4 * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
7 /* For the CLR_() macros */
11 #include "../builtin.h"
12 #include "../util/util.h"
13 #include <subcmd/parse-options.h>
14 #include "../util/cloexec.h"
29 #include <sys/resource.h>
31 #include <sys/prctl.h>
32 #include <sys/types.h>
33 #include <linux/time64.h>
39 # define RUSAGE_THREAD 1
43 * Regular printout to the terminal, supressed if -q is specified:
45 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
50 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
54 cpu_set_t bind_cpumask;
60 unsigned int loops_done;
66 pthread_mutex_t *process_lock;
69 /* Parameters set by options: */
72 /* Startup synchronization: */
73 bool serialize_startup;
79 /* Working set sizes: */
80 const char *mb_global_str;
81 const char *mb_proc_str;
82 const char *mb_proc_locked_str;
83 const char *mb_thread_str;
87 double mb_proc_locked;
90 /* Access patterns to the working set: */
94 bool data_zero_memset;
100 /* Working set initialization: */
112 long bytes_process_locked;
118 bool show_convergence;
119 bool measure_convergence;
125 /* Affinity options -C and -N: */
131 /* Global, read-writable area, accessible to all processes and threads: */
136 pthread_mutex_t startup_mutex;
137 int nr_tasks_started;
139 pthread_mutex_t startup_done_mutex;
141 pthread_mutex_t start_work_mutex;
142 int nr_tasks_working;
144 pthread_mutex_t stop_work_mutex;
147 struct thread_data *threads;
149 /* Convergence latency measurement: */
158 static struct global_info *g = NULL;
160 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
161 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
165 static const struct option options[] = {
166 OPT_INTEGER('p', "nr_proc" , &p0.nr_proc, "number of processes"),
167 OPT_INTEGER('t', "nr_threads" , &p0.nr_threads, "number of threads per process"),
169 OPT_STRING('G', "mb_global" , &p0.mb_global_str, "MB", "global memory (MBs)"),
170 OPT_STRING('P', "mb_proc" , &p0.mb_proc_str, "MB", "process memory (MBs)"),
171 OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
172 OPT_STRING('T', "mb_thread" , &p0.mb_thread_str, "MB", "thread memory (MBs)"),
174 OPT_UINTEGER('l', "nr_loops" , &p0.nr_loops, "max number of loops to run (default: unlimited)"),
175 OPT_UINTEGER('s', "nr_secs" , &p0.nr_secs, "max number of seconds to run (default: 5 secs)"),
176 OPT_UINTEGER('u', "usleep" , &p0.sleep_usecs, "usecs to sleep per loop iteration"),
178 OPT_BOOLEAN('R', "data_reads" , &p0.data_reads, "access the data via writes (can be mixed with -W)"),
179 OPT_BOOLEAN('W', "data_writes" , &p0.data_writes, "access the data via writes (can be mixed with -R)"),
180 OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards, "access the data backwards as well"),
181 OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
182 OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk, "access the data with random (32bit LFSR) walk"),
185 OPT_BOOLEAN('z', "init_zero" , &p0.init_zero, "bzero the initial allocations"),
186 OPT_BOOLEAN('I', "init_random" , &p0.init_random, "randomize the contents of the initial allocations"),
187 OPT_BOOLEAN('0', "init_cpu0" , &p0.init_cpu0, "do the initial allocations on CPU#0"),
188 OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs, "perturb thread 0/0 every X secs, to test convergence stability"),
190 OPT_INCR ('d', "show_details" , &p0.show_details, "Show details"),
191 OPT_INCR ('a', "all" , &p0.run_all, "Run all tests in the suite"),
192 OPT_INTEGER('H', "thp" , &p0.thp, "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
193 OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details"),
194 OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
195 OPT_BOOLEAN('q', "quiet" , &p0.show_quiet, "quiet mode"),
196 OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
198 /* Special option string parsing callbacks: */
199 OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
200 "bind the first N tasks to these specific cpus (the rest is unbound)",
202 OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
203 "bind the first N tasks to these specific memory nodes (the rest is unbound)",
208 static const char * const bench_numa_usage[] = {
209 "perf bench numa <options>",
213 static const char * const numa_usage[] = {
214 "perf bench numa mem [<options>]",
219 * To get number of numa nodes present.
221 static int nr_numa_nodes(void)
225 for (i = 0; i < g->p.nr_nodes; i++) {
226 if (numa_bitmask_isbitset(numa_nodes_ptr, i))
234 * To check if given numa node is present.
236 static int is_node_present(int node)
238 return numa_bitmask_isbitset(numa_nodes_ptr, node);
242 * To check given numa node has cpus.
244 static bool node_has_cpus(int node)
246 struct bitmask *cpu = numa_allocate_cpumask();
249 if (cpu && !numa_node_to_cpus(node, cpu)) {
250 for (i = 0; i < cpu->size; i++) {
251 if (numa_bitmask_isbitset(cpu, i))
256 return false; /* lets fall back to nocpus safely */
259 static cpu_set_t bind_to_cpu(int target_cpu)
261 cpu_set_t orig_mask, mask;
264 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
269 if (target_cpu == -1) {
272 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
275 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
276 CPU_SET(target_cpu, &mask);
279 ret = sched_setaffinity(0, sizeof(mask), &mask);
285 static cpu_set_t bind_to_node(int target_node)
287 int cpus_per_node = g->p.nr_cpus / nr_numa_nodes();
288 cpu_set_t orig_mask, mask;
292 BUG_ON(cpus_per_node * nr_numa_nodes() != g->p.nr_cpus);
293 BUG_ON(!cpus_per_node);
295 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
300 if (target_node == -1) {
301 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
304 int cpu_start = (target_node + 0) * cpus_per_node;
305 int cpu_stop = (target_node + 1) * cpus_per_node;
307 BUG_ON(cpu_stop > g->p.nr_cpus);
309 for (cpu = cpu_start; cpu < cpu_stop; cpu++)
313 ret = sched_setaffinity(0, sizeof(mask), &mask);
319 static void bind_to_cpumask(cpu_set_t mask)
323 ret = sched_setaffinity(0, sizeof(mask), &mask);
327 static void mempol_restore(void)
331 ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
336 static void bind_to_memnode(int node)
338 unsigned long nodemask;
344 BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
345 nodemask = 1L << node;
347 ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
348 dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
353 #define HPSIZE (2*1024*1024)
355 #define set_taskname(fmt...) \
359 snprintf(name, 20, fmt); \
360 prctl(PR_SET_NAME, name); \
363 static u8 *alloc_data(ssize_t bytes0, int map_flags,
364 int init_zero, int init_cpu0, int thp, int init_random)
374 /* Allocate and initialize all memory on CPU#0: */
376 int node = numa_node_of_cpu(0);
378 orig_mask = bind_to_node(node);
379 bind_to_memnode(node);
382 bytes = bytes0 + HPSIZE;
384 buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
385 BUG_ON(buf == (void *)-1);
387 if (map_flags == MAP_PRIVATE) {
389 ret = madvise(buf, bytes, MADV_HUGEPAGE);
390 if (ret && !g->print_once) {
392 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
396 ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
397 if (ret && !g->print_once) {
399 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
407 /* Initialize random contents, different in each word: */
409 u64 *wbuf = (void *)buf;
413 for (i = 0; i < bytes/8; i++)
418 /* Align to 2MB boundary: */
419 buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
421 /* Restore affinity: */
423 bind_to_cpumask(orig_mask);
430 static void free_data(void *data, ssize_t bytes)
437 ret = munmap(data, bytes);
442 * Create a shared memory buffer that can be shared between processes, zeroed:
444 static void * zalloc_shared_data(ssize_t bytes)
446 return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0, g->p.thp, g->p.init_random);
450 * Create a shared memory buffer that can be shared between processes:
452 static void * setup_shared_data(ssize_t bytes)
454 return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
458 * Allocate process-local memory - this will either be shared between
459 * threads of this process, or only be accessed by this thread:
461 static void * setup_private_data(ssize_t bytes)
463 return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
467 * Return a process-shared (global) mutex:
469 static void init_global_mutex(pthread_mutex_t *mutex)
471 pthread_mutexattr_t attr;
473 pthread_mutexattr_init(&attr);
474 pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
475 pthread_mutex_init(mutex, &attr);
478 static int parse_cpu_list(const char *arg)
480 p0.cpu_list_str = strdup(arg);
482 dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
487 static int parse_setup_cpu_list(void)
489 struct thread_data *td;
493 if (!g->p.cpu_list_str)
496 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
498 str0 = str = strdup(g->p.cpu_list_str);
503 tprintf("# binding tasks to CPUs:\n");
507 int bind_cpu, bind_cpu_0, bind_cpu_1;
508 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
513 tok = strsep(&str, ",");
517 tok_end = strstr(tok, "-");
519 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
521 /* Single CPU specified: */
522 bind_cpu_0 = bind_cpu_1 = atol(tok);
524 /* CPU range specified (for example: "5-11"): */
525 bind_cpu_0 = atol(tok);
526 bind_cpu_1 = atol(tok_end + 1);
530 tok_step = strstr(tok, "#");
532 step = atol(tok_step + 1);
533 BUG_ON(step <= 0 || step >= g->p.nr_cpus);
538 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
539 * where the _4 means the next 4 CPUs are allowed.
542 tok_len = strstr(tok, "_");
544 bind_len = atol(tok_len + 1);
545 BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
548 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
550 tok_mul = strstr(tok, "x");
552 mul = atol(tok_mul + 1);
556 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
558 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
559 printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
563 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
564 BUG_ON(bind_cpu_0 > bind_cpu_1);
566 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
569 for (i = 0; i < mul; i++) {
572 if (t >= g->p.nr_tasks) {
573 printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
581 tprintf("%2d/%d", bind_cpu, bind_len);
583 tprintf("%2d", bind_cpu);
586 CPU_ZERO(&td->bind_cpumask);
587 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
588 BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
589 CPU_SET(cpu, &td->bind_cpumask);
599 if (t < g->p.nr_tasks)
600 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
606 static int parse_cpus_opt(const struct option *opt __maybe_unused,
607 const char *arg, int unset __maybe_unused)
612 return parse_cpu_list(arg);
615 static int parse_node_list(const char *arg)
617 p0.node_list_str = strdup(arg);
619 dprintf("got NODE list: {%s}\n", p0.node_list_str);
624 static int parse_setup_node_list(void)
626 struct thread_data *td;
630 if (!g->p.node_list_str)
633 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
635 str0 = str = strdup(g->p.node_list_str);
640 tprintf("# binding tasks to NODEs:\n");
644 int bind_node, bind_node_0, bind_node_1;
645 char *tok, *tok_end, *tok_step, *tok_mul;
649 tok = strsep(&str, ",");
653 tok_end = strstr(tok, "-");
655 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
657 /* Single NODE specified: */
658 bind_node_0 = bind_node_1 = atol(tok);
660 /* NODE range specified (for example: "5-11"): */
661 bind_node_0 = atol(tok);
662 bind_node_1 = atol(tok_end + 1);
666 tok_step = strstr(tok, "#");
668 step = atol(tok_step + 1);
669 BUG_ON(step <= 0 || step >= g->p.nr_nodes);
672 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
674 tok_mul = strstr(tok, "x");
676 mul = atol(tok_mul + 1);
680 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
682 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
683 printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
687 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
688 BUG_ON(bind_node_0 > bind_node_1);
690 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
693 for (i = 0; i < mul; i++) {
694 if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
695 printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
701 tprintf(" %2d", bind_node);
703 tprintf(",%2d", bind_node);
705 td->bind_node = bind_node;
714 if (t < g->p.nr_tasks)
715 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
721 static int parse_nodes_opt(const struct option *opt __maybe_unused,
722 const char *arg, int unset __maybe_unused)
727 return parse_node_list(arg);
732 #define BIT(x) (1ul << x)
734 static inline uint32_t lfsr_32(uint32_t lfsr)
736 const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
737 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
741 * Make sure there's real data dependency to RAM (when read
742 * accesses are enabled), so the compiler, the CPU and the
743 * kernel (KSM, zero page, etc.) cannot optimize away RAM
746 static inline u64 access_data(u64 *data __attribute__((unused)), u64 val)
750 if (g->p.data_writes)
756 * The worker process does two types of work, a forwards going
757 * loop and a backwards going loop.
759 * We do this so that on multiprocessor systems we do not create
760 * a 'train' of processing, with highly synchronized processes,
761 * skewing the whole benchmark.
763 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
765 long words = bytes/sizeof(u64);
766 u64 *data = (void *)__data;
767 long chunk_0, chunk_1;
772 BUG_ON(!data && words);
773 BUG_ON(data && !words);
778 /* Very simple memset() work variant: */
779 if (g->p.data_zero_memset && !g->p.data_rand_walk) {
784 /* Spread out by PID/TID nr and by loop nr: */
785 chunk_0 = words/nr_max;
786 chunk_1 = words/g->p.nr_loops;
787 off = nr*chunk_0 + loop*chunk_1;
792 if (g->p.data_rand_walk) {
793 u32 lfsr = nr + loop + val;
796 for (i = 0; i < words/1024; i++) {
799 lfsr = lfsr_32(lfsr);
801 start = lfsr % words;
802 end = min(start + 1024, words-1);
804 if (g->p.data_zero_memset) {
805 bzero(data + start, (end-start) * sizeof(u64));
807 for (j = start; j < end; j++)
808 val = access_data(data + j, val);
811 } else if (!g->p.data_backwards || (nr + loop) & 1) {
817 /* Process data forwards: */
819 if (unlikely(d >= d1))
821 if (unlikely(d == d0))
824 val = access_data(d, val);
829 /* Process data backwards: */
835 /* Process data forwards: */
837 if (unlikely(d < data))
839 if (unlikely(d == d0))
842 val = access_data(d, val);
851 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
855 cpu = sched_getcpu();
857 g->threads[task_nr].curr_cpu = cpu;
858 prctl(0, bytes_worked);
861 #define MAX_NR_NODES 64
864 * Count the number of nodes a process's threads
867 * A count of 1 means that the process is compressed
868 * to a single node. A count of g->p.nr_nodes means it's
869 * spread out on the whole system.
871 static int count_process_nodes(int process_nr)
873 char node_present[MAX_NR_NODES] = { 0, };
877 for (t = 0; t < g->p.nr_threads; t++) {
878 struct thread_data *td;
882 task_nr = process_nr*g->p.nr_threads + t;
883 td = g->threads + task_nr;
885 node = numa_node_of_cpu(td->curr_cpu);
886 if (node < 0) /* curr_cpu was likely still -1 */
889 node_present[node] = 1;
894 for (n = 0; n < MAX_NR_NODES; n++)
895 nodes += node_present[n];
901 * Count the number of distinct process-threads a node contains.
903 * A count of 1 means that the node contains only a single
904 * process. If all nodes on the system contain at most one
905 * process then we are well-converged.
907 static int count_node_processes(int node)
912 for (p = 0; p < g->p.nr_proc; p++) {
913 for (t = 0; t < g->p.nr_threads; t++) {
914 struct thread_data *td;
918 task_nr = p*g->p.nr_threads + t;
919 td = g->threads + task_nr;
921 n = numa_node_of_cpu(td->curr_cpu);
932 static void calc_convergence_compression(int *strong)
934 unsigned int nodes_min, nodes_max;
940 for (p = 0; p < g->p.nr_proc; p++) {
941 unsigned int nodes = count_process_nodes(p);
948 nodes_min = min(nodes, nodes_min);
949 nodes_max = max(nodes, nodes_max);
952 /* Strong convergence: all threads compress on a single node: */
953 if (nodes_min == 1 && nodes_max == 1) {
957 tprintf(" {%d-%d}", nodes_min, nodes_max);
961 static void calc_convergence(double runtime_ns_max, double *convergence)
963 unsigned int loops_done_min, loops_done_max;
965 int nodes[MAX_NR_NODES];
976 if (!g->p.show_convergence && !g->p.measure_convergence)
979 for (node = 0; node < g->p.nr_nodes; node++)
985 for (t = 0; t < g->p.nr_tasks; t++) {
986 struct thread_data *td = g->threads + t;
987 unsigned int loops_done;
991 /* Not all threads have written it yet: */
995 node = numa_node_of_cpu(cpu);
999 loops_done = td->loops_done;
1000 loops_done_min = min(loops_done, loops_done_min);
1001 loops_done_max = max(loops_done, loops_done_max);
1005 nr_min = g->p.nr_tasks;
1008 for (node = 0; node < g->p.nr_nodes; node++) {
1009 if (!is_node_present(node))
1012 nr_min = min(nr, nr_min);
1013 nr_max = max(nr, nr_max);
1016 BUG_ON(nr_min > nr_max);
1018 BUG_ON(sum > g->p.nr_tasks);
1020 if (0 && (sum < g->p.nr_tasks))
1024 * Count the number of distinct process groups present
1025 * on nodes - when we are converged this will decrease
1030 for (node = 0; node < g->p.nr_nodes; node++) {
1033 if (!is_node_present(node))
1035 processes = count_node_processes(node);
1037 tprintf(" %2d/%-2d", nr, processes);
1039 process_groups += processes;
1042 distance = nr_max - nr_min;
1044 tprintf(" [%2d/%-2d]", distance, process_groups);
1046 tprintf(" l:%3d-%-3d (%3d)",
1047 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1049 if (loops_done_min && loops_done_max) {
1050 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1052 tprintf(" [%4.1f%%]", skew * 100.0);
1055 calc_convergence_compression(&strong);
1057 if (strong && process_groups == g->p.nr_proc) {
1058 if (!*convergence) {
1059 *convergence = runtime_ns_max;
1060 tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1061 if (g->p.measure_convergence) {
1062 g->all_converged = true;
1063 g->stop_work = true;
1068 tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1075 static void show_summary(double runtime_ns_max, int l, double *convergence)
1077 tprintf("\r # %5.1f%% [%.1f mins]",
1078 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1080 calc_convergence(runtime_ns_max, convergence);
1082 if (g->p.show_details >= 0)
1086 static void *worker_thread(void *__tdata)
1088 struct thread_data *td = __tdata;
1089 struct timeval start0, start, stop, diff;
1090 int process_nr = td->process_nr;
1091 int thread_nr = td->thread_nr;
1092 unsigned long last_perturbance;
1093 int task_nr = td->task_nr;
1094 int details = g->p.show_details;
1095 int first_task, last_task;
1096 double convergence = 0;
1098 double runtime_ns_max;
1102 u64 bytes_done, secs;
1105 struct rusage rusage;
1107 bind_to_cpumask(td->bind_cpumask);
1108 bind_to_memnode(td->bind_node);
1110 set_taskname("thread %d/%d", process_nr, thread_nr);
1112 global_data = g->data;
1113 process_data = td->process_data;
1114 thread_data = setup_private_data(g->p.bytes_thread);
1119 if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1123 if (process_nr == 0 && thread_nr == 0)
1127 printf("# thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1128 process_nr, thread_nr, global_data, process_data, thread_data);
1131 if (g->p.serialize_startup) {
1132 pthread_mutex_lock(&g->startup_mutex);
1133 g->nr_tasks_started++;
1134 pthread_mutex_unlock(&g->startup_mutex);
1136 /* Here we will wait for the main process to start us all at once: */
1137 pthread_mutex_lock(&g->start_work_mutex);
1138 g->nr_tasks_working++;
1140 /* Last one wake the main process: */
1141 if (g->nr_tasks_working == g->p.nr_tasks)
1142 pthread_mutex_unlock(&g->startup_done_mutex);
1144 pthread_mutex_unlock(&g->start_work_mutex);
1147 gettimeofday(&start0, NULL);
1149 start = stop = start0;
1150 last_perturbance = start.tv_sec;
1152 for (l = 0; l < g->p.nr_loops; l++) {
1158 val += do_work(global_data, g->p.bytes_global, process_nr, g->p.nr_proc, l, val);
1159 val += do_work(process_data, g->p.bytes_process, thread_nr, g->p.nr_threads, l, val);
1160 val += do_work(thread_data, g->p.bytes_thread, 0, 1, l, val);
1162 if (g->p.sleep_usecs) {
1163 pthread_mutex_lock(td->process_lock);
1164 usleep(g->p.sleep_usecs);
1165 pthread_mutex_unlock(td->process_lock);
1168 * Amount of work to be done under a process-global lock:
1170 if (g->p.bytes_process_locked) {
1171 pthread_mutex_lock(td->process_lock);
1172 val += do_work(process_data, g->p.bytes_process_locked, thread_nr, g->p.nr_threads, l, val);
1173 pthread_mutex_unlock(td->process_lock);
1176 work_done = g->p.bytes_global + g->p.bytes_process +
1177 g->p.bytes_process_locked + g->p.bytes_thread;
1179 update_curr_cpu(task_nr, work_done);
1180 bytes_done += work_done;
1182 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1187 gettimeofday(&stop, NULL);
1189 /* Check whether our max runtime timed out: */
1191 timersub(&stop, &start0, &diff);
1192 if ((u32)diff.tv_sec >= g->p.nr_secs) {
1193 g->stop_work = true;
1198 /* Update the summary at most once per second: */
1199 if (start.tv_sec == stop.tv_sec)
1203 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1204 * by migrating to CPU#0:
1206 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1207 cpu_set_t orig_mask;
1211 last_perturbance = stop.tv_sec;
1214 * Depending on where we are running, move into
1215 * the other half of the system, to create some
1218 this_cpu = g->threads[task_nr].curr_cpu;
1219 if (this_cpu < g->p.nr_cpus/2)
1220 target_cpu = g->p.nr_cpus-1;
1224 orig_mask = bind_to_cpu(target_cpu);
1226 /* Here we are running on the target CPU already */
1228 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1230 bind_to_cpumask(orig_mask);
1234 timersub(&stop, &start, &diff);
1235 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1236 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1239 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1240 process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1247 timersub(&stop, &start0, &diff);
1248 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1249 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1251 show_summary(runtime_ns_max, l, &convergence);
1254 gettimeofday(&stop, NULL);
1255 timersub(&stop, &start0, &diff);
1256 td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1257 td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1258 secs = td->runtime_ns / NSEC_PER_SEC;
1259 td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1261 getrusage(RUSAGE_THREAD, &rusage);
1262 td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1263 td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1264 td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1265 td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1267 free_data(thread_data, g->p.bytes_thread);
1269 pthread_mutex_lock(&g->stop_work_mutex);
1270 g->bytes_done += bytes_done;
1271 pthread_mutex_unlock(&g->stop_work_mutex);
1277 * A worker process starts a couple of threads:
1279 static void worker_process(int process_nr)
1281 pthread_mutex_t process_lock;
1282 struct thread_data *td;
1283 pthread_t *pthreads;
1289 pthread_mutex_init(&process_lock, NULL);
1290 set_taskname("process %d", process_nr);
1293 * Pick up the memory policy and the CPU binding of our first thread,
1294 * so that we initialize memory accordingly:
1296 task_nr = process_nr*g->p.nr_threads;
1297 td = g->threads + task_nr;
1299 bind_to_memnode(td->bind_node);
1300 bind_to_cpumask(td->bind_cpumask);
1302 pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1303 process_data = setup_private_data(g->p.bytes_process);
1305 if (g->p.show_details >= 3) {
1306 printf(" # process %2d global mem: %p, process mem: %p\n",
1307 process_nr, g->data, process_data);
1310 for (t = 0; t < g->p.nr_threads; t++) {
1311 task_nr = process_nr*g->p.nr_threads + t;
1312 td = g->threads + task_nr;
1314 td->process_data = process_data;
1315 td->process_nr = process_nr;
1317 td->task_nr = task_nr;
1320 td->process_lock = &process_lock;
1322 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1326 for (t = 0; t < g->p.nr_threads; t++) {
1327 ret = pthread_join(pthreads[t], NULL);
1331 free_data(process_data, g->p.bytes_process);
1335 static void print_summary(void)
1337 if (g->p.show_details < 0)
1341 printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1342 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1343 printf(" # %5dx %5ldMB global shared mem operations\n",
1344 g->p.nr_loops, g->p.bytes_global/1024/1024);
1345 printf(" # %5dx %5ldMB process shared mem operations\n",
1346 g->p.nr_loops, g->p.bytes_process/1024/1024);
1347 printf(" # %5dx %5ldMB thread local mem operations\n",
1348 g->p.nr_loops, g->p.bytes_thread/1024/1024);
1352 printf("\n ###\n"); fflush(stdout);
1355 static void init_thread_data(void)
1357 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1360 g->threads = zalloc_shared_data(size);
1362 for (t = 0; t < g->p.nr_tasks; t++) {
1363 struct thread_data *td = g->threads + t;
1366 /* Allow all nodes by default: */
1369 /* Allow all CPUs by default: */
1370 CPU_ZERO(&td->bind_cpumask);
1371 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1372 CPU_SET(cpu, &td->bind_cpumask);
1376 static void deinit_thread_data(void)
1378 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1380 free_data(g->threads, size);
1383 static int init(void)
1385 g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1387 /* Copy over options: */
1390 g->p.nr_cpus = numa_num_configured_cpus();
1392 g->p.nr_nodes = numa_max_node() + 1;
1394 /* char array in count_process_nodes(): */
1395 BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1397 if (g->p.show_quiet && !g->p.show_details)
1398 g->p.show_details = -1;
1400 /* Some memory should be specified: */
1401 if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1404 if (g->p.mb_global_str) {
1405 g->p.mb_global = atof(g->p.mb_global_str);
1406 BUG_ON(g->p.mb_global < 0);
1409 if (g->p.mb_proc_str) {
1410 g->p.mb_proc = atof(g->p.mb_proc_str);
1411 BUG_ON(g->p.mb_proc < 0);
1414 if (g->p.mb_proc_locked_str) {
1415 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1416 BUG_ON(g->p.mb_proc_locked < 0);
1417 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1420 if (g->p.mb_thread_str) {
1421 g->p.mb_thread = atof(g->p.mb_thread_str);
1422 BUG_ON(g->p.mb_thread < 0);
1425 BUG_ON(g->p.nr_threads <= 0);
1426 BUG_ON(g->p.nr_proc <= 0);
1428 g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1430 g->p.bytes_global = g->p.mb_global *1024L*1024L;
1431 g->p.bytes_process = g->p.mb_proc *1024L*1024L;
1432 g->p.bytes_process_locked = g->p.mb_proc_locked *1024L*1024L;
1433 g->p.bytes_thread = g->p.mb_thread *1024L*1024L;
1435 g->data = setup_shared_data(g->p.bytes_global);
1437 /* Startup serialization: */
1438 init_global_mutex(&g->start_work_mutex);
1439 init_global_mutex(&g->startup_mutex);
1440 init_global_mutex(&g->startup_done_mutex);
1441 init_global_mutex(&g->stop_work_mutex);
1446 if (parse_setup_cpu_list() || parse_setup_node_list())
1455 static void deinit(void)
1457 free_data(g->data, g->p.bytes_global);
1460 deinit_thread_data();
1462 free_data(g, sizeof(*g));
1467 * Print a short or long result, depending on the verbosity setting:
1469 static void print_res(const char *name, double val,
1470 const char *txt_unit, const char *txt_short, const char *txt_long)
1475 if (!g->p.show_quiet)
1476 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1478 printf(" %14.3f %s\n", val, txt_long);
1481 static int __bench_numa(const char *name)
1483 struct timeval start, stop, diff;
1484 u64 runtime_ns_min, runtime_ns_sum;
1485 pid_t *pids, pid, wpid;
1486 double delta_runtime;
1488 double runtime_sec_max;
1489 double runtime_sec_min;
1497 pids = zalloc(g->p.nr_proc * sizeof(*pids));
1500 /* All threads try to acquire it, this way we can wait for them to start up: */
1501 pthread_mutex_lock(&g->start_work_mutex);
1503 if (g->p.serialize_startup) {
1505 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1508 gettimeofday(&start, NULL);
1510 for (i = 0; i < g->p.nr_proc; i++) {
1512 dprintf(" # process %2d: PID %d\n", i, pid);
1516 /* Child process: */
1524 /* Wait for all the threads to start up: */
1525 while (g->nr_tasks_started != g->p.nr_tasks)
1526 usleep(USEC_PER_MSEC);
1528 BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1530 if (g->p.serialize_startup) {
1533 pthread_mutex_lock(&g->startup_done_mutex);
1535 /* This will start all threads: */
1536 pthread_mutex_unlock(&g->start_work_mutex);
1538 /* This mutex is locked - the last started thread will wake us: */
1539 pthread_mutex_lock(&g->startup_done_mutex);
1541 gettimeofday(&stop, NULL);
1543 timersub(&stop, &start, &diff);
1545 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1546 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1547 startup_sec /= NSEC_PER_SEC;
1549 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1553 pthread_mutex_unlock(&g->startup_done_mutex);
1555 gettimeofday(&start, NULL);
1558 /* Parent process: */
1561 for (i = 0; i < g->p.nr_proc; i++) {
1562 wpid = waitpid(pids[i], &wait_stat, 0);
1564 BUG_ON(!WIFEXITED(wait_stat));
1569 runtime_ns_min = -1LL;
1571 for (t = 0; t < g->p.nr_tasks; t++) {
1572 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1574 runtime_ns_sum += thread_runtime_ns;
1575 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1578 gettimeofday(&stop, NULL);
1579 timersub(&stop, &start, &diff);
1581 BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1583 tprintf("\n ###\n");
1586 runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1587 runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1588 runtime_sec_max /= NSEC_PER_SEC;
1590 runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1592 bytes = g->bytes_done;
1593 runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1595 if (g->p.measure_convergence) {
1596 print_res(name, runtime_sec_max,
1597 "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1600 print_res(name, runtime_sec_max,
1601 "secs,", "runtime-max/thread", "secs slowest (max) thread-runtime");
1603 print_res(name, runtime_sec_min,
1604 "secs,", "runtime-min/thread", "secs fastest (min) thread-runtime");
1606 print_res(name, runtime_avg,
1607 "secs,", "runtime-avg/thread", "secs average thread-runtime");
1609 delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1610 print_res(name, delta_runtime / runtime_sec_max * 100.0,
1611 "%,", "spread-runtime/thread", "% difference between max/avg runtime");
1613 print_res(name, bytes / g->p.nr_tasks / 1e9,
1614 "GB,", "data/thread", "GB data processed, per thread");
1616 print_res(name, bytes / 1e9,
1617 "GB,", "data-total", "GB data processed, total");
1619 print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1620 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1622 print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1623 "GB/sec,", "thread-speed", "GB/sec/thread speed");
1625 print_res(name, bytes / runtime_sec_max / 1e9,
1626 "GB/sec,", "total-speed", "GB/sec total speed");
1628 if (g->p.show_details >= 2) {
1629 char tname[14 + 2 * 10 + 1];
1630 struct thread_data *td;
1631 for (p = 0; p < g->p.nr_proc; p++) {
1632 for (t = 0; t < g->p.nr_threads; t++) {
1633 memset(tname, 0, sizeof(tname));
1634 td = g->threads + p*g->p.nr_threads + t;
1635 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1636 print_res(tname, td->speed_gbs,
1637 "GB/sec", "thread-speed", "GB/sec/thread speed");
1638 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1639 "secs", "thread-system-time", "system CPU time/thread");
1640 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1641 "secs", "thread-user-time", "user CPU time/thread");
1655 static int command_size(const char **argv)
1664 BUG_ON(size >= MAX_ARGS);
1669 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1673 printf("\n # Running %s \"perf bench numa", name);
1675 for (i = 0; i < argc; i++)
1676 printf(" %s", argv[i]);
1680 memset(p, 0, sizeof(*p));
1682 /* Initialize nonzero defaults: */
1684 p->serialize_startup = 1;
1685 p->data_reads = true;
1686 p->data_writes = true;
1687 p->data_backwards = true;
1688 p->data_rand_walk = true;
1690 p->init_random = true;
1691 p->mb_global_str = "1";
1695 p->run_all = argc == 1;
1698 static int run_bench_numa(const char *name, const char **argv)
1700 int argc = command_size(argv);
1702 init_params(&p0, name, argc, argv);
1703 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1707 if (__bench_numa(name))
1716 #define OPT_BW_RAM "-s", "20", "-zZq", "--thp", " 1", "--no-data_rand_walk"
1717 #define OPT_BW_RAM_NOTHP OPT_BW_RAM, "--thp", "-1"
1719 #define OPT_CONV "-s", "100", "-zZ0qcm", "--thp", " 1"
1720 #define OPT_CONV_NOTHP OPT_CONV, "--thp", "-1"
1722 #define OPT_BW "-s", "20", "-zZ0q", "--thp", " 1"
1723 #define OPT_BW_NOTHP OPT_BW, "--thp", "-1"
1726 * The built-in test-suite executed by "perf bench numa -a".
1728 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1730 static const char *tests[][MAX_ARGS] = {
1731 /* Basic single-stream NUMA bandwidth measurements: */
1732 { "RAM-bw-local,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1733 "-C" , "0", "-M", "0", OPT_BW_RAM },
1734 { "RAM-bw-local-NOTHP,",
1735 "mem", "-p", "1", "-t", "1", "-P", "1024",
1736 "-C" , "0", "-M", "0", OPT_BW_RAM_NOTHP },
1737 { "RAM-bw-remote,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1738 "-C" , "0", "-M", "1", OPT_BW_RAM },
1740 /* 2-stream NUMA bandwidth measurements: */
1741 { "RAM-bw-local-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1742 "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1743 { "RAM-bw-remote-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1744 "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1746 /* Cross-stream NUMA bandwidth measurement: */
1747 { "RAM-bw-cross,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1748 "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1750 /* Convergence latency measurements: */
1751 { " 1x3-convergence,", "mem", "-p", "1", "-t", "3", "-P", "512", OPT_CONV },
1752 { " 1x4-convergence,", "mem", "-p", "1", "-t", "4", "-P", "512", OPT_CONV },
1753 { " 1x6-convergence,", "mem", "-p", "1", "-t", "6", "-P", "1020", OPT_CONV },
1754 { " 2x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1755 { " 3x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1756 { " 4x4-convergence,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV },
1757 { " 4x4-convergence-NOTHP,",
1758 "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1759 { " 4x6-convergence,", "mem", "-p", "4", "-t", "6", "-P", "1020", OPT_CONV },
1760 { " 4x8-convergence,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_CONV },
1761 { " 8x4-convergence,", "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV },
1762 { " 8x4-convergence-NOTHP,",
1763 "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1764 { " 3x1-convergence,", "mem", "-p", "3", "-t", "1", "-P", "512", OPT_CONV },
1765 { " 4x1-convergence,", "mem", "-p", "4", "-t", "1", "-P", "512", OPT_CONV },
1766 { " 8x1-convergence,", "mem", "-p", "8", "-t", "1", "-P", "512", OPT_CONV },
1767 { "16x1-convergence,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_CONV },
1768 { "32x1-convergence,", "mem", "-p", "32", "-t", "1", "-P", "128", OPT_CONV },
1770 /* Various NUMA process/thread layout bandwidth measurements: */
1771 { " 2x1-bw-process,", "mem", "-p", "2", "-t", "1", "-P", "1024", OPT_BW },
1772 { " 3x1-bw-process,", "mem", "-p", "3", "-t", "1", "-P", "1024", OPT_BW },
1773 { " 4x1-bw-process,", "mem", "-p", "4", "-t", "1", "-P", "1024", OPT_BW },
1774 { " 8x1-bw-process,", "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW },
1775 { " 8x1-bw-process-NOTHP,",
1776 "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW_NOTHP },
1777 { "16x1-bw-process,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_BW },
1779 { " 4x1-bw-thread,", "mem", "-p", "1", "-t", "4", "-T", "256", OPT_BW },
1780 { " 8x1-bw-thread,", "mem", "-p", "1", "-t", "8", "-T", "256", OPT_BW },
1781 { "16x1-bw-thread,", "mem", "-p", "1", "-t", "16", "-T", "128", OPT_BW },
1782 { "32x1-bw-thread,", "mem", "-p", "1", "-t", "32", "-T", "64", OPT_BW },
1784 { " 2x3-bw-thread,", "mem", "-p", "2", "-t", "3", "-P", "512", OPT_BW },
1785 { " 4x4-bw-thread,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_BW },
1786 { " 4x6-bw-thread,", "mem", "-p", "4", "-t", "6", "-P", "512", OPT_BW },
1787 { " 4x8-bw-thread,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW },
1788 { " 4x8-bw-thread-NOTHP,",
1789 "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW_NOTHP },
1790 { " 3x3-bw-thread,", "mem", "-p", "3", "-t", "3", "-P", "512", OPT_BW },
1791 { " 5x5-bw-thread,", "mem", "-p", "5", "-t", "5", "-P", "512", OPT_BW },
1793 { "2x16-bw-thread,", "mem", "-p", "2", "-t", "16", "-P", "512", OPT_BW },
1794 { "1x32-bw-thread,", "mem", "-p", "1", "-t", "32", "-P", "2048", OPT_BW },
1796 { "numa02-bw,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW },
1797 { "numa02-bw-NOTHP,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW_NOTHP },
1798 { "numa01-bw-thread,", "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW },
1799 { "numa01-bw-thread-NOTHP,",
1800 "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW_NOTHP },
1803 static int bench_all(void)
1805 int nr = ARRAY_SIZE(tests);
1809 ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1812 for (i = 0; i < nr; i++) {
1813 run_bench_numa(tests[i][0], tests[i] + 1);
1821 int bench_numa(int argc, const char **argv, const char *prefix __maybe_unused)
1823 init_params(&p0, "main,", argc, argv);
1824 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1831 if (__bench_numa(NULL))
1837 usage_with_options(numa_usage, options);