1 // SPDX-License-Identifier: GPL-2.0
5 * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
9 /* For the CLR_() macros */
12 #include <subcmd/parse-options.h>
13 #include "../util/cloexec.h"
28 #include <sys/resource.h>
30 #include <sys/prctl.h>
31 #include <sys/types.h>
32 #include <linux/kernel.h>
33 #include <linux/time64.h>
34 #include <linux/numa.h>
35 #include <linux/zalloc.h>
41 # define RUSAGE_THREAD 1
45 * Regular printout to the terminal, supressed if -q is specified:
47 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
53 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
57 cpu_set_t bind_cpumask;
63 unsigned int loops_done;
69 pthread_mutex_t *process_lock;
72 /* Parameters set by options: */
75 /* Startup synchronization: */
76 bool serialize_startup;
82 /* Working set sizes: */
83 const char *mb_global_str;
84 const char *mb_proc_str;
85 const char *mb_proc_locked_str;
86 const char *mb_thread_str;
90 double mb_proc_locked;
93 /* Access patterns to the working set: */
97 bool data_zero_memset;
103 /* Working set initialization: */
115 long bytes_process_locked;
121 bool show_convergence;
122 bool measure_convergence;
128 /* Affinity options -C and -N: */
134 /* Global, read-writable area, accessible to all processes and threads: */
139 pthread_mutex_t startup_mutex;
140 int nr_tasks_started;
142 pthread_mutex_t startup_done_mutex;
144 pthread_mutex_t start_work_mutex;
145 int nr_tasks_working;
147 pthread_mutex_t stop_work_mutex;
150 struct thread_data *threads;
152 /* Convergence latency measurement: */
161 static struct global_info *g = NULL;
163 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
164 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
168 static const struct option options[] = {
169 OPT_INTEGER('p', "nr_proc" , &p0.nr_proc, "number of processes"),
170 OPT_INTEGER('t', "nr_threads" , &p0.nr_threads, "number of threads per process"),
172 OPT_STRING('G', "mb_global" , &p0.mb_global_str, "MB", "global memory (MBs)"),
173 OPT_STRING('P', "mb_proc" , &p0.mb_proc_str, "MB", "process memory (MBs)"),
174 OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
175 OPT_STRING('T', "mb_thread" , &p0.mb_thread_str, "MB", "thread memory (MBs)"),
177 OPT_UINTEGER('l', "nr_loops" , &p0.nr_loops, "max number of loops to run (default: unlimited)"),
178 OPT_UINTEGER('s', "nr_secs" , &p0.nr_secs, "max number of seconds to run (default: 5 secs)"),
179 OPT_UINTEGER('u', "usleep" , &p0.sleep_usecs, "usecs to sleep per loop iteration"),
181 OPT_BOOLEAN('R', "data_reads" , &p0.data_reads, "access the data via reads (can be mixed with -W)"),
182 OPT_BOOLEAN('W', "data_writes" , &p0.data_writes, "access the data via writes (can be mixed with -R)"),
183 OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards, "access the data backwards as well"),
184 OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
185 OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk, "access the data with random (32bit LFSR) walk"),
188 OPT_BOOLEAN('z', "init_zero" , &p0.init_zero, "bzero the initial allocations"),
189 OPT_BOOLEAN('I', "init_random" , &p0.init_random, "randomize the contents of the initial allocations"),
190 OPT_BOOLEAN('0', "init_cpu0" , &p0.init_cpu0, "do the initial allocations on CPU#0"),
191 OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs, "perturb thread 0/0 every X secs, to test convergence stability"),
193 OPT_INCR ('d', "show_details" , &p0.show_details, "Show details"),
194 OPT_INCR ('a', "all" , &p0.run_all, "Run all tests in the suite"),
195 OPT_INTEGER('H', "thp" , &p0.thp, "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
196 OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
197 "convergence is reached when each process (all its threads) is running on a single NUMA node."),
198 OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
199 OPT_BOOLEAN('q', "quiet" , &p0.show_quiet, "quiet mode"),
200 OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
202 /* Special option string parsing callbacks: */
203 OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
204 "bind the first N tasks to these specific cpus (the rest is unbound)",
206 OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
207 "bind the first N tasks to these specific memory nodes (the rest is unbound)",
212 static const char * const bench_numa_usage[] = {
213 "perf bench numa <options>",
217 static const char * const numa_usage[] = {
218 "perf bench numa mem [<options>]",
223 * To get number of numa nodes present.
225 static int nr_numa_nodes(void)
229 for (i = 0; i < g->p.nr_nodes; i++) {
230 if (numa_bitmask_isbitset(numa_nodes_ptr, i))
238 * To check if given numa node is present.
240 static int is_node_present(int node)
242 return numa_bitmask_isbitset(numa_nodes_ptr, node);
246 * To check given numa node has cpus.
248 static bool node_has_cpus(int node)
250 struct bitmask *cpu = numa_allocate_cpumask();
253 if (cpu && !numa_node_to_cpus(node, cpu)) {
254 for (i = 0; i < cpu->size; i++) {
255 if (numa_bitmask_isbitset(cpu, i))
260 return false; /* lets fall back to nocpus safely */
263 static cpu_set_t bind_to_cpu(int target_cpu)
265 cpu_set_t orig_mask, mask;
268 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
273 if (target_cpu == -1) {
276 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
279 BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
280 CPU_SET(target_cpu, &mask);
283 ret = sched_setaffinity(0, sizeof(mask), &mask);
289 static cpu_set_t bind_to_node(int target_node)
291 int cpus_per_node = g->p.nr_cpus / nr_numa_nodes();
292 cpu_set_t orig_mask, mask;
296 BUG_ON(cpus_per_node * nr_numa_nodes() != g->p.nr_cpus);
297 BUG_ON(!cpus_per_node);
299 ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
304 if (target_node == NUMA_NO_NODE) {
305 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
308 int cpu_start = (target_node + 0) * cpus_per_node;
309 int cpu_stop = (target_node + 1) * cpus_per_node;
311 BUG_ON(cpu_stop > g->p.nr_cpus);
313 for (cpu = cpu_start; cpu < cpu_stop; cpu++)
317 ret = sched_setaffinity(0, sizeof(mask), &mask);
323 static void bind_to_cpumask(cpu_set_t mask)
327 ret = sched_setaffinity(0, sizeof(mask), &mask);
331 static void mempol_restore(void)
335 ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
340 static void bind_to_memnode(int node)
342 unsigned long nodemask;
345 if (node == NUMA_NO_NODE)
348 BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask)*8);
349 nodemask = 1L << node;
351 ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
352 dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
357 #define HPSIZE (2*1024*1024)
359 #define set_taskname(fmt...) \
363 snprintf(name, 20, fmt); \
364 prctl(PR_SET_NAME, name); \
367 static u8 *alloc_data(ssize_t bytes0, int map_flags,
368 int init_zero, int init_cpu0, int thp, int init_random)
378 /* Allocate and initialize all memory on CPU#0: */
380 int node = numa_node_of_cpu(0);
382 orig_mask = bind_to_node(node);
383 bind_to_memnode(node);
386 bytes = bytes0 + HPSIZE;
388 buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
389 BUG_ON(buf == (void *)-1);
391 if (map_flags == MAP_PRIVATE) {
393 ret = madvise(buf, bytes, MADV_HUGEPAGE);
394 if (ret && !g->print_once) {
396 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
400 ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
401 if (ret && !g->print_once) {
403 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
411 /* Initialize random contents, different in each word: */
413 u64 *wbuf = (void *)buf;
417 for (i = 0; i < bytes/8; i++)
422 /* Align to 2MB boundary: */
423 buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
425 /* Restore affinity: */
427 bind_to_cpumask(orig_mask);
434 static void free_data(void *data, ssize_t bytes)
441 ret = munmap(data, bytes);
446 * Create a shared memory buffer that can be shared between processes, zeroed:
448 static void * zalloc_shared_data(ssize_t bytes)
450 return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0, g->p.thp, g->p.init_random);
454 * Create a shared memory buffer that can be shared between processes:
456 static void * setup_shared_data(ssize_t bytes)
458 return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
462 * Allocate process-local memory - this will either be shared between
463 * threads of this process, or only be accessed by this thread:
465 static void * setup_private_data(ssize_t bytes)
467 return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
471 * Return a process-shared (global) mutex:
473 static void init_global_mutex(pthread_mutex_t *mutex)
475 pthread_mutexattr_t attr;
477 pthread_mutexattr_init(&attr);
478 pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
479 pthread_mutex_init(mutex, &attr);
482 static int parse_cpu_list(const char *arg)
484 p0.cpu_list_str = strdup(arg);
486 dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
491 static int parse_setup_cpu_list(void)
493 struct thread_data *td;
497 if (!g->p.cpu_list_str)
500 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
502 str0 = str = strdup(g->p.cpu_list_str);
507 tprintf("# binding tasks to CPUs:\n");
511 int bind_cpu, bind_cpu_0, bind_cpu_1;
512 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
517 tok = strsep(&str, ",");
521 tok_end = strstr(tok, "-");
523 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
525 /* Single CPU specified: */
526 bind_cpu_0 = bind_cpu_1 = atol(tok);
528 /* CPU range specified (for example: "5-11"): */
529 bind_cpu_0 = atol(tok);
530 bind_cpu_1 = atol(tok_end + 1);
534 tok_step = strstr(tok, "#");
536 step = atol(tok_step + 1);
537 BUG_ON(step <= 0 || step >= g->p.nr_cpus);
542 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
543 * where the _4 means the next 4 CPUs are allowed.
546 tok_len = strstr(tok, "_");
548 bind_len = atol(tok_len + 1);
549 BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
552 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
554 tok_mul = strstr(tok, "x");
556 mul = atol(tok_mul + 1);
560 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
562 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
563 printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
567 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
568 BUG_ON(bind_cpu_0 > bind_cpu_1);
570 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
573 for (i = 0; i < mul; i++) {
576 if (t >= g->p.nr_tasks) {
577 printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
585 tprintf("%2d/%d", bind_cpu, bind_len);
587 tprintf("%2d", bind_cpu);
590 CPU_ZERO(&td->bind_cpumask);
591 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
592 BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
593 CPU_SET(cpu, &td->bind_cpumask);
603 if (t < g->p.nr_tasks)
604 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
610 static int parse_cpus_opt(const struct option *opt __maybe_unused,
611 const char *arg, int unset __maybe_unused)
616 return parse_cpu_list(arg);
619 static int parse_node_list(const char *arg)
621 p0.node_list_str = strdup(arg);
623 dprintf("got NODE list: {%s}\n", p0.node_list_str);
628 static int parse_setup_node_list(void)
630 struct thread_data *td;
634 if (!g->p.node_list_str)
637 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
639 str0 = str = strdup(g->p.node_list_str);
644 tprintf("# binding tasks to NODEs:\n");
648 int bind_node, bind_node_0, bind_node_1;
649 char *tok, *tok_end, *tok_step, *tok_mul;
653 tok = strsep(&str, ",");
657 tok_end = strstr(tok, "-");
659 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
661 /* Single NODE specified: */
662 bind_node_0 = bind_node_1 = atol(tok);
664 /* NODE range specified (for example: "5-11"): */
665 bind_node_0 = atol(tok);
666 bind_node_1 = atol(tok_end + 1);
670 tok_step = strstr(tok, "#");
672 step = atol(tok_step + 1);
673 BUG_ON(step <= 0 || step >= g->p.nr_nodes);
676 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
678 tok_mul = strstr(tok, "x");
680 mul = atol(tok_mul + 1);
684 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
686 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
687 printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
691 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
692 BUG_ON(bind_node_0 > bind_node_1);
694 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
697 for (i = 0; i < mul; i++) {
698 if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
699 printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
705 tprintf(" %2d", bind_node);
707 tprintf(",%2d", bind_node);
709 td->bind_node = bind_node;
718 if (t < g->p.nr_tasks)
719 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
725 static int parse_nodes_opt(const struct option *opt __maybe_unused,
726 const char *arg, int unset __maybe_unused)
731 return parse_node_list(arg);
736 #define BIT(x) (1ul << x)
738 static inline uint32_t lfsr_32(uint32_t lfsr)
740 const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
741 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
745 * Make sure there's real data dependency to RAM (when read
746 * accesses are enabled), so the compiler, the CPU and the
747 * kernel (KSM, zero page, etc.) cannot optimize away RAM
750 static inline u64 access_data(u64 *data, u64 val)
754 if (g->p.data_writes)
760 * The worker process does two types of work, a forwards going
761 * loop and a backwards going loop.
763 * We do this so that on multiprocessor systems we do not create
764 * a 'train' of processing, with highly synchronized processes,
765 * skewing the whole benchmark.
767 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
769 long words = bytes/sizeof(u64);
770 u64 *data = (void *)__data;
771 long chunk_0, chunk_1;
776 BUG_ON(!data && words);
777 BUG_ON(data && !words);
782 /* Very simple memset() work variant: */
783 if (g->p.data_zero_memset && !g->p.data_rand_walk) {
788 /* Spread out by PID/TID nr and by loop nr: */
789 chunk_0 = words/nr_max;
790 chunk_1 = words/g->p.nr_loops;
791 off = nr*chunk_0 + loop*chunk_1;
796 if (g->p.data_rand_walk) {
797 u32 lfsr = nr + loop + val;
800 for (i = 0; i < words/1024; i++) {
803 lfsr = lfsr_32(lfsr);
805 start = lfsr % words;
806 end = min(start + 1024, words-1);
808 if (g->p.data_zero_memset) {
809 bzero(data + start, (end-start) * sizeof(u64));
811 for (j = start; j < end; j++)
812 val = access_data(data + j, val);
815 } else if (!g->p.data_backwards || (nr + loop) & 1) {
821 /* Process data forwards: */
823 if (unlikely(d >= d1))
825 if (unlikely(d == d0))
828 val = access_data(d, val);
833 /* Process data backwards: */
839 /* Process data forwards: */
841 if (unlikely(d < data))
843 if (unlikely(d == d0))
846 val = access_data(d, val);
855 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
859 cpu = sched_getcpu();
861 g->threads[task_nr].curr_cpu = cpu;
862 prctl(0, bytes_worked);
865 #define MAX_NR_NODES 64
868 * Count the number of nodes a process's threads
871 * A count of 1 means that the process is compressed
872 * to a single node. A count of g->p.nr_nodes means it's
873 * spread out on the whole system.
875 static int count_process_nodes(int process_nr)
877 char node_present[MAX_NR_NODES] = { 0, };
881 for (t = 0; t < g->p.nr_threads; t++) {
882 struct thread_data *td;
886 task_nr = process_nr*g->p.nr_threads + t;
887 td = g->threads + task_nr;
889 node = numa_node_of_cpu(td->curr_cpu);
890 if (node < 0) /* curr_cpu was likely still -1 */
893 node_present[node] = 1;
898 for (n = 0; n < MAX_NR_NODES; n++)
899 nodes += node_present[n];
905 * Count the number of distinct process-threads a node contains.
907 * A count of 1 means that the node contains only a single
908 * process. If all nodes on the system contain at most one
909 * process then we are well-converged.
911 static int count_node_processes(int node)
916 for (p = 0; p < g->p.nr_proc; p++) {
917 for (t = 0; t < g->p.nr_threads; t++) {
918 struct thread_data *td;
922 task_nr = p*g->p.nr_threads + t;
923 td = g->threads + task_nr;
925 n = numa_node_of_cpu(td->curr_cpu);
936 static void calc_convergence_compression(int *strong)
938 unsigned int nodes_min, nodes_max;
944 for (p = 0; p < g->p.nr_proc; p++) {
945 unsigned int nodes = count_process_nodes(p);
952 nodes_min = min(nodes, nodes_min);
953 nodes_max = max(nodes, nodes_max);
956 /* Strong convergence: all threads compress on a single node: */
957 if (nodes_min == 1 && nodes_max == 1) {
961 tprintf(" {%d-%d}", nodes_min, nodes_max);
965 static void calc_convergence(double runtime_ns_max, double *convergence)
967 unsigned int loops_done_min, loops_done_max;
969 int nodes[MAX_NR_NODES];
980 if (!g->p.show_convergence && !g->p.measure_convergence)
983 for (node = 0; node < g->p.nr_nodes; node++)
989 for (t = 0; t < g->p.nr_tasks; t++) {
990 struct thread_data *td = g->threads + t;
991 unsigned int loops_done;
995 /* Not all threads have written it yet: */
999 node = numa_node_of_cpu(cpu);
1003 loops_done = td->loops_done;
1004 loops_done_min = min(loops_done, loops_done_min);
1005 loops_done_max = max(loops_done, loops_done_max);
1009 nr_min = g->p.nr_tasks;
1012 for (node = 0; node < g->p.nr_nodes; node++) {
1013 if (!is_node_present(node))
1016 nr_min = min(nr, nr_min);
1017 nr_max = max(nr, nr_max);
1020 BUG_ON(nr_min > nr_max);
1022 BUG_ON(sum > g->p.nr_tasks);
1024 if (0 && (sum < g->p.nr_tasks))
1028 * Count the number of distinct process groups present
1029 * on nodes - when we are converged this will decrease
1034 for (node = 0; node < g->p.nr_nodes; node++) {
1037 if (!is_node_present(node))
1039 processes = count_node_processes(node);
1041 tprintf(" %2d/%-2d", nr, processes);
1043 process_groups += processes;
1046 distance = nr_max - nr_min;
1048 tprintf(" [%2d/%-2d]", distance, process_groups);
1050 tprintf(" l:%3d-%-3d (%3d)",
1051 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1053 if (loops_done_min && loops_done_max) {
1054 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1056 tprintf(" [%4.1f%%]", skew * 100.0);
1059 calc_convergence_compression(&strong);
1061 if (strong && process_groups == g->p.nr_proc) {
1062 if (!*convergence) {
1063 *convergence = runtime_ns_max;
1064 tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1065 if (g->p.measure_convergence) {
1066 g->all_converged = true;
1067 g->stop_work = true;
1072 tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1079 static void show_summary(double runtime_ns_max, int l, double *convergence)
1081 tprintf("\r # %5.1f%% [%.1f mins]",
1082 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1084 calc_convergence(runtime_ns_max, convergence);
1086 if (g->p.show_details >= 0)
1090 static void *worker_thread(void *__tdata)
1092 struct thread_data *td = __tdata;
1093 struct timeval start0, start, stop, diff;
1094 int process_nr = td->process_nr;
1095 int thread_nr = td->thread_nr;
1096 unsigned long last_perturbance;
1097 int task_nr = td->task_nr;
1098 int details = g->p.show_details;
1099 int first_task, last_task;
1100 double convergence = 0;
1102 double runtime_ns_max;
1106 u64 bytes_done, secs;
1109 struct rusage rusage;
1111 bind_to_cpumask(td->bind_cpumask);
1112 bind_to_memnode(td->bind_node);
1114 set_taskname("thread %d/%d", process_nr, thread_nr);
1116 global_data = g->data;
1117 process_data = td->process_data;
1118 thread_data = setup_private_data(g->p.bytes_thread);
1123 if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1127 if (process_nr == 0 && thread_nr == 0)
1131 printf("# thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1132 process_nr, thread_nr, global_data, process_data, thread_data);
1135 if (g->p.serialize_startup) {
1136 pthread_mutex_lock(&g->startup_mutex);
1137 g->nr_tasks_started++;
1138 pthread_mutex_unlock(&g->startup_mutex);
1140 /* Here we will wait for the main process to start us all at once: */
1141 pthread_mutex_lock(&g->start_work_mutex);
1142 g->nr_tasks_working++;
1144 /* Last one wake the main process: */
1145 if (g->nr_tasks_working == g->p.nr_tasks)
1146 pthread_mutex_unlock(&g->startup_done_mutex);
1148 pthread_mutex_unlock(&g->start_work_mutex);
1151 gettimeofday(&start0, NULL);
1153 start = stop = start0;
1154 last_perturbance = start.tv_sec;
1156 for (l = 0; l < g->p.nr_loops; l++) {
1162 val += do_work(global_data, g->p.bytes_global, process_nr, g->p.nr_proc, l, val);
1163 val += do_work(process_data, g->p.bytes_process, thread_nr, g->p.nr_threads, l, val);
1164 val += do_work(thread_data, g->p.bytes_thread, 0, 1, l, val);
1166 if (g->p.sleep_usecs) {
1167 pthread_mutex_lock(td->process_lock);
1168 usleep(g->p.sleep_usecs);
1169 pthread_mutex_unlock(td->process_lock);
1172 * Amount of work to be done under a process-global lock:
1174 if (g->p.bytes_process_locked) {
1175 pthread_mutex_lock(td->process_lock);
1176 val += do_work(process_data, g->p.bytes_process_locked, thread_nr, g->p.nr_threads, l, val);
1177 pthread_mutex_unlock(td->process_lock);
1180 work_done = g->p.bytes_global + g->p.bytes_process +
1181 g->p.bytes_process_locked + g->p.bytes_thread;
1183 update_curr_cpu(task_nr, work_done);
1184 bytes_done += work_done;
1186 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1191 gettimeofday(&stop, NULL);
1193 /* Check whether our max runtime timed out: */
1195 timersub(&stop, &start0, &diff);
1196 if ((u32)diff.tv_sec >= g->p.nr_secs) {
1197 g->stop_work = true;
1202 /* Update the summary at most once per second: */
1203 if (start.tv_sec == stop.tv_sec)
1207 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1208 * by migrating to CPU#0:
1210 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1211 cpu_set_t orig_mask;
1215 last_perturbance = stop.tv_sec;
1218 * Depending on where we are running, move into
1219 * the other half of the system, to create some
1222 this_cpu = g->threads[task_nr].curr_cpu;
1223 if (this_cpu < g->p.nr_cpus/2)
1224 target_cpu = g->p.nr_cpus-1;
1228 orig_mask = bind_to_cpu(target_cpu);
1230 /* Here we are running on the target CPU already */
1232 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1234 bind_to_cpumask(orig_mask);
1238 timersub(&stop, &start, &diff);
1239 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1240 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1243 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1244 process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1251 timersub(&stop, &start0, &diff);
1252 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1253 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1255 show_summary(runtime_ns_max, l, &convergence);
1258 gettimeofday(&stop, NULL);
1259 timersub(&stop, &start0, &diff);
1260 td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1261 td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1262 secs = td->runtime_ns / NSEC_PER_SEC;
1263 td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1265 getrusage(RUSAGE_THREAD, &rusage);
1266 td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1267 td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1268 td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1269 td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1271 free_data(thread_data, g->p.bytes_thread);
1273 pthread_mutex_lock(&g->stop_work_mutex);
1274 g->bytes_done += bytes_done;
1275 pthread_mutex_unlock(&g->stop_work_mutex);
1281 * A worker process starts a couple of threads:
1283 static void worker_process(int process_nr)
1285 pthread_mutex_t process_lock;
1286 struct thread_data *td;
1287 pthread_t *pthreads;
1293 pthread_mutex_init(&process_lock, NULL);
1294 set_taskname("process %d", process_nr);
1297 * Pick up the memory policy and the CPU binding of our first thread,
1298 * so that we initialize memory accordingly:
1300 task_nr = process_nr*g->p.nr_threads;
1301 td = g->threads + task_nr;
1303 bind_to_memnode(td->bind_node);
1304 bind_to_cpumask(td->bind_cpumask);
1306 pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1307 process_data = setup_private_data(g->p.bytes_process);
1309 if (g->p.show_details >= 3) {
1310 printf(" # process %2d global mem: %p, process mem: %p\n",
1311 process_nr, g->data, process_data);
1314 for (t = 0; t < g->p.nr_threads; t++) {
1315 task_nr = process_nr*g->p.nr_threads + t;
1316 td = g->threads + task_nr;
1318 td->process_data = process_data;
1319 td->process_nr = process_nr;
1321 td->task_nr = task_nr;
1324 td->process_lock = &process_lock;
1326 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1330 for (t = 0; t < g->p.nr_threads; t++) {
1331 ret = pthread_join(pthreads[t], NULL);
1335 free_data(process_data, g->p.bytes_process);
1339 static void print_summary(void)
1341 if (g->p.show_details < 0)
1345 printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1346 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1347 printf(" # %5dx %5ldMB global shared mem operations\n",
1348 g->p.nr_loops, g->p.bytes_global/1024/1024);
1349 printf(" # %5dx %5ldMB process shared mem operations\n",
1350 g->p.nr_loops, g->p.bytes_process/1024/1024);
1351 printf(" # %5dx %5ldMB thread local mem operations\n",
1352 g->p.nr_loops, g->p.bytes_thread/1024/1024);
1356 printf("\n ###\n"); fflush(stdout);
1359 static void init_thread_data(void)
1361 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1364 g->threads = zalloc_shared_data(size);
1366 for (t = 0; t < g->p.nr_tasks; t++) {
1367 struct thread_data *td = g->threads + t;
1370 /* Allow all nodes by default: */
1371 td->bind_node = NUMA_NO_NODE;
1373 /* Allow all CPUs by default: */
1374 CPU_ZERO(&td->bind_cpumask);
1375 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1376 CPU_SET(cpu, &td->bind_cpumask);
1380 static void deinit_thread_data(void)
1382 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1384 free_data(g->threads, size);
1387 static int init(void)
1389 g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1391 /* Copy over options: */
1394 g->p.nr_cpus = numa_num_configured_cpus();
1396 g->p.nr_nodes = numa_max_node() + 1;
1398 /* char array in count_process_nodes(): */
1399 BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
1401 if (g->p.show_quiet && !g->p.show_details)
1402 g->p.show_details = -1;
1404 /* Some memory should be specified: */
1405 if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1408 if (g->p.mb_global_str) {
1409 g->p.mb_global = atof(g->p.mb_global_str);
1410 BUG_ON(g->p.mb_global < 0);
1413 if (g->p.mb_proc_str) {
1414 g->p.mb_proc = atof(g->p.mb_proc_str);
1415 BUG_ON(g->p.mb_proc < 0);
1418 if (g->p.mb_proc_locked_str) {
1419 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1420 BUG_ON(g->p.mb_proc_locked < 0);
1421 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1424 if (g->p.mb_thread_str) {
1425 g->p.mb_thread = atof(g->p.mb_thread_str);
1426 BUG_ON(g->p.mb_thread < 0);
1429 BUG_ON(g->p.nr_threads <= 0);
1430 BUG_ON(g->p.nr_proc <= 0);
1432 g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1434 g->p.bytes_global = g->p.mb_global *1024L*1024L;
1435 g->p.bytes_process = g->p.mb_proc *1024L*1024L;
1436 g->p.bytes_process_locked = g->p.mb_proc_locked *1024L*1024L;
1437 g->p.bytes_thread = g->p.mb_thread *1024L*1024L;
1439 g->data = setup_shared_data(g->p.bytes_global);
1441 /* Startup serialization: */
1442 init_global_mutex(&g->start_work_mutex);
1443 init_global_mutex(&g->startup_mutex);
1444 init_global_mutex(&g->startup_done_mutex);
1445 init_global_mutex(&g->stop_work_mutex);
1450 if (parse_setup_cpu_list() || parse_setup_node_list())
1459 static void deinit(void)
1461 free_data(g->data, g->p.bytes_global);
1464 deinit_thread_data();
1466 free_data(g, sizeof(*g));
1471 * Print a short or long result, depending on the verbosity setting:
1473 static void print_res(const char *name, double val,
1474 const char *txt_unit, const char *txt_short, const char *txt_long)
1479 if (!g->p.show_quiet)
1480 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1482 printf(" %14.3f %s\n", val, txt_long);
1485 static int __bench_numa(const char *name)
1487 struct timeval start, stop, diff;
1488 u64 runtime_ns_min, runtime_ns_sum;
1489 pid_t *pids, pid, wpid;
1490 double delta_runtime;
1492 double runtime_sec_max;
1493 double runtime_sec_min;
1501 pids = zalloc(g->p.nr_proc * sizeof(*pids));
1504 /* All threads try to acquire it, this way we can wait for them to start up: */
1505 pthread_mutex_lock(&g->start_work_mutex);
1507 if (g->p.serialize_startup) {
1509 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1512 gettimeofday(&start, NULL);
1514 for (i = 0; i < g->p.nr_proc; i++) {
1516 dprintf(" # process %2d: PID %d\n", i, pid);
1520 /* Child process: */
1528 /* Wait for all the threads to start up: */
1529 while (g->nr_tasks_started != g->p.nr_tasks)
1530 usleep(USEC_PER_MSEC);
1532 BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
1534 if (g->p.serialize_startup) {
1537 pthread_mutex_lock(&g->startup_done_mutex);
1539 /* This will start all threads: */
1540 pthread_mutex_unlock(&g->start_work_mutex);
1542 /* This mutex is locked - the last started thread will wake us: */
1543 pthread_mutex_lock(&g->startup_done_mutex);
1545 gettimeofday(&stop, NULL);
1547 timersub(&stop, &start, &diff);
1549 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1550 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1551 startup_sec /= NSEC_PER_SEC;
1553 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1557 pthread_mutex_unlock(&g->startup_done_mutex);
1559 gettimeofday(&start, NULL);
1562 /* Parent process: */
1565 for (i = 0; i < g->p.nr_proc; i++) {
1566 wpid = waitpid(pids[i], &wait_stat, 0);
1568 BUG_ON(!WIFEXITED(wait_stat));
1573 runtime_ns_min = -1LL;
1575 for (t = 0; t < g->p.nr_tasks; t++) {
1576 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1578 runtime_ns_sum += thread_runtime_ns;
1579 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1582 gettimeofday(&stop, NULL);
1583 timersub(&stop, &start, &diff);
1585 BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1587 tprintf("\n ###\n");
1590 runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1591 runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1592 runtime_sec_max /= NSEC_PER_SEC;
1594 runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1596 bytes = g->bytes_done;
1597 runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1599 if (g->p.measure_convergence) {
1600 print_res(name, runtime_sec_max,
1601 "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1604 print_res(name, runtime_sec_max,
1605 "secs,", "runtime-max/thread", "secs slowest (max) thread-runtime");
1607 print_res(name, runtime_sec_min,
1608 "secs,", "runtime-min/thread", "secs fastest (min) thread-runtime");
1610 print_res(name, runtime_avg,
1611 "secs,", "runtime-avg/thread", "secs average thread-runtime");
1613 delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1614 print_res(name, delta_runtime / runtime_sec_max * 100.0,
1615 "%,", "spread-runtime/thread", "% difference between max/avg runtime");
1617 print_res(name, bytes / g->p.nr_tasks / 1e9,
1618 "GB,", "data/thread", "GB data processed, per thread");
1620 print_res(name, bytes / 1e9,
1621 "GB,", "data-total", "GB data processed, total");
1623 print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1624 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1626 print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1627 "GB/sec,", "thread-speed", "GB/sec/thread speed");
1629 print_res(name, bytes / runtime_sec_max / 1e9,
1630 "GB/sec,", "total-speed", "GB/sec total speed");
1632 if (g->p.show_details >= 2) {
1633 char tname[14 + 2 * 10 + 1];
1634 struct thread_data *td;
1635 for (p = 0; p < g->p.nr_proc; p++) {
1636 for (t = 0; t < g->p.nr_threads; t++) {
1637 memset(tname, 0, sizeof(tname));
1638 td = g->threads + p*g->p.nr_threads + t;
1639 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1640 print_res(tname, td->speed_gbs,
1641 "GB/sec", "thread-speed", "GB/sec/thread speed");
1642 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1643 "secs", "thread-system-time", "system CPU time/thread");
1644 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1645 "secs", "thread-user-time", "user CPU time/thread");
1659 static int command_size(const char **argv)
1668 BUG_ON(size >= MAX_ARGS);
1673 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1677 printf("\n # Running %s \"perf bench numa", name);
1679 for (i = 0; i < argc; i++)
1680 printf(" %s", argv[i]);
1684 memset(p, 0, sizeof(*p));
1686 /* Initialize nonzero defaults: */
1688 p->serialize_startup = 1;
1689 p->data_reads = true;
1690 p->data_writes = true;
1691 p->data_backwards = true;
1692 p->data_rand_walk = true;
1694 p->init_random = true;
1695 p->mb_global_str = "1";
1699 p->run_all = argc == 1;
1702 static int run_bench_numa(const char *name, const char **argv)
1704 int argc = command_size(argv);
1706 init_params(&p0, name, argc, argv);
1707 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1711 if (__bench_numa(name))
1720 #define OPT_BW_RAM "-s", "20", "-zZq", "--thp", " 1", "--no-data_rand_walk"
1721 #define OPT_BW_RAM_NOTHP OPT_BW_RAM, "--thp", "-1"
1723 #define OPT_CONV "-s", "100", "-zZ0qcm", "--thp", " 1"
1724 #define OPT_CONV_NOTHP OPT_CONV, "--thp", "-1"
1726 #define OPT_BW "-s", "20", "-zZ0q", "--thp", " 1"
1727 #define OPT_BW_NOTHP OPT_BW, "--thp", "-1"
1730 * The built-in test-suite executed by "perf bench numa -a".
1732 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1734 static const char *tests[][MAX_ARGS] = {
1735 /* Basic single-stream NUMA bandwidth measurements: */
1736 { "RAM-bw-local,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1737 "-C" , "0", "-M", "0", OPT_BW_RAM },
1738 { "RAM-bw-local-NOTHP,",
1739 "mem", "-p", "1", "-t", "1", "-P", "1024",
1740 "-C" , "0", "-M", "0", OPT_BW_RAM_NOTHP },
1741 { "RAM-bw-remote,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1742 "-C" , "0", "-M", "1", OPT_BW_RAM },
1744 /* 2-stream NUMA bandwidth measurements: */
1745 { "RAM-bw-local-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1746 "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1747 { "RAM-bw-remote-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1748 "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1750 /* Cross-stream NUMA bandwidth measurement: */
1751 { "RAM-bw-cross,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1752 "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1754 /* Convergence latency measurements: */
1755 { " 1x3-convergence,", "mem", "-p", "1", "-t", "3", "-P", "512", OPT_CONV },
1756 { " 1x4-convergence,", "mem", "-p", "1", "-t", "4", "-P", "512", OPT_CONV },
1757 { " 1x6-convergence,", "mem", "-p", "1", "-t", "6", "-P", "1020", OPT_CONV },
1758 { " 2x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1759 { " 3x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1760 { " 4x4-convergence,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV },
1761 { " 4x4-convergence-NOTHP,",
1762 "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1763 { " 4x6-convergence,", "mem", "-p", "4", "-t", "6", "-P", "1020", OPT_CONV },
1764 { " 4x8-convergence,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_CONV },
1765 { " 8x4-convergence,", "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV },
1766 { " 8x4-convergence-NOTHP,",
1767 "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1768 { " 3x1-convergence,", "mem", "-p", "3", "-t", "1", "-P", "512", OPT_CONV },
1769 { " 4x1-convergence,", "mem", "-p", "4", "-t", "1", "-P", "512", OPT_CONV },
1770 { " 8x1-convergence,", "mem", "-p", "8", "-t", "1", "-P", "512", OPT_CONV },
1771 { "16x1-convergence,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_CONV },
1772 { "32x1-convergence,", "mem", "-p", "32", "-t", "1", "-P", "128", OPT_CONV },
1774 /* Various NUMA process/thread layout bandwidth measurements: */
1775 { " 2x1-bw-process,", "mem", "-p", "2", "-t", "1", "-P", "1024", OPT_BW },
1776 { " 3x1-bw-process,", "mem", "-p", "3", "-t", "1", "-P", "1024", OPT_BW },
1777 { " 4x1-bw-process,", "mem", "-p", "4", "-t", "1", "-P", "1024", OPT_BW },
1778 { " 8x1-bw-process,", "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW },
1779 { " 8x1-bw-process-NOTHP,",
1780 "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW_NOTHP },
1781 { "16x1-bw-process,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_BW },
1783 { " 4x1-bw-thread,", "mem", "-p", "1", "-t", "4", "-T", "256", OPT_BW },
1784 { " 8x1-bw-thread,", "mem", "-p", "1", "-t", "8", "-T", "256", OPT_BW },
1785 { "16x1-bw-thread,", "mem", "-p", "1", "-t", "16", "-T", "128", OPT_BW },
1786 { "32x1-bw-thread,", "mem", "-p", "1", "-t", "32", "-T", "64", OPT_BW },
1788 { " 2x3-bw-thread,", "mem", "-p", "2", "-t", "3", "-P", "512", OPT_BW },
1789 { " 4x4-bw-thread,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_BW },
1790 { " 4x6-bw-thread,", "mem", "-p", "4", "-t", "6", "-P", "512", OPT_BW },
1791 { " 4x8-bw-thread,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW },
1792 { " 4x8-bw-thread-NOTHP,",
1793 "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW_NOTHP },
1794 { " 3x3-bw-thread,", "mem", "-p", "3", "-t", "3", "-P", "512", OPT_BW },
1795 { " 5x5-bw-thread,", "mem", "-p", "5", "-t", "5", "-P", "512", OPT_BW },
1797 { "2x16-bw-thread,", "mem", "-p", "2", "-t", "16", "-P", "512", OPT_BW },
1798 { "1x32-bw-thread,", "mem", "-p", "1", "-t", "32", "-P", "2048", OPT_BW },
1800 { "numa02-bw,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW },
1801 { "numa02-bw-NOTHP,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW_NOTHP },
1802 { "numa01-bw-thread,", "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW },
1803 { "numa01-bw-thread-NOTHP,",
1804 "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW_NOTHP },
1807 static int bench_all(void)
1809 int nr = ARRAY_SIZE(tests);
1813 ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1816 for (i = 0; i < nr; i++) {
1817 run_bench_numa(tests[i][0], tests[i] + 1);
1825 int bench_numa(int argc, const char **argv)
1827 init_params(&p0, "main,", argc, argv);
1828 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1835 if (__bench_numa(NULL))
1841 usage_with_options(numa_usage, options);