1 // SPDX-License-Identifier: GPL-2.0
5 * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
10 #include <subcmd/parse-options.h>
11 #include "../util/cloexec.h"
27 #include <sys/resource.h>
29 #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"
38 #include "../util/mutex.h"
39 #include <api/fs/fs.h>
44 # define RUSAGE_THREAD 1
48 * Regular printout to the terminal, suppressed if -q is specified:
50 #define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
56 #define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
60 cpu_set_t *bind_cpumask;
66 unsigned int loops_done;
72 struct mutex *process_lock;
75 /* Parameters set by options: */
78 /* Startup synchronization: */
79 bool serialize_startup;
85 /* Working set sizes: */
86 const char *mb_global_str;
87 const char *mb_proc_str;
88 const char *mb_proc_locked_str;
89 const char *mb_thread_str;
93 double mb_proc_locked;
96 /* Access patterns to the working set: */
100 bool data_zero_memset;
106 /* Working set initialization: */
118 long bytes_process_locked;
123 bool show_convergence;
124 bool measure_convergence;
130 /* Affinity options -C and -N: */
136 /* Global, read-writable area, accessible to all processes and threads: */
141 struct mutex startup_mutex;
142 struct cond startup_cond;
143 int nr_tasks_started;
145 struct mutex start_work_mutex;
146 struct cond start_work_cond;
147 int nr_tasks_working;
150 struct mutex stop_work_mutex;
153 struct thread_data *threads;
155 /* Convergence latency measurement: */
164 static struct global_info *g = NULL;
166 static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
167 static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
171 static const struct option options[] = {
172 OPT_INTEGER('p', "nr_proc" , &p0.nr_proc, "number of processes"),
173 OPT_INTEGER('t', "nr_threads" , &p0.nr_threads, "number of threads per process"),
175 OPT_STRING('G', "mb_global" , &p0.mb_global_str, "MB", "global memory (MBs)"),
176 OPT_STRING('P', "mb_proc" , &p0.mb_proc_str, "MB", "process memory (MBs)"),
177 OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
178 OPT_STRING('T', "mb_thread" , &p0.mb_thread_str, "MB", "thread memory (MBs)"),
180 OPT_UINTEGER('l', "nr_loops" , &p0.nr_loops, "max number of loops to run (default: unlimited)"),
181 OPT_UINTEGER('s', "nr_secs" , &p0.nr_secs, "max number of seconds to run (default: 5 secs)"),
182 OPT_UINTEGER('u', "usleep" , &p0.sleep_usecs, "usecs to sleep per loop iteration"),
184 OPT_BOOLEAN('R', "data_reads" , &p0.data_reads, "access the data via reads (can be mixed with -W)"),
185 OPT_BOOLEAN('W', "data_writes" , &p0.data_writes, "access the data via writes (can be mixed with -R)"),
186 OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards, "access the data backwards as well"),
187 OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
188 OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk, "access the data with random (32bit LFSR) walk"),
191 OPT_BOOLEAN('z', "init_zero" , &p0.init_zero, "bzero the initial allocations"),
192 OPT_BOOLEAN('I', "init_random" , &p0.init_random, "randomize the contents of the initial allocations"),
193 OPT_BOOLEAN('0', "init_cpu0" , &p0.init_cpu0, "do the initial allocations on CPU#0"),
194 OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs, "perturb thread 0/0 every X secs, to test convergence stability"),
196 OPT_INCR ('d', "show_details" , &p0.show_details, "Show details"),
197 OPT_INCR ('a', "all" , &p0.run_all, "Run all tests in the suite"),
198 OPT_INTEGER('H', "thp" , &p0.thp, "MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
199 OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details, "
200 "convergence is reached when each process (all its threads) is running on a single NUMA node."),
201 OPT_BOOLEAN('m', "measure_convergence", &p0.measure_convergence, "measure convergence latency"),
202 OPT_BOOLEAN('q', "quiet" , &quiet,
203 "quiet mode (do not show any warnings or messages)"),
204 OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
206 /* Special option string parsing callbacks: */
207 OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
208 "bind the first N tasks to these specific cpus (the rest is unbound)",
210 OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
211 "bind the first N tasks to these specific memory nodes (the rest is unbound)",
216 static const char * const bench_numa_usage[] = {
217 "perf bench numa <options>",
221 static const char * const numa_usage[] = {
222 "perf bench numa mem [<options>]",
227 * To get number of numa nodes present.
229 static int nr_numa_nodes(void)
233 for (i = 0; i < g->p.nr_nodes; i++) {
234 if (numa_bitmask_isbitset(numa_nodes_ptr, i))
242 * To check if given numa node is present.
244 static int is_node_present(int node)
246 return numa_bitmask_isbitset(numa_nodes_ptr, node);
250 * To check given numa node has cpus.
252 static bool node_has_cpus(int node)
254 struct bitmask *cpumask = numa_allocate_cpumask();
255 bool ret = false; /* fall back to nocpus */
259 if (!numa_node_to_cpus(node, cpumask)) {
260 for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
261 if (numa_bitmask_isbitset(cpumask, cpu)) {
267 numa_free_cpumask(cpumask);
272 static cpu_set_t *bind_to_cpu(int target_cpu)
274 int nrcpus = numa_num_possible_cpus();
275 cpu_set_t *orig_mask, *mask;
278 orig_mask = CPU_ALLOC(nrcpus);
280 size = CPU_ALLOC_SIZE(nrcpus);
281 CPU_ZERO_S(size, orig_mask);
283 if (sched_getaffinity(0, size, orig_mask))
286 mask = CPU_ALLOC(nrcpus);
290 CPU_ZERO_S(size, mask);
292 if (target_cpu == -1) {
295 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
296 CPU_SET_S(cpu, size, mask);
298 if (target_cpu < 0 || target_cpu >= g->p.nr_cpus)
301 CPU_SET_S(target_cpu, size, mask);
304 if (sched_setaffinity(0, size, mask))
314 /* BUG_ON due to failure in allocation of orig_mask/mask */
319 static cpu_set_t *bind_to_node(int target_node)
321 int nrcpus = numa_num_possible_cpus();
323 cpu_set_t *orig_mask, *mask;
326 orig_mask = CPU_ALLOC(nrcpus);
328 size = CPU_ALLOC_SIZE(nrcpus);
329 CPU_ZERO_S(size, orig_mask);
331 if (sched_getaffinity(0, size, orig_mask))
334 mask = CPU_ALLOC(nrcpus);
338 CPU_ZERO_S(size, mask);
340 if (target_node == NUMA_NO_NODE) {
341 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
342 CPU_SET_S(cpu, size, mask);
344 struct bitmask *cpumask = numa_allocate_cpumask();
349 if (!numa_node_to_cpus(target_node, cpumask)) {
350 for (cpu = 0; cpu < (int)cpumask->size; cpu++) {
351 if (numa_bitmask_isbitset(cpumask, cpu))
352 CPU_SET_S(cpu, size, mask);
355 numa_free_cpumask(cpumask);
358 if (sched_setaffinity(0, size, mask))
368 /* BUG_ON due to failure in allocation of orig_mask/mask */
373 static void bind_to_cpumask(cpu_set_t *mask)
376 size_t size = CPU_ALLOC_SIZE(numa_num_possible_cpus());
378 ret = sched_setaffinity(0, size, mask);
385 static void mempol_restore(void)
389 ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
394 static void bind_to_memnode(int node)
396 struct bitmask *node_mask;
399 if (node == NUMA_NO_NODE)
402 node_mask = numa_allocate_nodemask();
405 numa_bitmask_clearall(node_mask);
406 numa_bitmask_setbit(node_mask, node);
408 ret = set_mempolicy(MPOL_BIND, node_mask->maskp, node_mask->size + 1);
409 dprintf("binding to node %d, mask: %016lx => %d\n", node, *node_mask->maskp, ret);
411 numa_bitmask_free(node_mask);
415 #define HPSIZE (2*1024*1024)
417 #define set_taskname(fmt...) \
421 snprintf(name, 20, fmt); \
422 prctl(PR_SET_NAME, name); \
425 static u8 *alloc_data(ssize_t bytes0, int map_flags,
426 int init_zero, int init_cpu0, int thp, int init_random)
428 cpu_set_t *orig_mask = NULL;
436 /* Allocate and initialize all memory on CPU#0: */
438 int node = numa_node_of_cpu(0);
440 orig_mask = bind_to_node(node);
441 bind_to_memnode(node);
444 bytes = bytes0 + HPSIZE;
446 buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
447 BUG_ON(buf == (void *)-1);
449 if (map_flags == MAP_PRIVATE) {
451 ret = madvise(buf, bytes, MADV_HUGEPAGE);
452 if (ret && !g->print_once) {
454 printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
458 ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
459 if (ret && !g->print_once) {
461 printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
469 /* Initialize random contents, different in each word: */
471 u64 *wbuf = (void *)buf;
475 for (i = 0; i < bytes/8; i++)
480 /* Align to 2MB boundary: */
481 buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
483 /* Restore affinity: */
485 bind_to_cpumask(orig_mask);
493 static void free_data(void *data, ssize_t bytes)
500 ret = munmap(data, bytes);
505 * Create a shared memory buffer that can be shared between processes, zeroed:
507 static void * zalloc_shared_data(ssize_t bytes)
509 return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0, g->p.thp, g->p.init_random);
513 * Create a shared memory buffer that can be shared between processes:
515 static void * setup_shared_data(ssize_t bytes)
517 return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
521 * Allocate process-local memory - this will either be shared between
522 * threads of this process, or only be accessed by this thread:
524 static void * setup_private_data(ssize_t bytes)
526 return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0, g->p.thp, g->p.init_random);
529 static int parse_cpu_list(const char *arg)
531 p0.cpu_list_str = strdup(arg);
533 dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
539 * Check whether a CPU is online
542 * 1 -> if CPU is online
543 * 0 -> if CPU is offline
546 static int is_cpu_online(unsigned int cpu)
554 snprintf(buf, sizeof(buf),
555 "/sys/devices/system/cpu/cpu%d", cpu);
556 if (stat(buf, &statbuf) != 0)
560 * Check if /sys/devices/system/cpu/cpux/online file
561 * exists. Some cases cpu0 won't have online file since
562 * it is not expected to be turned off generally.
563 * In kernels without CONFIG_HOTPLUG_CPU, this
566 snprintf(buf, sizeof(buf),
567 "/sys/devices/system/cpu/cpu%d/online", cpu);
568 if (stat(buf, &statbuf) != 0)
572 * Read online file using sysfs__read_str.
573 * If read or open fails, return -1.
574 * If read succeeds, return value from file
575 * which gets stored in "str"
577 snprintf(buf, sizeof(buf),
578 "devices/system/cpu/cpu%d/online", cpu);
580 if (sysfs__read_str(buf, &str, &strlen) < 0)
589 static int parse_setup_cpu_list(void)
591 struct thread_data *td;
595 if (!g->p.cpu_list_str)
598 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
600 str0 = str = strdup(g->p.cpu_list_str);
605 tprintf("# binding tasks to CPUs:\n");
609 int bind_cpu, bind_cpu_0, bind_cpu_1;
610 char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
615 tok = strsep(&str, ",");
619 tok_end = strstr(tok, "-");
621 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
623 /* Single CPU specified: */
624 bind_cpu_0 = bind_cpu_1 = atol(tok);
626 /* CPU range specified (for example: "5-11"): */
627 bind_cpu_0 = atol(tok);
628 bind_cpu_1 = atol(tok_end + 1);
632 tok_step = strstr(tok, "#");
634 step = atol(tok_step + 1);
635 BUG_ON(step <= 0 || step >= g->p.nr_cpus);
640 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
641 * where the _4 means the next 4 CPUs are allowed.
644 tok_len = strstr(tok, "_");
646 bind_len = atol(tok_len + 1);
647 BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
650 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
652 tok_mul = strstr(tok, "x");
654 mul = atol(tok_mul + 1);
658 dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
660 if (bind_cpu_0 >= g->p.nr_cpus || bind_cpu_1 >= g->p.nr_cpus) {
661 printf("\nTest not applicable, system has only %d CPUs.\n", g->p.nr_cpus);
665 if (is_cpu_online(bind_cpu_0) != 1 || is_cpu_online(bind_cpu_1) != 1) {
666 printf("\nTest not applicable, bind_cpu_0 or bind_cpu_1 is offline\n");
670 BUG_ON(bind_cpu_0 < 0 || bind_cpu_1 < 0);
671 BUG_ON(bind_cpu_0 > bind_cpu_1);
673 for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
674 size_t size = CPU_ALLOC_SIZE(g->p.nr_cpus);
677 for (i = 0; i < mul; i++) {
680 if (t >= g->p.nr_tasks) {
681 printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
689 tprintf("%2d/%d", bind_cpu, bind_len);
691 tprintf("%2d", bind_cpu);
694 td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
695 BUG_ON(!td->bind_cpumask);
696 CPU_ZERO_S(size, td->bind_cpumask);
697 for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
698 if (cpu < 0 || cpu >= g->p.nr_cpus) {
699 CPU_FREE(td->bind_cpumask);
702 CPU_SET_S(cpu, size, td->bind_cpumask);
712 if (t < g->p.nr_tasks)
713 printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
719 static int parse_cpus_opt(const struct option *opt __maybe_unused,
720 const char *arg, int unset __maybe_unused)
725 return parse_cpu_list(arg);
728 static int parse_node_list(const char *arg)
730 p0.node_list_str = strdup(arg);
732 dprintf("got NODE list: {%s}\n", p0.node_list_str);
737 static int parse_setup_node_list(void)
739 struct thread_data *td;
743 if (!g->p.node_list_str)
746 dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
748 str0 = str = strdup(g->p.node_list_str);
753 tprintf("# binding tasks to NODEs:\n");
757 int bind_node, bind_node_0, bind_node_1;
758 char *tok, *tok_end, *tok_step, *tok_mul;
762 tok = strsep(&str, ",");
766 tok_end = strstr(tok, "-");
768 dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
770 /* Single NODE specified: */
771 bind_node_0 = bind_node_1 = atol(tok);
773 /* NODE range specified (for example: "5-11"): */
774 bind_node_0 = atol(tok);
775 bind_node_1 = atol(tok_end + 1);
779 tok_step = strstr(tok, "#");
781 step = atol(tok_step + 1);
782 BUG_ON(step <= 0 || step >= g->p.nr_nodes);
785 /* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
787 tok_mul = strstr(tok, "x");
789 mul = atol(tok_mul + 1);
793 dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
795 if (bind_node_0 >= g->p.nr_nodes || bind_node_1 >= g->p.nr_nodes) {
796 printf("\nTest not applicable, system has only %d nodes.\n", g->p.nr_nodes);
800 BUG_ON(bind_node_0 < 0 || bind_node_1 < 0);
801 BUG_ON(bind_node_0 > bind_node_1);
803 for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
806 for (i = 0; i < mul; i++) {
807 if (t >= g->p.nr_tasks || !node_has_cpus(bind_node)) {
808 printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
814 tprintf(" %2d", bind_node);
816 tprintf(",%2d", bind_node);
818 td->bind_node = bind_node;
827 if (t < g->p.nr_tasks)
828 printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
834 static int parse_nodes_opt(const struct option *opt __maybe_unused,
835 const char *arg, int unset __maybe_unused)
840 return parse_node_list(arg);
843 static inline uint32_t lfsr_32(uint32_t lfsr)
845 const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
846 return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
850 * Make sure there's real data dependency to RAM (when read
851 * accesses are enabled), so the compiler, the CPU and the
852 * kernel (KSM, zero page, etc.) cannot optimize away RAM
855 static inline u64 access_data(u64 *data, u64 val)
859 if (g->p.data_writes)
865 * The worker process does two types of work, a forwards going
866 * loop and a backwards going loop.
868 * We do this so that on multiprocessor systems we do not create
869 * a 'train' of processing, with highly synchronized processes,
870 * skewing the whole benchmark.
872 static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
874 long words = bytes/sizeof(u64);
875 u64 *data = (void *)__data;
876 long chunk_0, chunk_1;
881 BUG_ON(!data && words);
882 BUG_ON(data && !words);
887 /* Very simple memset() work variant: */
888 if (g->p.data_zero_memset && !g->p.data_rand_walk) {
893 /* Spread out by PID/TID nr and by loop nr: */
894 chunk_0 = words/nr_max;
895 chunk_1 = words/g->p.nr_loops;
896 off = nr*chunk_0 + loop*chunk_1;
901 if (g->p.data_rand_walk) {
902 u32 lfsr = nr + loop + val;
905 for (i = 0; i < words/1024; i++) {
908 lfsr = lfsr_32(lfsr);
910 start = lfsr % words;
911 end = min(start + 1024, words-1);
913 if (g->p.data_zero_memset) {
914 bzero(data + start, (end-start) * sizeof(u64));
916 for (j = start; j < end; j++)
917 val = access_data(data + j, val);
920 } else if (!g->p.data_backwards || (nr + loop) & 1) {
921 /* Process data forwards: */
928 if (unlikely(d >= d1))
930 if (unlikely(d == d0))
933 val = access_data(d, val);
938 /* Process data backwards: */
945 if (unlikely(d < data))
947 if (unlikely(d == d0))
950 val = access_data(d, val);
959 static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
963 cpu = sched_getcpu();
965 g->threads[task_nr].curr_cpu = cpu;
966 prctl(0, bytes_worked);
970 * Count the number of nodes a process's threads
973 * A count of 1 means that the process is compressed
974 * to a single node. A count of g->p.nr_nodes means it's
975 * spread out on the whole system.
977 static int count_process_nodes(int process_nr)
983 node_present = (char *)malloc(g->p.nr_nodes * sizeof(char));
984 BUG_ON(!node_present);
985 for (nodes = 0; nodes < g->p.nr_nodes; nodes++)
986 node_present[nodes] = 0;
988 for (t = 0; t < g->p.nr_threads; t++) {
989 struct thread_data *td;
993 task_nr = process_nr*g->p.nr_threads + t;
994 td = g->threads + task_nr;
996 node = numa_node_of_cpu(td->curr_cpu);
997 if (node < 0) /* curr_cpu was likely still -1 */ {
1002 node_present[node] = 1;
1007 for (n = 0; n < g->p.nr_nodes; n++)
1008 nodes += node_present[n];
1015 * Count the number of distinct process-threads a node contains.
1017 * A count of 1 means that the node contains only a single
1018 * process. If all nodes on the system contain at most one
1019 * process then we are well-converged.
1021 static int count_node_processes(int node)
1026 for (p = 0; p < g->p.nr_proc; p++) {
1027 for (t = 0; t < g->p.nr_threads; t++) {
1028 struct thread_data *td;
1032 task_nr = p*g->p.nr_threads + t;
1033 td = g->threads + task_nr;
1035 n = numa_node_of_cpu(td->curr_cpu);
1046 static void calc_convergence_compression(int *strong)
1048 unsigned int nodes_min, nodes_max;
1054 for (p = 0; p < g->p.nr_proc; p++) {
1055 unsigned int nodes = count_process_nodes(p);
1062 nodes_min = min(nodes, nodes_min);
1063 nodes_max = max(nodes, nodes_max);
1066 /* Strong convergence: all threads compress on a single node: */
1067 if (nodes_min == 1 && nodes_max == 1) {
1071 tprintf(" {%d-%d}", nodes_min, nodes_max);
1075 static void calc_convergence(double runtime_ns_max, double *convergence)
1077 unsigned int loops_done_min, loops_done_max;
1090 if (!g->p.show_convergence && !g->p.measure_convergence)
1093 nodes = (int *)malloc(g->p.nr_nodes * sizeof(int));
1095 for (node = 0; node < g->p.nr_nodes; node++)
1098 loops_done_min = -1;
1101 for (t = 0; t < g->p.nr_tasks; t++) {
1102 struct thread_data *td = g->threads + t;
1103 unsigned int loops_done;
1107 /* Not all threads have written it yet: */
1111 node = numa_node_of_cpu(cpu);
1115 loops_done = td->loops_done;
1116 loops_done_min = min(loops_done, loops_done_min);
1117 loops_done_max = max(loops_done, loops_done_max);
1121 nr_min = g->p.nr_tasks;
1124 for (node = 0; node < g->p.nr_nodes; node++) {
1125 if (!is_node_present(node))
1128 nr_min = min(nr, nr_min);
1129 nr_max = max(nr, nr_max);
1132 BUG_ON(nr_min > nr_max);
1134 BUG_ON(sum > g->p.nr_tasks);
1136 if (0 && (sum < g->p.nr_tasks)) {
1142 * Count the number of distinct process groups present
1143 * on nodes - when we are converged this will decrease
1148 for (node = 0; node < g->p.nr_nodes; node++) {
1151 if (!is_node_present(node))
1153 processes = count_node_processes(node);
1155 tprintf(" %2d/%-2d", nr, processes);
1157 process_groups += processes;
1160 distance = nr_max - nr_min;
1162 tprintf(" [%2d/%-2d]", distance, process_groups);
1164 tprintf(" l:%3d-%-3d (%3d)",
1165 loops_done_min, loops_done_max, loops_done_max-loops_done_min);
1167 if (loops_done_min && loops_done_max) {
1168 double skew = 1.0 - (double)loops_done_min/loops_done_max;
1170 tprintf(" [%4.1f%%]", skew * 100.0);
1173 calc_convergence_compression(&strong);
1175 if (strong && process_groups == g->p.nr_proc) {
1176 if (!*convergence) {
1177 *convergence = runtime_ns_max;
1178 tprintf(" (%6.1fs converged)\n", *convergence / NSEC_PER_SEC);
1179 if (g->p.measure_convergence) {
1180 g->all_converged = true;
1181 g->stop_work = true;
1186 tprintf(" (%6.1fs de-converged)", runtime_ns_max / NSEC_PER_SEC);
1195 static void show_summary(double runtime_ns_max, int l, double *convergence)
1197 tprintf("\r # %5.1f%% [%.1f mins]",
1198 (double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max / NSEC_PER_SEC / 60.0);
1200 calc_convergence(runtime_ns_max, convergence);
1202 if (g->p.show_details >= 0)
1206 static void *worker_thread(void *__tdata)
1208 struct thread_data *td = __tdata;
1209 struct timeval start0, start, stop, diff;
1210 int process_nr = td->process_nr;
1211 int thread_nr = td->thread_nr;
1212 unsigned long last_perturbance;
1213 int task_nr = td->task_nr;
1214 int details = g->p.show_details;
1215 int first_task, last_task;
1216 double convergence = 0;
1218 double runtime_ns_max;
1222 u64 bytes_done, secs;
1225 struct rusage rusage;
1227 bind_to_cpumask(td->bind_cpumask);
1228 bind_to_memnode(td->bind_node);
1230 set_taskname("thread %d/%d", process_nr, thread_nr);
1232 global_data = g->data;
1233 process_data = td->process_data;
1234 thread_data = setup_private_data(g->p.bytes_thread);
1239 if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
1243 if (process_nr == 0 && thread_nr == 0)
1247 printf("# thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
1248 process_nr, thread_nr, global_data, process_data, thread_data);
1251 if (g->p.serialize_startup) {
1252 mutex_lock(&g->startup_mutex);
1253 g->nr_tasks_started++;
1254 /* The last thread wakes the main process. */
1255 if (g->nr_tasks_started == g->p.nr_tasks)
1256 cond_signal(&g->startup_cond);
1258 mutex_unlock(&g->startup_mutex);
1260 /* Here we will wait for the main process to start us all at once: */
1261 mutex_lock(&g->start_work_mutex);
1262 g->start_work = false;
1263 g->nr_tasks_working++;
1264 while (!g->start_work)
1265 cond_wait(&g->start_work_cond, &g->start_work_mutex);
1267 mutex_unlock(&g->start_work_mutex);
1270 gettimeofday(&start0, NULL);
1272 start = stop = start0;
1273 last_perturbance = start.tv_sec;
1275 for (l = 0; l < g->p.nr_loops; l++) {
1281 val += do_work(global_data, g->p.bytes_global, process_nr, g->p.nr_proc, l, val);
1282 val += do_work(process_data, g->p.bytes_process, thread_nr, g->p.nr_threads, l, val);
1283 val += do_work(thread_data, g->p.bytes_thread, 0, 1, l, val);
1285 if (g->p.sleep_usecs) {
1286 mutex_lock(td->process_lock);
1287 usleep(g->p.sleep_usecs);
1288 mutex_unlock(td->process_lock);
1291 * Amount of work to be done under a process-global lock:
1293 if (g->p.bytes_process_locked) {
1294 mutex_lock(td->process_lock);
1295 val += do_work(process_data, g->p.bytes_process_locked, thread_nr, g->p.nr_threads, l, val);
1296 mutex_unlock(td->process_lock);
1299 work_done = g->p.bytes_global + g->p.bytes_process +
1300 g->p.bytes_process_locked + g->p.bytes_thread;
1302 update_curr_cpu(task_nr, work_done);
1303 bytes_done += work_done;
1305 if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
1310 gettimeofday(&stop, NULL);
1312 /* Check whether our max runtime timed out: */
1314 timersub(&stop, &start0, &diff);
1315 if ((u32)diff.tv_sec >= g->p.nr_secs) {
1316 g->stop_work = true;
1321 /* Update the summary at most once per second: */
1322 if (start.tv_sec == stop.tv_sec)
1326 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
1327 * by migrating to CPU#0:
1329 if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
1330 cpu_set_t *orig_mask;
1334 last_perturbance = stop.tv_sec;
1337 * Depending on where we are running, move into
1338 * the other half of the system, to create some
1341 this_cpu = g->threads[task_nr].curr_cpu;
1342 if (this_cpu < g->p.nr_cpus/2)
1343 target_cpu = g->p.nr_cpus-1;
1347 orig_mask = bind_to_cpu(target_cpu);
1349 /* Here we are running on the target CPU already */
1351 printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
1353 bind_to_cpumask(orig_mask);
1354 CPU_FREE(orig_mask);
1358 timersub(&stop, &start, &diff);
1359 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1360 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1363 printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016"PRIx64"]\n",
1364 process_nr, thread_nr, runtime_ns_max / bytes_done, val);
1371 timersub(&stop, &start0, &diff);
1372 runtime_ns_max = diff.tv_sec * NSEC_PER_SEC;
1373 runtime_ns_max += diff.tv_usec * NSEC_PER_USEC;
1375 show_summary(runtime_ns_max, l, &convergence);
1378 gettimeofday(&stop, NULL);
1379 timersub(&stop, &start0, &diff);
1380 td->runtime_ns = diff.tv_sec * NSEC_PER_SEC;
1381 td->runtime_ns += diff.tv_usec * NSEC_PER_USEC;
1382 secs = td->runtime_ns / NSEC_PER_SEC;
1383 td->speed_gbs = secs ? bytes_done / secs / 1e9 : 0;
1385 getrusage(RUSAGE_THREAD, &rusage);
1386 td->system_time_ns = rusage.ru_stime.tv_sec * NSEC_PER_SEC;
1387 td->system_time_ns += rusage.ru_stime.tv_usec * NSEC_PER_USEC;
1388 td->user_time_ns = rusage.ru_utime.tv_sec * NSEC_PER_SEC;
1389 td->user_time_ns += rusage.ru_utime.tv_usec * NSEC_PER_USEC;
1391 free_data(thread_data, g->p.bytes_thread);
1393 mutex_lock(&g->stop_work_mutex);
1394 g->bytes_done += bytes_done;
1395 mutex_unlock(&g->stop_work_mutex);
1401 * A worker process starts a couple of threads:
1403 static void worker_process(int process_nr)
1405 struct mutex process_lock;
1406 struct thread_data *td;
1407 pthread_t *pthreads;
1413 mutex_init(&process_lock);
1414 set_taskname("process %d", process_nr);
1417 * Pick up the memory policy and the CPU binding of our first thread,
1418 * so that we initialize memory accordingly:
1420 task_nr = process_nr*g->p.nr_threads;
1421 td = g->threads + task_nr;
1423 bind_to_memnode(td->bind_node);
1424 bind_to_cpumask(td->bind_cpumask);
1426 pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
1427 process_data = setup_private_data(g->p.bytes_process);
1429 if (g->p.show_details >= 3) {
1430 printf(" # process %2d global mem: %p, process mem: %p\n",
1431 process_nr, g->data, process_data);
1434 for (t = 0; t < g->p.nr_threads; t++) {
1435 task_nr = process_nr*g->p.nr_threads + t;
1436 td = g->threads + task_nr;
1438 td->process_data = process_data;
1439 td->process_nr = process_nr;
1441 td->task_nr = task_nr;
1444 td->process_lock = &process_lock;
1446 ret = pthread_create(pthreads + t, NULL, worker_thread, td);
1450 for (t = 0; t < g->p.nr_threads; t++) {
1451 ret = pthread_join(pthreads[t], NULL);
1455 free_data(process_data, g->p.bytes_process);
1459 static void print_summary(void)
1461 if (g->p.show_details < 0)
1465 printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
1466 g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", nr_numa_nodes(), g->p.nr_cpus);
1467 printf(" # %5dx %5ldMB global shared mem operations\n",
1468 g->p.nr_loops, g->p.bytes_global/1024/1024);
1469 printf(" # %5dx %5ldMB process shared mem operations\n",
1470 g->p.nr_loops, g->p.bytes_process/1024/1024);
1471 printf(" # %5dx %5ldMB thread local mem operations\n",
1472 g->p.nr_loops, g->p.bytes_thread/1024/1024);
1476 printf("\n ###\n"); fflush(stdout);
1479 static void init_thread_data(void)
1481 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1484 g->threads = zalloc_shared_data(size);
1486 for (t = 0; t < g->p.nr_tasks; t++) {
1487 struct thread_data *td = g->threads + t;
1488 size_t cpuset_size = CPU_ALLOC_SIZE(g->p.nr_cpus);
1491 /* Allow all nodes by default: */
1492 td->bind_node = NUMA_NO_NODE;
1494 /* Allow all CPUs by default: */
1495 td->bind_cpumask = CPU_ALLOC(g->p.nr_cpus);
1496 BUG_ON(!td->bind_cpumask);
1497 CPU_ZERO_S(cpuset_size, td->bind_cpumask);
1498 for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
1499 CPU_SET_S(cpu, cpuset_size, td->bind_cpumask);
1503 static void deinit_thread_data(void)
1505 ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
1508 /* Free the bind_cpumask allocated for thread_data */
1509 for (t = 0; t < g->p.nr_tasks; t++) {
1510 struct thread_data *td = g->threads + t;
1511 CPU_FREE(td->bind_cpumask);
1514 free_data(g->threads, size);
1517 static int init(void)
1519 g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
1521 /* Copy over options: */
1524 g->p.nr_cpus = numa_num_configured_cpus();
1526 g->p.nr_nodes = numa_max_node() + 1;
1528 /* char array in count_process_nodes(): */
1529 BUG_ON(g->p.nr_nodes < 0);
1531 if (quiet && !g->p.show_details)
1532 g->p.show_details = -1;
1534 /* Some memory should be specified: */
1535 if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
1538 if (g->p.mb_global_str) {
1539 g->p.mb_global = atof(g->p.mb_global_str);
1540 BUG_ON(g->p.mb_global < 0);
1543 if (g->p.mb_proc_str) {
1544 g->p.mb_proc = atof(g->p.mb_proc_str);
1545 BUG_ON(g->p.mb_proc < 0);
1548 if (g->p.mb_proc_locked_str) {
1549 g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
1550 BUG_ON(g->p.mb_proc_locked < 0);
1551 BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
1554 if (g->p.mb_thread_str) {
1555 g->p.mb_thread = atof(g->p.mb_thread_str);
1556 BUG_ON(g->p.mb_thread < 0);
1559 BUG_ON(g->p.nr_threads <= 0);
1560 BUG_ON(g->p.nr_proc <= 0);
1562 g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
1564 g->p.bytes_global = g->p.mb_global *1024L*1024L;
1565 g->p.bytes_process = g->p.mb_proc *1024L*1024L;
1566 g->p.bytes_process_locked = g->p.mb_proc_locked *1024L*1024L;
1567 g->p.bytes_thread = g->p.mb_thread *1024L*1024L;
1569 g->data = setup_shared_data(g->p.bytes_global);
1571 /* Startup serialization: */
1572 mutex_init_pshared(&g->start_work_mutex);
1573 cond_init_pshared(&g->start_work_cond);
1574 mutex_init_pshared(&g->startup_mutex);
1575 cond_init_pshared(&g->startup_cond);
1576 mutex_init_pshared(&g->stop_work_mutex);
1581 if (parse_setup_cpu_list() || parse_setup_node_list())
1590 static void deinit(void)
1592 free_data(g->data, g->p.bytes_global);
1595 deinit_thread_data();
1597 free_data(g, sizeof(*g));
1602 * Print a short or long result, depending on the verbosity setting:
1604 static void print_res(const char *name, double val,
1605 const char *txt_unit, const char *txt_short, const char *txt_long)
1611 printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
1613 printf(" %14.3f %s\n", val, txt_long);
1616 static int __bench_numa(const char *name)
1618 struct timeval start, stop, diff;
1619 u64 runtime_ns_min, runtime_ns_sum;
1620 pid_t *pids, pid, wpid;
1621 double delta_runtime;
1623 double runtime_sec_max;
1624 double runtime_sec_min;
1632 pids = zalloc(g->p.nr_proc * sizeof(*pids));
1635 if (g->p.serialize_startup) {
1637 tprintf(" # Startup synchronization: ..."); fflush(stdout);
1640 gettimeofday(&start, NULL);
1642 for (i = 0; i < g->p.nr_proc; i++) {
1644 dprintf(" # process %2d: PID %d\n", i, pid);
1648 /* Child process: */
1657 if (g->p.serialize_startup) {
1658 bool threads_ready = false;
1662 * Wait for all the threads to start up. The last thread will
1663 * signal this process.
1665 mutex_lock(&g->startup_mutex);
1666 while (g->nr_tasks_started != g->p.nr_tasks)
1667 cond_wait(&g->startup_cond, &g->startup_mutex);
1669 mutex_unlock(&g->startup_mutex);
1671 /* Wait for all threads to be at the start_work_cond. */
1672 while (!threads_ready) {
1673 mutex_lock(&g->start_work_mutex);
1674 threads_ready = (g->nr_tasks_working == g->p.nr_tasks);
1675 mutex_unlock(&g->start_work_mutex);
1680 gettimeofday(&stop, NULL);
1682 timersub(&stop, &start, &diff);
1684 startup_sec = diff.tv_sec * NSEC_PER_SEC;
1685 startup_sec += diff.tv_usec * NSEC_PER_USEC;
1686 startup_sec /= NSEC_PER_SEC;
1688 tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
1692 /* Start all threads running. */
1693 mutex_lock(&g->start_work_mutex);
1694 g->start_work = true;
1695 mutex_unlock(&g->start_work_mutex);
1696 cond_broadcast(&g->start_work_cond);
1698 gettimeofday(&start, NULL);
1701 /* Parent process: */
1704 for (i = 0; i < g->p.nr_proc; i++) {
1705 wpid = waitpid(pids[i], &wait_stat, 0);
1707 BUG_ON(!WIFEXITED(wait_stat));
1712 runtime_ns_min = -1LL;
1714 for (t = 0; t < g->p.nr_tasks; t++) {
1715 u64 thread_runtime_ns = g->threads[t].runtime_ns;
1717 runtime_ns_sum += thread_runtime_ns;
1718 runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
1721 gettimeofday(&stop, NULL);
1722 timersub(&stop, &start, &diff);
1724 BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
1726 tprintf("\n ###\n");
1729 runtime_sec_max = diff.tv_sec * NSEC_PER_SEC;
1730 runtime_sec_max += diff.tv_usec * NSEC_PER_USEC;
1731 runtime_sec_max /= NSEC_PER_SEC;
1733 runtime_sec_min = runtime_ns_min / NSEC_PER_SEC;
1735 bytes = g->bytes_done;
1736 runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / NSEC_PER_SEC;
1738 if (g->p.measure_convergence) {
1739 print_res(name, runtime_sec_max,
1740 "secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
1743 print_res(name, runtime_sec_max,
1744 "secs,", "runtime-max/thread", "secs slowest (max) thread-runtime");
1746 print_res(name, runtime_sec_min,
1747 "secs,", "runtime-min/thread", "secs fastest (min) thread-runtime");
1749 print_res(name, runtime_avg,
1750 "secs,", "runtime-avg/thread", "secs average thread-runtime");
1752 delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
1753 print_res(name, delta_runtime / runtime_sec_max * 100.0,
1754 "%,", "spread-runtime/thread", "% difference between max/avg runtime");
1756 print_res(name, bytes / g->p.nr_tasks / 1e9,
1757 "GB,", "data/thread", "GB data processed, per thread");
1759 print_res(name, bytes / 1e9,
1760 "GB,", "data-total", "GB data processed, total");
1762 print_res(name, runtime_sec_max * NSEC_PER_SEC / (bytes / g->p.nr_tasks),
1763 "nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
1765 print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
1766 "GB/sec,", "thread-speed", "GB/sec/thread speed");
1768 print_res(name, bytes / runtime_sec_max / 1e9,
1769 "GB/sec,", "total-speed", "GB/sec total speed");
1771 if (g->p.show_details >= 2) {
1772 char tname[14 + 2 * 11 + 1];
1773 struct thread_data *td;
1774 for (p = 0; p < g->p.nr_proc; p++) {
1775 for (t = 0; t < g->p.nr_threads; t++) {
1776 memset(tname, 0, sizeof(tname));
1777 td = g->threads + p*g->p.nr_threads + t;
1778 snprintf(tname, sizeof(tname), "process%d:thread%d", p, t);
1779 print_res(tname, td->speed_gbs,
1780 "GB/sec", "thread-speed", "GB/sec/thread speed");
1781 print_res(tname, td->system_time_ns / NSEC_PER_SEC,
1782 "secs", "thread-system-time", "system CPU time/thread");
1783 print_res(tname, td->user_time_ns / NSEC_PER_SEC,
1784 "secs", "thread-user-time", "user CPU time/thread");
1798 static int command_size(const char **argv)
1807 BUG_ON(size >= MAX_ARGS);
1812 static void init_params(struct params *p, const char *name, int argc, const char **argv)
1816 printf("\n # Running %s \"perf bench numa", name);
1818 for (i = 0; i < argc; i++)
1819 printf(" %s", argv[i]);
1823 memset(p, 0, sizeof(*p));
1825 /* Initialize nonzero defaults: */
1827 p->serialize_startup = 1;
1828 p->data_reads = true;
1829 p->data_writes = true;
1830 p->data_backwards = true;
1831 p->data_rand_walk = true;
1833 p->init_random = true;
1834 p->mb_global_str = "1";
1838 p->run_all = argc == 1;
1841 static int run_bench_numa(const char *name, const char **argv)
1843 int argc = command_size(argv);
1845 init_params(&p0, name, argc, argv);
1846 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1850 if (__bench_numa(name))
1859 #define OPT_BW_RAM "-s", "20", "-zZq", "--thp", " 1", "--no-data_rand_walk"
1860 #define OPT_BW_RAM_NOTHP OPT_BW_RAM, "--thp", "-1"
1862 #define OPT_CONV "-s", "100", "-zZ0qcm", "--thp", " 1"
1863 #define OPT_CONV_NOTHP OPT_CONV, "--thp", "-1"
1865 #define OPT_BW "-s", "20", "-zZ0q", "--thp", " 1"
1866 #define OPT_BW_NOTHP OPT_BW, "--thp", "-1"
1869 * The built-in test-suite executed by "perf bench numa -a".
1871 * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
1873 static const char *tests[][MAX_ARGS] = {
1874 /* Basic single-stream NUMA bandwidth measurements: */
1875 { "RAM-bw-local,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1876 "-C" , "0", "-M", "0", OPT_BW_RAM },
1877 { "RAM-bw-local-NOTHP,",
1878 "mem", "-p", "1", "-t", "1", "-P", "1024",
1879 "-C" , "0", "-M", "0", OPT_BW_RAM_NOTHP },
1880 { "RAM-bw-remote,", "mem", "-p", "1", "-t", "1", "-P", "1024",
1881 "-C" , "0", "-M", "1", OPT_BW_RAM },
1883 /* 2-stream NUMA bandwidth measurements: */
1884 { "RAM-bw-local-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1885 "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
1886 { "RAM-bw-remote-2x,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1887 "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
1889 /* Cross-stream NUMA bandwidth measurement: */
1890 { "RAM-bw-cross,", "mem", "-p", "2", "-t", "1", "-P", "1024",
1891 "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
1893 /* Convergence latency measurements: */
1894 { " 1x3-convergence,", "mem", "-p", "1", "-t", "3", "-P", "512", OPT_CONV },
1895 { " 1x4-convergence,", "mem", "-p", "1", "-t", "4", "-P", "512", OPT_CONV },
1896 { " 1x6-convergence,", "mem", "-p", "1", "-t", "6", "-P", "1020", OPT_CONV },
1897 { " 2x3-convergence,", "mem", "-p", "2", "-t", "3", "-P", "1020", OPT_CONV },
1898 { " 3x3-convergence,", "mem", "-p", "3", "-t", "3", "-P", "1020", OPT_CONV },
1899 { " 4x4-convergence,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV },
1900 { " 4x4-convergence-NOTHP,",
1901 "mem", "-p", "4", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1902 { " 4x6-convergence,", "mem", "-p", "4", "-t", "6", "-P", "1020", OPT_CONV },
1903 { " 4x8-convergence,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_CONV },
1904 { " 8x4-convergence,", "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV },
1905 { " 8x4-convergence-NOTHP,",
1906 "mem", "-p", "8", "-t", "4", "-P", "512", OPT_CONV_NOTHP },
1907 { " 3x1-convergence,", "mem", "-p", "3", "-t", "1", "-P", "512", OPT_CONV },
1908 { " 4x1-convergence,", "mem", "-p", "4", "-t", "1", "-P", "512", OPT_CONV },
1909 { " 8x1-convergence,", "mem", "-p", "8", "-t", "1", "-P", "512", OPT_CONV },
1910 { "16x1-convergence,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_CONV },
1911 { "32x1-convergence,", "mem", "-p", "32", "-t", "1", "-P", "128", OPT_CONV },
1913 /* Various NUMA process/thread layout bandwidth measurements: */
1914 { " 2x1-bw-process,", "mem", "-p", "2", "-t", "1", "-P", "1024", OPT_BW },
1915 { " 3x1-bw-process,", "mem", "-p", "3", "-t", "1", "-P", "1024", OPT_BW },
1916 { " 4x1-bw-process,", "mem", "-p", "4", "-t", "1", "-P", "1024", OPT_BW },
1917 { " 8x1-bw-process,", "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW },
1918 { " 8x1-bw-process-NOTHP,",
1919 "mem", "-p", "8", "-t", "1", "-P", " 512", OPT_BW_NOTHP },
1920 { "16x1-bw-process,", "mem", "-p", "16", "-t", "1", "-P", "256", OPT_BW },
1922 { " 1x4-bw-thread,", "mem", "-p", "1", "-t", "4", "-T", "256", OPT_BW },
1923 { " 1x8-bw-thread,", "mem", "-p", "1", "-t", "8", "-T", "256", OPT_BW },
1924 { "1x16-bw-thread,", "mem", "-p", "1", "-t", "16", "-T", "128", OPT_BW },
1925 { "1x32-bw-thread,", "mem", "-p", "1", "-t", "32", "-T", "64", OPT_BW },
1927 { " 2x3-bw-process,", "mem", "-p", "2", "-t", "3", "-P", "512", OPT_BW },
1928 { " 4x4-bw-process,", "mem", "-p", "4", "-t", "4", "-P", "512", OPT_BW },
1929 { " 4x6-bw-process,", "mem", "-p", "4", "-t", "6", "-P", "512", OPT_BW },
1930 { " 4x8-bw-process,", "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW },
1931 { " 4x8-bw-process-NOTHP,",
1932 "mem", "-p", "4", "-t", "8", "-P", "512", OPT_BW_NOTHP },
1933 { " 3x3-bw-process,", "mem", "-p", "3", "-t", "3", "-P", "512", OPT_BW },
1934 { " 5x5-bw-process,", "mem", "-p", "5", "-t", "5", "-P", "512", OPT_BW },
1936 { "2x16-bw-process,", "mem", "-p", "2", "-t", "16", "-P", "512", OPT_BW },
1937 { "1x32-bw-process,", "mem", "-p", "1", "-t", "32", "-P", "2048", OPT_BW },
1939 { "numa02-bw,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW },
1940 { "numa02-bw-NOTHP,", "mem", "-p", "1", "-t", "32", "-T", "32", OPT_BW_NOTHP },
1941 { "numa01-bw-thread,", "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW },
1942 { "numa01-bw-thread-NOTHP,",
1943 "mem", "-p", "2", "-t", "16", "-T", "192", OPT_BW_NOTHP },
1946 static int bench_all(void)
1948 int nr = ARRAY_SIZE(tests);
1952 ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
1955 for (i = 0; i < nr; i++) {
1956 run_bench_numa(tests[i][0], tests[i] + 1);
1964 int bench_numa(int argc, const char **argv)
1966 init_params(&p0, "main,", argc, argv);
1967 argc = parse_options(argc, argv, options, bench_numa_usage, 0);
1974 if (__bench_numa(NULL))
1980 usage_with_options(numa_usage, options);