-data_rand_walk"
1721#define OPT_BW_RAM_NOTHP OPT_BW_RAM, "--thp", "-1"
1722
1723#define OPT_CONV "-s", "100", "-zZ0qcm", "--thp", " 1"
1724#define OPT_CONV_NOTHP OPT_CONV, "--thp", "-1"
1725
1726#define OPT_BW "-s", "20", "-zZ0q", "--thp", " 1"
1727#define OPT_BW_NOTHP OPT_BW, "--thp", "-1"
1728
1729/*
1730 * The built-in test-suite executed by "perf bench numa -a".
1731 *
1732 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1733 */
1734static const char *tests[][MAX_ARGS] = {
1735 /* Basic single-stream NUMA bandwidth measurements: */
1736 { "RAM-bw-local,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1737 "-C" , "0", "-M", "0", OPT_BW_RAM },
1738 { "RAM-bw-local-NOTHP,",
1739 "mem", "-p", "1", "-t", "1", "-P", "1024",
1740 "-C" , "0", "-M", "0", OPT_BW_RAM_NOTHP },
1741 { "RAM-bw-remote,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1742 "-C" , "0", "-M", "1", OPT_BW_RAM },
1743
1744 /* 2-stream NUMA bandwidth measurements: */
1745 { "RAM-bw-local-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1746 "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1747 { "RAM-bw-remote-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1748 "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1749
1750 /* Cross-stream NUMA bandwidth measurement: */
1751 { "RAM-bw-cross,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1752 "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1753
1754 /* Convergence latency measurements: */
1755 { " 1x3-convergence,", "mem", "-p", "1", "-t", "3", "-P", "512", OPT_CONV },
1756 { " 1x4-convergence,", "mem", "-p", "1", "-t", "4", "-P", "512", OPT_CONV },
1757 { " 1x6-convergence,", "mem", "-p", "1", "-t", "6", "-P", "1020", OPT_CONV },
1758 { " 2x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1759 { " 3x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1760 { " 4x4-convergence,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV },
1761 { " 4x4-convergence-NOTHP,",
1762 "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1763 { " 4x6-convergence,", "mem", "-p", "4", "-t", "6", "-P", "1020", OPT_CONV },
1764 { " 4x8-convergence,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_CONV },
1765 { " 8x4-convergence,", "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV },
1766 { " 8x4-convergence-NOTHP,",
1767 "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1768 { " 3x1-convergence,", "mem", "-p", "3", "-t", "1", "-P", "512", OPT_CONV },
1769 { " 4x1-convergence,", "mem", "-p", "4", "-t", "1", "-P", "512", OPT_CONV },
1770 { " 8x1-convergence,", "mem", "-p", "8", "-t", "1", "-P", "512", OPT_CONV },
1771 { "16x1-convergence,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_CONV },
1772 { "32x1-convergence,", "mem", "-p", "32", "-t", "1", "-P", "128", OPT_CONV },
1773
1774 /* Various NUMA process/thread layout bandwidth measurements: */
1775 { " 2x1-bw-process,", "mem", "-p", "2", "-t", "1", "-P", "1024", OPT_BW },
1776 { " 3x1-bw-process,", "mem", "-p", "3", "-t", "1", "-P", "1024", OPT_BW },
1777 { " 4x1-bw-process,", "mem", "-p", "4", "-t", "1", "-P", "1024", OPT_BW },
1778 { " 8x1-bw-process,", "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW },
1779 { " 8x1-bw-process-NOTHP,",
1780 "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW_NOTHP },
1781 { "16x1-bw-process,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_BW },
1782
1783 { " 4x1-bw-thread,", "mem", "-p", "1", "-t", "4", "-T", "256", OPT_BW },
1784 { " 8x1-bw-thread,", "mem", "-p", "1", "-t", "8", "-T", "256", OPT_BW },
1785 { "16x1-bw-thread,", "mem", "-p", "1", "-t", "16", "-T", "128", OPT_BW },
1786 { "32x1-bw-thread,", "mem", "-p", "1", "-t", "32", "-T", "64", OPT_BW },
1787
1788 { " 2x3-bw-thread,", "mem", "-p", "2", "-t", "3", "-P", "512", OPT_BW },
1789 { " 4x4-bw-thread,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_BW },
1790 { " 4x6-bw-thread,", "mem", "-p", "4", "-t", "6", "-P", "512", OPT_BW },
1791 { " 4x8-bw-thread,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW },
1792 { " 4x8-bw-thread-NOTHP,",
1793 "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW_NOTHP },
1794 { " 3x3-bw-thread,", "mem", "-p", "3", "-t", "3", "-P", "512", OPT_BW },
1795 { " 5x5-bw-thread,", "mem", "-p", "5", "-t", "5", "-P", "512", OPT_BW },
1796
1797 { "2x16-bw-thread,", "mem", "-p", "2", "-t", "16", "-P", "512", OPT_BW },
1798 { "1x32-bw-thread,", "mem", "-p", "1", "-t", "32", "-P", "2048", OPT_BW },
1799
1800 { "numa02-bw,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW },
1801 { "numa02-bw-NOTHP,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW_NOTHP },
1802 { "numa01-bw-thread,", "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW },
1803 { "numa01-bw-thread-NOTHP,",
1804 "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW_NOTHP },
1805};
1806
1807static int bench_all(void)
1808{
1809 int nr = ARRAY_SIZE(tests);
1810 int ret;
1811 int i;
1812
1813 ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1814 BUG_ON(ret < 0);
1815
1816 for (i = 0; i < nr; i++) {
1817 run_bench_numa(tests[i][0], tests[i] + 1);
1818 }
1819
1820 printf("\n");
1821
1822 return 0;
1823}
1824
1825int bench_numa(int argc, const char **argv)
1826{
1827 init_params(&p0, "main,", argc, argv);
1828 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1829 if (argc)
1830 goto err;
1831
1832 if (p0.run_all)
1833 return bench_all();
1834
1835 if (__bench_numa(NULL))
1836 goto err;
1837
1838 return 0;
1839
1840err:
1841 usage_with_options(numa_usage, options);
1842 return -1;
1843}
1/*
2 * numa.c
3 *
4 * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
5 */
6
7#include "../perf.h"
8#include "../builtin.h"
9#include "../util/util.h"
10#include <subcmd/parse-options.h>
11#include "../util/cloexec.h"
12
13#include "bench.h"
14
15#include <errno.h>
16#include <sched.h>
17#include <stdio.h>
18#include <assert.h>
19#include <malloc.h>
20#include <signal.h>
21#include <stdlib.h>
22#include <string.h>
23#include <unistd.h>
24#include <pthread.h>
25#include <sys/mman.h>
26#include <sys/time.h>
27#include <sys/resource.h>
28#include <sys/wait.h>
29#include <sys/prctl.h>
30#include <sys/types.h>
31
32#include <numa.h>
33#include <numaif.h>
34
35/*
36 * Regular printout to the terminal, supressed if -q is specified:
37 */
38#define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
39
40/*
41 * Debug printf:
42 */
43#define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
44
45struct thread_data {
46 int curr_cpu;
47 cpu_set_t bind_cpumask;
48 int bind_node;
49 u8 *process_data;
50 int process_nr;
51 int thread_nr;
52 int task_nr;
53 unsigned int loops_done;
54 u64 val;
55 u64 runtime_ns;
56 u64 system_time_ns;
57 u64 user_time_ns;
58 double speed_gbs;
59 pthread_mutex_t *process_lock;
60};
61
62/* Parameters set by options: */
63
64struct params {
65 /* Startup synchronization: */
66 bool serialize_startup;
67
68 /* Task hierarchy: */
69 int nr_proc;
70 int nr_threads;
71
72 /* Working set sizes: */
73 const char *mb_global_str;
74 const char *mb_proc_str;
75 const char *mb_proc_locked_str;
76 const char *mb_thread_str;
77
78 double mb_global;
79 double mb_proc;
80 double mb_proc_locked;
81 double mb_thread;
82
83 /* Access patterns to the working set: */
84 bool data_reads;
85 bool data_writes;
86 bool data_backwards;
87 bool data_zero_memset;
88 bool data_rand_walk;
89 u32 nr_loops;
90 u32 nr_secs;
91 u32 sleep_usecs;
92
93 /* Working set initialization: */
94 bool init_zero;
95 bool init_random;
96 bool init_cpu0;
97
98 /* Misc options: */
99 int show_details;
100 int run_all;
101 int thp;
102
103 long bytes_global;
104 long bytes_process;
105 long bytes_process_locked;
106 long bytes_thread;
107
108 int nr_tasks;
109 bool show_quiet;
110
111 bool show_convergence;
112 bool measure_convergence;
113
114 int perturb_secs;
115 int nr_cpus;
116 int nr_nodes;
117
118 /* Affinity options -C and -N: */
119 char *cpu_list_str;
120 char *node_list_str;
121};
122
123
124/* Global, read-writable area, accessible to all processes and threads: */
125
126struct global_info {
127 u8 *data;
128
129 pthread_mutex_t startup_mutex;
130 int nr_tasks_started;
131
132 pthread_mutex_t startup_done_mutex;
133
134 pthread_mutex_t start_work_mutex;
135 int nr_tasks_working;
136
137 pthread_mutex_t stop_work_mutex;
138 u64 bytes_done;
139
140 struct thread_data *threads;
141
142 /* Convergence latency measurement: */
143 bool all_converged;
144 bool stop_work;
145
146 int print_once;
147
148 struct params p;
149};
150
151static struct global_info *g = NULL;
152
153static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
154static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
155
156struct params p0;
157
158static const struct option options[] = {
159 OPT_INTEGER('p', "nr_proc" , &p0.nr_proc, "number of processes"),
160 OPT_INTEGER('t', "nr_threads" , &p0.nr_threads, "number of threads per process"),
161
162 OPT_STRING('G', "mb_global" , &p0.mb_global_str, "MB", "global memory (MBs)"),
163 OPT_STRING('P', "mb_proc" , &p0.mb_proc_str, "MB", "process memory (MBs)"),
164 OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
165 OPT_STRING('T', "mb_thread" , &p0.mb_thread_str, "MB", "thread memory (MBs)"),
166
167 OPT_UINTEGER('l', "nr_loops" , &p0.nr_loops, "max number of loops to run (default: unlimited)"),
168 OPT_UINTEGER('s', "nr_secs" , &p0.nr_secs, "max number of seconds to run (default: 5 secs)"),
169 OPT_UINTEGER('u', "usleep" , &p0.sleep_usecs, "usecs to sleep per loop iteration"),
170
171 OPT_BOOLEAN('R', "data_reads" , &p0.data_reads, "access the data via writes (can be mixed with -W)"),
172 OPT_BOOLEAN('W', "data_writes" , &p0.data_writes, "access the data via writes (can be mixed with -R)"),
173 OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards, "access the data backwards as well"),
174 OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
175 OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk, "access the data with random (32bit LFSR) walk"),
176
177
178 OPT_BOOLEAN('z', "init_zero" , &p0.init_zero, "bzero the initial allocations"),
179 OPT_BOOLEAN('I', "init_random" , &p0.init_random, "randomize the contents of the initial allocations"),
180 OPT_BOOLEAN('0', "init_cpu0" , &p0.init_cpu0, "do the initial allocations on CPU#0"),
181 OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs, "perturb thread 0/0 every X secs, to test convergence stability"),
182
183 OPT_INCR ('d', "show_details" , &p0.show_details, "Show details"),
184 OPT_INCR ('a', "all" , &p0.run_all, "Run all tests in the suite"),
185 OPT_INTEGER('H', "thp" , &p0.thp, "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
186 OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details"),
187 OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
188 OPT_BOOLEAN('q', "quiet" , &p0.show_quiet, "quiet mode"),
189 OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
190
191 /* Special option string parsing callbacks: */
192 OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
193 "bind the first N tasks to these specific cpus (the rest is unbound)",
194 parse_cpus_opt),
195 OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
196 "bind the first N tasks to these specific memory nodes (the rest is unbound)",
197 parse_nodes_opt),
198 OPT_END()
199};
200
201static const char * const bench_numa_usage[] = {
202 "perf bench numa <options>",
203 NULL
204};
205
206static const char * const numa_usage[] = {
207 "perf bench numa mem [<options>]",
208 NULL
209};
210
211static cpu_set_t bind_to_cpu(int target_cpu)
212{
213 cpu_set_t orig_mask, mask;
214 int ret;
215
216 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
217 BUG_ON(ret);
218
219 CPU_ZERO(&mask);
220
221 if (target_cpu == -1) {
222 int cpu;
223
224 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
225 CPU_SET(cpu, &mask);
226 } else {
227 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
228 CPU_SET(target_cpu, &mask);
229 }
230
231 ret = sched_setaffinity(0, sizeof(mask), &mask);
232 BUG_ON(ret);
233
234 return orig_mask;
235}
236
237static cpu_set_t bind_to_node(int target_node)
238{
239 int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
240 cpu_set_t orig_mask, mask;
241 int cpu;
242 int ret;
243
244 BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
245 BUG_ON(!cpus_per_node);
246
247 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
248 BUG_ON(ret);
249
250 CPU_ZERO(&mask);
251
252 if (target_node == -1) {
253 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
254 CPU_SET(cpu, &mask);
255 } else {
256 int cpu_start = (target_node + 0) * cpus_per_node;
257 int cpu_stop = (target_node + 1) * cpus_per_node;
258
259 BUG_ON(cpu_stop > g->p.nr_cpus);
260
261 for (cpu = cpu_start; cpu < cpu_stop; cpu++)
262 CPU_SET(cpu, &mask);
263 }
264
265 ret = sched_setaffinity(0, sizeof(mask), &mask);
266 BUG_ON(ret);
267
268 return orig_mask;
269}
270
271static void bind_to_cpumask(cpu_set_t mask)
272{
273 int ret;
274
275 ret = sched_setaffinity(0, sizeof(mask), &mask);
276 BUG_ON(ret);
277}
278
279static void mempol_restore(void)
280{
281 int ret;
282
283 ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
284
285 BUG_ON(ret);
286}
287
288static void bind_to_memnode(int node)
289{
290 unsigned long nodemask;
291 int ret;
292
293 if (node == -1)
294 return;
295
296 BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
297 nodemask = 1L << node;
298
299 ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
300 dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
301
302 BUG_ON(ret);
303}
304
305#define HPSIZE (2*1024*1024)
306
307#define set_taskname(fmt...) \
308do { \
309 char name[20]; \
310 \
311 snprintf(name, 20, fmt); \
312 prctl(PR_SET_NAME, name); \
313} while (0)
314
315static u8 *alloc_data(ssize_t bytes0, int map_flags,
316 int init_zero, int init_cpu0, int thp, int init_random)
317{
318 cpu_set_t orig_mask;
319 ssize_t bytes;
320 u8 *buf;
321 int ret;
322
323 if (!bytes0)
324 return NULL;
325
326 /* Allocate and initialize all memory on CPU#0: */
327 if (init_cpu0) {
328 orig_mask = bind_to_node(0);
329 bind_to_memnode(0);
330 }
331
332 bytes = bytes0 + HPSIZE;
333
334 buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
335 BUG_ON(buf == (void *)-1);
336
337 if (map_flags == MAP_PRIVATE) {
338 if (thp > 0) {
339 ret = madvise(buf, bytes, MADV_HUGEPAGE);
340 if (ret && !g->print_once) {
341 g->print_once = 1;
342 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
343 }
344 }
345 if (thp < 0) {
346 ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
347 if (ret && !g->print_once) {
348 g->print_once = 1;
349 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
350 }
351 }
352 }
353
354 if (init_zero) {
355 bzero(buf, bytes);
356 } else {
357 /* Initialize random contents, different in each word: */
358 if (init_random) {
359 u64 *wbuf = (void *)buf;
360 long off = rand();
361 long i;
362
363 for (i = 0; i < bytes/8; i++)
364 wbuf[i] = i + off;
365 }
366 }
367
368 /* Align to 2MB boundary: */
369 buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
370
371 /* Restore affinity: */
372 if (init_cpu0) {
373 bind_to_cpumask(orig_mask);
374 mempol_restore();
375 }
376
377 return buf;
378}
379
380static void free_data(void *data, ssize_t bytes)
381{
382 int ret;
383
384 if (!data)
385 return;
386
387 ret = munmap(data, bytes);
388 BUG_ON(ret);
389}
390
391/*
392 * Create a shared memory buffer that can be shared between processes, zeroed:
393 */
394static void * zalloc_shared_data(ssize_t bytes)
395{
396 return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0, g->p.thp, g->p.init_random);
397}
398
399/*
400 * Create a shared memory buffer that can be shared between processes:
401 */
402static void * setup_shared_data(ssize_t bytes)
403{
404 return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
405}
406
407/*
408 * Allocate process-local memory - this will either be shared between
409 * threads of this process, or only be accessed by this thread:
410 */
411static void * setup_private_data(ssize_t bytes)
412{
413 return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
414}
415
416/*
417 * Return a process-shared (global) mutex:
418 */
419static void init_global_mutex(pthread_mutex_t *mutex)
420{
421 pthread_mutexattr_t attr;
422
423 pthread_mutexattr_init(&attr);
424 pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
425 pthread_mutex_init(mutex, &attr);
426}
427
428static int parse_cpu_list(const char *arg)
429{
430 p0.cpu_list_str = strdup(arg);
431
432 dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
433
434 return 0;
435}
436
437static int parse_setup_cpu_list(void)
438{
439 struct thread_data *td;
440 char *str0, *str;
441 int t;
442
443 if (!g->p.cpu_list_str)
444 return 0;
445
446 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
447
448 str0 = str = strdup(g->p.cpu_list_str);
449 t = 0;
450
451 BUG_ON(!str);
452
453 tprintf("# binding tasks to CPUs:\n");
454 tprintf("# ");
455
456 while (true) {
457 int bind_cpu, bind_cpu_0, bind_cpu_1;
458 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
459 int bind_len;
460 int step;
461 int mul;
462
463 tok = strsep(&str, ",");
464 if (!tok)
465 break;
466
467 tok_end = strstr(tok, "-");
468
469 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
470 if (!tok_end) {
471 /* Single CPU specified: */
472 bind_cpu_0 = bind_cpu_1 = atol(tok);
473 } else {
474 /* CPU range specified (for example: "5-11"): */
475 bind_cpu_0 = atol(tok);
476 bind_cpu_1 = atol(tok_end + 1);
477 }
478
479 step = 1;
480 tok_step = strstr(tok, "#");
481 if (tok_step) {
482 step = atol(tok_step + 1);
483 BUG_ON(step <= 0 || step >= g->p.nr_cpus);
484 }
485
486 /*
487 * Mask length.
488 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
489 * where the _4 means the next 4 CPUs are allowed.
490 */
491 bind_len = 1;
492 tok_len = strstr(tok, "_");
493 if (tok_len) {
494 bind_len = atol(tok_len + 1);
495 BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
496 }
497
498 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
499 mul = 1;
500 tok_mul = strstr(tok, "x");
501 if (tok_mul) {
502 mul = atol(tok_mul + 1);
503 BUG_ON(mul <= 0);
504 }
505
506 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
507
508 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
509 printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
510 return -1;
511 }
512
513 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
514 BUG_ON(bind_cpu_0 > bind_cpu_1);
515
516 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
517 int i;
518
519 for (i = 0; i < mul; i++) {
520 int cpu;
521
522 if (t >= g->p.nr_tasks) {
523 printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
524 goto out;
525 }
526 td = g->threads + t;
527
528 if (t)
529 tprintf(",");
530 if (bind_len > 1) {
531 tprintf("%2d/%d", bind_cpu, bind_len);
532 } else {
533 tprintf("%2d", bind_cpu);
534 }
535
536 CPU_ZERO(&td->bind_cpumask);
537 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
538 BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
539 CPU_SET(cpu, &td->bind_cpumask);
540 }
541 t++;
542 }
543 }
544 }
545out:
546
547 tprintf("\n");
548
549 if (t < g->p.nr_tasks)
550 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
551
552 free(str0);
553 return 0;
554}
555
556static int parse_cpus_opt(const struct option *opt __maybe_unused,
557 const char *arg, int unset __maybe_unused)
558{
559 if (!arg)
560 return -1;
561
562 return parse_cpu_list(arg);
563}
564
565static int parse_node_list(const char *arg)
566{
567 p0.node_list_str = strdup(arg);
568
569 dprintf("got NODE list: {%s}\n", p0.node_list_str);
570
571 return 0;
572}
573
574static int parse_setup_node_list(void)
575{
576 struct thread_data *td;
577 char *str0, *str;
578 int t;
579
580 if (!g->p.node_list_str)
581 return 0;
582
583 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
584
585 str0 = str = strdup(g->p.node_list_str);
586 t = 0;
587
588 BUG_ON(!str);
589
590 tprintf("# binding tasks to NODEs:\n");
591 tprintf("# ");
592
593 while (true) {
594 int bind_node, bind_node_0, bind_node_1;
595 char *tok, *tok_end, *tok_step, *tok_mul;
596 int step;
597 int mul;
598
599 tok = strsep(&str, ",");
600 if (!tok)
601 break;
602
603 tok_end = strstr(tok, "-");
604
605 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
606 if (!tok_end) {
607 /* Single NODE specified: */
608 bind_node_0 = bind_node_1 = atol(tok);
609 } else {
610 /* NODE range specified (for example: "5-11"): */
611 bind_node_0 = atol(tok);
612 bind_node_1 = atol(tok_end + 1);
613 }
614
615 step = 1;
616 tok_step = strstr(tok, "#");
617 if (tok_step) {
618 step = atol(tok_step + 1);
619 BUG_ON(step <= 0 || step >= g->p.nr_nodes);
620 }
621
622 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
623 mul = 1;
624 tok_mul = strstr(tok, "x");
625 if (tok_mul) {
626 mul = atol(tok_mul + 1);
627 BUG_ON(mul <= 0);
628 }
629
630 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
631
632 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
633 printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
634 return -1;
635 }
636
637 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
638 BUG_ON(bind_node_0 > bind_node_1);
639
640 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
641 int i;
642
643 for (i = 0; i < mul; i++) {
644 if (t >= g->p.nr_tasks) {
645 printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
646 goto out;
647 }
648 td = g->threads + t;
649
650 if (!t)
651 tprintf(" %2d", bind_node);
652 else
653 tprintf(",%2d", bind_node);
654
655 td->bind_node = bind_node;
656 t++;
657 }
658 }
659 }
660out:
661
662 tprintf("\n");
663
664 if (t < g->p.nr_tasks)
665 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
666
667 free(str0);
668 return 0;
669}
670
671static int parse_nodes_opt(const struct option *opt __maybe_unused,
672 const char *arg, int unset __maybe_unused)
673{
674 if (!arg)
675 return -1;
676
677 return parse_node_list(arg);
678
679 return 0;
680}
681
682#define BIT(x) (1ul << x)
683
684static inline uint32_t lfsr_32(uint32_t lfsr)
685{
686 const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
687 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
688}
689
690/*
691 * Make sure there's real data dependency to RAM (when read
692 * accesses are enabled), so the compiler, the CPU and the
693 * kernel (KSM, zero page, etc.) cannot optimize away RAM
694 * accesses:
695 */
696static inline u64 access_data(u64 *data __attribute__((unused)), u64 val)
697{
698 if (g->p.data_reads)
699 val += *data;
700 if (g->p.data_writes)
701 *data = val + 1;
702 return val;
703}
704
705/*
706 * The worker process does two types of work, a forwards going
707 * loop and a backwards going loop.
708 *
709 * We do this so that on multiprocessor systems we do not create
710 * a 'train' of processing, with highly synchronized processes,
711 * skewing the whole benchmark.
712 */
713static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
714{
715 long words = bytes/sizeof(u64);
716 u64 *data = (void *)__data;
717 long chunk_0, chunk_1;
718 u64 *d0, *d, *d1;
719 long off;
720 long i;
721
722 BUG_ON(!data && words);
723 BUG_ON(data && !words);
724
725 if (!data)
726 return val;
727
728 /* Very simple memset() work variant: */
729 if (g->p.data_zero_memset && !g->p.data_rand_walk) {
730 bzero(data, bytes);
731 return val;
732 }
733
734 /* Spread out by PID/TID nr and by loop nr: */
735 chunk_0 = words/nr_max;
736 chunk_1 = words/g->p.nr_loops;
737 off = nr*chunk_0 + loop*chunk_1;
738
739 while (off >= words)
740 off -= words;
741
742 if (g->p.data_rand_walk) {
743 u32 lfsr = nr + loop + val;
744 int j;
745
746 for (i = 0; i < words/1024; i++) {
747 long start, end;
748
749 lfsr = lfsr_32(lfsr);
750
751 start = lfsr % words;
752 end = min(start + 1024, words-1);
753
754 if (g->p.data_zero_memset) {
755 bzero(data + start, (end-start) * sizeof(u64));
756 } else {
757 for (j = start; j < end; j++)
758 val = access_data(data + j, val);
759 }
760 }
761 } else if (!g->p.data_backwards || (nr + loop) & 1) {
762
763 d0 = data + off;
764 d = data + off + 1;
765 d1 = data + words;
766
767 /* Process data forwards: */
768 for (;;) {
769 if (unlikely(d >= d1))
770 d = data;
771 if (unlikely(d == d0))
772 break;
773
774 val = access_data(d, val);
775
776 d++;
777 }
778 } else {
779 /* Process data backwards: */
780
781 d0 = data + off;
782 d = data + off - 1;
783 d1 = data + words;
784
785 /* Process data forwards: */
786 for (;;) {
787 if (unlikely(d < data))
788 d = data + words-1;
789 if (unlikely(d == d0))
790 break;
791
792 val = access_data(d, val);
793
794 d--;
795 }
796 }
797
798 return val;
799}
800
801static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
802{
803 unsigned int cpu;
804
805 cpu = sched_getcpu();
806
807 g->threads[task_nr].curr_cpu = cpu;
808 prctl(0, bytes_worked);
809}
810
811#define MAX_NR_NODES 64
812
813/*
814 * Count the number of nodes a process's threads
815 * are spread out on.
816 *
817 * A count of 1 means that the process is compressed
818 * to a single node. A count of g->p.nr_nodes means it's
819 * spread out on the whole system.
820 */
821static int count_process_nodes(int process_nr)
822{
823 char node_present[MAX_NR_NODES] = { 0, };
824 int nodes;
825 int n, t;
826
827 for (t = 0; t < g->p.nr_threads; t++) {
828 struct thread_data *td;
829 int task_nr;
830 int node;
831
832 task_nr = process_nr*g->p.nr_threads + t;
833 td = g->threads + task_nr;
834
835 node = numa_node_of_cpu(td->curr_cpu);
836 if (node < 0) /* curr_cpu was likely still -1 */
837 return 0;
838
839 node_present[node] = 1;
840 }
841
842 nodes = 0;
843
844 for (n = 0; n < MAX_NR_NODES; n++)
845 nodes += node_present[n];
846
847 return nodes;
848}
849
850/*
851 * Count the number of distinct process-threads a node contains.
852 *
853 * A count of 1 means that the node contains only a single
854 * process. If all nodes on the system contain at most one
855 * process then we are well-converged.
856 */
857static int count_node_processes(int node)
858{
859 int processes = 0;
860 int t, p;
861
862 for (p = 0; p < g->p.nr_proc; p++) {
863 for (t = 0; t < g->p.nr_threads; t++) {
864 struct thread_data *td;
865 int task_nr;
866 int n;
867
868 task_nr = p*g->p.nr_threads + t;
869 td = g->threads + task_nr;
870
871 n = numa_node_of_cpu(td->curr_cpu);
872 if (n == node) {
873 processes++;
874 break;
875 }
876 }
877 }
878
879 return processes;
880}
881
882static void calc_convergence_compression(int *strong)
883{
884 unsigned int nodes_min, nodes_max;
885 int p;
886
887 nodes_min = -1;
888 nodes_max = 0;
889
890 for (p = 0; p < g->p.nr_proc; p++) {
891 unsigned int nodes = count_process_nodes(p);
892
893 if (!nodes) {
894 *strong = 0;
895 return;
896 }
897
898 nodes_min = min(nodes, nodes_min);
899 nodes_max = max(nodes, nodes_max);
900 }
901
902 /* Strong convergence: all threads compress on a single node: */
903 if (nodes_min == 1 && nodes_max == 1) {
904 *strong = 1;
905 } else {
906 *strong = 0;
907 tprintf(" {%d-%d}", nodes_min, nodes_max);
908 }
909}
910
911static void calc_convergence(double runtime_ns_max, double *convergence)
912{
913 unsigned int loops_done_min, loops_done_max;
914 int process_groups;
915 int nodes[MAX_NR_NODES];
916 int distance;
917 int nr_min;
918 int nr_max;
919 int strong;
920 int sum;
921 int nr;
922 int node;
923 int cpu;
924 int t;
925
926 if (!g->p.show_convergence && !g->p.measure_convergence)
927 return;
928
929 for (node = 0; node < g->p.nr_nodes; node++)
930 nodes[node] = 0;
931
932 loops_done_min = -1;
933 loops_done_max = 0;
934
935 for (t = 0; t < g->p.nr_tasks; t++) {
936 struct thread_data *td = g->threads + t;
937 unsigned int loops_done;
938
939 cpu = td->curr_cpu;
940
941 /* Not all threads have written it yet: */
942 if (cpu < 0)
943 continue;
944
945 node = numa_node_of_cpu(cpu);
946
947 nodes[node]++;
948
949 loops_done = td->loops_done;
950 loops_done_min = min(loops_done, loops_done_min);
951 loops_done_max = max(loops_done, loops_done_max);
952 }
953
954 nr_max = 0;
955 nr_min = g->p.nr_tasks;
956 sum = 0;
957
958 for (node = 0; node < g->p.nr_nodes; node++) {
959 nr = nodes[node];
960 nr_min = min(nr, nr_min);
961 nr_max = max(nr, nr_max);
962 sum += nr;
963 }
964 BUG_ON(nr_min > nr_max);
965
966 BUG_ON(sum > g->p.nr_tasks);
967
968 if (0 && (sum < g->p.nr_tasks))
969 return;
970
971 /*
972 * Count the number of distinct process groups present
973 * on nodes - when we are converged this will decrease
974 * to g->p.nr_proc:
975 */
976 process_groups = 0;
977
978 for (node = 0; node < g->p.nr_nodes; node++) {
979 int processes = count_node_processes(node);
980
981 nr = nodes[node];
982 tprintf(" %2d/%-2d", nr, processes);
983
984 process_groups += processes;
985 }
986
987 distance = nr_max - nr_min;
988
989 tprintf(" [%2d/%-2d]", distance, process_groups);
990
991 tprintf(" l:%3d-%-3d (%3d)",
992 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
993
994 if (loops_done_min && loops_done_max) {
995 double skew = 1.0 - (double)loops_done_min/loops_done_max;
996
997 tprintf(" [%4.1f%%]", skew * 100.0);
998 }
999
1000 calc_convergence_compression(&strong);
1001
1002 if (strong && process_groups == g->p.nr_proc) {
1003 if (!*convergence) {
1004 *convergence = runtime_ns_max;
1005 tprintf(" (%6.1fs converged)\n", *convergence/1e9);
1006 if (g->p.measure_convergence) {
1007 g->all_converged = true;
1008 g->stop_work = true;
1009 }
1010 }
1011 } else {
1012 if (*convergence) {
1013 tprintf(" (%6.1fs de-converged)", runtime_ns_max/1e9);
1014 *convergence = 0;
1015 }
1016 tprintf("\n");
1017 }
1018}
1019
1020static void show_summary(double runtime_ns_max, int l, double *convergence)
1021{
1022 tprintf("\r # %5.1f%% [%.1f mins]",
1023 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max/1e9 / 60.0);
1024
1025 calc_convergence(runtime_ns_max, convergence);
1026
1027 if (g->p.show_details >= 0)
1028 fflush(stdout);
1029}
1030
1031static void *worker_thread(void *__tdata)
1032{
1033 struct thread_data *td = __tdata;
1034 struct timeval start0, start, stop, diff;
1035 int process_nr = td->process_nr;
1036 int thread_nr = td->thread_nr;
1037 unsigned long last_perturbance;
1038 int task_nr = td->task_nr;
1039 int details = g->p.show_details;
1040 int first_task, last_task;
1041 double convergence = 0;
1042 u64 val = td->val;
1043 double runtime_ns_max;
1044 u8 *global_data;
1045 u8 *process_data;
1046 u8 *thread_data;
1047 u64 bytes_done;
1048 long work_done;
1049 u32 l;
1050 struct rusage rusage;
1051
1052 bind_to_cpumask(td->bind_cpumask);
1053 bind_to_memnode(td->bind_node);
1054
1055 set_taskname("thread %d/%d", process_nr, thread_nr);
1056
1057 global_data = g->data;
1058 process_data = td->process_data;
1059 thread_data = setup_private_data(g->p.bytes_thread);
1060
1061 bytes_done = 0;
1062
1063 last_task = 0;
1064 if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1065 last_task = 1;
1066
1067 first_task = 0;
1068 if (process_nr == 0 && thread_nr == 0)
1069 first_task = 1;
1070
1071 if (details >= 2) {
1072 printf("# thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1073 process_nr, thread_nr, global_data, process_data, thread_data);
1074 }
1075
1076 if (g->p.serialize_startup) {
1077 pthread_mutex_lock(&g->startup_mutex);
1078 g->nr_tasks_started++;
1079 pthread_mutex_unlock(&g->startup_mutex);
1080
1081 /* Here we will wait for the main process to start us all at once: */
1082 pthread_mutex_lock(&g->start_work_mutex);
1083 g->nr_tasks_working++;
1084
1085 /* Last one wake the main process: */
1086 if (g->nr_tasks_working == g->p.nr_tasks)
1087 pthread_mutex_unlock(&g->startup_done_mutex);
1088
1089 pthread_mutex_unlock(&g->start_work_mutex);
1090 }
1091
1092 gettimeofday(&start0, NULL);
1093
1094 start = stop = start0;
1095 last_perturbance = start.tv_sec;
1096
1097 for (l = 0; l < g->p.nr_loops; l++) {
1098 start = stop;
1099
1100 if (g->stop_work)
1101 break;
1102
1103 val += do_work(global_data, g->p.bytes_global, process_nr, g->p.nr_proc, l, val);
1104 val += do_work(process_data, g->p.bytes_process, thread_nr, g->p.nr_threads, l, val);
1105 val += do_work(thread_data, g->p.bytes_thread, 0, 1, l, val);
1106
1107 if (g->p.sleep_usecs) {
1108 pthread_mutex_lock(td->process_lock);
1109 usleep(g->p.sleep_usecs);
1110 pthread_mutex_unlock(td->process_lock);
1111 }
1112 /*
1113 * Amount of work to be done under a process-global lock:
1114 */
1115 if (g->p.bytes_process_locked) {
1116 pthread_mutex_lock(td->process_lock);
1117 val += do_work(process_data, g->p.bytes_process_locked, thread_nr, g->p.nr_threads, l, val);
1118 pthread_mutex_unlock(td->process_lock);
1119 }
1120
1121 work_done = g->p.bytes_global + g->p.bytes_process +
1122 g->p.bytes_process_locked + g->p.bytes_thread;
1123
1124 update_curr_cpu(task_nr, work_done);
1125 bytes_done += work_done;
1126
1127 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1128 continue;
1129
1130 td->loops_done = l;
1131
1132 gettimeofday(&stop, NULL);
1133
1134 /* Check whether our max runtime timed out: */
1135 if (g->p.nr_secs) {
1136 timersub(&stop, &start0, &diff);
1137 if ((u32)diff.tv_sec >= g->p.nr_secs) {
1138 g->stop_work = true;
1139 break;
1140 }
1141 }
1142
1143 /* Update the summary at most once per second: */
1144 if (start.tv_sec == stop.tv_sec)
1145 continue;
1146
1147 /*
1148 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1149 * by migrating to CPU#0:
1150 */
1151 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1152 cpu_set_t orig_mask;
1153 int target_cpu;
1154 int this_cpu;
1155
1156 last_perturbance = stop.tv_sec;
1157
1158 /*
1159 * Depending on where we are running, move into
1160 * the other half of the system, to create some
1161 * real disturbance:
1162 */
1163 this_cpu = g->threads[task_nr].curr_cpu;
1164 if (this_cpu < g->p.nr_cpus/2)
1165 target_cpu = g->p.nr_cpus-1;
1166 else
1167 target_cpu = 0;
1168
1169 orig_mask = bind_to_cpu(target_cpu);
1170
1171 /* Here we are running on the target CPU already */
1172 if (details >= 1)
1173 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1174
1175 bind_to_cpumask(orig_mask);
1176 }
1177
1178 if (details >= 3) {
1179 timersub(&stop, &start, &diff);
1180 runtime_ns_max = diff.tv_sec * 1000000000;
1181 runtime_ns_max += diff.tv_usec * 1000;
1182
1183 if (details >= 0) {
1184 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1185 process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1186 }
1187 fflush(stdout);
1188 }
1189 if (!last_task)
1190 continue;
1191
1192 timersub(&stop, &start0, &diff);
1193 runtime_ns_max = diff.tv_sec * 1000000000ULL;
1194 runtime_ns_max += diff.tv_usec * 1000ULL;
1195
1196 show_summary(runtime_ns_max, l, &convergence);
1197 }
1198
1199 gettimeofday(&stop, NULL);
1200 timersub(&stop, &start0, &diff);
1201 td->runtime_ns = diff.tv_sec * 1000000000ULL;
1202 td->runtime_ns += diff.tv_usec * 1000ULL;
1203 td->speed_gbs = bytes_done / (td->runtime_ns / 1e9) / 1e9;
1204
1205 getrusage(RUSAGE_THREAD, &rusage);
1206 td->system_time_ns = rusage.ru_stime.tv_sec * 1000000000ULL;
1207 td->system_time_ns += rusage.ru_stime.tv_usec * 1000ULL;
1208 td->user_time_ns = rusage.ru_utime.tv_sec * 1000000000ULL;
1209 td->user_time_ns += rusage.ru_utime.tv_usec * 1000ULL;
1210
1211 free_data(thread_data, g->p.bytes_thread);
1212
1213 pthread_mutex_lock(&g->stop_work_mutex);
1214 g->bytes_done += bytes_done;
1215 pthread_mutex_unlock(&g->stop_work_mutex);
1216
1217 return NULL;
1218}
1219
1220/*
1221 * A worker process starts a couple of threads:
1222 */
1223static void worker_process(int process_nr)
1224{
1225 pthread_mutex_t process_lock;
1226 struct thread_data *td;
1227 pthread_t *pthreads;
1228 u8 *process_data;
1229 int task_nr;
1230 int ret;
1231 int t;
1232
1233 pthread_mutex_init(&process_lock, NULL);
1234 set_taskname("process %d", process_nr);
1235
1236 /*
1237 * Pick up the memory policy and the CPU binding of our first thread,
1238 * so that we initialize memory accordingly:
1239 */
1240 task_nr = process_nr*g->p.nr_threads;
1241 td = g->threads + task_nr;
1242
1243 bind_to_memnode(td->bind_node);
1244 bind_to_cpumask(td->bind_cpumask);
1245
1246 pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1247 process_data = setup_private_data(g->p.bytes_process);
1248
1249 if (g->p.show_details >= 3) {
1250 printf(" # process %2d global mem: %p, process mem: %p\n",
1251 process_nr, g->data, process_data);
1252 }
1253
1254 for (t = 0; t < g->p.nr_threads; t++) {
1255 task_nr = process_nr*g->p.nr_threads + t;
1256 td = g->threads + task_nr;
1257
1258 td->process_data = process_data;
1259 td->process_nr = process_nr;
1260 td->thread_nr = t;
1261 td->task_nr = task_nr;
1262 td->val = rand();
1263 td->curr_cpu = -1;
1264 td->process_lock = &process_lock;
1265
1266 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1267 BUG_ON(ret);
1268 }
1269
1270 for (t = 0; t < g->p.nr_threads; t++) {
1271 ret = pthread_join(pthreads[t], NULL);
1272 BUG_ON(ret);
1273 }
1274
1275 free_data(process_data, g->p.bytes_process);
1276 free(pthreads);
1277}
1278
1279static void print_summary(void)
1280{
1281 if (g->p.show_details < 0)
1282 return;
1283
1284 printf("\n ###\n");
1285 printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1286 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1287 printf(" # %5dx %5ldMB global shared mem operations\n",
1288 g->p.nr_loops, g->p.bytes_global/1024/1024);
1289 printf(" # %5dx %5ldMB process shared mem operations\n",
1290 g->p.nr_loops, g->p.bytes_process/1024/1024);
1291 printf(" # %5dx %5ldMB thread local mem operations\n",
1292 g->p.nr_loops, g->p.bytes_thread/1024/1024);
1293
1294 printf(" ###\n");
1295
1296 printf("\n ###\n"); fflush(stdout);
1297}
1298
1299static void init_thread_data(void)
1300{
1301 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1302 int t;
1303
1304 g->threads = zalloc_shared_data(size);
1305
1306 for (t = 0; t < g->p.nr_tasks; t++) {
1307 struct thread_data *td = g->threads + t;
1308 int cpu;
1309
1310 /* Allow all nodes by default: */
1311 td->bind_node = -1;
1312
1313 /* Allow all CPUs by default: */
1314 CPU_ZERO(&td->bind_cpumask);
1315 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1316 CPU_SET(cpu, &td->bind_cpumask);
1317 }
1318}
1319
1320static void deinit_thread_data(void)
1321{
1322 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1323
1324 free_data(g->threads, size);
1325}
1326
1327static int init(void)
1328{
1329 g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1330
1331 /* Copy over options: */
1332 g->p = p0;
1333
1334 g->p.nr_cpus = numa_num_configured_cpus();
1335
1336 g->p.nr_nodes = numa_max_node() + 1;
1337
1338 /* char array in count_process_nodes(): */
1339 BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1340
1341 if (g->p.show_quiet && !g->p.show_details)
1342 g->p.show_details = -1;
1343
1344 /* Some memory should be specified: */
1345 if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1346 return -1;
1347
1348 if (g->p.mb_global_str) {
1349 g->p.mb_global = atof(g->p.mb_global_str);
1350 BUG_ON(g->p.mb_global < 0);
1351 }
1352
1353 if (g->p.mb_proc_str) {
1354 g->p.mb_proc = atof(g->p.mb_proc_str);
1355 BUG_ON(g->p.mb_proc < 0);
1356 }
1357
1358 if (g->p.mb_proc_locked_str) {
1359 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1360 BUG_ON(g->p.mb_proc_locked < 0);
1361 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1362 }
1363
1364 if (g->p.mb_thread_str) {
1365 g->p.mb_thread = atof(g->p.mb_thread_str);
1366 BUG_ON(g->p.mb_thread < 0);
1367 }
1368
1369 BUG_ON(g->p.nr_threads <= 0);
1370 BUG_ON(g->p.nr_proc <= 0);
1371
1372 g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1373
1374 g->p.bytes_global = g->p.mb_global *1024L*1024L;
1375 g->p.bytes_process = g->p.mb_proc *1024L*1024L;
1376 g->p.bytes_process_locked = g->p.mb_proc_locked *1024L*1024L;
1377 g->p.bytes_thread = g->p.mb_thread *1024L*1024L;
1378
1379 g->data = setup_shared_data(g->p.bytes_global);
1380
1381 /* Startup serialization: */
1382 init_global_mutex(&g->start_work_mutex);
1383 init_global_mutex(&g->startup_mutex);
1384 init_global_mutex(&g->startup_done_mutex);
1385 init_global_mutex(&g->stop_work_mutex);
1386
1387 init_thread_data();
1388
1389 tprintf("#\n");
1390 if (parse_setup_cpu_list() || parse_setup_node_list())
1391 return -1;
1392 tprintf("#\n");
1393
1394 print_summary();
1395
1396 return 0;
1397}
1398
1399static void deinit(void)
1400{
1401 free_data(g->data, g->p.bytes_global);
1402 g->data = NULL;
1403
1404 deinit_thread_data();
1405
1406 free_data(g, sizeof(*g));
1407 g = NULL;
1408}
1409
1410/*
1411 * Print a short or long result, depending on the verbosity setting:
1412 */
1413static void print_res(const char *name, double val,
1414 const char *txt_unit, const char *txt_short, const char *txt_long)
1415{
1416 if (!name)
1417 name = "main,";
1418
1419 if (!g->p.show_quiet)
1420 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1421 else
1422 printf(" %14.3f %s\n", val, txt_long);
1423}
1424
1425static int __bench_numa(const char *name)
1426{
1427 struct timeval start, stop, diff;
1428 u64 runtime_ns_min, runtime_ns_sum;
1429 pid_t *pids, pid, wpid;
1430 double delta_runtime;
1431 double runtime_avg;
1432 double runtime_sec_max;
1433 double runtime_sec_min;
1434 int wait_stat;
1435 double bytes;
1436 int i, t, p;
1437
1438 if (init())
1439 return -1;
1440
1441 pids = zalloc(g->p.nr_proc * sizeof(*pids));
1442 pid = -1;
1443
1444 /* All threads try to acquire it, this way we can wait for them to start up: */
1445 pthread_mutex_lock(&g->start_work_mutex);
1446
1447 if (g->p.serialize_startup) {
1448 tprintf(" #\n");
1449 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1450 }
1451
1452 gettimeofday(&start, NULL);
1453
1454 for (i = 0; i < g->p.nr_proc; i++) {
1455 pid = fork();
1456 dprintf(" # process %2d: PID %d\n", i, pid);
1457
1458 BUG_ON(pid < 0);
1459 if (!pid) {
1460 /* Child process: */
1461 worker_process(i);
1462
1463 exit(0);
1464 }
1465 pids[i] = pid;
1466
1467 }
1468 /* Wait for all the threads to start up: */
1469 while (g->nr_tasks_started != g->p.nr_tasks)
1470 usleep(1000);
1471
1472 BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1473
1474 if (g->p.serialize_startup) {
1475 double startup_sec;
1476
1477 pthread_mutex_lock(&g->startup_done_mutex);
1478
1479 /* This will start all threads: */
1480 pthread_mutex_unlock(&g->start_work_mutex);
1481
1482 /* This mutex is locked - the last started thread will wake us: */
1483 pthread_mutex_lock(&g->startup_done_mutex);
1484
1485 gettimeofday(&stop, NULL);
1486
1487 timersub(&stop, &start, &diff);
1488
1489 startup_sec = diff.tv_sec * 1000000000.0;
1490 startup_sec += diff.tv_usec * 1000.0;
1491 startup_sec /= 1e9;
1492
1493 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1494 tprintf(" #\n");
1495
1496 start = stop;
1497 pthread_mutex_unlock(&g->startup_done_mutex);
1498 } else {
1499 gettimeofday(&start, NULL);
1500 }
1501
1502 /* Parent process: */
1503
1504
1505 for (i = 0; i < g->p.nr_proc; i++) {
1506 wpid = waitpid(pids[i], &wait_stat, 0);
1507 BUG_ON(wpid < 0);
1508 BUG_ON(!WIFEXITED(wait_stat));
1509
1510 }
1511
1512 runtime_ns_sum = 0;
1513 runtime_ns_min = -1LL;
1514
1515 for (t = 0; t < g->p.nr_tasks; t++) {
1516 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1517
1518 runtime_ns_sum += thread_runtime_ns;
1519 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1520 }
1521
1522 gettimeofday(&stop, NULL);
1523 timersub(&stop, &start, &diff);
1524
1525 BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1526
1527 tprintf("\n ###\n");
1528 tprintf("\n");
1529
1530 runtime_sec_max = diff.tv_sec * 1000000000.0;
1531 runtime_sec_max += diff.tv_usec * 1000.0;
1532 runtime_sec_max /= 1e9;
1533
1534 runtime_sec_min = runtime_ns_min/1e9;
1535
1536 bytes = g->bytes_done;
1537 runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / 1e9;
1538
1539 if (g->p.measure_convergence) {
1540 print_res(name, runtime_sec_max,
1541 "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1542 }
1543
1544 print_res(name, runtime_sec_max,
1545 "secs,", "runtime-max/thread", "secs slowest (max) thread-runtime");
1546
1547 print_res(name, runtime_sec_min,
1548 "secs,", "runtime-min/thread", "secs fastest (min) thread-runtime");
1549
1550 print_res(name, runtime_avg,
1551 "secs,", "runtime-avg/thread", "secs average thread-runtime");
1552
1553 delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1554 print_res(name, delta_runtime / runtime_sec_max * 100.0,
1555 "%,", "spread-runtime/thread", "% difference between max/avg runtime");
1556
1557 print_res(name, bytes / g->p.nr_tasks / 1e9,
1558 "GB,", "data/thread", "GB data processed, per thread");
1559
1560 print_res(name, bytes / 1e9,
1561 "GB,", "data-total", "GB data processed, total");
1562
1563 print_res(name, runtime_sec_max * 1e9 / (bytes / g->p.nr_tasks),
1564 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1565
1566 print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1567 "GB/sec,", "thread-speed", "GB/sec/thread speed");
1568
1569 print_res(name, bytes / runtime_sec_max / 1e9,
1570 "GB/sec,", "total-speed", "GB/sec total speed");
1571
1572 if (g->p.show_details >= 2) {
1573 char tname[32];
1574 struct thread_data *td;
1575 for (p = 0; p < g->p.nr_proc; p++) {
1576 for (t = 0; t < g->p.nr_threads; t++) {
1577 memset(tname, 0, 32);
1578 td = g->threads + p*g->p.nr_threads + t;
1579 snprintf(tname, 32, "process%d:thread%d", p, t);
1580 print_res(tname, td->speed_gbs,
1581 "GB/sec", "thread-speed", "GB/sec/thread speed");
1582 print_res(tname, td->system_time_ns / 1e9,
1583 "secs", "thread-system-time", "system CPU time/thread");
1584 print_res(tname, td->user_time_ns / 1e9,
1585 "secs", "thread-user-time", "user CPU time/thread");
1586 }
1587 }
1588 }
1589
1590 free(pids);
1591
1592 deinit();
1593
1594 return 0;
1595}
1596
1597#define MAX_ARGS 50
1598
1599static int command_size(const char **argv)
1600{
1601 int size = 0;
1602
1603 while (*argv) {
1604 size++;
1605 argv++;
1606 }
1607
1608 BUG_ON(size >= MAX_ARGS);
1609
1610 return size;
1611}
1612
1613static void init_params(struct params *p, const char *name, int argc, const char **argv)
1614{
1615 int i;
1616
1617 printf("\n # Running %s \"perf bench numa", name);
1618
1619 for (i = 0; i < argc; i++)
1620 printf(" %s", argv[i]);
1621
1622 printf("\"\n");
1623
1624 memset(p, 0, sizeof(*p));
1625
1626 /* Initialize nonzero defaults: */
1627
1628 p->serialize_startup = 1;
1629 p->data_reads = true;
1630 p->data_writes = true;
1631 p->data_backwards = true;
1632 p->data_rand_walk = true;
1633 p->nr_loops = -1;
1634 p->init_random = true;
1635 p->mb_global_str = "1";
1636 p->nr_proc = 1;
1637 p->nr_threads = 1;
1638 p->nr_secs = 5;
1639 p->run_all = argc == 1;
1640}
1641
1642static int run_bench_numa(const char *name, const char **argv)
1643{
1644 int argc = command_size(argv);
1645
1646 init_params(&p0, name, argc, argv);
1647 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1648 if (argc)
1649 goto err;
1650
1651 if (__bench_numa(name))
1652 goto err;
1653
1654 return 0;
1655
1656err:
1657 return -1;
1658}
1659
1660#define OPT_BW_RAM "-s", "20", "-zZq", "--thp", " 1", "--no-data_rand_walk"
1661#define OPT_BW_RAM_NOTHP OPT_BW_RAM, "--thp", "-1"
1662
1663#define OPT_CONV "-s", "100", "-zZ0qcm", "--thp", " 1"
1664#define OPT_CONV_NOTHP OPT_CONV, "--thp", "-1"
1665
1666#define OPT_BW "-s", "20", "-zZ0q", "--thp", " 1"
1667#define OPT_BW_NOTHP OPT_BW, "--thp", "-1"
1668
1669/*
1670 * The built-in test-suite executed by "perf bench numa -a".
1671 *
1672 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1673 */
1674static const char *tests[][MAX_ARGS] = {
1675 /* Basic single-stream NUMA bandwidth measurements: */
1676 { "RAM-bw-local,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1677 "-C" , "0", "-M", "0", OPT_BW_RAM },
1678 { "RAM-bw-local-NOTHP,",
1679 "mem", "-p", "1", "-t", "1", "-P", "1024",
1680 "-C" , "0", "-M", "0", OPT_BW_RAM_NOTHP },
1681 { "RAM-bw-remote,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1682 "-C" , "0", "-M", "1", OPT_BW_RAM },
1683
1684 /* 2-stream NUMA bandwidth measurements: */
1685 { "RAM-bw-local-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1686 "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1687 { "RAM-bw-remote-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1688 "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1689
1690 /* Cross-stream NUMA bandwidth measurement: */
1691 { "RAM-bw-cross,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1692 "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1693
1694 /* Convergence latency measurements: */
1695 { " 1x3-convergence,", "mem", "-p", "1", "-t", "3", "-P", "512", OPT_CONV },
1696 { " 1x4-convergence,", "mem", "-p", "1", "-t", "4", "-P", "512", OPT_CONV },
1697 { " 1x6-convergence,", "mem", "-p", "1", "-t", "6", "-P", "1020", OPT_CONV },
1698 { " 2x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1699 { " 3x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1700 { " 4x4-convergence,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV },
1701 { " 4x4-convergence-NOTHP,",
1702 "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1703 { " 4x6-convergence,", "mem", "-p", "4", "-t", "6", "-P", "1020", OPT_CONV },
1704 { " 4x8-convergence,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_CONV },
1705 { " 8x4-convergence,", "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV },
1706 { " 8x4-convergence-NOTHP,",
1707 "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1708 { " 3x1-convergence,", "mem", "-p", "3", "-t", "1", "-P", "512", OPT_CONV },
1709 { " 4x1-convergence,", "mem", "-p", "4", "-t", "1", "-P", "512", OPT_CONV },
1710 { " 8x1-convergence,", "mem", "-p", "8", "-t", "1", "-P", "512", OPT_CONV },
1711 { "16x1-convergence,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_CONV },
1712 { "32x1-convergence,", "mem", "-p", "32", "-t", "1", "-P", "128", OPT_CONV },
1713
1714 /* Various NUMA process/thread layout bandwidth measurements: */
1715 { " 2x1-bw-process,", "mem", "-p", "2", "-t", "1", "-P", "1024", OPT_BW },
1716 { " 3x1-bw-process,", "mem", "-p", "3", "-t", "1", "-P", "1024", OPT_BW },
1717 { " 4x1-bw-process,", "mem", "-p", "4", "-t", "1", "-P", "1024", OPT_BW },
1718 { " 8x1-bw-process,", "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW },
1719 { " 8x1-bw-process-NOTHP,",
1720 "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW_NOTHP },
1721 { "16x1-bw-process,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_BW },
1722
1723 { " 4x1-bw-thread,", "mem", "-p", "1", "-t", "4", "-T", "256", OPT_BW },
1724 { " 8x1-bw-thread,", "mem", "-p", "1", "-t", "8", "-T", "256", OPT_BW },
1725 { "16x1-bw-thread,", "mem", "-p", "1", "-t", "16", "-T", "128", OPT_BW },
1726 { "32x1-bw-thread,", "mem", "-p", "1", "-t", "32", "-T", "64", OPT_BW },
1727
1728 { " 2x3-bw-thread,", "mem", "-p", "2", "-t", "3", "-P", "512", OPT_BW },
1729 { " 4x4-bw-thread,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_BW },
1730 { " 4x6-bw-thread,", "mem", "-p", "4", "-t", "6", "-P", "512", OPT_BW },
1731 { " 4x8-bw-thread,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW },
1732 { " 4x8-bw-thread-NOTHP,",
1733 "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW_NOTHP },
1734 { " 3x3-bw-thread,", "mem", "-p", "3", "-t", "3", "-P", "512", OPT_BW },
1735 { " 5x5-bw-thread,", "mem", "-p", "5", "-t", "5", "-P", "512", OPT_BW },
1736
1737 { "2x16-bw-thread,", "mem", "-p", "2", "-t", "16", "-P", "512", OPT_BW },
1738 { "1x32-bw-thread,", "mem", "-p", "1", "-t", "32", "-P", "2048", OPT_BW },
1739
1740 { "numa02-bw,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW },
1741 { "numa02-bw-NOTHP,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW_NOTHP },
1742 { "numa01-bw-thread,", "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW },
1743 { "numa01-bw-thread-NOTHP,",
1744 "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW_NOTHP },
1745};
1746
1747static int bench_all(void)
1748{
1749 int nr = ARRAY_SIZE(tests);
1750 int ret;
1751 int i;
1752
1753 ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1754 BUG_ON(ret < 0);
1755
1756 for (i = 0; i < nr; i++) {
1757 run_bench_numa(tests[i][0], tests[i] + 1);
1758 }
1759
1760 printf("\n");
1761
1762 return 0;
1763}
1764
1765int bench_numa(int argc, const char **argv, const char *prefix __maybe_unused)
1766{
1767 init_params(&p0, "main,", argc, argv);
1768 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1769 if (argc)
1770 goto err;
1771
1772 if (p0.run_all)
1773 return bench_all();
1774
1775 if (__bench_numa(NULL))
1776 goto err;
1777
1778 return 0;
1779
1780err:
1781 usage_with_options(numa_usage, options);
1782 return -1;
1783}