4 * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
8 /* For the CLR_() macros */
12 #include "../builtin.h"
13 #include "../util/util.h"
14 #include <subcmd/parse-options.h>
15 #include "../util/cloexec.h"
30 #include <sys/resource.h>
32 #include <sys/prctl.h>
33 #include <sys/types.h>
34 #include <linux/kernel.h>
35 #include <linux/time64.h>
41 * Regular printout to the terminal, supressed if -q is specified:
43 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
49 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
53 cpu_set_t bind_cpumask;
59 unsigned int loops_done;
65 pthread_mutex_t *process_lock;
68 /* Parameters set by options: */
71 /* Startup synchronization: */
72 bool serialize_startup;
78 /* Working set sizes: */
79 const char *mb_global_str;
80 const char *mb_proc_str;
81 const char *mb_proc_locked_str;
82 const char *mb_thread_str;
86 double mb_proc_locked;
89 /* Access patterns to the working set: */
93 bool data_zero_memset;
99 /* Working set initialization: */
111 long bytes_process_locked;
117 bool show_convergence;
118 bool measure_convergence;
124 /* Affinity options -C and -N: */
130 /* Global, read-writable area, accessible to all processes and threads: */
135 pthread_mutex_t startup_mutex;
136 int nr_tasks_started;
138 pthread_mutex_t startup_done_mutex;
140 pthread_mutex_t start_work_mutex;
141 int nr_tasks_working;
143 pthread_mutex_t stop_work_mutex;
146 struct thread_data *threads;
148 /* Convergence latency measurement: */
157 static struct global_info *g = NULL;
159 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
160 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
164 static const struct option options[] = {
165 OPT_INTEGER('p', "nr_proc" , &p0.nr_proc, "number of processes"),
166 OPT_INTEGER('t', "nr_threads" , &p0.nr_threads, "number of threads per process"),
168 OPT_STRING('G', "mb_global" , &p0.mb_global_str, "MB", "global memory (MBs)"),
169 OPT_STRING('P', "mb_proc" , &p0.mb_proc_str, "MB", "process memory (MBs)"),
170 OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
171 OPT_STRING('T', "mb_thread" , &p0.mb_thread_str, "MB", "thread memory (MBs)"),
173 OPT_UINTEGER('l', "nr_loops" , &p0.nr_loops, "max number of loops to run (default: unlimited)"),
174 OPT_UINTEGER('s', "nr_secs" , &p0.nr_secs, "max number of seconds to run (default: 5 secs)"),
175 OPT_UINTEGER('u', "usleep" , &p0.sleep_usecs, "usecs to sleep per loop iteration"),
177 OPT_BOOLEAN('R', "data_reads" , &p0.data_reads, "access the data via writes (can be mixed with -W)"),
178 OPT_BOOLEAN('W', "data_writes" , &p0.data_writes, "access the data via writes (can be mixed with -R)"),
179 OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards, "access the data backwards as well"),
180 OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
181 OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk, "access the data with random (32bit LFSR) walk"),
184 OPT_BOOLEAN('z', "init_zero" , &p0.init_zero, "bzero the initial allocations"),
185 OPT_BOOLEAN('I', "init_random" , &p0.init_random, "randomize the contents of the initial allocations"),
186 OPT_BOOLEAN('0', "init_cpu0" , &p0.init_cpu0, "do the initial allocations on CPU#0"),
187 OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs, "perturb thread 0/0 every X secs, to test convergence stability"),
189 OPT_INCR ('d', "show_details" , &p0.show_details, "Show details"),
190 OPT_INCR ('a', "all" , &p0.run_all, "Run all tests in the suite"),
191 OPT_INTEGER('H', "thp" , &p0.thp, "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
192 OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
193 "convergence is reached when each process (all its threads) is running on a single NUMA node."),
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>]",
218 static cpu_set_t bind_to_cpu(int target_cpu)
220 cpu_set_t orig_mask, mask;
223 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
228 if (target_cpu == -1) {
231 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
234 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
235 CPU_SET(target_cpu, &mask);
238 ret = sched_setaffinity(0, sizeof(mask), &mask);
244 static cpu_set_t bind_to_node(int target_node)
246 int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
247 cpu_set_t orig_mask, mask;
251 BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
252 BUG_ON(!cpus_per_node);
254 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
259 if (target_node == -1) {
260 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
263 int cpu_start = (target_node + 0) * cpus_per_node;
264 int cpu_stop = (target_node + 1) * cpus_per_node;
266 BUG_ON(cpu_stop > g->p.nr_cpus);
268 for (cpu = cpu_start; cpu < cpu_stop; cpu++)
272 ret = sched_setaffinity(0, sizeof(mask), &mask);
278 static void bind_to_cpumask(cpu_set_t mask)
282 ret = sched_setaffinity(0, sizeof(mask), &mask);
286 static void mempol_restore(void)
290 ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
295 static void bind_to_memnode(int node)
297 unsigned long nodemask;
303 BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
304 nodemask = 1L << node;
306 ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
307 dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
312 #define HPSIZE (2*1024*1024)
314 #define set_taskname(fmt...) \
318 snprintf(name, 20, fmt); \
319 prctl(PR_SET_NAME, name); \
322 static u8 *alloc_data(ssize_t bytes0, int map_flags,
323 int init_zero, int init_cpu0, int thp, int init_random)
333 /* Allocate and initialize all memory on CPU#0: */
335 orig_mask = bind_to_node(0);
339 bytes = bytes0 + HPSIZE;
341 buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
342 BUG_ON(buf == (void *)-1);
344 if (map_flags == MAP_PRIVATE) {
346 ret = madvise(buf, bytes, MADV_HUGEPAGE);
347 if (ret && !g->print_once) {
349 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
353 ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
354 if (ret && !g->print_once) {
356 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
364 /* Initialize random contents, different in each word: */
366 u64 *wbuf = (void *)buf;
370 for (i = 0; i < bytes/8; i++)
375 /* Align to 2MB boundary: */
376 buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
378 /* Restore affinity: */
380 bind_to_cpumask(orig_mask);
387 static void free_data(void *data, ssize_t bytes)
394 ret = munmap(data, bytes);
399 * Create a shared memory buffer that can be shared between processes, zeroed:
401 static void * zalloc_shared_data(ssize_t bytes)
403 return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0, g->p.thp, g->p.init_random);
407 * Create a shared memory buffer that can be shared between processes:
409 static void * setup_shared_data(ssize_t bytes)
411 return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
415 * Allocate process-local memory - this will either be shared between
416 * threads of this process, or only be accessed by this thread:
418 static void * setup_private_data(ssize_t bytes)
420 return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
424 * Return a process-shared (global) mutex:
426 static void init_global_mutex(pthread_mutex_t *mutex)
428 pthread_mutexattr_t attr;
430 pthread_mutexattr_init(&attr);
431 pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
432 pthread_mutex_init(mutex, &attr);
435 static int parse_cpu_list(const char *arg)
437 p0.cpu_list_str = strdup(arg);
439 dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
444 static int parse_setup_cpu_list(void)
446 struct thread_data *td;
450 if (!g->p.cpu_list_str)
453 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
455 str0 = str = strdup(g->p.cpu_list_str);
460 tprintf("# binding tasks to CPUs:\n");
464 int bind_cpu, bind_cpu_0, bind_cpu_1;
465 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
470 tok = strsep(&str, ",");
474 tok_end = strstr(tok, "-");
476 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
478 /* Single CPU specified: */
479 bind_cpu_0 = bind_cpu_1 = atol(tok);
481 /* CPU range specified (for example: "5-11"): */
482 bind_cpu_0 = atol(tok);
483 bind_cpu_1 = atol(tok_end + 1);
487 tok_step = strstr(tok, "#");
489 step = atol(tok_step + 1);
490 BUG_ON(step <= 0 || step >= g->p.nr_cpus);
495 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
496 * where the _4 means the next 4 CPUs are allowed.
499 tok_len = strstr(tok, "_");
501 bind_len = atol(tok_len + 1);
502 BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
505 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
507 tok_mul = strstr(tok, "x");
509 mul = atol(tok_mul + 1);
513 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
515 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
516 printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
520 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
521 BUG_ON(bind_cpu_0 > bind_cpu_1);
523 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
526 for (i = 0; i < mul; i++) {
529 if (t >= g->p.nr_tasks) {
530 printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
538 tprintf("%2d/%d", bind_cpu, bind_len);
540 tprintf("%2d", bind_cpu);
543 CPU_ZERO(&td->bind_cpumask);
544 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
545 BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
546 CPU_SET(cpu, &td->bind_cpumask);
556 if (t < g->p.nr_tasks)
557 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
563 static int parse_cpus_opt(const struct option *opt __maybe_unused,
564 const char *arg, int unset __maybe_unused)
569 return parse_cpu_list(arg);
572 static int parse_node_list(const char *arg)
574 p0.node_list_str = strdup(arg);
576 dprintf("got NODE list: {%s}\n", p0.node_list_str);
581 static int parse_setup_node_list(void)
583 struct thread_data *td;
587 if (!g->p.node_list_str)
590 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
592 str0 = str = strdup(g->p.node_list_str);
597 tprintf("# binding tasks to NODEs:\n");
601 int bind_node, bind_node_0, bind_node_1;
602 char *tok, *tok_end, *tok_step, *tok_mul;
606 tok = strsep(&str, ",");
610 tok_end = strstr(tok, "-");
612 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
614 /* Single NODE specified: */
615 bind_node_0 = bind_node_1 = atol(tok);
617 /* NODE range specified (for example: "5-11"): */
618 bind_node_0 = atol(tok);
619 bind_node_1 = atol(tok_end + 1);
623 tok_step = strstr(tok, "#");
625 step = atol(tok_step + 1);
626 BUG_ON(step <= 0 || step >= g->p.nr_nodes);
629 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
631 tok_mul = strstr(tok, "x");
633 mul = atol(tok_mul + 1);
637 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
639 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
640 printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
644 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
645 BUG_ON(bind_node_0 > bind_node_1);
647 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
650 for (i = 0; i < mul; i++) {
651 if (t >= g->p.nr_tasks) {
652 printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
658 tprintf(" %2d", bind_node);
660 tprintf(",%2d", bind_node);
662 td->bind_node = bind_node;
671 if (t < g->p.nr_tasks)
672 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
678 static int parse_nodes_opt(const struct option *opt __maybe_unused,
679 const char *arg, int unset __maybe_unused)
684 return parse_node_list(arg);
689 #define BIT(x) (1ul << x)
691 static inline uint32_t lfsr_32(uint32_t lfsr)
693 const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
694 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
698 * Make sure there's real data dependency to RAM (when read
699 * accesses are enabled), so the compiler, the CPU and the
700 * kernel (KSM, zero page, etc.) cannot optimize away RAM
703 static inline u64 access_data(u64 *data, u64 val)
707 if (g->p.data_writes)
713 * The worker process does two types of work, a forwards going
714 * loop and a backwards going loop.
716 * We do this so that on multiprocessor systems we do not create
717 * a 'train' of processing, with highly synchronized processes,
718 * skewing the whole benchmark.
720 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
722 long words = bytes/sizeof(u64);
723 u64 *data = (void *)__data;
724 long chunk_0, chunk_1;
729 BUG_ON(!data && words);
730 BUG_ON(data && !words);
735 /* Very simple memset() work variant: */
736 if (g->p.data_zero_memset && !g->p.data_rand_walk) {
741 /* Spread out by PID/TID nr and by loop nr: */
742 chunk_0 = words/nr_max;
743 chunk_1 = words/g->p.nr_loops;
744 off = nr*chunk_0 + loop*chunk_1;
749 if (g->p.data_rand_walk) {
750 u32 lfsr = nr + loop + val;
753 for (i = 0; i < words/1024; i++) {
756 lfsr = lfsr_32(lfsr);
758 start = lfsr % words;
759 end = min(start + 1024, words-1);
761 if (g->p.data_zero_memset) {
762 bzero(data + start, (end-start) * sizeof(u64));
764 for (j = start; j < end; j++)
765 val = access_data(data + j, val);
768 } else if (!g->p.data_backwards || (nr + loop) & 1) {
774 /* Process data forwards: */
776 if (unlikely(d >= d1))
778 if (unlikely(d == d0))
781 val = access_data(d, val);
786 /* Process data backwards: */
792 /* Process data forwards: */
794 if (unlikely(d < data))
796 if (unlikely(d == d0))
799 val = access_data(d, val);
808 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
812 cpu = sched_getcpu();
814 g->threads[task_nr].curr_cpu = cpu;
815 prctl(0, bytes_worked);
818 #define MAX_NR_NODES 64
821 * Count the number of nodes a process's threads
824 * A count of 1 means that the process is compressed
825 * to a single node. A count of g->p.nr_nodes means it's
826 * spread out on the whole system.
828 static int count_process_nodes(int process_nr)
830 char node_present[MAX_NR_NODES] = { 0, };
834 for (t = 0; t < g->p.nr_threads; t++) {
835 struct thread_data *td;
839 task_nr = process_nr*g->p.nr_threads + t;
840 td = g->threads + task_nr;
842 node = numa_node_of_cpu(td->curr_cpu);
843 if (node < 0) /* curr_cpu was likely still -1 */
846 node_present[node] = 1;
851 for (n = 0; n < MAX_NR_NODES; n++)
852 nodes += node_present[n];
858 * Count the number of distinct process-threads a node contains.
860 * A count of 1 means that the node contains only a single
861 * process. If all nodes on the system contain at most one
862 * process then we are well-converged.
864 static int count_node_processes(int node)
869 for (p = 0; p < g->p.nr_proc; p++) {
870 for (t = 0; t < g->p.nr_threads; t++) {
871 struct thread_data *td;
875 task_nr = p*g->p.nr_threads + t;
876 td = g->threads + task_nr;
878 n = numa_node_of_cpu(td->curr_cpu);
889 static void calc_convergence_compression(int *strong)
891 unsigned int nodes_min, nodes_max;
897 for (p = 0; p < g->p.nr_proc; p++) {
898 unsigned int nodes = count_process_nodes(p);
905 nodes_min = min(nodes, nodes_min);
906 nodes_max = max(nodes, nodes_max);
909 /* Strong convergence: all threads compress on a single node: */
910 if (nodes_min == 1 && nodes_max == 1) {
914 tprintf(" {%d-%d}", nodes_min, nodes_max);
918 static void calc_convergence(double runtime_ns_max, double *convergence)
920 unsigned int loops_done_min, loops_done_max;
922 int nodes[MAX_NR_NODES];
933 if (!g->p.show_convergence && !g->p.measure_convergence)
936 for (node = 0; node < g->p.nr_nodes; node++)
942 for (t = 0; t < g->p.nr_tasks; t++) {
943 struct thread_data *td = g->threads + t;
944 unsigned int loops_done;
948 /* Not all threads have written it yet: */
952 node = numa_node_of_cpu(cpu);
956 loops_done = td->loops_done;
957 loops_done_min = min(loops_done, loops_done_min);
958 loops_done_max = max(loops_done, loops_done_max);
962 nr_min = g->p.nr_tasks;
965 for (node = 0; node < g->p.nr_nodes; node++) {
967 nr_min = min(nr, nr_min);
968 nr_max = max(nr, nr_max);
971 BUG_ON(nr_min > nr_max);
973 BUG_ON(sum > g->p.nr_tasks);
975 if (0 && (sum < g->p.nr_tasks))
979 * Count the number of distinct process groups present
980 * on nodes - when we are converged this will decrease
985 for (node = 0; node < g->p.nr_nodes; node++) {
986 int processes = count_node_processes(node);
989 tprintf(" %2d/%-2d", nr, processes);
991 process_groups += processes;
994 distance = nr_max - nr_min;
996 tprintf(" [%2d/%-2d]", distance, process_groups);
998 tprintf(" l:%3d-%-3d (%3d)",
999 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1001 if (loops_done_min && loops_done_max) {
1002 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1004 tprintf(" [%4.1f%%]", skew * 100.0);
1007 calc_convergence_compression(&strong);
1009 if (strong && process_groups == g->p.nr_proc) {
1010 if (!*convergence) {
1011 *convergence = runtime_ns_max;
1012 tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1013 if (g->p.measure_convergence) {
1014 g->all_converged = true;
1015 g->stop_work = true;
1020 tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1027 static void show_summary(double runtime_ns_max, int l, double *convergence)
1029 tprintf("\r # %5.1f%% [%.1f mins]",
1030 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1032 calc_convergence(runtime_ns_max, convergence);
1034 if (g->p.show_details >= 0)
1038 static void *worker_thread(void *__tdata)
1040 struct thread_data *td = __tdata;
1041 struct timeval start0, start, stop, diff;
1042 int process_nr = td->process_nr;
1043 int thread_nr = td->thread_nr;
1044 unsigned long last_perturbance;
1045 int task_nr = td->task_nr;
1046 int details = g->p.show_details;
1047 int first_task, last_task;
1048 double convergence = 0;
1050 double runtime_ns_max;
1057 struct rusage rusage;
1059 bind_to_cpumask(td->bind_cpumask);
1060 bind_to_memnode(td->bind_node);
1062 set_taskname("thread %d/%d", process_nr, thread_nr);
1064 global_data = g->data;
1065 process_data = td->process_data;
1066 thread_data = setup_private_data(g->p.bytes_thread);
1071 if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1075 if (process_nr == 0 && thread_nr == 0)
1079 printf("# thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1080 process_nr, thread_nr, global_data, process_data, thread_data);
1083 if (g->p.serialize_startup) {
1084 pthread_mutex_lock(&g->startup_mutex);
1085 g->nr_tasks_started++;
1086 pthread_mutex_unlock(&g->startup_mutex);
1088 /* Here we will wait for the main process to start us all at once: */
1089 pthread_mutex_lock(&g->start_work_mutex);
1090 g->nr_tasks_working++;
1092 /* Last one wake the main process: */
1093 if (g->nr_tasks_working == g->p.nr_tasks)
1094 pthread_mutex_unlock(&g->startup_done_mutex);
1096 pthread_mutex_unlock(&g->start_work_mutex);
1099 gettimeofday(&start0, NULL);
1101 start = stop = start0;
1102 last_perturbance = start.tv_sec;
1104 for (l = 0; l < g->p.nr_loops; l++) {
1110 val += do_work(global_data, g->p.bytes_global, process_nr, g->p.nr_proc, l, val);
1111 val += do_work(process_data, g->p.bytes_process, thread_nr, g->p.nr_threads, l, val);
1112 val += do_work(thread_data, g->p.bytes_thread, 0, 1, l, val);
1114 if (g->p.sleep_usecs) {
1115 pthread_mutex_lock(td->process_lock);
1116 usleep(g->p.sleep_usecs);
1117 pthread_mutex_unlock(td->process_lock);
1120 * Amount of work to be done under a process-global lock:
1122 if (g->p.bytes_process_locked) {
1123 pthread_mutex_lock(td->process_lock);
1124 val += do_work(process_data, g->p.bytes_process_locked, thread_nr, g->p.nr_threads, l, val);
1125 pthread_mutex_unlock(td->process_lock);
1128 work_done = g->p.bytes_global + g->p.bytes_process +
1129 g->p.bytes_process_locked + g->p.bytes_thread;
1131 update_curr_cpu(task_nr, work_done);
1132 bytes_done += work_done;
1134 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1139 gettimeofday(&stop, NULL);
1141 /* Check whether our max runtime timed out: */
1143 timersub(&stop, &start0, &diff);
1144 if ((u32)diff.tv_sec >= g->p.nr_secs) {
1145 g->stop_work = true;
1150 /* Update the summary at most once per second: */
1151 if (start.tv_sec == stop.tv_sec)
1155 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1156 * by migrating to CPU#0:
1158 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1159 cpu_set_t orig_mask;
1163 last_perturbance = stop.tv_sec;
1166 * Depending on where we are running, move into
1167 * the other half of the system, to create some
1170 this_cpu = g->threads[task_nr].curr_cpu;
1171 if (this_cpu < g->p.nr_cpus/2)
1172 target_cpu = g->p.nr_cpus-1;
1176 orig_mask = bind_to_cpu(target_cpu);
1178 /* Here we are running on the target CPU already */
1180 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1182 bind_to_cpumask(orig_mask);
1186 timersub(&stop, &start, &diff);
1187 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1188 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1191 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1192 process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1199 timersub(&stop, &start0, &diff);
1200 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1201 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1203 show_summary(runtime_ns_max, l, &convergence);
1206 gettimeofday(&stop, NULL);
1207 timersub(&stop, &start0, &diff);
1208 td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1209 td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1210 td->speed_gbs = bytes_done / (td->runtime_ns / NSEC_PER_SEC) / 1e9;
1212 getrusage(RUSAGE_THREAD, &rusage);
1213 td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1214 td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1215 td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1216 td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1218 free_data(thread_data, g->p.bytes_thread);
1220 pthread_mutex_lock(&g->stop_work_mutex);
1221 g->bytes_done += bytes_done;
1222 pthread_mutex_unlock(&g->stop_work_mutex);
1228 * A worker process starts a couple of threads:
1230 static void worker_process(int process_nr)
1232 pthread_mutex_t process_lock;
1233 struct thread_data *td;
1234 pthread_t *pthreads;
1240 pthread_mutex_init(&process_lock, NULL);
1241 set_taskname("process %d", process_nr);
1244 * Pick up the memory policy and the CPU binding of our first thread,
1245 * so that we initialize memory accordingly:
1247 task_nr = process_nr*g->p.nr_threads;
1248 td = g->threads + task_nr;
1250 bind_to_memnode(td->bind_node);
1251 bind_to_cpumask(td->bind_cpumask);
1253 pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1254 process_data = setup_private_data(g->p.bytes_process);
1256 if (g->p.show_details >= 3) {
1257 printf(" # process %2d global mem: %p, process mem: %p\n",
1258 process_nr, g->data, process_data);
1261 for (t = 0; t < g->p.nr_threads; t++) {
1262 task_nr = process_nr*g->p.nr_threads + t;
1263 td = g->threads + task_nr;
1265 td->process_data = process_data;
1266 td->process_nr = process_nr;
1268 td->task_nr = task_nr;
1271 td->process_lock = &process_lock;
1273 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1277 for (t = 0; t < g->p.nr_threads; t++) {
1278 ret = pthread_join(pthreads[t], NULL);
1282 free_data(process_data, g->p.bytes_process);
1286 static void print_summary(void)
1288 if (g->p.show_details < 0)
1292 printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1293 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
1294 printf(" # %5dx %5ldMB global shared mem operations\n",
1295 g->p.nr_loops, g->p.bytes_global/1024/1024);
1296 printf(" # %5dx %5ldMB process shared mem operations\n",
1297 g->p.nr_loops, g->p.bytes_process/1024/1024);
1298 printf(" # %5dx %5ldMB thread local mem operations\n",
1299 g->p.nr_loops, g->p.bytes_thread/1024/1024);
1303 printf("\n ###\n"); fflush(stdout);
1306 static void init_thread_data(void)
1308 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1311 g->threads = zalloc_shared_data(size);
1313 for (t = 0; t < g->p.nr_tasks; t++) {
1314 struct thread_data *td = g->threads + t;
1317 /* Allow all nodes by default: */
1320 /* Allow all CPUs by default: */
1321 CPU_ZERO(&td->bind_cpumask);
1322 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1323 CPU_SET(cpu, &td->bind_cpumask);
1327 static void deinit_thread_data(void)
1329 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1331 free_data(g->threads, size);
1334 static int init(void)
1336 g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1338 /* Copy over options: */
1341 g->p.nr_cpus = numa_num_configured_cpus();
1343 g->p.nr_nodes = numa_max_node() + 1;
1345 /* char array in count_process_nodes(): */
1346 BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1348 if (g->p.show_quiet && !g->p.show_details)
1349 g->p.show_details = -1;
1351 /* Some memory should be specified: */
1352 if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1355 if (g->p.mb_global_str) {
1356 g->p.mb_global = atof(g->p.mb_global_str);
1357 BUG_ON(g->p.mb_global < 0);
1360 if (g->p.mb_proc_str) {
1361 g->p.mb_proc = atof(g->p.mb_proc_str);
1362 BUG_ON(g->p.mb_proc < 0);
1365 if (g->p.mb_proc_locked_str) {
1366 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1367 BUG_ON(g->p.mb_proc_locked < 0);
1368 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1371 if (g->p.mb_thread_str) {
1372 g->p.mb_thread = atof(g->p.mb_thread_str);
1373 BUG_ON(g->p.mb_thread < 0);
1376 BUG_ON(g->p.nr_threads <= 0);
1377 BUG_ON(g->p.nr_proc <= 0);
1379 g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1381 g->p.bytes_global = g->p.mb_global *1024L*1024L;
1382 g->p.bytes_process = g->p.mb_proc *1024L*1024L;
1383 g->p.bytes_process_locked = g->p.mb_proc_locked *1024L*1024L;
1384 g->p.bytes_thread = g->p.mb_thread *1024L*1024L;
1386 g->data = setup_shared_data(g->p.bytes_global);
1388 /* Startup serialization: */
1389 init_global_mutex(&g->start_work_mutex);
1390 init_global_mutex(&g->startup_mutex);
1391 init_global_mutex(&g->startup_done_mutex);
1392 init_global_mutex(&g->stop_work_mutex);
1397 if (parse_setup_cpu_list() || parse_setup_node_list())
1406 static void deinit(void)
1408 free_data(g->data, g->p.bytes_global);
1411 deinit_thread_data();
1413 free_data(g, sizeof(*g));
1418 * Print a short or long result, depending on the verbosity setting:
1420 static void print_res(const char *name, double val,
1421 const char *txt_unit, const char *txt_short, const char *txt_long)
1426 if (!g->p.show_quiet)
1427 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1429 printf(" %14.3f %s\n", val, txt_long);
1432 static int __bench_numa(const char *name)
1434 struct timeval start, stop, diff;
1435 u64 runtime_ns_min, runtime_ns_sum;
1436 pid_t *pids, pid, wpid;
1437 double delta_runtime;
1439 double runtime_sec_max;
1440 double runtime_sec_min;
1448 pids = zalloc(g->p.nr_proc * sizeof(*pids));
1451 /* All threads try to acquire it, this way we can wait for them to start up: */
1452 pthread_mutex_lock(&g->start_work_mutex);
1454 if (g->p.serialize_startup) {
1456 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1459 gettimeofday(&start, NULL);
1461 for (i = 0; i < g->p.nr_proc; i++) {
1463 dprintf(" # process %2d: PID %d\n", i, pid);
1467 /* Child process: */
1475 /* Wait for all the threads to start up: */
1476 while (g->nr_tasks_started != g->p.nr_tasks)
1477 usleep(USEC_PER_MSEC);
1479 BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1481 if (g->p.serialize_startup) {
1484 pthread_mutex_lock(&g->startup_done_mutex);
1486 /* This will start all threads: */
1487 pthread_mutex_unlock(&g->start_work_mutex);
1489 /* This mutex is locked - the last started thread will wake us: */
1490 pthread_mutex_lock(&g->startup_done_mutex);
1492 gettimeofday(&stop, NULL);
1494 timersub(&stop, &start, &diff);
1496 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1497 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1498 startup_sec /= NSEC_PER_SEC;
1500 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1504 pthread_mutex_unlock(&g->startup_done_mutex);
1506 gettimeofday(&start, NULL);
1509 /* Parent process: */
1512 for (i = 0; i < g->p.nr_proc; i++) {
1513 wpid = waitpid(pids[i], &wait_stat, 0);
1515 BUG_ON(!WIFEXITED(wait_stat));
1520 runtime_ns_min = -1LL;
1522 for (t = 0; t < g->p.nr_tasks; t++) {
1523 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1525 runtime_ns_sum += thread_runtime_ns;
1526 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1529 gettimeofday(&stop, NULL);
1530 timersub(&stop, &start, &diff);
1532 BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1534 tprintf("\n ###\n");
1537 runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1538 runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1539 runtime_sec_max /= NSEC_PER_SEC;
1541 runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1543 bytes = g->bytes_done;
1544 runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1546 if (g->p.measure_convergence) {
1547 print_res(name, runtime_sec_max,
1548 "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1551 print_res(name, runtime_sec_max,
1552 "secs,", "runtime-max/thread", "secs slowest (max) thread-runtime");
1554 print_res(name, runtime_sec_min,
1555 "secs,", "runtime-min/thread", "secs fastest (min) thread-runtime");
1557 print_res(name, runtime_avg,
1558 "secs,", "runtime-avg/thread", "secs average thread-runtime");
1560 delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1561 print_res(name, delta_runtime / runtime_sec_max * 100.0,
1562 "%,", "spread-runtime/thread", "% difference between max/avg runtime");
1564 print_res(name, bytes / g->p.nr_tasks / 1e9,
1565 "GB,", "data/thread", "GB data processed, per thread");
1567 print_res(name, bytes / 1e9,
1568 "GB,", "data-total", "GB data processed, total");
1570 print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1571 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1573 print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1574 "GB/sec,", "thread-speed", "GB/sec/thread speed");
1576 print_res(name, bytes / runtime_sec_max / 1e9,
1577 "GB/sec,", "total-speed", "GB/sec total speed");
1579 if (g->p.show_details >= 2) {
1580 char tname[14 + 2 * 10 + 1];
1581 struct thread_data *td;
1582 for (p = 0; p < g->p.nr_proc; p++) {
1583 for (t = 0; t < g->p.nr_threads; t++) {
1584 memset(tname, 0, sizeof(tname));
1585 td = g->threads + p*g->p.nr_threads + t;
1586 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1587 print_res(tname, td->speed_gbs,
1588 "GB/sec", "thread-speed", "GB/sec/thread speed");
1589 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1590 "secs", "thread-system-time", "system CPU time/thread");
1591 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1592 "secs", "thread-user-time", "user CPU time/thread");
1606 static int command_size(const char **argv)
1615 BUG_ON(size >= MAX_ARGS);
1620 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1624 printf("\n # Running %s \"perf bench numa", name);
1626 for (i = 0; i < argc; i++)
1627 printf(" %s", argv[i]);
1631 memset(p, 0, sizeof(*p));
1633 /* Initialize nonzero defaults: */
1635 p->serialize_startup = 1;
1636 p->data_reads = true;
1637 p->data_writes = true;
1638 p->data_backwards = true;
1639 p->data_rand_walk = true;
1641 p->init_random = true;
1642 p->mb_global_str = "1";
1646 p->run_all = argc == 1;
1649 static int run_bench_numa(const char *name, const char **argv)
1651 int argc = command_size(argv);
1653 init_params(&p0, name, argc, argv);
1654 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1658 if (__bench_numa(name))
1667 #define OPT_BW_RAM "-s", "20", "-zZq", "--thp", " 1", "--no-data_rand_walk"
1668 #define OPT_BW_RAM_NOTHP OPT_BW_RAM, "--thp", "-1"
1670 #define OPT_CONV "-s", "100", "-zZ0qcm", "--thp", " 1"
1671 #define OPT_CONV_NOTHP OPT_CONV, "--thp", "-1"
1673 #define OPT_BW "-s", "20", "-zZ0q", "--thp", " 1"
1674 #define OPT_BW_NOTHP OPT_BW, "--thp", "-1"
1677 * The built-in test-suite executed by "perf bench numa -a".
1679 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1681 static const char *tests[][MAX_ARGS] = {
1682 /* Basic single-stream NUMA bandwidth measurements: */
1683 { "RAM-bw-local,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1684 "-C" , "0", "-M", "0", OPT_BW_RAM },
1685 { "RAM-bw-local-NOTHP,",
1686 "mem", "-p", "1", "-t", "1", "-P", "1024",
1687 "-C" , "0", "-M", "0", OPT_BW_RAM_NOTHP },
1688 { "RAM-bw-remote,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1689 "-C" , "0", "-M", "1", OPT_BW_RAM },
1691 /* 2-stream NUMA bandwidth measurements: */
1692 { "RAM-bw-local-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1693 "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1694 { "RAM-bw-remote-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1695 "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1697 /* Cross-stream NUMA bandwidth measurement: */
1698 { "RAM-bw-cross,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1699 "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1701 /* Convergence latency measurements: */
1702 { " 1x3-convergence,", "mem", "-p", "1", "-t", "3", "-P", "512", OPT_CONV },
1703 { " 1x4-convergence,", "mem", "-p", "1", "-t", "4", "-P", "512", OPT_CONV },
1704 { " 1x6-convergence,", "mem", "-p", "1", "-t", "6", "-P", "1020", OPT_CONV },
1705 { " 2x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1706 { " 3x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1707 { " 4x4-convergence,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV },
1708 { " 4x4-convergence-NOTHP,",
1709 "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1710 { " 4x6-convergence,", "mem", "-p", "4", "-t", "6", "-P", "1020", OPT_CONV },
1711 { " 4x8-convergence,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_CONV },
1712 { " 8x4-convergence,", "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV },
1713 { " 8x4-convergence-NOTHP,",
1714 "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1715 { " 3x1-convergence,", "mem", "-p", "3", "-t", "1", "-P", "512", OPT_CONV },
1716 { " 4x1-convergence,", "mem", "-p", "4", "-t", "1", "-P", "512", OPT_CONV },
1717 { " 8x1-convergence,", "mem", "-p", "8", "-t", "1", "-P", "512", OPT_CONV },
1718 { "16x1-convergence,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_CONV },
1719 { "32x1-convergence,", "mem", "-p", "32", "-t", "1", "-P", "128", OPT_CONV },
1721 /* Various NUMA process/thread layout bandwidth measurements: */
1722 { " 2x1-bw-process,", "mem", "-p", "2", "-t", "1", "-P", "1024", OPT_BW },
1723 { " 3x1-bw-process,", "mem", "-p", "3", "-t", "1", "-P", "1024", OPT_BW },
1724 { " 4x1-bw-process,", "mem", "-p", "4", "-t", "1", "-P", "1024", OPT_BW },
1725 { " 8x1-bw-process,", "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW },
1726 { " 8x1-bw-process-NOTHP,",
1727 "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW_NOTHP },
1728 { "16x1-bw-process,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_BW },
1730 { " 4x1-bw-thread,", "mem", "-p", "1", "-t", "4", "-T", "256", OPT_BW },
1731 { " 8x1-bw-thread,", "mem", "-p", "1", "-t", "8", "-T", "256", OPT_BW },
1732 { "16x1-bw-thread,", "mem", "-p", "1", "-t", "16", "-T", "128", OPT_BW },
1733 { "32x1-bw-thread,", "mem", "-p", "1", "-t", "32", "-T", "64", OPT_BW },
1735 { " 2x3-bw-thread,", "mem", "-p", "2", "-t", "3", "-P", "512", OPT_BW },
1736 { " 4x4-bw-thread,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_BW },
1737 { " 4x6-bw-thread,", "mem", "-p", "4", "-t", "6", "-P", "512", OPT_BW },
1738 { " 4x8-bw-thread,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW },
1739 { " 4x8-bw-thread-NOTHP,",
1740 "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW_NOTHP },
1741 { " 3x3-bw-thread,", "mem", "-p", "3", "-t", "3", "-P", "512", OPT_BW },
1742 { " 5x5-bw-thread,", "mem", "-p", "5", "-t", "5", "-P", "512", OPT_BW },
1744 { "2x16-bw-thread,", "mem", "-p", "2", "-t", "16", "-P", "512", OPT_BW },
1745 { "1x32-bw-thread,", "mem", "-p", "1", "-t", "32", "-P", "2048", OPT_BW },
1747 { "numa02-bw,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW },
1748 { "numa02-bw-NOTHP,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW_NOTHP },
1749 { "numa01-bw-thread,", "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW },
1750 { "numa01-bw-thread-NOTHP,",
1751 "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW_NOTHP },
1754 static int bench_all(void)
1756 int nr = ARRAY_SIZE(tests);
1760 ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1763 for (i = 0; i < nr; i++) {
1764 run_bench_numa(tests[i][0], tests[i] + 1);
1772 int bench_numa(int argc, const char **argv)
1774 init_params(&p0, "main,", argc, argv);
1775 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1782 if (__bench_numa(NULL))
1788 usage_with_options(numa_usage, options);