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>
37 #include "../util/header.h"
42 # define RUSAGE_THREAD 1
46 * Regular printout to the terminal, suppressed if -q is specified:
48 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
54 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
58 cpu_set_t *bind_cpumask;
64 unsigned int loops_done;
70 pthread_mutex_t *process_lock;
73 /* Parameters set by options: */
76 /* Startup synchronization: */
77 bool serialize_startup;
83 /* Working set sizes: */
84 const char *mb_global_str;
85 const char *mb_proc_str;
86 const char *mb_proc_locked_str;
87 const char *mb_thread_str;
91 double mb_proc_locked;
94 /* Access patterns to the working set: */
98 bool data_zero_memset;
104 /* Working set initialization: */
116 long bytes_process_locked;
122 bool show_convergence;
123 bool measure_convergence;
129 /* Affinity options -C and -N: */
135 /* Global, read-writable area, accessible to all processes and threads: */
140 pthread_mutex_t startup_mutex;
141 pthread_cond_t startup_cond;
142 int nr_tasks_started;
144 pthread_mutex_t start_work_mutex;
145 pthread_cond_t start_work_cond;
146 int nr_tasks_working;
149 pthread_mutex_t stop_work_mutex;
152 struct thread_data *threads;
154 /* Convergence latency measurement: */
163 static struct global_info *g = NULL;
165 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
166 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
170 static const struct option options[] = {
171 OPT_INTEGER('p', "nr_proc" , &p0.nr_proc, "number of processes"),
172 OPT_INTEGER('t', "nr_threads" , &p0.nr_threads, "number of threads per process"),
174 OPT_STRING('G', "mb_global" , &p0.mb_global_str, "MB", "global memory (MBs)"),
175 OPT_STRING('P', "mb_proc" , &p0.mb_proc_str, "MB", "process memory (MBs)"),
176 OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
177 OPT_STRING('T', "mb_thread" , &p0.mb_thread_str, "MB", "thread memory (MBs)"),
179 OPT_UINTEGER('l', "nr_loops" , &p0.nr_loops, "max number of loops to run (default: unlimited)"),
180 OPT_UINTEGER('s', "nr_secs" , &p0.nr_secs, "max number of seconds to run (default: 5 secs)"),
181 OPT_UINTEGER('u', "usleep" , &p0.sleep_usecs, "usecs to sleep per loop iteration"),
183 OPT_BOOLEAN('R', "data_reads" , &p0.data_reads, "access the data via reads (can be mixed with -W)"),
184 OPT_BOOLEAN('W', "data_writes" , &p0.data_writes, "access the data via writes (can be mixed with -R)"),
185 OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards, "access the data backwards as well"),
186 OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
187 OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk, "access the data with random (32bit LFSR) walk"),
190 OPT_BOOLEAN('z', "init_zero" , &p0.init_zero, "bzero the initial allocations"),
191 OPT_BOOLEAN('I', "init_random" , &p0.init_random, "randomize the contents of the initial allocations"),
192 OPT_BOOLEAN('0', "init_cpu0" , &p0.init_cpu0, "do the initial allocations on CPU#0"),
193 OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs, "perturb thread 0/0 every X secs, to test convergence stability"),
195 OPT_INCR ('d', "show_details" , &p0.show_details, "Show details"),
196 OPT_INCR ('a', "all" , &p0.run_all, "Run all tests in the suite"),
197 OPT_INTEGER('H', "thp" , &p0.thp, "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
198 OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
199 "convergence is reached when each process (all its threads) is running on a single NUMA node."),
200 OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
201 OPT_BOOLEAN('q', "quiet" , &p0.show_quiet, "quiet mode"),
202 OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
204 /* Special option string parsing callbacks: */
205 OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
206 "bind the first N tasks to these specific cpus (the rest is unbound)",
208 OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
209 "bind the first N tasks to these specific memory nodes (the rest is unbound)",
214 static const char * const bench_numa_usage[] = {
215 "perf bench numa <options>",
219 static const char * const numa_usage[] = {
220 "perf bench numa mem [<options>]",
225 * To get number of numa nodes present.
227 static int nr_numa_nodes(void)
231 for (i = 0; i < g->p.nr_nodes; i++) {
232 if (numa_bitmask_isbitset(numa_nodes_ptr, i))
240 * To check if given numa node is present.
242 static int is_node_present(int node)
244 return numa_bitmask_isbitset(numa_nodes_ptr, node);
248 * To check given numa node has cpus.
250 static bool node_has_cpus(int node)
252 struct bitmask *cpumask = numa_allocate_cpumask();
253 bool ret = false; /* fall back to nocpus */
257 if (!numa_node_to_cpus(node, cpumask)) {
258 for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
259 if (numa_bitmask_isbitset(cpumask, cpu)) {
265 numa_free_cpumask(cpumask);
270 static cpu_set_t *bind_to_cpu(int target_cpu)
272 int nrcpus = numa_num_possible_cpus();
273 cpu_set_t *orig_mask, *mask;
276 orig_mask = CPU_ALLOC(nrcpus);
278 size = CPU_ALLOC_SIZE(nrcpus);
279 CPU_ZERO_S(size, orig_mask);
281 if (sched_getaffinity(0, size, orig_mask))
284 mask = CPU_ALLOC(nrcpus);
288 CPU_ZERO_S(size, mask);
290 if (target_cpu == -1) {
293 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
294 CPU_SET_S(cpu, size, mask);
296 if (target_cpu < 0 || target_cpu >= g->p.nr_cpus)
299 CPU_SET_S(target_cpu, size, mask);
302 if (sched_setaffinity(0, size, mask))
312 /* BUG_ON due to failure in allocation of orig_mask/mask */
317 static cpu_set_t *bind_to_node(int target_node)
319 int nrcpus = numa_num_possible_cpus();
321 cpu_set_t *orig_mask, *mask;
324 orig_mask = CPU_ALLOC(nrcpus);
326 size = CPU_ALLOC_SIZE(nrcpus);
327 CPU_ZERO_S(size, orig_mask);
329 if (sched_getaffinity(0, size, orig_mask))
332 mask = CPU_ALLOC(nrcpus);
336 CPU_ZERO_S(size, mask);
338 if (target_node == NUMA_NO_NODE) {
339 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
340 CPU_SET_S(cpu, size, mask);
342 struct bitmask *cpumask = numa_allocate_cpumask();
347 if (!numa_node_to_cpus(target_node, cpumask)) {
348 for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
349 if (numa_bitmask_isbitset(cpumask, cpu))
350 CPU_SET_S(cpu, size, mask);
353 numa_free_cpumask(cpumask);
356 if (sched_setaffinity(0, size, mask))
366 /* BUG_ON due to failure in allocation of orig_mask/mask */
371 static void bind_to_cpumask(cpu_set_t *mask)
374 size_t size = CPU_ALLOC_SIZE(numa_num_possible_cpus());
376 ret = sched_setaffinity(0, size, mask);
383 static void mempol_restore(void)
387 ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
392 static void bind_to_memnode(int node)
394 struct bitmask *node_mask;
397 if (node == NUMA_NO_NODE)
400 node_mask = numa_allocate_nodemask();
403 numa_bitmask_clearall(node_mask);
404 numa_bitmask_setbit(node_mask, node);
406 ret = set_mempolicy(MPOL_BIND, node_mask->maskp, node_mask->size + 1);
407 dprintf("binding to node %d, mask: %016lx => %d\n", node, *node_mask->maskp, ret);
409 numa_bitmask_free(node_mask);
413 #define HPSIZE (2*1024*1024)
415 #define set_taskname(fmt...) \
419 snprintf(name, 20, fmt); \
420 prctl(PR_SET_NAME, name); \
423 static u8 *alloc_data(ssize_t bytes0, int map_flags,
424 int init_zero, int init_cpu0, int thp, int init_random)
426 cpu_set_t *orig_mask = NULL;
434 /* Allocate and initialize all memory on CPU#0: */
436 int node = numa_node_of_cpu(0);
438 orig_mask = bind_to_node(node);
439 bind_to_memnode(node);
442 bytes = bytes0 + HPSIZE;
444 buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
445 BUG_ON(buf == (void *)-1);
447 if (map_flags == MAP_PRIVATE) {
449 ret = madvise(buf, bytes, MADV_HUGEPAGE);
450 if (ret && !g->print_once) {
452 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
456 ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
457 if (ret && !g->print_once) {
459 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
467 /* Initialize random contents, different in each word: */
469 u64 *wbuf = (void *)buf;
473 for (i = 0; i < bytes/8; i++)
478 /* Align to 2MB boundary: */
479 buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
481 /* Restore affinity: */
483 bind_to_cpumask(orig_mask);
491 static void free_data(void *data, ssize_t bytes)
498 ret = munmap(data, bytes);
503 * Create a shared memory buffer that can be shared between processes, zeroed:
505 static void * zalloc_shared_data(ssize_t bytes)
507 return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0, g->p.thp, g->p.init_random);
511 * Create a shared memory buffer that can be shared between processes:
513 static void * setup_shared_data(ssize_t bytes)
515 return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
519 * Allocate process-local memory - this will either be shared between
520 * threads of this process, or only be accessed by this thread:
522 static void * setup_private_data(ssize_t bytes)
524 return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
528 * Return a process-shared (global) mutex:
530 static void init_global_mutex(pthread_mutex_t *mutex)
532 pthread_mutexattr_t attr;
534 pthread_mutexattr_init(&attr);
535 pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
536 pthread_mutex_init(mutex, &attr);
540 * Return a process-shared (global) condition variable:
542 static void init_global_cond(pthread_cond_t *cond)
544 pthread_condattr_t attr;
546 pthread_condattr_init(&attr);
547 pthread_condattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
548 pthread_cond_init(cond, &attr);
551 static int parse_cpu_list(const char *arg)
553 p0.cpu_list_str = strdup(arg);
555 dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
560 static int parse_setup_cpu_list(void)
562 struct thread_data *td;
566 if (!g->p.cpu_list_str)
569 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
571 str0 = str = strdup(g->p.cpu_list_str);
576 tprintf("# binding tasks to CPUs:\n");
580 int bind_cpu, bind_cpu_0, bind_cpu_1;
581 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
586 tok = strsep(&str, ",");
590 tok_end = strstr(tok, "-");
592 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
594 /* Single CPU specified: */
595 bind_cpu_0 = bind_cpu_1 = atol(tok);
597 /* CPU range specified (for example: "5-11"): */
598 bind_cpu_0 = atol(tok);
599 bind_cpu_1 = atol(tok_end + 1);
603 tok_step = strstr(tok, "#");
605 step = atol(tok_step + 1);
606 BUG_ON(step <= 0 || step >= g->p.nr_cpus);
611 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
612 * where the _4 means the next 4 CPUs are allowed.
615 tok_len = strstr(tok, "_");
617 bind_len = atol(tok_len + 1);
618 BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
621 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
623 tok_mul = strstr(tok, "x");
625 mul = atol(tok_mul + 1);
629 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
631 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
632 printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
636 if (is_cpu_online(bind_cpu_0) != 1 || is_cpu_online(bind_cpu_1) != 1) {
637 printf("\nTest not applicable, bind_cpu_0 or bind_cpu_1 is offline\n");
641 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
642 BUG_ON(bind_cpu_0 > bind_cpu_1);
644 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
645 size_t size = CPU_ALLOC_SIZE(g->p.nr_cpus);
648 for (i = 0; i < mul; i++) {
651 if (t >= g->p.nr_tasks) {
652 printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
660 tprintf("%2d/%d", bind_cpu, bind_len);
662 tprintf("%2d", bind_cpu);
665 td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
666 BUG_ON(!td->bind_cpumask);
667 CPU_ZERO_S(size, td->bind_cpumask);
668 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
669 if (cpu < 0 || cpu >= g->p.nr_cpus) {
670 CPU_FREE(td->bind_cpumask);
673 CPU_SET_S(cpu, size, td->bind_cpumask);
683 if (t < g->p.nr_tasks)
684 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
690 static int parse_cpus_opt(const struct option *opt __maybe_unused,
691 const char *arg, int unset __maybe_unused)
696 return parse_cpu_list(arg);
699 static int parse_node_list(const char *arg)
701 p0.node_list_str = strdup(arg);
703 dprintf("got NODE list: {%s}\n", p0.node_list_str);
708 static int parse_setup_node_list(void)
710 struct thread_data *td;
714 if (!g->p.node_list_str)
717 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
719 str0 = str = strdup(g->p.node_list_str);
724 tprintf("# binding tasks to NODEs:\n");
728 int bind_node, bind_node_0, bind_node_1;
729 char *tok, *tok_end, *tok_step, *tok_mul;
733 tok = strsep(&str, ",");
737 tok_end = strstr(tok, "-");
739 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
741 /* Single NODE specified: */
742 bind_node_0 = bind_node_1 = atol(tok);
744 /* NODE range specified (for example: "5-11"): */
745 bind_node_0 = atol(tok);
746 bind_node_1 = atol(tok_end + 1);
750 tok_step = strstr(tok, "#");
752 step = atol(tok_step + 1);
753 BUG_ON(step <= 0 || step >= g->p.nr_nodes);
756 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
758 tok_mul = strstr(tok, "x");
760 mul = atol(tok_mul + 1);
764 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
766 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
767 printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
771 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
772 BUG_ON(bind_node_0 > bind_node_1);
774 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
777 for (i = 0; i < mul; i++) {
778 if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
779 printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
785 tprintf(" %2d", bind_node);
787 tprintf(",%2d", bind_node);
789 td->bind_node = bind_node;
798 if (t < g->p.nr_tasks)
799 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
805 static int parse_nodes_opt(const struct option *opt __maybe_unused,
806 const char *arg, int unset __maybe_unused)
811 return parse_node_list(arg);
814 static inline uint32_t lfsr_32(uint32_t lfsr)
816 const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
817 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
821 * Make sure there's real data dependency to RAM (when read
822 * accesses are enabled), so the compiler, the CPU and the
823 * kernel (KSM, zero page, etc.) cannot optimize away RAM
826 static inline u64 access_data(u64 *data, u64 val)
830 if (g->p.data_writes)
836 * The worker process does two types of work, a forwards going
837 * loop and a backwards going loop.
839 * We do this so that on multiprocessor systems we do not create
840 * a 'train' of processing, with highly synchronized processes,
841 * skewing the whole benchmark.
843 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
845 long words = bytes/sizeof(u64);
846 u64 *data = (void *)__data;
847 long chunk_0, chunk_1;
852 BUG_ON(!data && words);
853 BUG_ON(data && !words);
858 /* Very simple memset() work variant: */
859 if (g->p.data_zero_memset && !g->p.data_rand_walk) {
864 /* Spread out by PID/TID nr and by loop nr: */
865 chunk_0 = words/nr_max;
866 chunk_1 = words/g->p.nr_loops;
867 off = nr*chunk_0 + loop*chunk_1;
872 if (g->p.data_rand_walk) {
873 u32 lfsr = nr + loop + val;
876 for (i = 0; i < words/1024; i++) {
879 lfsr = lfsr_32(lfsr);
881 start = lfsr % words;
882 end = min(start + 1024, words-1);
884 if (g->p.data_zero_memset) {
885 bzero(data + start, (end-start) * sizeof(u64));
887 for (j = start; j < end; j++)
888 val = access_data(data + j, val);
891 } else if (!g->p.data_backwards || (nr + loop) & 1) {
892 /* Process data forwards: */
899 if (unlikely(d >= d1))
901 if (unlikely(d == d0))
904 val = access_data(d, val);
909 /* Process data backwards: */
916 if (unlikely(d < data))
918 if (unlikely(d == d0))
921 val = access_data(d, val);
930 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
934 cpu = sched_getcpu();
936 g->threads[task_nr].curr_cpu = cpu;
937 prctl(0, bytes_worked);
941 * Count the number of nodes a process's threads
944 * A count of 1 means that the process is compressed
945 * to a single node. A count of g->p.nr_nodes means it's
946 * spread out on the whole system.
948 static int count_process_nodes(int process_nr)
954 node_present = (char *)malloc(g->p.nr_nodes * sizeof(char));
955 BUG_ON(!node_present);
956 for (nodes = 0; nodes < g->p.nr_nodes; nodes++)
957 node_present[nodes] = 0;
959 for (t = 0; t < g->p.nr_threads; t++) {
960 struct thread_data *td;
964 task_nr = process_nr*g->p.nr_threads + t;
965 td = g->threads + task_nr;
967 node = numa_node_of_cpu(td->curr_cpu);
968 if (node < 0) /* curr_cpu was likely still -1 */ {
973 node_present[node] = 1;
978 for (n = 0; n < g->p.nr_nodes; n++)
979 nodes += node_present[n];
986 * Count the number of distinct process-threads a node contains.
988 * A count of 1 means that the node contains only a single
989 * process. If all nodes on the system contain at most one
990 * process then we are well-converged.
992 static int count_node_processes(int node)
997 for (p = 0; p < g->p.nr_proc; p++) {
998 for (t = 0; t < g->p.nr_threads; t++) {
999 struct thread_data *td;
1003 task_nr = p*g->p.nr_threads + t;
1004 td = g->threads + task_nr;
1006 n = numa_node_of_cpu(td->curr_cpu);
1017 static void calc_convergence_compression(int *strong)
1019 unsigned int nodes_min, nodes_max;
1025 for (p = 0; p < g->p.nr_proc; p++) {
1026 unsigned int nodes = count_process_nodes(p);
1033 nodes_min = min(nodes, nodes_min);
1034 nodes_max = max(nodes, nodes_max);
1037 /* Strong convergence: all threads compress on a single node: */
1038 if (nodes_min == 1 && nodes_max == 1) {
1042 tprintf(" {%d-%d}", nodes_min, nodes_max);
1046 static void calc_convergence(double runtime_ns_max, double *convergence)
1048 unsigned int loops_done_min, loops_done_max;
1061 if (!g->p.show_convergence && !g->p.measure_convergence)
1064 nodes = (int *)malloc(g->p.nr_nodes * sizeof(int));
1066 for (node = 0; node < g->p.nr_nodes; node++)
1069 loops_done_min = -1;
1072 for (t = 0; t < g->p.nr_tasks; t++) {
1073 struct thread_data *td = g->threads + t;
1074 unsigned int loops_done;
1078 /* Not all threads have written it yet: */
1082 node = numa_node_of_cpu(cpu);
1086 loops_done = td->loops_done;
1087 loops_done_min = min(loops_done, loops_done_min);
1088 loops_done_max = max(loops_done, loops_done_max);
1092 nr_min = g->p.nr_tasks;
1095 for (node = 0; node < g->p.nr_nodes; node++) {
1096 if (!is_node_present(node))
1099 nr_min = min(nr, nr_min);
1100 nr_max = max(nr, nr_max);
1103 BUG_ON(nr_min > nr_max);
1105 BUG_ON(sum > g->p.nr_tasks);
1107 if (0 && (sum < g->p.nr_tasks)) {
1113 * Count the number of distinct process groups present
1114 * on nodes - when we are converged this will decrease
1119 for (node = 0; node < g->p.nr_nodes; node++) {
1122 if (!is_node_present(node))
1124 processes = count_node_processes(node);
1126 tprintf(" %2d/%-2d", nr, processes);
1128 process_groups += processes;
1131 distance = nr_max - nr_min;
1133 tprintf(" [%2d/%-2d]", distance, process_groups);
1135 tprintf(" l:%3d-%-3d (%3d)",
1136 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1138 if (loops_done_min && loops_done_max) {
1139 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1141 tprintf(" [%4.1f%%]", skew * 100.0);
1144 calc_convergence_compression(&strong);
1146 if (strong && process_groups == g->p.nr_proc) {
1147 if (!*convergence) {
1148 *convergence = runtime_ns_max;
1149 tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1150 if (g->p.measure_convergence) {
1151 g->all_converged = true;
1152 g->stop_work = true;
1157 tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1166 static void show_summary(double runtime_ns_max, int l, double *convergence)
1168 tprintf("\r # %5.1f%% [%.1f mins]",
1169 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1171 calc_convergence(runtime_ns_max, convergence);
1173 if (g->p.show_details >= 0)
1177 static void *worker_thread(void *__tdata)
1179 struct thread_data *td = __tdata;
1180 struct timeval start0, start, stop, diff;
1181 int process_nr = td->process_nr;
1182 int thread_nr = td->thread_nr;
1183 unsigned long last_perturbance;
1184 int task_nr = td->task_nr;
1185 int details = g->p.show_details;
1186 int first_task, last_task;
1187 double convergence = 0;
1189 double runtime_ns_max;
1193 u64 bytes_done, secs;
1196 struct rusage rusage;
1198 bind_to_cpumask(td->bind_cpumask);
1199 bind_to_memnode(td->bind_node);
1201 set_taskname("thread %d/%d", process_nr, thread_nr);
1203 global_data = g->data;
1204 process_data = td->process_data;
1205 thread_data = setup_private_data(g->p.bytes_thread);
1210 if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1214 if (process_nr == 0 && thread_nr == 0)
1218 printf("# thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1219 process_nr, thread_nr, global_data, process_data, thread_data);
1222 if (g->p.serialize_startup) {
1223 pthread_mutex_lock(&g->startup_mutex);
1224 g->nr_tasks_started++;
1225 /* The last thread wakes the main process. */
1226 if (g->nr_tasks_started == g->p.nr_tasks)
1227 pthread_cond_signal(&g->startup_cond);
1229 pthread_mutex_unlock(&g->startup_mutex);
1231 /* Here we will wait for the main process to start us all at once: */
1232 pthread_mutex_lock(&g->start_work_mutex);
1233 g->start_work = false;
1234 g->nr_tasks_working++;
1235 while (!g->start_work)
1236 pthread_cond_wait(&g->start_work_cond, &g->start_work_mutex);
1238 pthread_mutex_unlock(&g->start_work_mutex);
1241 gettimeofday(&start0, NULL);
1243 start = stop = start0;
1244 last_perturbance = start.tv_sec;
1246 for (l = 0; l < g->p.nr_loops; l++) {
1252 val += do_work(global_data, g->p.bytes_global, process_nr, g->p.nr_proc, l, val);
1253 val += do_work(process_data, g->p.bytes_process, thread_nr, g->p.nr_threads, l, val);
1254 val += do_work(thread_data, g->p.bytes_thread, 0, 1, l, val);
1256 if (g->p.sleep_usecs) {
1257 pthread_mutex_lock(td->process_lock);
1258 usleep(g->p.sleep_usecs);
1259 pthread_mutex_unlock(td->process_lock);
1262 * Amount of work to be done under a process-global lock:
1264 if (g->p.bytes_process_locked) {
1265 pthread_mutex_lock(td->process_lock);
1266 val += do_work(process_data, g->p.bytes_process_locked, thread_nr, g->p.nr_threads, l, val);
1267 pthread_mutex_unlock(td->process_lock);
1270 work_done = g->p.bytes_global + g->p.bytes_process +
1271 g->p.bytes_process_locked + g->p.bytes_thread;
1273 update_curr_cpu(task_nr, work_done);
1274 bytes_done += work_done;
1276 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1281 gettimeofday(&stop, NULL);
1283 /* Check whether our max runtime timed out: */
1285 timersub(&stop, &start0, &diff);
1286 if ((u32)diff.tv_sec >= g->p.nr_secs) {
1287 g->stop_work = true;
1292 /* Update the summary at most once per second: */
1293 if (start.tv_sec == stop.tv_sec)
1297 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1298 * by migrating to CPU#0:
1300 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1301 cpu_set_t *orig_mask;
1305 last_perturbance = stop.tv_sec;
1308 * Depending on where we are running, move into
1309 * the other half of the system, to create some
1312 this_cpu = g->threads[task_nr].curr_cpu;
1313 if (this_cpu < g->p.nr_cpus/2)
1314 target_cpu = g->p.nr_cpus-1;
1318 orig_mask = bind_to_cpu(target_cpu);
1320 /* Here we are running on the target CPU already */
1322 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1324 bind_to_cpumask(orig_mask);
1325 CPU_FREE(orig_mask);
1329 timersub(&stop, &start, &diff);
1330 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1331 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1334 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1335 process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1342 timersub(&stop, &start0, &diff);
1343 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1344 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1346 show_summary(runtime_ns_max, l, &convergence);
1349 gettimeofday(&stop, NULL);
1350 timersub(&stop, &start0, &diff);
1351 td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1352 td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1353 secs = td->runtime_ns / NSEC_PER_SEC;
1354 td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1356 getrusage(RUSAGE_THREAD, &rusage);
1357 td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1358 td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1359 td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1360 td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1362 free_data(thread_data, g->p.bytes_thread);
1364 pthread_mutex_lock(&g->stop_work_mutex);
1365 g->bytes_done += bytes_done;
1366 pthread_mutex_unlock(&g->stop_work_mutex);
1372 * A worker process starts a couple of threads:
1374 static void worker_process(int process_nr)
1376 pthread_mutex_t process_lock;
1377 struct thread_data *td;
1378 pthread_t *pthreads;
1384 pthread_mutex_init(&process_lock, NULL);
1385 set_taskname("process %d", process_nr);
1388 * Pick up the memory policy and the CPU binding of our first thread,
1389 * so that we initialize memory accordingly:
1391 task_nr = process_nr*g->p.nr_threads;
1392 td = g->threads + task_nr;
1394 bind_to_memnode(td->bind_node);
1395 bind_to_cpumask(td->bind_cpumask);
1397 pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1398 process_data = setup_private_data(g->p.bytes_process);
1400 if (g->p.show_details >= 3) {
1401 printf(" # process %2d global mem: %p, process mem: %p\n",
1402 process_nr, g->data, process_data);
1405 for (t = 0; t < g->p.nr_threads; t++) {
1406 task_nr = process_nr*g->p.nr_threads + t;
1407 td = g->threads + task_nr;
1409 td->process_data = process_data;
1410 td->process_nr = process_nr;
1412 td->task_nr = task_nr;
1415 td->process_lock = &process_lock;
1417 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1421 for (t = 0; t < g->p.nr_threads; t++) {
1422 ret = pthread_join(pthreads[t], NULL);
1426 free_data(process_data, g->p.bytes_process);
1430 static void print_summary(void)
1432 if (g->p.show_details < 0)
1436 printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1437 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1438 printf(" # %5dx %5ldMB global shared mem operations\n",
1439 g->p.nr_loops, g->p.bytes_global/1024/1024);
1440 printf(" # %5dx %5ldMB process shared mem operations\n",
1441 g->p.nr_loops, g->p.bytes_process/1024/1024);
1442 printf(" # %5dx %5ldMB thread local mem operations\n",
1443 g->p.nr_loops, g->p.bytes_thread/1024/1024);
1447 printf("\n ###\n"); fflush(stdout);
1450 static void init_thread_data(void)
1452 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1455 g->threads = zalloc_shared_data(size);
1457 for (t = 0; t < g->p.nr_tasks; t++) {
1458 struct thread_data *td = g->threads + t;
1459 size_t cpuset_size = CPU_ALLOC_SIZE(g->p.nr_cpus);
1462 /* Allow all nodes by default: */
1463 td->bind_node = NUMA_NO_NODE;
1465 /* Allow all CPUs by default: */
1466 td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
1467 BUG_ON(!td->bind_cpumask);
1468 CPU_ZERO_S(cpuset_size, td->bind_cpumask);
1469 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1470 CPU_SET_S(cpu, cpuset_size, td->bind_cpumask);
1474 static void deinit_thread_data(void)
1476 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1479 /* Free the bind_cpumask allocated for thread_data */
1480 for (t = 0; t < g->p.nr_tasks; t++) {
1481 struct thread_data *td = g->threads + t;
1482 CPU_FREE(td->bind_cpumask);
1485 free_data(g->threads, size);
1488 static int init(void)
1490 g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1492 /* Copy over options: */
1495 g->p.nr_cpus = numa_num_configured_cpus();
1497 g->p.nr_nodes = numa_max_node() + 1;
1499 /* char array in count_process_nodes(): */
1500 BUG_ON(g->p.nr_nodes < 0);
1502 if (g->p.show_quiet && !g->p.show_details)
1503 g->p.show_details = -1;
1505 /* Some memory should be specified: */
1506 if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1509 if (g->p.mb_global_str) {
1510 g->p.mb_global = atof(g->p.mb_global_str);
1511 BUG_ON(g->p.mb_global < 0);
1514 if (g->p.mb_proc_str) {
1515 g->p.mb_proc = atof(g->p.mb_proc_str);
1516 BUG_ON(g->p.mb_proc < 0);
1519 if (g->p.mb_proc_locked_str) {
1520 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1521 BUG_ON(g->p.mb_proc_locked < 0);
1522 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1525 if (g->p.mb_thread_str) {
1526 g->p.mb_thread = atof(g->p.mb_thread_str);
1527 BUG_ON(g->p.mb_thread < 0);
1530 BUG_ON(g->p.nr_threads <= 0);
1531 BUG_ON(g->p.nr_proc <= 0);
1533 g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1535 g->p.bytes_global = g->p.mb_global *1024L*1024L;
1536 g->p.bytes_process = g->p.mb_proc *1024L*1024L;
1537 g->p.bytes_process_locked = g->p.mb_proc_locked *1024L*1024L;
1538 g->p.bytes_thread = g->p.mb_thread *1024L*1024L;
1540 g->data = setup_shared_data(g->p.bytes_global);
1542 /* Startup serialization: */
1543 init_global_mutex(&g->start_work_mutex);
1544 init_global_cond(&g->start_work_cond);
1545 init_global_mutex(&g->startup_mutex);
1546 init_global_cond(&g->startup_cond);
1547 init_global_mutex(&g->stop_work_mutex);
1552 if (parse_setup_cpu_list() || parse_setup_node_list())
1561 static void deinit(void)
1563 free_data(g->data, g->p.bytes_global);
1566 deinit_thread_data();
1568 free_data(g, sizeof(*g));
1573 * Print a short or long result, depending on the verbosity setting:
1575 static void print_res(const char *name, double val,
1576 const char *txt_unit, const char *txt_short, const char *txt_long)
1581 if (!g->p.show_quiet)
1582 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1584 printf(" %14.3f %s\n", val, txt_long);
1587 static int __bench_numa(const char *name)
1589 struct timeval start, stop, diff;
1590 u64 runtime_ns_min, runtime_ns_sum;
1591 pid_t *pids, pid, wpid;
1592 double delta_runtime;
1594 double runtime_sec_max;
1595 double runtime_sec_min;
1603 pids = zalloc(g->p.nr_proc * sizeof(*pids));
1606 if (g->p.serialize_startup) {
1608 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1611 gettimeofday(&start, NULL);
1613 for (i = 0; i < g->p.nr_proc; i++) {
1615 dprintf(" # process %2d: PID %d\n", i, pid);
1619 /* Child process: */
1628 if (g->p.serialize_startup) {
1629 bool threads_ready = false;
1633 * Wait for all the threads to start up. The last thread will
1634 * signal this process.
1636 pthread_mutex_lock(&g->startup_mutex);
1637 while (g->nr_tasks_started != g->p.nr_tasks)
1638 pthread_cond_wait(&g->startup_cond, &g->startup_mutex);
1640 pthread_mutex_unlock(&g->startup_mutex);
1642 /* Wait for all threads to be at the start_work_cond. */
1643 while (!threads_ready) {
1644 pthread_mutex_lock(&g->start_work_mutex);
1645 threads_ready = (g->nr_tasks_working == g->p.nr_tasks);
1646 pthread_mutex_unlock(&g->start_work_mutex);
1651 gettimeofday(&stop, NULL);
1653 timersub(&stop, &start, &diff);
1655 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1656 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1657 startup_sec /= NSEC_PER_SEC;
1659 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1663 /* Start all threads running. */
1664 pthread_mutex_lock(&g->start_work_mutex);
1665 g->start_work = true;
1666 pthread_mutex_unlock(&g->start_work_mutex);
1667 pthread_cond_broadcast(&g->start_work_cond);
1669 gettimeofday(&start, NULL);
1672 /* Parent process: */
1675 for (i = 0; i < g->p.nr_proc; i++) {
1676 wpid = waitpid(pids[i], &wait_stat, 0);
1678 BUG_ON(!WIFEXITED(wait_stat));
1683 runtime_ns_min = -1LL;
1685 for (t = 0; t < g->p.nr_tasks; t++) {
1686 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1688 runtime_ns_sum += thread_runtime_ns;
1689 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1692 gettimeofday(&stop, NULL);
1693 timersub(&stop, &start, &diff);
1695 BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1697 tprintf("\n ###\n");
1700 runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1701 runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1702 runtime_sec_max /= NSEC_PER_SEC;
1704 runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1706 bytes = g->bytes_done;
1707 runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1709 if (g->p.measure_convergence) {
1710 print_res(name, runtime_sec_max,
1711 "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1714 print_res(name, runtime_sec_max,
1715 "secs,", "runtime-max/thread", "secs slowest (max) thread-runtime");
1717 print_res(name, runtime_sec_min,
1718 "secs,", "runtime-min/thread", "secs fastest (min) thread-runtime");
1720 print_res(name, runtime_avg,
1721 "secs,", "runtime-avg/thread", "secs average thread-runtime");
1723 delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1724 print_res(name, delta_runtime / runtime_sec_max * 100.0,
1725 "%,", "spread-runtime/thread", "% difference between max/avg runtime");
1727 print_res(name, bytes / g->p.nr_tasks / 1e9,
1728 "GB,", "data/thread", "GB data processed, per thread");
1730 print_res(name, bytes / 1e9,
1731 "GB,", "data-total", "GB data processed, total");
1733 print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1734 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1736 print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1737 "GB/sec,", "thread-speed", "GB/sec/thread speed");
1739 print_res(name, bytes / runtime_sec_max / 1e9,
1740 "GB/sec,", "total-speed", "GB/sec total speed");
1742 if (g->p.show_details >= 2) {
1743 char tname[14 + 2 * 11 + 1];
1744 struct thread_data *td;
1745 for (p = 0; p < g->p.nr_proc; p++) {
1746 for (t = 0; t < g->p.nr_threads; t++) {
1747 memset(tname, 0, sizeof(tname));
1748 td = g->threads + p*g->p.nr_threads + t;
1749 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1750 print_res(tname, td->speed_gbs,
1751 "GB/sec", "thread-speed", "GB/sec/thread speed");
1752 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1753 "secs", "thread-system-time", "system CPU time/thread");
1754 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1755 "secs", "thread-user-time", "user CPU time/thread");
1769 static int command_size(const char **argv)
1778 BUG_ON(size >= MAX_ARGS);
1783 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1787 printf("\n # Running %s \"perf bench numa", name);
1789 for (i = 0; i < argc; i++)
1790 printf(" %s", argv[i]);
1794 memset(p, 0, sizeof(*p));
1796 /* Initialize nonzero defaults: */
1798 p->serialize_startup = 1;
1799 p->data_reads = true;
1800 p->data_writes = true;
1801 p->data_backwards = true;
1802 p->data_rand_walk = true;
1804 p->init_random = true;
1805 p->mb_global_str = "1";
1809 p->run_all = argc == 1;
1812 static int run_bench_numa(const char *name, const char **argv)
1814 int argc = command_size(argv);
1816 init_params(&p0, name, argc, argv);
1817 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1821 if (__bench_numa(name))
1830 #define OPT_BW_RAM "-s", "20", "-zZq", "--thp", " 1", "--no-data_rand_walk"
1831 #define OPT_BW_RAM_NOTHP OPT_BW_RAM, "--thp", "-1"
1833 #define OPT_CONV "-s", "100", "-zZ0qcm", "--thp", " 1"
1834 #define OPT_CONV_NOTHP OPT_CONV, "--thp", "-1"
1836 #define OPT_BW "-s", "20", "-zZ0q", "--thp", " 1"
1837 #define OPT_BW_NOTHP OPT_BW, "--thp", "-1"
1840 * The built-in test-suite executed by "perf bench numa -a".
1842 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1844 static const char *tests[][MAX_ARGS] = {
1845 /* Basic single-stream NUMA bandwidth measurements: */
1846 { "RAM-bw-local,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1847 "-C" , "0", "-M", "0", OPT_BW_RAM },
1848 { "RAM-bw-local-NOTHP,",
1849 "mem", "-p", "1", "-t", "1", "-P", "1024",
1850 "-C" , "0", "-M", "0", OPT_BW_RAM_NOTHP },
1851 { "RAM-bw-remote,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1852 "-C" , "0", "-M", "1", OPT_BW_RAM },
1854 /* 2-stream NUMA bandwidth measurements: */
1855 { "RAM-bw-local-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1856 "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1857 { "RAM-bw-remote-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1858 "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1860 /* Cross-stream NUMA bandwidth measurement: */
1861 { "RAM-bw-cross,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1862 "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1864 /* Convergence latency measurements: */
1865 { " 1x3-convergence,", "mem", "-p", "1", "-t", "3", "-P", "512", OPT_CONV },
1866 { " 1x4-convergence,", "mem", "-p", "1", "-t", "4", "-P", "512", OPT_CONV },
1867 { " 1x6-convergence,", "mem", "-p", "1", "-t", "6", "-P", "1020", OPT_CONV },
1868 { " 2x3-convergence,", "mem", "-p", "2", "-t", "3", "-P", "1020", OPT_CONV },
1869 { " 3x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1870 { " 4x4-convergence,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV },
1871 { " 4x4-convergence-NOTHP,",
1872 "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1873 { " 4x6-convergence,", "mem", "-p", "4", "-t", "6", "-P", "1020", OPT_CONV },
1874 { " 4x8-convergence,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_CONV },
1875 { " 8x4-convergence,", "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV },
1876 { " 8x4-convergence-NOTHP,",
1877 "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1878 { " 3x1-convergence,", "mem", "-p", "3", "-t", "1", "-P", "512", OPT_CONV },
1879 { " 4x1-convergence,", "mem", "-p", "4", "-t", "1", "-P", "512", OPT_CONV },
1880 { " 8x1-convergence,", "mem", "-p", "8", "-t", "1", "-P", "512", OPT_CONV },
1881 { "16x1-convergence,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_CONV },
1882 { "32x1-convergence,", "mem", "-p", "32", "-t", "1", "-P", "128", OPT_CONV },
1884 /* Various NUMA process/thread layout bandwidth measurements: */
1885 { " 2x1-bw-process,", "mem", "-p", "2", "-t", "1", "-P", "1024", OPT_BW },
1886 { " 3x1-bw-process,", "mem", "-p", "3", "-t", "1", "-P", "1024", OPT_BW },
1887 { " 4x1-bw-process,", "mem", "-p", "4", "-t", "1", "-P", "1024", OPT_BW },
1888 { " 8x1-bw-process,", "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW },
1889 { " 8x1-bw-process-NOTHP,",
1890 "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW_NOTHP },
1891 { "16x1-bw-process,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_BW },
1893 { " 1x4-bw-thread,", "mem", "-p", "1", "-t", "4", "-T", "256", OPT_BW },
1894 { " 1x8-bw-thread,", "mem", "-p", "1", "-t", "8", "-T", "256", OPT_BW },
1895 { "1x16-bw-thread,", "mem", "-p", "1", "-t", "16", "-T", "128", OPT_BW },
1896 { "1x32-bw-thread,", "mem", "-p", "1", "-t", "32", "-T", "64", OPT_BW },
1898 { " 2x3-bw-process,", "mem", "-p", "2", "-t", "3", "-P", "512", OPT_BW },
1899 { " 4x4-bw-process,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_BW },
1900 { " 4x6-bw-process,", "mem", "-p", "4", "-t", "6", "-P", "512", OPT_BW },
1901 { " 4x8-bw-process,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW },
1902 { " 4x8-bw-process-NOTHP,",
1903 "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW_NOTHP },
1904 { " 3x3-bw-process,", "mem", "-p", "3", "-t", "3", "-P", "512", OPT_BW },
1905 { " 5x5-bw-process,", "mem", "-p", "5", "-t", "5", "-P", "512", OPT_BW },
1907 { "2x16-bw-process,", "mem", "-p", "2", "-t", "16", "-P", "512", OPT_BW },
1908 { "1x32-bw-process,", "mem", "-p", "1", "-t", "32", "-P", "2048", OPT_BW },
1910 { "numa02-bw,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW },
1911 { "numa02-bw-NOTHP,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW_NOTHP },
1912 { "numa01-bw-thread,", "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW },
1913 { "numa01-bw-thread-NOTHP,",
1914 "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW_NOTHP },
1917 static int bench_all(void)
1919 int nr = ARRAY_SIZE(tests);
1923 ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1926 for (i = 0; i < nr; i++) {
1927 run_bench_numa(tests[i][0], tests[i] + 1);
1935 int bench_numa(int argc, const char **argv)
1937 init_params(&p0, "main,", argc, argv);
1938 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1945 if (__bench_numa(NULL))
1951 usage_with_options(numa_usage, options);