1 // SPDX-License-Identifier: GPL-2.0
5 * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
9 /* For the CLR_() macros */
13 #include "../builtin.h"
14 #include "../util/util.h"
15 #include <subcmd/parse-options.h>
16 #include "../util/cloexec.h"
31 #include <sys/resource.h>
33 #include <sys/prctl.h>
34 #include <sys/types.h>
35 #include <linux/kernel.h>
36 #include <linux/time64.h>
42 * Regular printout to the terminal, supressed if -q is specified:
44 #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 "convergence is reached when each process (all its threads) is running on a single NUMA node."),
195 OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
196 OPT_BOOLEAN('q', "quiet" , &p0.show_quiet, "quiet mode"),
197 OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
199 /* Special option string parsing callbacks: */
200 OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
201 "bind the first N tasks to these specific cpus (the rest is unbound)",
203 OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
204 "bind the first N tasks to these specific memory nodes (the rest is unbound)",
209 static const char * const bench_numa_usage[] = {
210 "perf bench numa <options>",
214 static const char * const numa_usage[] = {
215 "perf bench numa mem [<options>]",
219 static cpu_set_t bind_to_cpu(int target_cpu)
221 cpu_set_t orig_mask, mask;
224 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
229 if (target_cpu == -1) {
232 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
235 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
236 CPU_SET(target_cpu, &mask);
239 ret = sched_setaffinity(0, sizeof(mask), &mask);
245 static cpu_set_t bind_to_node(int target_node)
247 int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
248 cpu_set_t orig_mask, mask;
252 BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
253 BUG_ON(!cpus_per_node);
255 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
260 if (target_node == -1) {
261 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
264 int cpu_start = (target_node + 0) * cpus_per_node;
265 int cpu_stop = (target_node + 1) * cpus_per_node;
267 BUG_ON(cpu_stop > g->p.nr_cpus);
269 for (cpu = cpu_start; cpu < cpu_stop; cpu++)
273 ret = sched_setaffinity(0, sizeof(mask), &mask);
279 static void bind_to_cpumask(cpu_set_t mask)
283 ret = sched_setaffinity(0, sizeof(mask), &mask);
287 static void mempol_restore(void)
291 ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
296 static void bind_to_memnode(int node)
298 unsigned long nodemask;
304 BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
305 nodemask = 1L << node;
307 ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
308 dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
313 #define HPSIZE (2*1024*1024)
315 #define set_taskname(fmt...) \
319 snprintf(name, 20, fmt); \
320 prctl(PR_SET_NAME, name); \
323 static u8 *alloc_data(ssize_t bytes0, int map_flags,
324 int init_zero, int init_cpu0, int thp, int init_random)
334 /* Allocate and initialize all memory on CPU#0: */
336 orig_mask = bind_to_node(0);
340 bytes = bytes0 + HPSIZE;
342 buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
343 BUG_ON(buf == (void *)-1);
345 if (map_flags == MAP_PRIVATE) {
347 ret = madvise(buf, bytes, MADV_HUGEPAGE);
348 if (ret && !g->print_once) {
350 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
354 ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
355 if (ret && !g->print_once) {
357 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
365 /* Initialize random contents, different in each word: */
367 u64 *wbuf = (void *)buf;
371 for (i = 0; i < bytes/8; i++)
376 /* Align to 2MB boundary: */
377 buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
379 /* Restore affinity: */
381 bind_to_cpumask(orig_mask);
388 static void free_data(void *data, ssize_t bytes)
395 ret = munmap(data, bytes);
400 * Create a shared memory buffer that can be shared between processes, zeroed:
402 static void * zalloc_shared_data(ssize_t bytes)
404 return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0, g->p.thp, g->p.init_random);
408 * Create a shared memory buffer that can be shared between processes:
410 static void * setup_shared_data(ssize_t bytes)
412 return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
416 * Allocate process-local memory - this will either be shared between
417 * threads of this process, or only be accessed by this thread:
419 static void * setup_private_data(ssize_t bytes)
421 return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
425 * Return a process-shared (global) mutex:
427 static void init_global_mutex(pthread_mutex_t *mutex)
429 pthread_mutexattr_t attr;
431 pthread_mutexattr_init(&attr);
432 pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
433 pthread_mutex_init(mutex, &attr);
436 static int parse_cpu_list(const char *arg)
438 p0.cpu_list_str = strdup(arg);
440 dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
445 static int parse_setup_cpu_list(void)
447 struct thread_data *td;
451 if (!g->p.cpu_list_str)
454 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
456 str0 = str = strdup(g->p.cpu_list_str);
461 tprintf("# binding tasks to CPUs:\n");
465 int bind_cpu, bind_cpu_0, bind_cpu_1;
466 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
471 tok = strsep(&str, ",");
475 tok_end = strstr(tok, "-");
477 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
479 /* Single CPU specified: */
480 bind_cpu_0 = bind_cpu_1 = atol(tok);
482 /* CPU range specified (for example: "5-11"): */
483 bind_cpu_0 = atol(tok);
484 bind_cpu_1 = atol(tok_end + 1);
488 tok_step = strstr(tok, "#");
490 step = atol(tok_step + 1);
491 BUG_ON(step <= 0 || step >= g->p.nr_cpus);
496 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
497 * where the _4 means the next 4 CPUs are allowed.
500 tok_len = strstr(tok, "_");
502 bind_len = atol(tok_len + 1);
503 BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
506 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
508 tok_mul = strstr(tok, "x");
510 mul = atol(tok_mul + 1);
514 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
516 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
517 printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
521 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
522 BUG_ON(bind_cpu_0 > bind_cpu_1);
524 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
527 for (i = 0; i < mul; i++) {
530 if (t >= g->p.nr_tasks) {
531 printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
539 tprintf("%2d/%d", bind_cpu, bind_len);
541 tprintf("%2d", bind_cpu);
544 CPU_ZERO(&td->bind_cpumask);
545 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
546 BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
547 CPU_SET(cpu, &td->bind_cpumask);
557 if (t < g->p.nr_tasks)
558 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
564 static int parse_cpus_opt(const struct option *opt __maybe_unused,
565 const char *arg, int unset __maybe_unused)
570 return parse_cpu_list(arg);
573 static int parse_node_list(const char *arg)
575 p0.node_list_str = strdup(arg);
577 dprintf("got NODE list: {%s}\n", p0.node_list_str);
582 static int parse_setup_node_list(void)
584 struct thread_data *td;
588 if (!g->p.node_list_str)
591 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
593 str0 = str = strdup(g->p.node_list_str);
598 tprintf("# binding tasks to NODEs:\n");
602 int bind_node, bind_node_0, bind_node_1;
603 char *tok, *tok_end, *tok_step, *tok_mul;
607 tok = strsep(&str, ",");
611 tok_end = strstr(tok, "-");
613 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
615 /* Single NODE specified: */
616 bind_node_0 = bind_node_1 = atol(tok);
618 /* NODE range specified (for example: "5-11"): */
619 bind_node_0 = atol(tok);
620 bind_node_1 = atol(tok_end + 1);
624 tok_step = strstr(tok, "#");
626 step = atol(tok_step + 1);
627 BUG_ON(step <= 0 || step >= g->p.nr_nodes);
630 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
632 tok_mul = strstr(tok, "x");
634 mul = atol(tok_mul + 1);
638 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
640 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
641 printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
645 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
646 BUG_ON(bind_node_0 > bind_node_1);
648 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
651 for (i = 0; i < mul; i++) {
652 if (t >= g->p.nr_tasks) {
653 printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
659 tprintf(" %2d", bind_node);
661 tprintf(",%2d", bind_node);
663 td->bind_node = bind_node;
672 if (t < g->p.nr_tasks)
673 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
679 static int parse_nodes_opt(const struct option *opt __maybe_unused,
680 const char *arg, int unset __maybe_unused)
685 return parse_node_list(arg);
690 #define BIT(x) (1ul << x)
692 static inline uint32_t lfsr_32(uint32_t lfsr)
694 const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
695 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
699 * Make sure there's real data dependency to RAM (when read
700 * accesses are enabled), so the compiler, the CPU and the
701 * kernel (KSM, zero page, etc.) cannot optimize away RAM
704 static inline u64 access_data(u64 *data, u64 val)
708 if (g->p.data_writes)
714 * The worker process does two types of work, a forwards going
715 * loop and a backwards going loop.
717 * We do this so that on multiprocessor systems we do not create
718 * a 'train' of processing, with highly synchronized processes,
719 * skewing the whole benchmark.
721 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
723 long words = bytes/sizeof(u64);
724 u64 *data = (void *)__data;
725 long chunk_0, chunk_1;
730 BUG_ON(!data && words);
731 BUG_ON(data && !words);
736 /* Very simple memset() work variant: */
737 if (g->p.data_zero_memset && !g->p.data_rand_walk) {
742 /* Spread out by PID/TID nr and by loop nr: */
743 chunk_0 = words/nr_max;
744 chunk_1 = words/g->p.nr_loops;
745 off = nr*chunk_0 + loop*chunk_1;
750 if (g->p.data_rand_walk) {
751 u32 lfsr = nr + loop + val;
754 for (i = 0; i < words/1024; i++) {
757 lfsr = lfsr_32(lfsr);
759 start = lfsr % words;
760 end = min(start + 1024, words-1);
762 if (g->p.data_zero_memset) {
763 bzero(data + start, (end-start) * sizeof(u64));
765 for (j = start; j < end; j++)
766 val = access_data(data + j, val);
769 } else if (!g->p.data_backwards || (nr + loop) & 1) {
775 /* Process data forwards: */
777 if (unlikely(d >= d1))
779 if (unlikely(d == d0))
782 val = access_data(d, val);
787 /* Process data backwards: */
793 /* Process data forwards: */
795 if (unlikely(d < data))
797 if (unlikely(d == d0))
800 val = access_data(d, val);
809 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
813 cpu = sched_getcpu();
815 g->threads[task_nr].curr_cpu = cpu;
816 prctl(0, bytes_worked);
819 #define MAX_NR_NODES 64
822 * Count the number of nodes a process's threads
825 * A count of 1 means that the process is compressed
826 * to a single node. A count of g->p.nr_nodes means it's
827 * spread out on the whole system.
829 static int count_process_nodes(int process_nr)
831 char node_present[MAX_NR_NODES] = { 0, };
835 for (t = 0; t < g->p.nr_threads; t++) {
836 struct thread_data *td;
840 task_nr = process_nr*g->p.nr_threads + t;
841 td = g->threads + task_nr;
843 node = numa_node_of_cpu(td->curr_cpu);
844 if (node < 0) /* curr_cpu was likely still -1 */
847 node_present[node] = 1;
852 for (n = 0; n < MAX_NR_NODES; n++)
853 nodes += node_present[n];
859 * Count the number of distinct process-threads a node contains.
861 * A count of 1 means that the node contains only a single
862 * process. If all nodes on the system contain at most one
863 * process then we are well-converged.
865 static int count_node_processes(int node)
870 for (p = 0; p < g->p.nr_proc; p++) {
871 for (t = 0; t < g->p.nr_threads; t++) {
872 struct thread_data *td;
876 task_nr = p*g->p.nr_threads + t;
877 td = g->threads + task_nr;
879 n = numa_node_of_cpu(td->curr_cpu);
890 static void calc_convergence_compression(int *strong)
892 unsigned int nodes_min, nodes_max;
898 for (p = 0; p < g->p.nr_proc; p++) {
899 unsigned int nodes = count_process_nodes(p);
906 nodes_min = min(nodes, nodes_min);
907 nodes_max = max(nodes, nodes_max);
910 /* Strong convergence: all threads compress on a single node: */
911 if (nodes_min == 1 && nodes_max == 1) {
915 tprintf(" {%d-%d}", nodes_min, nodes_max);
919 static void calc_convergence(double runtime_ns_max, double *convergence)
921 unsigned int loops_done_min, loops_done_max;
923 int nodes[MAX_NR_NODES];
934 if (!g->p.show_convergence && !g->p.measure_convergence)
937 for (node = 0; node < g->p.nr_nodes; node++)
943 for (t = 0; t < g->p.nr_tasks; t++) {
944 struct thread_data *td = g->threads + t;
945 unsigned int loops_done;
949 /* Not all threads have written it yet: */
953 node = numa_node_of_cpu(cpu);
957 loops_done = td->loops_done;
958 loops_done_min = min(loops_done, loops_done_min);
959 loops_done_max = max(loops_done, loops_done_max);
963 nr_min = g->p.nr_tasks;
966 for (node = 0; node < g->p.nr_nodes; node++) {
968 nr_min = min(nr, nr_min);
969 nr_max = max(nr, nr_max);
972 BUG_ON(nr_min > nr_max);
974 BUG_ON(sum > g->p.nr_tasks);
976 if (0 && (sum < g->p.nr_tasks))
980 * Count the number of distinct process groups present
981 * on nodes - when we are converged this will decrease
986 for (node = 0; node < g->p.nr_nodes; node++) {
987 int processes = count_node_processes(node);
990 tprintf(" %2d/%-2d", nr, processes);
992 process_groups += processes;
995 distance = nr_max - nr_min;
997 tprintf(" [%2d/%-2d]", distance, process_groups);
999 tprintf(" l:%3d-%-3d (%3d)",
1000 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1002 if (loops_done_min && loops_done_max) {
1003 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1005 tprintf(" [%4.1f%%]", skew * 100.0);
1008 calc_convergence_compression(&strong);
1010 if (strong && process_groups == g->p.nr_proc) {
1011 if (!*convergence) {
1012 *convergence = runtime_ns_max;
1013 tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1014 if (g->p.measure_convergence) {
1015 g->all_converged = true;
1016 g->stop_work = true;
1021 tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1028 static void show_summary(double runtime_ns_max, int l, double *convergence)
1030 tprintf("\r # %5.1f%% [%.1f mins]",
1031 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1033 calc_convergence(runtime_ns_max, convergence);
1035 if (g->p.show_details >= 0)
1039 static void *worker_thread(void *__tdata)
1041 struct thread_data *td = __tdata;
1042 struct timeval start0, start, stop, diff;
1043 int process_nr = td->process_nr;
1044 int thread_nr = td->thread_nr;
1045 unsigned long last_perturbance;
1046 int task_nr = td->task_nr;
1047 int details = g->p.show_details;
1048 int first_task, last_task;
1049 double convergence = 0;
1051 double runtime_ns_max;
1058 struct rusage rusage;
1060 bind_to_cpumask(td->bind_cpumask);
1061 bind_to_memnode(td->bind_node);
1063 set_taskname("thread %d/%d", process_nr, thread_nr);
1065 global_data = g->data;
1066 process_data = td->process_data;
1067 thread_data = setup_private_data(g->p.bytes_thread);
1072 if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1076 if (process_nr == 0 && thread_nr == 0)
1080 printf("# thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1081 process_nr, thread_nr, global_data, process_data, thread_data);
1084 if (g->p.serialize_startup) {
1085 pthread_mutex_lock(&g->startup_mutex);
1086 g->nr_tasks_started++;
1087 pthread_mutex_unlock(&g->startup_mutex);
1089 /* Here we will wait for the main process to start us all at once: */
1090 pthread_mutex_lock(&g->start_work_mutex);
1091 g->nr_tasks_working++;
1093 /* Last one wake the main process: */
1094 if (g->nr_tasks_working == g->p.nr_tasks)
1095 pthread_mutex_unlock(&g->startup_done_mutex);
1097 pthread_mutex_unlock(&g->start_work_mutex);
1100 gettimeofday(&start0, NULL);
1102 start = stop = start0;
1103 last_perturbance = start.tv_sec;
1105 for (l = 0; l < g->p.nr_loops; l++) {
1111 val += do_work(global_data, g->p.bytes_global, process_nr, g->p.nr_proc, l, val);
1112 val += do_work(process_data, g->p.bytes_process, thread_nr, g->p.nr_threads, l, val);
1113 val += do_work(thread_data, g->p.bytes_thread, 0, 1, l, val);
1115 if (g->p.sleep_usecs) {
1116 pthread_mutex_lock(td->process_lock);
1117 usleep(g->p.sleep_usecs);
1118 pthread_mutex_unlock(td->process_lock);
1121 * Amount of work to be done under a process-global lock:
1123 if (g->p.bytes_process_locked) {
1124 pthread_mutex_lock(td->process_lock);
1125 val += do_work(process_data, g->p.bytes_process_locked, thread_nr, g->p.nr_threads, l, val);
1126 pthread_mutex_unlock(td->process_lock);
1129 work_done = g->p.bytes_global + g->p.bytes_process +
1130 g->p.bytes_process_locked + g->p.bytes_thread;
1132 update_curr_cpu(task_nr, work_done);
1133 bytes_done += work_done;
1135 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1140 gettimeofday(&stop, NULL);
1142 /* Check whether our max runtime timed out: */
1144 timersub(&stop, &start0, &diff);
1145 if ((u32)diff.tv_sec >= g->p.nr_secs) {
1146 g->stop_work = true;
1151 /* Update the summary at most once per second: */
1152 if (start.tv_sec == stop.tv_sec)
1156 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1157 * by migrating to CPU#0:
1159 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1160 cpu_set_t orig_mask;
1164 last_perturbance = stop.tv_sec;
1167 * Depending on where we are running, move into
1168 * the other half of the system, to create some
1171 this_cpu = g->threads[task_nr].curr_cpu;
1172 if (this_cpu < g->p.nr_cpus/2)
1173 target_cpu = g->p.nr_cpus-1;
1177 orig_mask = bind_to_cpu(target_cpu);
1179 /* Here we are running on the target CPU already */
1181 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1183 bind_to_cpumask(orig_mask);
1187 timersub(&stop, &start, &diff);
1188 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1189 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1192 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1193 process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1200 timersub(&stop, &start0, &diff);
1201 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1202 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1204 show_summary(runtime_ns_max, l, &convergence);
1207 gettimeofday(&stop, NULL);
1208 timersub(&stop, &start0, &diff);
1209 td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1210 td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1211 td->speed_gbs = bytes_done / (td->runtime_ns / NSEC_PER_SEC) / 1e9;
1213 getrusage(RUSAGE_THREAD, &rusage);
1214 td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1215 td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1216 td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1217 td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1219 free_data(thread_data, g->p.bytes_thread);
1221 pthread_mutex_lock(&g->stop_work_mutex);
1222 g->bytes_done += bytes_done;
1223 pthread_mutex_unlock(&g->stop_work_mutex);
1229 * A worker process starts a couple of threads:
1231 static void worker_process(int process_nr)
1233 pthread_mutex_t process_lock;
1234 struct thread_data *td;
1235 pthread_t *pthreads;
1241 pthread_mutex_init(&process_lock, NULL);
1242 set_taskname("process %d", process_nr);
1245 * Pick up the memory policy and the CPU binding of our first thread,
1246 * so that we initialize memory accordingly:
1248 task_nr = process_nr*g->p.nr_threads;
1249 td = g->threads + task_nr;
1251 bind_to_memnode(td->bind_node);
1252 bind_to_cpumask(td->bind_cpumask);
1254 pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1255 process_data = setup_private_data(g->p.bytes_process);
1257 if (g->p.show_details >= 3) {
1258 printf(" # process %2d global mem: %p, process mem: %p\n",
1259 process_nr, g->data, process_data);
1262 for (t = 0; t < g->p.nr_threads; t++) {
1263 task_nr = process_nr*g->p.nr_threads + t;
1264 td = g->threads + task_nr;
1266 td->process_data = process_data;
1267 td->process_nr = process_nr;
1269 td->task_nr = task_nr;
1272 td->process_lock = &process_lock;
1274 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1278 for (t = 0; t < g->p.nr_threads; t++) {
1279 ret = pthread_join(pthreads[t], NULL);
1283 free_data(process_data, g->p.bytes_process);
1287 static void print_summary(void)
1289 if (g->p.show_details < 0)
1293 printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1294 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1295 printf(" # %5dx %5ldMB global shared mem operations\n",
1296 g->p.nr_loops, g->p.bytes_global/1024/1024);
1297 printf(" # %5dx %5ldMB process shared mem operations\n",
1298 g->p.nr_loops, g->p.bytes_process/1024/1024);
1299 printf(" # %5dx %5ldMB thread local mem operations\n",
1300 g->p.nr_loops, g->p.bytes_thread/1024/1024);
1304 printf("\n ###\n"); fflush(stdout);
1307 static void init_thread_data(void)
1309 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1312 g->threads = zalloc_shared_data(size);
1314 for (t = 0; t < g->p.nr_tasks; t++) {
1315 struct thread_data *td = g->threads + t;
1318 /* Allow all nodes by default: */
1321 /* Allow all CPUs by default: */
1322 CPU_ZERO(&td->bind_cpumask);
1323 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1324 CPU_SET(cpu, &td->bind_cpumask);
1328 static void deinit_thread_data(void)
1330 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1332 free_data(g->threads, size);
1335 static int init(void)
1337 g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1339 /* Copy over options: */
1342 g->p.nr_cpus = numa_num_configured_cpus();
1344 g->p.nr_nodes = numa_max_node() + 1;
1346 /* char array in count_process_nodes(): */
1347 BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1349 if (g->p.show_quiet && !g->p.show_details)
1350 g->p.show_details = -1;
1352 /* Some memory should be specified: */
1353 if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1356 if (g->p.mb_global_str) {
1357 g->p.mb_global = atof(g->p.mb_global_str);
1358 BUG_ON(g->p.mb_global < 0);
1361 if (g->p.mb_proc_str) {
1362 g->p.mb_proc = atof(g->p.mb_proc_str);
1363 BUG_ON(g->p.mb_proc < 0);
1366 if (g->p.mb_proc_locked_str) {
1367 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1368 BUG_ON(g->p.mb_proc_locked < 0);
1369 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1372 if (g->p.mb_thread_str) {
1373 g->p.mb_thread = atof(g->p.mb_thread_str);
1374 BUG_ON(g->p.mb_thread < 0);
1377 BUG_ON(g->p.nr_threads <= 0);
1378 BUG_ON(g->p.nr_proc <= 0);
1380 g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1382 g->p.bytes_global = g->p.mb_global *1024L*1024L;
1383 g->p.bytes_process = g->p.mb_proc *1024L*1024L;
1384 g->p.bytes_process_locked = g->p.mb_proc_locked *1024L*1024L;
1385 g->p.bytes_thread = g->p.mb_thread *1024L*1024L;
1387 g->data = setup_shared_data(g->p.bytes_global);
1389 /* Startup serialization: */
1390 init_global_mutex(&g->start_work_mutex);
1391 init_global_mutex(&g->startup_mutex);
1392 init_global_mutex(&g->startup_done_mutex);
1393 init_global_mutex(&g->stop_work_mutex);
1398 if (parse_setup_cpu_list() || parse_setup_node_list())
1407 static void deinit(void)
1409 free_data(g->data, g->p.bytes_global);
1412 deinit_thread_data();
1414 free_data(g, sizeof(*g));
1419 * Print a short or long result, depending on the verbosity setting:
1421 static void print_res(const char *name, double val,
1422 const char *txt_unit, const char *txt_short, const char *txt_long)
1427 if (!g->p.show_quiet)
1428 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1430 printf(" %14.3f %s\n", val, txt_long);
1433 static int __bench_numa(const char *name)
1435 struct timeval start, stop, diff;
1436 u64 runtime_ns_min, runtime_ns_sum;
1437 pid_t *pids, pid, wpid;
1438 double delta_runtime;
1440 double runtime_sec_max;
1441 double runtime_sec_min;
1449 pids = zalloc(g->p.nr_proc * sizeof(*pids));
1452 /* All threads try to acquire it, this way we can wait for them to start up: */
1453 pthread_mutex_lock(&g->start_work_mutex);
1455 if (g->p.serialize_startup) {
1457 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1460 gettimeofday(&start, NULL);
1462 for (i = 0; i < g->p.nr_proc; i++) {
1464 dprintf(" # process %2d: PID %d\n", i, pid);
1468 /* Child process: */
1476 /* Wait for all the threads to start up: */
1477 while (g->nr_tasks_started != g->p.nr_tasks)
1478 usleep(USEC_PER_MSEC);
1480 BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1482 if (g->p.serialize_startup) {
1485 pthread_mutex_lock(&g->startup_done_mutex);
1487 /* This will start all threads: */
1488 pthread_mutex_unlock(&g->start_work_mutex);
1490 /* This mutex is locked - the last started thread will wake us: */
1491 pthread_mutex_lock(&g->startup_done_mutex);
1493 gettimeofday(&stop, NULL);
1495 timersub(&stop, &start, &diff);
1497 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1498 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1499 startup_sec /= NSEC_PER_SEC;
1501 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1505 pthread_mutex_unlock(&g->startup_done_mutex);
1507 gettimeofday(&start, NULL);
1510 /* Parent process: */
1513 for (i = 0; i < g->p.nr_proc; i++) {
1514 wpid = waitpid(pids[i], &wait_stat, 0);
1516 BUG_ON(!WIFEXITED(wait_stat));
1521 runtime_ns_min = -1LL;
1523 for (t = 0; t < g->p.nr_tasks; t++) {
1524 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1526 runtime_ns_sum += thread_runtime_ns;
1527 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1530 gettimeofday(&stop, NULL);
1531 timersub(&stop, &start, &diff);
1533 BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1535 tprintf("\n ###\n");
1538 runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1539 runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1540 runtime_sec_max /= NSEC_PER_SEC;
1542 runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1544 bytes = g->bytes_done;
1545 runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1547 if (g->p.measure_convergence) {
1548 print_res(name, runtime_sec_max,
1549 "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1552 print_res(name, runtime_sec_max,
1553 "secs,", "runtime-max/thread", "secs slowest (max) thread-runtime");
1555 print_res(name, runtime_sec_min,
1556 "secs,", "runtime-min/thread", "secs fastest (min) thread-runtime");
1558 print_res(name, runtime_avg,
1559 "secs,", "runtime-avg/thread", "secs average thread-runtime");
1561 delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1562 print_res(name, delta_runtime / runtime_sec_max * 100.0,
1563 "%,", "spread-runtime/thread", "% difference between max/avg runtime");
1565 print_res(name, bytes / g->p.nr_tasks / 1e9,
1566 "GB,", "data/thread", "GB data processed, per thread");
1568 print_res(name, bytes / 1e9,
1569 "GB,", "data-total", "GB data processed, total");
1571 print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1572 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1574 print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1575 "GB/sec,", "thread-speed", "GB/sec/thread speed");
1577 print_res(name, bytes / runtime_sec_max / 1e9,
1578 "GB/sec,", "total-speed", "GB/sec total speed");
1580 if (g->p.show_details >= 2) {
1581 char tname[14 + 2 * 10 + 1];
1582 struct thread_data *td;
1583 for (p = 0; p < g->p.nr_proc; p++) {
1584 for (t = 0; t < g->p.nr_threads; t++) {
1585 memset(tname, 0, sizeof(tname));
1586 td = g->threads + p*g->p.nr_threads + t;
1587 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1588 print_res(tname, td->speed_gbs,
1589 "GB/sec", "thread-speed", "GB/sec/thread speed");
1590 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1591 "secs", "thread-system-time", "system CPU time/thread");
1592 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1593 "secs", "thread-user-time", "user CPU time/thread");
1607 static int command_size(const char **argv)
1616 BUG_ON(size >= MAX_ARGS);
1621 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1625 printf("\n # Running %s \"perf bench numa", name);
1627 for (i = 0; i < argc; i++)
1628 printf(" %s", argv[i]);
1632 memset(p, 0, sizeof(*p));
1634 /* Initialize nonzero defaults: */
1636 p->serialize_startup = 1;
1637 p->data_reads = true;
1638 p->data_writes = true;
1639 p->data_backwards = true;
1640 p->data_rand_walk = true;
1642 p->init_random = true;
1643 p->mb_global_str = "1";
1647 p->run_all = argc == 1;
1650 static int run_bench_numa(const char *name, const char **argv)
1652 int argc = command_size(argv);
1654 init_params(&p0, name, argc, argv);
1655 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1659 if (__bench_numa(name))
1668 #define OPT_BW_RAM "-s", "20", "-zZq", "--thp", " 1", "--no-data_rand_walk"
1669 #define OPT_BW_RAM_NOTHP OPT_BW_RAM, "--thp", "-1"
1671 #define OPT_CONV "-s", "100", "-zZ0qcm", "--thp", " 1"
1672 #define OPT_CONV_NOTHP OPT_CONV, "--thp", "-1"
1674 #define OPT_BW "-s", "20", "-zZ0q", "--thp", " 1"
1675 #define OPT_BW_NOTHP OPT_BW, "--thp", "-1"
1678 * The built-in test-suite executed by "perf bench numa -a".
1680 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1682 static const char *tests[][MAX_ARGS] = {
1683 /* Basic single-stream NUMA bandwidth measurements: */
1684 { "RAM-bw-local,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1685 "-C" , "0", "-M", "0", OPT_BW_RAM },
1686 { "RAM-bw-local-NOTHP,",
1687 "mem", "-p", "1", "-t", "1", "-P", "1024",
1688 "-C" , "0", "-M", "0", OPT_BW_RAM_NOTHP },
1689 { "RAM-bw-remote,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1690 "-C" , "0", "-M", "1", OPT_BW_RAM },
1692 /* 2-stream NUMA bandwidth measurements: */
1693 { "RAM-bw-local-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1694 "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1695 { "RAM-bw-remote-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1696 "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1698 /* Cross-stream NUMA bandwidth measurement: */
1699 { "RAM-bw-cross,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1700 "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1702 /* Convergence latency measurements: */
1703 { " 1x3-convergence,", "mem", "-p", "1", "-t", "3", "-P", "512", OPT_CONV },
1704 { " 1x4-convergence,", "mem", "-p", "1", "-t", "4", "-P", "512", OPT_CONV },
1705 { " 1x6-convergence,", "mem", "-p", "1", "-t", "6", "-P", "1020", OPT_CONV },
1706 { " 2x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1707 { " 3x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1708 { " 4x4-convergence,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV },
1709 { " 4x4-convergence-NOTHP,",
1710 "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1711 { " 4x6-convergence,", "mem", "-p", "4", "-t", "6", "-P", "1020", OPT_CONV },
1712 { " 4x8-convergence,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_CONV },
1713 { " 8x4-convergence,", "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV },
1714 { " 8x4-convergence-NOTHP,",
1715 "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1716 { " 3x1-convergence,", "mem", "-p", "3", "-t", "1", "-P", "512", OPT_CONV },
1717 { " 4x1-convergence,", "mem", "-p", "4", "-t", "1", "-P", "512", OPT_CONV },
1718 { " 8x1-convergence,", "mem", "-p", "8", "-t", "1", "-P", "512", OPT_CONV },
1719 { "16x1-convergence,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_CONV },
1720 { "32x1-convergence,", "mem", "-p", "32", "-t", "1", "-P", "128", OPT_CONV },
1722 /* Various NUMA process/thread layout bandwidth measurements: */
1723 { " 2x1-bw-process,", "mem", "-p", "2", "-t", "1", "-P", "1024", OPT_BW },
1724 { " 3x1-bw-process,", "mem", "-p", "3", "-t", "1", "-P", "1024", OPT_BW },
1725 { " 4x1-bw-process,", "mem", "-p", "4", "-t", "1", "-P", "1024", OPT_BW },
1726 { " 8x1-bw-process,", "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW },
1727 { " 8x1-bw-process-NOTHP,",
1728 "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW_NOTHP },
1729 { "16x1-bw-process,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_BW },
1731 { " 4x1-bw-thread,", "mem", "-p", "1", "-t", "4", "-T", "256", OPT_BW },
1732 { " 8x1-bw-thread,", "mem", "-p", "1", "-t", "8", "-T", "256", OPT_BW },
1733 { "16x1-bw-thread,", "mem", "-p", "1", "-t", "16", "-T", "128", OPT_BW },
1734 { "32x1-bw-thread,", "mem", "-p", "1", "-t", "32", "-T", "64", OPT_BW },
1736 { " 2x3-bw-thread,", "mem", "-p", "2", "-t", "3", "-P", "512", OPT_BW },
1737 { " 4x4-bw-thread,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_BW },
1738 { " 4x6-bw-thread,", "mem", "-p", "4", "-t", "6", "-P", "512", OPT_BW },
1739 { " 4x8-bw-thread,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW },
1740 { " 4x8-bw-thread-NOTHP,",
1741 "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW_NOTHP },
1742 { " 3x3-bw-thread,", "mem", "-p", "3", "-t", "3", "-P", "512", OPT_BW },
1743 { " 5x5-bw-thread,", "mem", "-p", "5", "-t", "5", "-P", "512", OPT_BW },
1745 { "2x16-bw-thread,", "mem", "-p", "2", "-t", "16", "-P", "512", OPT_BW },
1746 { "1x32-bw-thread,", "mem", "-p", "1", "-t", "32", "-P", "2048", OPT_BW },
1748 { "numa02-bw,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW },
1749 { "numa02-bw-NOTHP,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW_NOTHP },
1750 { "numa01-bw-thread,", "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW },
1751 { "numa01-bw-thread-NOTHP,",
1752 "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW_NOTHP },
1755 static int bench_all(void)
1757 int nr = ARRAY_SIZE(tests);
1761 ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1764 for (i = 0; i < nr; i++) {
1765 run_bench_numa(tests[i][0], tests[i] + 1);
1773 int bench_numa(int argc, const char **argv)
1775 init_params(&p0, "main,", argc, argv);
1776 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1783 if (__bench_numa(NULL))
1789 usage_with_options(numa_usage, options);