perf: Add 'perf bench numa mem' NUMA performance measurement suite

Add a suite of NUMA performance benchmarks.

The goal was simulate the behavior and access patterns of real NUMA
workloads, via a wide range of parameters, so this tool goes well
beyond simple bzero() measurements that most NUMA micro-benchmarks use:

 - It processes the data and creates a chain of data dependencies,
   like a real workload would. Neither the compiler, nor the
   kernel (via KSM and other optimizations) nor the CPU can
   eliminate parts of the workload.

 - It randomizes the initial state and also randomizes the target
   addresses of the processing - it's not a simple forward scan
   of addresses.

 - It provides flexible options to set process, thread and memory
   relationship information: -G sets "global" memory shared between
   all test processes, -P sets "process" memory shared by all
   threads of a process and -T sets "thread" private memory.

 - There's a NUMA convergence monitoring and convergence latency
   measurement option via -c and -m.

 - Micro-sleeps and synchronization can be injected to provoke lock
   contention and scheduling, via the -u and -S options. This simulates
   IO and contention.

 - The -x option instructs the workload to 'perturb' itself artificially
   every N seconds, by moving to the first and last CPU of the system
   periodically. This way the stability of convergence equilibrium and
   the number of steps taken for the scheduler to reach equilibrium again
   can be measured.

 - The amount of work can be specified via the -l loop count, and/or
   via a -s seconds-timeout value.

 - CPU and node memory binding options, to test hard binding scenarios.
   THP can be turned on and off via madvise() calls.

 - Live reporting of convergence progress in an 'at glance' output format.
   Printing of convergence and deconvergence events.

The 'perf bench numa mem -a' option will start an array of about 30
individual tests that will each output such measurements:

 # Running  5x5-bw-thread, "perf bench numa mem -p 5 -t 5 -P 512 -s 20 -zZ0q --thp  1"
  5x5-bw-thread,                         20.276, secs,           runtime-max/thread
  5x5-bw-thread,                         20.004, secs,           runtime-min/thread
  5x5-bw-thread,                         20.155, secs,           runtime-avg/thread
  5x5-bw-thread,                          0.671, %,              spread-runtime/thread
  5x5-bw-thread,                         21.153, GB,             data/thread
  5x5-bw-thread,                        528.818, GB,             data-total
  5x5-bw-thread,                          0.959, nsecs,          runtime/byte/thread
  5x5-bw-thread,                          1.043, GB/sec,         thread-speed
  5x5-bw-thread,                         26.081, GB/sec,         total-speed

See the help text and the code for more details.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>

4 files changed, 1747 insertions, 1 deletions

diff --git a/tools/perf/Makefile b/tools/perf/Makefile
index a84021abb3fe..b62dbc0d974a 100644
--- a/tools/perf/Makefile
+++ b/tools/perf/Makefile
@@ -103,7 +103,7 @@ ifdef PARSER_DEBUG
 endif
 
 CFLAGS = -fno-omit-frame-pointer -ggdb3 -funwind-tables -Wall -Wextra -std=gnu99 $(CFLAGS_WERROR) $(CFLAGS_OPTIMIZE) $(EXTRA_WARNINGS) $(EXTRA_CFLAGS) $(PARSER_DEBUG_CFLAGS)
-EXTLIBS = -lpthread -lrt -lelf -lm
+EXTLIBS = -lpthread -lrt -lelf -lm -lnuma
 ALL_CFLAGS = $(CFLAGS) -D_LARGEFILE64_SOURCE -D_FILE_OFFSET_BITS=64 -D_GNU_SOURCE
 ALL_LDFLAGS = $(LDFLAGS)
 STRIP ?= strip
@@ -492,6 +492,7 @@ LIB_OBJS += $(OUTPUT)tests/python-use.o
 BUILTIN_OBJS += $(OUTPUT)builtin-annotate.o
 BUILTIN_OBJS += $(OUTPUT)builtin-bench.o
 # Benchmark modules
+BUILTIN_OBJS += $(OUTPUT)bench/numa.o
 BUILTIN_OBJS += $(OUTPUT)bench/sched-messaging.o
 BUILTIN_OBJS += $(OUTPUT)bench/sched-pipe.o
 ifeq ($(RAW_ARCH),x86_64)
diff --git a/tools/perf/bench/bench.h b/tools/perf/bench/bench.h
index 8f89998eeaf4..a5223e6a7b43 100644
--- a/tools/perf/bench/bench.h
+++ b/tools/perf/bench/bench.h
@@ -1,6 +1,7 @@
 #ifndef BENCH_H
 #define BENCH_H
 
+extern int bench_numa(int argc, const char **argv, const char *prefix);
 extern int bench_sched_messaging(int argc, const char **argv, const char *prefix);
 extern int bench_sched_pipe(int argc, const char **argv, const char *prefix);
 extern int bench_mem_memcpy(int argc, const char **argv,
diff --git a/tools/perf/bench/numa.c b/tools/perf/bench/numa.c
new file mode 100644
index 000000000000..30d1c3225b46
--- /dev/null
+++ b/tools/perf/bench/numa.c
@@ -0,0 +1,1731 @@
+/*
+ * numa.c
+ *
+ * numa: Simulate NUMA-sensitive workload and measure their NUMA performance
+ */
+
+#include "../perf.h"
+#include "../builtin.h"
+#include "../util/util.h"
+#include "../util/parse-options.h"
+
+#include "bench.h"
+
+#include <errno.h>
+#include <sched.h>
+#include <stdio.h>
+#include <assert.h>
+#include <malloc.h>
+#include <signal.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <pthread.h>
+#include <sys/mman.h>
+#include <sys/time.h>
+#include <sys/wait.h>
+#include <sys/prctl.h>
+#include <sys/types.h>
+
+#include <numa.h>
+#include <numaif.h>
+
+/*
+ * Regular printout to the terminal, supressed if -q is specified:
+ */
+#define tprintf(x...) do { if (g && g->p.show_details >= 0) printf(x); } while (0)
+
+/*
+ * Debug printf:
+ */
+#define dprintf(x...) do { if (g && g->p.show_details >= 1) printf(x); } while (0)
+
+struct thread_data {
+	int			curr_cpu;
+	cpu_set_t		bind_cpumask;
+	int			bind_node;
+	u8			*process_data;
+	int			process_nr;
+	int			thread_nr;
+	int			task_nr;
+	unsigned int		loops_done;
+	u64			val;
+	u64			runtime_ns;
+	pthread_mutex_t		*process_lock;
+};
+
+/* Parameters set by options: */
+
+struct params {
+	/* Startup synchronization: */
+	bool			serialize_startup;
+
+	/* Task hierarchy: */
+	int			nr_proc;
+	int			nr_threads;
+
+	/* Working set sizes: */
+	const char		*mb_global_str;
+	const char		*mb_proc_str;
+	const char		*mb_proc_locked_str;
+	const char		*mb_thread_str;
+
+	double			mb_global;
+	double			mb_proc;
+	double			mb_proc_locked;
+	double			mb_thread;
+
+	/* Access patterns to the working set: */
+	bool			data_reads;
+	bool			data_writes;
+	bool			data_backwards;
+	bool			data_zero_memset;
+	bool			data_rand_walk;
+	u32			nr_loops;
+	u32			nr_secs;
+	u32			sleep_usecs;
+
+	/* Working set initialization: */
+	bool			init_zero;
+	bool			init_random;
+	bool			init_cpu0;
+
+	/* Misc options: */
+	int			show_details;
+	int			run_all;
+	int			thp;
+
+	long			bytes_global;
+	long			bytes_process;
+	long			bytes_process_locked;
+	long			bytes_thread;
+
+	int			nr_tasks;
+	bool			show_quiet;
+
+	bool			show_convergence;
+	bool			measure_convergence;
+
+	int			perturb_secs;
+	int			nr_cpus;
+	int			nr_nodes;
+
+	/* Affinity options -C and -N: */
+	char			*cpu_list_str;
+	char			*node_list_str;
+};
+
+
+/* Global, read-writable area, accessible to all processes and threads: */
+
+struct global_info {
+	u8			*data;
+
+	pthread_mutex_t		startup_mutex;
+	int			nr_tasks_started;
+
+	pthread_mutex_t		startup_done_mutex;
+
+	pthread_mutex_t		start_work_mutex;
+	int			nr_tasks_working;
+
+	pthread_mutex_t		stop_work_mutex;
+	u64			bytes_done;
+
+	struct thread_data	*threads;
+
+	/* Convergence latency measurement: */
+	bool			all_converged;
+	bool			stop_work;
+
+	int			print_once;
+
+	struct params		p;
+};
+
+static struct global_info	*g = NULL;
+
+static int parse_cpus_opt(const struct option *opt, const char *arg, int unset);
+static int parse_nodes_opt(const struct option *opt, const char *arg, int unset);
+
+struct params p0;
+
+static const struct option options[] = {
+	OPT_INTEGER('p', "nr_proc"	, &p0.nr_proc,		"number of processes"),
+	OPT_INTEGER('t', "nr_threads"	, &p0.nr_threads,	"number of threads per process"),
+
+	OPT_STRING('G', "mb_global"	, &p0.mb_global_str,	"MB", "global  memory (MBs)"),
+	OPT_STRING('P', "mb_proc"	, &p0.mb_proc_str,	"MB", "process memory (MBs)"),
+	OPT_STRING('L', "mb_proc_locked", &p0.mb_proc_locked_str,"MB", "process serialized/locked memory access (MBs), <= process_memory"),
+	OPT_STRING('T', "mb_thread"	, &p0.mb_thread_str,	"MB", "thread  memory (MBs)"),
+
+	OPT_UINTEGER('l', "nr_loops"	, &p0.nr_loops,		"max number of loops to run"),
+	OPT_UINTEGER('s', "nr_secs"	, &p0.nr_secs,		"max number of seconds to run"),
+	OPT_UINTEGER('u', "usleep"	, &p0.sleep_usecs,	"usecs to sleep per loop iteration"),
+
+	OPT_BOOLEAN('R', "data_reads"	, &p0.data_reads,	"access the data via writes (can be mixed with -W)"),
+	OPT_BOOLEAN('W', "data_writes"	, &p0.data_writes,	"access the data via writes (can be mixed with -R)"),
+	OPT_BOOLEAN('B', "data_backwards", &p0.data_backwards,	"access the data backwards as well"),
+	OPT_BOOLEAN('Z', "data_zero_memset", &p0.data_zero_memset,"access the data via glibc bzero only"),
+	OPT_BOOLEAN('r', "data_rand_walk", &p0.data_rand_walk,	"access the data with random (32bit LFSR) walk"),
+
+
+	OPT_BOOLEAN('z', "init_zero"	, &p0.init_zero,	"bzero the initial allocations"),
+	OPT_BOOLEAN('I', "init_random"	, &p0.init_random,	"randomize the contents of the initial allocations"),
+	OPT_BOOLEAN('0', "init_cpu0"	, &p0.init_cpu0,	"do the initial allocations on CPU#0"),
+	OPT_INTEGER('x', "perturb_secs", &p0.perturb_secs,	"perturb thread 0/0 every X secs, to test convergence stability"),
+
+	OPT_INCR   ('d', "show_details"	, &p0.show_details,	"Show details"),
+	OPT_INCR   ('a', "all"		, &p0.run_all,		"Run all tests in the suite"),
+	OPT_INTEGER('H', "thp"		, &p0.thp,		"MADV_NOHUGEPAGE < 0 < MADV_HUGEPAGE"),
+	OPT_BOOLEAN('c', "show_convergence", &p0.show_convergence, "show convergence details"),
+	OPT_BOOLEAN('m', "measure_convergence",	&p0.measure_convergence, "measure convergence latency"),
+	OPT_BOOLEAN('q', "quiet"	, &p0.show_quiet,	"bzero the initial allocations"),
+	OPT_BOOLEAN('S', "serialize-startup", &p0.serialize_startup,"serialize thread startup"),
+
+	/* Special option string parsing callbacks: */
+        OPT_CALLBACK('C', "cpus", NULL, "cpu[,cpu2,...cpuN]",
+			"bind the first N tasks to these specific cpus (the rest is unbound)",
+			parse_cpus_opt),
+        OPT_CALLBACK('M', "memnodes", NULL, "node[,node2,...nodeN]",
+			"bind the first N tasks to these specific memory nodes (the rest is unbound)",
+			parse_nodes_opt),
+	OPT_END()
+};
+
+static const char * const bench_numa_usage[] = {
+	"perf bench numa <options>",
+	NULL
+};
+
+static const char * const numa_usage[] = {
+	"perf bench numa mem [<options>]",
+	NULL
+};
+
+static cpu_set_t bind_to_cpu(int target_cpu)
+{
+	cpu_set_t orig_mask, mask;
+	int ret;
+
+	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
+	BUG_ON(ret);
+
+	CPU_ZERO(&mask);
+
+	if (target_cpu == -1) {
+		int cpu;
+
+		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
+			CPU_SET(cpu, &mask);
+	} else {
+		BUG_ON(target_cpu < 0 || target_cpu >= g->p.nr_cpus);
+		CPU_SET(target_cpu, &mask);
+	}
+
+	ret = sched_setaffinity(0, sizeof(mask), &mask);
+	BUG_ON(ret);
+
+	return orig_mask;
+}
+
+static cpu_set_t bind_to_node(int target_node)
+{
+	int cpus_per_node = g->p.nr_cpus/g->p.nr_nodes;
+	cpu_set_t orig_mask, mask;
+	int cpu;
+	int ret;
+
+	BUG_ON(cpus_per_node*g->p.nr_nodes != g->p.nr_cpus);
+	BUG_ON(!cpus_per_node);
+
+	ret = sched_getaffinity(0, sizeof(orig_mask), &orig_mask);
+	BUG_ON(ret);
+
+	CPU_ZERO(&mask);
+
+	if (target_node == -1) {
+		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
+			CPU_SET(cpu, &mask);
+	} else {
+		int cpu_start = (target_node + 0) * cpus_per_node;
+		int cpu_stop  = (target_node + 1) * cpus_per_node;
+
+		BUG_ON(cpu_stop > g->p.nr_cpus);
+
+		for (cpu = cpu_start; cpu < cpu_stop; cpu++)
+			CPU_SET(cpu, &mask);
+	}
+
+	ret = sched_setaffinity(0, sizeof(mask), &mask);
+	BUG_ON(ret);
+
+	return orig_mask;
+}
+
+static void bind_to_cpumask(cpu_set_t mask)
+{
+	int ret;
+
+	ret = sched_setaffinity(0, sizeof(mask), &mask);
+	BUG_ON(ret);
+}
+
+static void mempol_restore(void)
+{
+	int ret;
+
+	ret = set_mempolicy(MPOL_DEFAULT, NULL, g->p.nr_nodes-1);
+
+	BUG_ON(ret);
+}
+
+static void bind_to_memnode(int node)
+{
+	unsigned long nodemask;
+	int ret;
+
+	if (node == -1)
+		return;
+
+	BUG_ON(g->p.nr_nodes > (int)sizeof(nodemask));
+	nodemask = 1L << node;
+
+	ret = set_mempolicy(MPOL_BIND, &nodemask, sizeof(nodemask)*8);
+	dprintf("binding to node %d, mask: %016lx => %d\n", node, nodemask, ret);
+
+	BUG_ON(ret);
+}
+
+#define HPSIZE (2*1024*1024)
+
+#define set_taskname(fmt...)				\
+do {							\
+	char name[20];					\
+							\
+	snprintf(name, 20, fmt);			\
+	prctl(PR_SET_NAME, name);			\
+} while (0)
+
+static u8 *alloc_data(ssize_t bytes0, int map_flags,
+		      int init_zero, int init_cpu0, int thp, int init_random)
+{
+	cpu_set_t orig_mask;
+	ssize_t bytes;
+	u8 *buf;
+	int ret;
+
+	if (!bytes0)
+		return NULL;
+
+	/* Allocate and initialize all memory on CPU#0: */
+	if (init_cpu0) {
+		orig_mask = bind_to_node(0);
+		bind_to_memnode(0);
+	}
+
+	bytes = bytes0 + HPSIZE;
+
+	buf = (void *)mmap(0, bytes, PROT_READ|PROT_WRITE, MAP_ANON|map_flags, -1, 0);
+	BUG_ON(buf == (void *)-1);
+
+	if (map_flags == MAP_PRIVATE) {
+		if (thp > 0) {
+			ret = madvise(buf, bytes, MADV_HUGEPAGE);
+			if (ret && !g->print_once) {
+				g->print_once = 1;
+				printf("WARNING: Could not enable THP - do: 'echo madvise > /sys/kernel/mm/transparent_hugepage/enabled'\n");
+			}
+		}
+		if (thp < 0) {
+			ret = madvise(buf, bytes, MADV_NOHUGEPAGE);
+			if (ret && !g->print_once) {
+				g->print_once = 1;
+				printf("WARNING: Could not disable THP: run a CONFIG_TRANSPARENT_HUGEPAGE kernel?\n");
+			}
+		}
+	}
+
+	if (init_zero) {
+		bzero(buf, bytes);
+	} else {
+		/* Initialize random contents, different in each word: */
+		if (init_random) {
+			u64 *wbuf = (void *)buf;
+			long off = rand();
+			long i;
+
+			for (i = 0; i < bytes/8; i++)
+				wbuf[i] = i + off;
+		}
+	}
+
+	/* Align to 2MB boundary: */
+	buf = (void *)(((unsigned long)buf + HPSIZE-1) & ~(HPSIZE-1));
+
+	/* Restore affinity: */
+	if (init_cpu0) {
+		bind_to_cpumask(orig_mask);
+		mempol_restore();
+	}
+
+	return buf;
+}
+
+static void free_data(void *data, ssize_t bytes)
+{
+	int ret;
+
+	if (!data)
+		return;
+
+	ret = munmap(data, bytes);
+	BUG_ON(ret);
+}
+
+/*
+ * Create a shared memory buffer that can be shared between processes, zeroed:
+ */
+static void * zalloc_shared_data(ssize_t bytes)
+{
+	return alloc_data(bytes, MAP_SHARED, 1, g->p.init_cpu0,  g->p.thp, g->p.init_random);
+}
+
+/*
+ * Create a shared memory buffer that can be shared between processes:
+ */
+static void * setup_shared_data(ssize_t bytes)
+{
+	return alloc_data(bytes, MAP_SHARED, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
+}
+
+/*
+ * Allocate process-local memory - this will either be shared between
+ * threads of this process, or only be accessed by this thread:
+ */
+static void * setup_private_data(ssize_t bytes)
+{
+	return alloc_data(bytes, MAP_PRIVATE, 0, g->p.init_cpu0,  g->p.thp, g->p.init_random);
+}
+
+/*
+ * Return a process-shared (global) mutex:
+ */
+static void init_global_mutex(pthread_mutex_t *mutex)
+{
+	pthread_mutexattr_t attr;
+
+	pthread_mutexattr_init(&attr);
+	pthread_mutexattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
+	pthread_mutex_init(mutex, &attr);
+}
+
+static int parse_cpu_list(const char *arg)
+{
+	p0.cpu_list_str = strdup(arg);
+
+	dprintf("got CPU list: {%s}\n", p0.cpu_list_str);
+
+	return 0;
+}
+
+static void parse_setup_cpu_list(void)
+{
+	struct thread_data *td;
+	char *str0, *str;
+	int t;
+
+	if (!g->p.cpu_list_str)
+		return;
+
+	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
+
+	str0 = str = strdup(g->p.cpu_list_str);
+	t = 0;
+
+	BUG_ON(!str);
+
+	tprintf("# binding tasks to CPUs:\n");
+	tprintf("#  ");
+
+	while (true) {
+		int bind_cpu, bind_cpu_0, bind_cpu_1;
+		char *tok, *tok_end, *tok_step, *tok_len, *tok_mul;
+		int bind_len;
+		int step;
+		int mul;
+
+		tok = strsep(&str, ",");
+		if (!tok)
+			break;
+
+		tok_end = strstr(tok, "-");
+
+		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
+		if (!tok_end) {
+			/* Single CPU specified: */
+			bind_cpu_0 = bind_cpu_1 = atol(tok);
+		} else {
+			/* CPU range specified (for example: "5-11"): */
+			bind_cpu_0 = atol(tok);
+			bind_cpu_1 = atol(tok_end + 1);
+		}
+
+		step = 1;
+		tok_step = strstr(tok, "#");
+		if (tok_step) {
+			step = atol(tok_step + 1);
+			BUG_ON(step <= 0 || step >= g->p.nr_cpus);
+		}
+
+		/*
+		 * Mask length.
+		 * Eg: "--cpus 8_4-16#4" means: '--cpus 8_4,12_4,16_4',
+		 * where the _4 means the next 4 CPUs are allowed.
+		 */
+		bind_len = 1;
+		tok_len = strstr(tok, "_");
+		if (tok_len) {
+			bind_len = atol(tok_len + 1);
+			BUG_ON(bind_len <= 0 || bind_len > g->p.nr_cpus);
+		}
+
+		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
+		mul = 1;
+		tok_mul = strstr(tok, "x");
+		if (tok_mul) {
+			mul = atol(tok_mul + 1);
+			BUG_ON(mul <= 0);
+		}
+
+		dprintf("CPUs: %d_%d-%d#%dx%d\n", bind_cpu_0, bind_len, bind_cpu_1, step, mul);
+
+		BUG_ON(bind_cpu_0 < 0 || bind_cpu_0 >= g->p.nr_cpus);
+		BUG_ON(bind_cpu_1 < 0 || bind_cpu_1 >= g->p.nr_cpus);
+		BUG_ON(bind_cpu_0 > bind_cpu_1);
+
+		for (bind_cpu = bind_cpu_0; bind_cpu <= bind_cpu_1; bind_cpu += step) {
+			int i;
+
+			for (i = 0; i < mul; i++) {
+				int cpu;
+
+				if (t >= g->p.nr_tasks) {
+					printf("\n# NOTE: ignoring bind CPUs starting at CPU#%d\n #", bind_cpu);
+					goto out;
+				}
+				td = g->threads + t;
+
+				if (t)
+					tprintf(",");
+				if (bind_len > 1) {
+					tprintf("%2d/%d", bind_cpu, bind_len);
+				} else {
+					tprintf("%2d", bind_cpu);
+				}
+
+				CPU_ZERO(&td->bind_cpumask);
+				for (cpu = bind_cpu; cpu < bind_cpu+bind_len; cpu++) {
+					BUG_ON(cpu < 0 || cpu >= g->p.nr_cpus);
+					CPU_SET(cpu, &td->bind_cpumask);
+				}
+				t++;
+			}
+		}
+	}
+out:
+
+	tprintf("\n");
+
+	if (t < g->p.nr_tasks)
+		printf("# NOTE: %d tasks bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
+
+	free(str0);
+}
+
+static int parse_cpus_opt(const struct option *opt __maybe_unused,
+			  const char *arg, int unset __maybe_unused)
+{
+	if (!arg)
+		return -1;
+
+	return parse_cpu_list(arg);
+}
+
+static int parse_node_list(const char *arg)
+{
+	p0.node_list_str = strdup(arg);
+
+	dprintf("got NODE list: {%s}\n", p0.node_list_str);
+
+	return 0;
+}
+
+static void parse_setup_node_list(void)
+{
+	struct thread_data *td;
+	char *str0, *str;
+	int t;
+
+	if (!g->p.node_list_str)
+		return;
+
+	dprintf("g->p.nr_tasks: %d\n", g->p.nr_tasks);
+
+	str0 = str = strdup(g->p.node_list_str);
+	t = 0;
+
+	BUG_ON(!str);
+
+	tprintf("# binding tasks to NODEs:\n");
+	tprintf("# ");
+
+	while (true) {
+		int bind_node, bind_node_0, bind_node_1;
+		char *tok, *tok_end, *tok_step, *tok_mul;
+		int step;
+		int mul;
+
+		tok = strsep(&str, ",");
+		if (!tok)
+			break;
+
+		tok_end = strstr(tok, "-");
+
+		dprintf("\ntoken: {%s}, end: {%s}\n", tok, tok_end);
+		if (!tok_end) {
+			/* Single NODE specified: */
+			bind_node_0 = bind_node_1 = atol(tok);
+		} else {
+			/* NODE range specified (for example: "5-11"): */
+			bind_node_0 = atol(tok);
+			bind_node_1 = atol(tok_end + 1);
+		}
+
+		step = 1;
+		tok_step = strstr(tok, "#");
+		if (tok_step) {
+			step = atol(tok_step + 1);
+			BUG_ON(step <= 0 || step >= g->p.nr_nodes);
+		}
+
+		/* Multiplicator shortcut, "0x8" is a shortcut for: "0,0,0,0,0,0,0,0" */
+		mul = 1;
+		tok_mul = strstr(tok, "x");
+		if (tok_mul) {
+			mul = atol(tok_mul + 1);
+			BUG_ON(mul <= 0);
+		}
+
+		dprintf("NODEs: %d-%d #%d\n", bind_node_0, bind_node_1, step);
+
+		BUG_ON(bind_node_0 < 0 || bind_node_0 >= g->p.nr_nodes);
+		BUG_ON(bind_node_1 < 0 || bind_node_1 >= g->p.nr_nodes);
+		BUG_ON(bind_node_0 > bind_node_1);
+
+		for (bind_node = bind_node_0; bind_node <= bind_node_1; bind_node += step) {
+			int i;
+
+			for (i = 0; i < mul; i++) {
+				if (t >= g->p.nr_tasks) {
+					printf("\n# NOTE: ignoring bind NODEs starting at NODE#%d\n", bind_node);
+					goto out;
+				}
+				td = g->threads + t;
+
+				if (!t)
+					tprintf(" %2d", bind_node);
+				else
+					tprintf(",%2d", bind_node);
+
+				td->bind_node = bind_node;
+				t++;
+			}
+		}
+	}
+out:
+
+	tprintf("\n");
+
+	if (t < g->p.nr_tasks)
+		printf("# NOTE: %d tasks mem-bound, %d tasks unbound\n", t, g->p.nr_tasks - t);
+
+	free(str0);
+}
+
+static int parse_nodes_opt(const struct option *opt __maybe_unused,
+			  const char *arg, int unset __maybe_unused)
+{
+	if (!arg)
+		return -1;
+
+	return parse_node_list(arg);
+
+	return 0;
+}
+
+#define BIT(x) (1ul << x)
+
+static inline uint32_t lfsr_32(uint32_t lfsr)
+{
+	const uint32_t taps = BIT(1) | BIT(5) | BIT(6) | BIT(31);
+	return (lfsr>>1) ^ ((0x0u - (lfsr & 0x1u)) & taps);
+}
+
+/*
+ * Make sure there's real data dependency to RAM (when read
+ * accesses are enabled), so the compiler, the CPU and the
+ * kernel (KSM, zero page, etc.) cannot optimize away RAM
+ * accesses:
+ */
+static inline u64 access_data(u64 *data __attribute__((unused)), u64 val)
+{
+	if (g->p.data_reads)
+		val += *data;
+	if (g->p.data_writes)
+		*data = val + 1;
+	return val;
+}
+
+/*
+ * The worker process does two types of work, a forwards going
+ * loop and a backwards going loop.
+ *
+ * We do this so that on multiprocessor systems we do not create
+ * a 'train' of processing, with highly synchronized processes,
+ * skewing the whole benchmark.
+ */
+static u64 do_work(u8 *__data, long bytes, int nr, int nr_max, int loop, u64 val)
+{
+	long words = bytes/sizeof(u64);
+	u64 *data = (void *)__data;
+	long chunk_0, chunk_1;
+	u64 *d0, *d, *d1;
+	long off;
+	long i;
+
+	BUG_ON(!data && words);
+	BUG_ON(data && !words);
+
+	if (!data)
+		return val;
+
+	/* Very simple memset() work variant: */
+	if (g->p.data_zero_memset && !g->p.data_rand_walk) {
+		bzero(data, bytes);
+		return val;
+	}
+
+	/* Spread out by PID/TID nr and by loop nr: */
+	chunk_0 = words/nr_max;
+	chunk_1 = words/g->p.nr_loops;
+	off = nr*chunk_0 + loop*chunk_1;
+
+	while (off >= words)
+		off -= words;
+
+	if (g->p.data_rand_walk) {
+		u32 lfsr = nr + loop + val;
+		int j;
+
+		for (i = 0; i < words/1024; i++) {
+			long start, end;
+
+			lfsr = lfsr_32(lfsr);
+
+			start = lfsr % words;
+			end = min(start + 1024, words-1);
+
+			if (g->p.data_zero_memset) {
+				bzero(data + start, (end-start) * sizeof(u64));
+			} else {
+				for (j = start; j < end; j++)
+					val = access_data(data + j, val);
+			}
+		}
+	} else if (!g->p.data_backwards || (nr + loop) & 1) {
+
+		d0 = data + off;
+		d  = data + off + 1;
+		d1 = data + words;
+
+		/* Process data forwards: */
+		for (;;) {
+			if (unlikely(d >= d1))
+				d = data;
+			if (unlikely(d == d0))
+				break;
+
+			val = access_data(d, val);
+
+			d++;
+		}
+	} else {
+		/* Process data backwards: */
+
+		d0 = data + off;
+		d  = data + off - 1;
+		d1 = data + words;
+
+		/* Process data forwards: */
+		for (;;) {
+			if (unlikely(d < data))
+				d = data + words-1;
+			if (unlikely(d == d0))
+				break;
+
+			val = access_data(d, val);
+
+			d--;
+		}
+	}
+
+	return val;
+}
+
+static void update_curr_cpu(int task_nr, unsigned long bytes_worked)
+{
+	unsigned int cpu;
+
+	cpu = sched_getcpu();
+
+	g->threads[task_nr].curr_cpu = cpu;
+	prctl(0, bytes_worked);
+}
+
+#define MAX_NR_NODES	64
+
+/*
+ * Count the number of nodes a process's threads
+ * are spread out on.
+ *
+ * A count of 1 means that the process is compressed
+ * to a single node. A count of g->p.nr_nodes means it's
+ * spread out on the whole system.
+ */
+static int count_process_nodes(int process_nr)
+{
+	char node_present[MAX_NR_NODES] = { 0, };
+	int nodes;
+	int n, t;
+
+	for (t = 0; t < g->p.nr_threads; t++) {
+		struct thread_data *td;
+		int task_nr;
+		int node;
+
+		task_nr = process_nr*g->p.nr_threads + t;
+		td = g->threads + task_nr;
+
+		node = numa_node_of_cpu(td->curr_cpu);
+		node_present[node] = 1;
+	}
+
+	nodes = 0;
+
+	for (n = 0; n < MAX_NR_NODES; n++)
+		nodes += node_present[n];
+
+	return nodes;
+}
+
+/*
+ * Count the number of distinct process-threads a node contains.
+ *
+ * A count of 1 means that the node contains only a single
+ * process. If all nodes on the system contain at most one
+ * process then we are well-converged.
+ */
+static int count_node_processes(int node)
+{
+	int processes = 0;
+	int t, p;
+
+	for (p = 0; p < g->p.nr_proc; p++) {
+		for (t = 0; t < g->p.nr_threads; t++) {
+			struct thread_data *td;
+			int task_nr;
+			int n;
+
+			task_nr = p*g->p.nr_threads + t;
+			td = g->threads + task_nr;
+
+			n = numa_node_of_cpu(td->curr_cpu);
+			if (n == node) {
+				processes++;
+				break;
+			}
+		}
+	}
+
+	return processes;
+}
+
+static void calc_convergence_compression(int *strong)
+{
+	unsigned int nodes_min, nodes_max;
+	int p;
+
+	nodes_min = -1;
+	nodes_max =  0;
+
+	for (p = 0; p < g->p.nr_proc; p++) {
+		unsigned int nodes = count_process_nodes(p);
+
+		nodes_min = min(nodes, nodes_min);
+		nodes_max = max(nodes, nodes_max);
+	}
+
+	/* Strong convergence: all threads compress on a single node: */
+	if (nodes_min == 1 && nodes_max == 1) {
+		*strong = 1;
+	} else {
+		*strong = 0;
+		tprintf(" {%d-%d}", nodes_min, nodes_max);
+	}
+}
+
+static void calc_convergence(double runtime_ns_max, double *convergence)
+{
+	unsigned int loops_done_min, loops_done_max;
+	int process_groups;
+	int nodes[MAX_NR_NODES];
+	int distance;
+	int nr_min;
+	int nr_max;
+	int strong;
+	int sum;
+	int nr;
+	int node;
+	int cpu;
+	int t;
+
+	if (!g->p.show_convergence && !g->p.measure_convergence)
+		return;
+
+	for (node = 0; node < g->p.nr_nodes; node++)
+		nodes[node] = 0;
+
+	loops_done_min = -1;
+	loops_done_max = 0;
+
+	for (t = 0; t < g->p.nr_tasks; t++) {
+		struct thread_data *td = g->threads + t;
+		unsigned int loops_done;
+
+		cpu = td->curr_cpu;
+
+		/* Not all threads have written it yet: */
+		if (cpu < 0)
+			continue;
+
+		node = numa_node_of_cpu(cpu);
+
+		nodes[node]++;
+
+		loops_done = td->loops_done;
+		loops_done_min = min(loops_done, loops_done_min);
+		loops_done_max = max(loops_done, loops_done_max);
+	}
+
+	nr_max = 0;
+	nr_min = g->p.nr_tasks;
+	sum = 0;
+
+	for (node = 0; node < g->p.nr_nodes; node++) {
+		nr = nodes[node];
+		nr_min = min(nr, nr_min);
+		nr_max = max(nr, nr_max);
+		sum += nr;
+	}
+	BUG_ON(nr_min > nr_max);
+
+	BUG_ON(sum > g->p.nr_tasks);
+
+	if (0 && (sum < g->p.nr_tasks))
+		return;
+
+	/*
+	 * Count the number of distinct process groups present
+	 * on nodes - when we are converged this will decrease
+	 * to g->p.nr_proc:
+	 */
+	process_groups = 0;
+
+	for (node = 0; node < g->p.nr_nodes; node++) {
+		int processes = count_node_processes(node);
+
+		nr = nodes[node];
+		tprintf(" %2d/%-2d", nr, processes);
+
+		process_groups += processes;
+	}
+
+	distance = nr_max - nr_min;
+
+	tprintf(" [%2d/%-2d]", distance, process_groups);
+
+	tprintf(" l:%3d-%-3d (%3d)",
+		loops_done_min, loops_done_max, loops_done_max-loops_done_min);
+
+	if (loops_done_min && loops_done_max) {
+		double skew = 1.0 - (double)loops_done_min/loops_done_max;
+
+		tprintf(" [%4.1f%%]", skew * 100.0);
+	}
+
+	calc_convergence_compression(&strong);
+
+	if (strong && process_groups == g->p.nr_proc) {
+		if (!*convergence) {
+			*convergence = runtime_ns_max;
+			tprintf(" (%6.1fs converged)\n", *convergence/1e9);
+			if (g->p.measure_convergence) {
+				g->all_converged = true;
+				g->stop_work = true;
+			}
+		}
+	} else {
+		if (*convergence) {
+			tprintf(" (%6.1fs de-converged)", runtime_ns_max/1e9);
+			*convergence = 0;
+		}
+		tprintf("\n");
+	}
+}
+
+static void show_summary(double runtime_ns_max, int l, double *convergence)
+{
+	tprintf("\r #  %5.1f%%  [%.1f mins]",
+		(double)(l+1)/g->p.nr_loops*100.0, runtime_ns_max/1e9 / 60.0);
+
+	calc_convergence(runtime_ns_max, convergence);
+
+	if (g->p.show_details >= 0)
+		fflush(stdout);
+}
+
+static void *worker_thread(void *__tdata)
+{
+	struct thread_data *td = __tdata;
+	struct timeval start0, start, stop, diff;
+	int process_nr = td->process_nr;
+	int thread_nr = td->thread_nr;
+	unsigned long last_perturbance;
+	int task_nr = td->task_nr;
+	int details = g->p.show_details;
+	int first_task, last_task;
+	double convergence = 0;
+	u64 val = td->val;
+	double runtime_ns_max;
+	u8 *global_data;
+	u8 *process_data;
+	u8 *thread_data;
+	u64 bytes_done;
+	long work_done;
+	u32 l;
+
+	bind_to_cpumask(td->bind_cpumask);
+	bind_to_memnode(td->bind_node);
+
+	set_taskname("thread %d/%d", process_nr, thread_nr);
+
+	global_data = g->data;
+	process_data = td->process_data;
+	thread_data = setup_private_data(g->p.bytes_thread);
+
+	bytes_done = 0;
+
+	last_task = 0;
+	if (process_nr == g->p.nr_proc-1 && thread_nr == g->p.nr_threads-1)
+		last_task = 1;
+
+	first_task = 0;
+	if (process_nr == 0 && thread_nr == 0)
+		first_task = 1;
+
+	if (details >= 2) {
+		printf("#  thread %2d / %2d global mem: %p, process mem: %p, thread mem: %p\n",
+			process_nr, thread_nr, global_data, process_data, thread_data);
+	}
+
+	if (g->p.serialize_startup) {
+		pthread_mutex_lock(&g->startup_mutex);
+		g->nr_tasks_started++;
+		pthread_mutex_unlock(&g->startup_mutex);
+
+		/* Here we will wait for the main process to start us all at once: */
+		pthread_mutex_lock(&g->start_work_mutex);
+		g->nr_tasks_working++;
+
+		/* Last one wake the main process: */
+		if (g->nr_tasks_working == g->p.nr_tasks)
+			pthread_mutex_unlock(&g->startup_done_mutex);
+
+		pthread_mutex_unlock(&g->start_work_mutex);
+	}
+
+	gettimeofday(&start0, NULL);
+
+	start = stop = start0;
+	last_perturbance = start.tv_sec;
+
+	for (l = 0; l < g->p.nr_loops; l++) {
+		start = stop;
+
+		if (g->stop_work)
+			break;
+
+		val += do_work(global_data,  g->p.bytes_global,  process_nr, g->p.nr_proc,	l, val);
+		val += do_work(process_data, g->p.bytes_process, thread_nr,  g->p.nr_threads,	l, val);
+		val += do_work(thread_data,  g->p.bytes_thread,  0,          1,		l, val);
+
+		if (g->p.sleep_usecs) {
+			pthread_mutex_lock(td->process_lock);
+			usleep(g->p.sleep_usecs);
+			pthread_mutex_unlock(td->process_lock);
+		}
+		/*
+		 * Amount of work to be done under a process-global lock:
+		 */
+		if (g->p.bytes_process_locked) {
+			pthread_mutex_lock(td->process_lock);
+			val += do_work(process_data, g->p.bytes_process_locked, thread_nr,  g->p.nr_threads,	l, val);
+			pthread_mutex_unlock(td->process_lock);
+		}
+
+		work_done = g->p.bytes_global + g->p.bytes_process +
+			    g->p.bytes_process_locked + g->p.bytes_thread;
+
+		update_curr_cpu(task_nr, work_done);
+		bytes_done += work_done;
+
+		if (details < 0 && !g->p.perturb_secs && !g->p.measure_convergence && !g->p.nr_secs)
+			continue;
+
+		td->loops_done = l;
+
+		gettimeofday(&stop, NULL);
+
+		/* Check whether our max runtime timed out: */
+		if (g->p.nr_secs) {
+			timersub(&stop, &start0, &diff);
+			if (diff.tv_sec >= g->p.nr_secs) {
+				g->stop_work = true;
+				break;
+			}
+		}
+
+		/* Update the summary at most once per second: */
+		if (start.tv_sec == stop.tv_sec)
+			continue;
+
+		/*
+		 * Perturb the first task's equilibrium every g->p.perturb_secs seconds,
+		 * by migrating to CPU#0:
+		 */
+		if (first_task && g->p.perturb_secs && (int)(stop.tv_sec - last_perturbance) >= g->p.perturb_secs) {
+			cpu_set_t orig_mask;
+			int target_cpu;
+			int this_cpu;
+
+			last_perturbance = stop.tv_sec;
+
+			/*
+			 * Depending on where we are running, move into
+			 * the other half of the system, to create some
+			 * real disturbance:
+			 */
+			this_cpu = g->threads[task_nr].curr_cpu;
+			if (this_cpu < g->p.nr_cpus/2)
+				target_cpu = g->p.nr_cpus-1;
+			else
+				target_cpu = 0;
+
+			orig_mask = bind_to_cpu(target_cpu);
+
+			/* Here we are running on the target CPU already */
+			if (details >= 1)
+				printf(" (injecting perturbalance, moved to CPU#%d)\n", target_cpu);
+
+			bind_to_cpumask(orig_mask);
+		}
+
+		if (details >= 3) {
+			timersub(&stop, &start, &diff);
+			runtime_ns_max = diff.tv_sec * 1000000000;
+			runtime_ns_max += diff.tv_usec * 1000;
+
+			if (details >= 0) {
+				printf(" #%2d / %2d: %14.2lf nsecs/op [val: %016lx]\n",
+					process_nr, thread_nr, runtime_ns_max / bytes_done, val);
+			}
+			fflush(stdout);
+		}
+		if (!last_task)
+			continue;
+
+		timersub(&stop, &start0, &diff);
+		runtime_ns_max = diff.tv_sec * 1000000000ULL;
+		runtime_ns_max += diff.tv_usec * 1000ULL;
+
+		show_summary(runtime_ns_max, l, &convergence);
+	}
+
+	gettimeofday(&stop, NULL);
+	timersub(&stop, &start0, &diff);
+	td->runtime_ns = diff.tv_sec * 1000000000ULL;
+	td->runtime_ns += diff.tv_usec * 1000ULL;
+
+	free_data(thread_data, g->p.bytes_thread);
+
+	pthread_mutex_lock(&g->stop_work_mutex);
+	g->bytes_done += bytes_done;
+	pthread_mutex_unlock(&g->stop_work_mutex);
+
+	return NULL;
+}
+
+/*
+ * A worker process starts a couple of threads:
+ */
+static void worker_process(int process_nr)
+{
+	pthread_mutex_t process_lock;
+	struct thread_data *td;
+	pthread_t *pthreads;
+	u8 *process_data;
+	int task_nr;
+	int ret;
+	int t;
+
+	pthread_mutex_init(&process_lock, NULL);
+	set_taskname("process %d", process_nr);
+
+	/*
+	 * Pick up the memory policy and the CPU binding of our first thread,
+	 * so that we initialize memory accordingly:
+	 */
+	task_nr = process_nr*g->p.nr_threads;
+	td = g->threads + task_nr;
+
+	bind_to_memnode(td->bind_node);
+	bind_to_cpumask(td->bind_cpumask);
+
+	pthreads = zalloc(g->p.nr_threads * sizeof(pthread_t));
+	process_data = setup_private_data(g->p.bytes_process);
+
+	if (g->p.show_details >= 3) {
+		printf(" # process %2d global mem: %p, process mem: %p\n",
+			process_nr, g->data, process_data);
+	}
+
+	for (t = 0; t < g->p.nr_threads; t++) {
+		task_nr = process_nr*g->p.nr_threads + t;
+		td = g->threads + task_nr;
+
+		td->process_data = process_data;
+		td->process_nr   = process_nr;
+		td->thread_nr    = t;
+		td->task_nr	 = task_nr;
+		td->val          = rand();
+		td->curr_cpu	 = -1;
+		td->process_lock = &process_lock;
+
+		ret = pthread_create(pthreads + t, NULL, worker_thread, td);
+		BUG_ON(ret);
+	}
+
+	for (t = 0; t < g->p.nr_threads; t++) {
+                ret = pthread_join(pthreads[t], NULL);
+		BUG_ON(ret);
+	}
+
+	free_data(process_data, g->p.bytes_process);
+	free(pthreads);
+}
+
+static void print_summary(void)
+{
+	if (g->p.show_details < 0)
+		return;
+
+	printf("\n ###\n");
+	printf(" # %d %s will execute (on %d nodes, %d CPUs):\n",
+		g->p.nr_tasks, g->p.nr_tasks == 1 ? "task" : "tasks", g->p.nr_nodes, g->p.nr_cpus);
+	printf(" #      %5dx %5ldMB global  shared mem operations\n",
+			g->p.nr_loops, g->p.bytes_global/1024/1024);
+	printf(" #      %5dx %5ldMB process shared mem operations\n",
+			g->p.nr_loops, g->p.bytes_process/1024/1024);
+	printf(" #      %5dx %5ldMB thread  local  mem operations\n",
+			g->p.nr_loops, g->p.bytes_thread/1024/1024);
+
+	printf(" ###\n");
+
+	printf("\n ###\n"); fflush(stdout);
+}
+
+static void init_thread_data(void)
+{
+	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
+	int t;
+
+	g->threads = zalloc_shared_data(size);
+
+	for (t = 0; t < g->p.nr_tasks; t++) {
+		struct thread_data *td = g->threads + t;
+		int cpu;
+
+		/* Allow all nodes by default: */
+		td->bind_node = -1;
+
+		/* Allow all CPUs by default: */
+		CPU_ZERO(&td->bind_cpumask);
+		for (cpu = 0; cpu < g->p.nr_cpus; cpu++)
+			CPU_SET(cpu, &td->bind_cpumask);
+	}
+}
+
+static void deinit_thread_data(void)
+{
+	ssize_t size = sizeof(*g->threads)*g->p.nr_tasks;
+
+	free_data(g->threads, size);
+}
+
+static int init(void)
+{
+	g = (void *)alloc_data(sizeof(*g), MAP_SHARED, 1, 0, 0 /* THP */, 0);
+
+	/* Copy over options: */
+	g->p = p0;
+
+	g->p.nr_cpus = numa_num_configured_cpus();
+
+	g->p.nr_nodes = numa_max_node() + 1;
+
+	/* char array in count_process_nodes(): */
+	BUG_ON(g->p.nr_nodes > MAX_NR_NODES || g->p.nr_nodes < 0);
+
+	if (g->p.show_quiet && !g->p.show_details)
+		g->p.show_details = -1;
+
+	/* Some memory should be specified: */
+	if (!g->p.mb_global_str && !g->p.mb_proc_str && !g->p.mb_thread_str)
+		return -1;
+
+	if (g->p.mb_global_str) {
+		g->p.mb_global = atof(g->p.mb_global_str);
+		BUG_ON(g->p.mb_global < 0);
+	}
+
+	if (g->p.mb_proc_str) {
+		g->p.mb_proc = atof(g->p.mb_proc_str);
+		BUG_ON(g->p.mb_proc < 0);
+	}
+
+	if (g->p.mb_proc_locked_str) {
+		g->p.mb_proc_locked = atof(g->p.mb_proc_locked_str);
+		BUG_ON(g->p.mb_proc_locked < 0);
+		BUG_ON(g->p.mb_proc_locked > g->p.mb_proc);
+	}
+
+	if (g->p.mb_thread_str) {
+		g->p.mb_thread = atof(g->p.mb_thread_str);
+		BUG_ON(g->p.mb_thread < 0);
+	}
+
+	BUG_ON(g->p.nr_threads <= 0);
+	BUG_ON(g->p.nr_proc <= 0);
+
+	g->p.nr_tasks = g->p.nr_proc*g->p.nr_threads;
+
+	g->p.bytes_global		= g->p.mb_global	*1024L*1024L;
+	g->p.bytes_process		= g->p.mb_proc		*1024L*1024L;
+	g->p.bytes_process_locked	= g->p.mb_proc_locked	*1024L*1024L;
+	g->p.bytes_thread		= g->p.mb_thread	*1024L*1024L;
+
+	g->data = setup_shared_data(g->p.bytes_global);
+
+	/* Startup serialization: */
+	init_global_mutex(&g->start_work_mutex);
+	init_global_mutex(&g->startup_mutex);
+	init_global_mutex(&g->startup_done_mutex);
+	init_global_mutex(&g->stop_work_mutex);
+
+	init_thread_data();
+
+	tprintf("#\n");
+	parse_setup_cpu_list();
+	parse_setup_node_list();
+	tprintf("#\n");
+
+	print_summary();
+
+	return 0;
+}
+
+static void deinit(void)
+{
+	free_data(g->data, g->p.bytes_global);
+	g->data = NULL;
+
+	deinit_thread_data();
+
+	free_data(g, sizeof(*g));
+	g = NULL;
+}
+
+/*
+ * Print a short or long result, depending on the verbosity setting:
+ */
+static void print_res(const char *name, double val,
+		      const char *txt_unit, const char *txt_short, const char *txt_long)
+{
+	if (!name)
+		name = "main,";
+
+	if (g->p.show_quiet)
+		printf(" %-30s %15.3f, %-15s %s\n", name, val, txt_unit, txt_short);
+	else
+		printf(" %14.3f %s\n", val, txt_long);
+}
+
+static int __bench_numa(const char *name)
+{
+	struct timeval start, stop, diff;
+	u64 runtime_ns_min, runtime_ns_sum;
+	pid_t *pids, pid, wpid;
+	double delta_runtime;
+	double runtime_avg;
+	double runtime_sec_max;
+	double runtime_sec_min;
+	int wait_stat;
+	double bytes;
+	int i, t;
+
+	if (init())
+		return -1;
+
+	pids = zalloc(g->p.nr_proc * sizeof(*pids));
+	pid = -1;
+
+	/* All threads try to acquire it, this way we can wait for them to start up: */
+	pthread_mutex_lock(&g->start_work_mutex);
+
+	if (g->p.serialize_startup) {
+		tprintf(" #\n");
+		tprintf(" # Startup synchronization: ..."); fflush(stdout);
+	}
+
+	gettimeofday(&start, NULL);
+
+	for (i = 0; i < g->p.nr_proc; i++) {
+		pid = fork();
+		dprintf(" # process %2d: PID %d\n", i, pid);
+
+		BUG_ON(pid < 0);
+		if (!pid) {
+			/* Child process: */
+			worker_process(i);
+
+			exit(0);
+		}
+		pids[i] = pid;
+
+	}
+	/* Wait for all the threads to start up: */
+	while (g->nr_tasks_started != g->p.nr_tasks)
+		usleep(1000);
+
+	BUG_ON(g->nr_tasks_started != g->p.nr_tasks);
+
+	if (g->p.serialize_startup) {
+		double startup_sec;
+
+		pthread_mutex_lock(&g->startup_done_mutex);
+
+		/* This will start all threads: */
+		pthread_mutex_unlock(&g->start_work_mutex);
+
+		/* This mutex is locked - the last started thread will wake us: */
+		pthread_mutex_lock(&g->startup_done_mutex);
+
+		gettimeofday(&stop, NULL);
+
+		timersub(&stop, &start, &diff);
+
+		startup_sec = diff.tv_sec * 1000000000.0;
+		startup_sec += diff.tv_usec * 1000.0;
+		startup_sec /= 1e9;
+
+		tprintf(" threads initialized in %.6f seconds.\n", startup_sec);
+		tprintf(" #\n");
+
+		start = stop;
+		pthread_mutex_unlock(&g->startup_done_mutex);
+	} else {
+		gettimeofday(&start, NULL);
+	}
+
+	/* Parent process: */
+
+
+	for (i = 0; i < g->p.nr_proc; i++) {
+		wpid = waitpid(pids[i], &wait_stat, 0);
+		BUG_ON(wpid < 0);
+		BUG_ON(!WIFEXITED(wait_stat));
+
+	}
+
+	runtime_ns_sum = 0;
+	runtime_ns_min = -1LL;
+
+	for (t = 0; t < g->p.nr_tasks; t++) {
+		u64 thread_runtime_ns = g->threads[t].runtime_ns;
+
+		runtime_ns_sum += thread_runtime_ns;
+		runtime_ns_min = min(thread_runtime_ns, runtime_ns_min);
+	}
+
+	gettimeofday(&stop, NULL);
+	timersub(&stop, &start, &diff);
+
+	BUG_ON(bench_format != BENCH_FORMAT_DEFAULT);
+
+	tprintf("\n ###\n");
+	tprintf("\n");
+
+	runtime_sec_max = diff.tv_sec * 1000000000.0;
+	runtime_sec_max += diff.tv_usec * 1000.0;
+	runtime_sec_max /= 1e9;
+
+	runtime_sec_min = runtime_ns_min/1e9;
+
+	bytes = g->bytes_done;
+	runtime_avg = (double)runtime_ns_sum / g->p.nr_tasks / 1e9;
+
+	if (g->p.measure_convergence) {
+		print_res(name, runtime_sec_max,
+			"secs,", "NUMA-convergence-latency", "secs latency to NUMA-converge");
+	}
+
+	print_res(name, runtime_sec_max,
+		"secs,", "runtime-max/thread",	"secs slowest (max) thread-runtime");
+
+	print_res(name, runtime_sec_min,
+		"secs,", "runtime-min/thread",	"secs fastest (min) thread-runtime");
+
+	print_res(name, runtime_avg,
+		"secs,", "runtime-avg/thread",	"secs average thread-runtime");
+
+	delta_runtime = (runtime_sec_max - runtime_sec_min)/2.0;
+	print_res(name, delta_runtime / runtime_sec_max * 100.0,
+		"%,", "spread-runtime/thread",	"% difference between max/avg runtime");
+
+	print_res(name, bytes / g->p.nr_tasks / 1e9,
+		"GB,", "data/thread",		"GB data processed, per thread");
+
+	print_res(name, bytes / 1e9,
+		"GB,", "data-total",		"GB data processed, total");
+
+	print_res(name, runtime_sec_max * 1e9 / (bytes / g->p.nr_tasks),
+		"nsecs,", "runtime/byte/thread","nsecs/byte/thread runtime");
+
+	print_res(name, bytes / g->p.nr_tasks / 1e9 / runtime_sec_max,
+		"GB/sec,", "thread-speed",	"GB/sec/thread speed");
+
+	print_res(name, bytes / runtime_sec_max / 1e9,
+		"GB/sec,", "total-speed",	"GB/sec total speed");
+
+	free(pids);
+
+	deinit();
+
+	return 0;
+}
+
+#define MAX_ARGS 50
+
+static int command_size(const char **argv)
+{
+	int size = 0;
+
+	while (*argv) {
+		size++;
+		argv++;
+	}
+
+	BUG_ON(size >= MAX_ARGS);
+
+	return size;
+}
+
+static void init_params(struct params *p, const char *name, int argc, const char **argv)
+{
+	int i;
+
+	printf("\n # Running %s \"perf bench numa", name);
+
+	for (i = 0; i < argc; i++)
+		printf(" %s", argv[i]);
+
+	printf("\"\n");
+
+	memset(p, 0, sizeof(*p));
+
+	/* Initialize nonzero defaults: */
+
+	p->serialize_startup		= 1;
+	p->data_reads			= true;
+	p->data_writes			= true;
+	p->data_backwards		= true;
+	p->data_rand_walk		= true;
+	p->nr_loops			= -1;
+	p->init_random			= true;
+}
+
+static int run_bench_numa(const char *name, const char **argv)
+{
+	int argc = command_size(argv);
+
+	init_params(&p0, name, argc, argv);
+	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
+	if (argc)
+		goto err;
+
+	if (__bench_numa(name))
+		goto err;
+
+	return 0;
+
+err:
+	usage_with_options(numa_usage, options);
+	return -1;
+}
+
+#define OPT_BW_RAM		"-s",  "20", "-zZq",    "--thp", " 1", "--no-data_rand_walk"
+#define OPT_BW_RAM_NOTHP	OPT_BW_RAM,		"--thp", "-1"
+
+#define OPT_CONV		"-s", "100", "-zZ0qcm", "--thp", " 1"
+#define OPT_CONV_NOTHP		OPT_CONV,		"--thp", "-1"
+
+#define OPT_BW			"-s",  "20", "-zZ0q",   "--thp", " 1"
+#define OPT_BW_NOTHP		OPT_BW,			"--thp", "-1"
+
+/*
+ * The built-in test-suite executed by "perf bench numa -a".
+ *
+ * (A minimum of 4 nodes and 16 GB of RAM is recommended.)
+ */
+static const char *tests[][MAX_ARGS] = {
+   /* Basic single-stream NUMA bandwidth measurements: */
+   { "RAM-bw-local,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
+			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM },
+   { "RAM-bw-local-NOTHP,",
+			  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
+			  "-C" ,   "0", "-M",   "0", OPT_BW_RAM_NOTHP },
+   { "RAM-bw-remote,",	  "mem",  "-p",  "1",  "-t",  "1", "-P", "1024",
+			  "-C" ,   "0", "-M",   "1", OPT_BW_RAM },
+
+   /* 2-stream NUMA bandwidth measurements: */
+   { "RAM-bw-local-2x,",  "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
+			   "-C", "0,2", "-M", "0x2", OPT_BW_RAM },
+   { "RAM-bw-remote-2x,", "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
+		 	   "-C", "0,2", "-M", "1x2", OPT_BW_RAM },
+
+   /* Cross-stream NUMA bandwidth measurement: */
+   { "RAM-bw-cross,",     "mem",  "-p",  "2",  "-t",  "1", "-P", "1024",
+		 	   "-C", "0,8", "-M", "1,0", OPT_BW_RAM },
+
+   /* Convergence latency measurements: */
+   { " 1x3-convergence,", "mem",  "-p",  "1", "-t",  "3", "-P",  "512", OPT_CONV },
+   { " 1x4-convergence,", "mem",  "-p",  "1", "-t",  "4", "-P",  "512", OPT_CONV },
+   { " 1x6-convergence,", "mem",  "-p",  "1", "-t",  "6", "-P", "1020", OPT_CONV },
+   { " 2x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
+   { " 3x3-convergence,", "mem",  "-p",  "3", "-t",  "3", "-P", "1020", OPT_CONV },
+   { " 4x4-convergence,", "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV },
+   { " 4x4-convergence-NOTHP,",
+			  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
+   { " 4x6-convergence,", "mem",  "-p",  "4", "-t",  "6", "-P", "1020", OPT_CONV },
+   { " 4x8-convergence,", "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_CONV },
+   { " 8x4-convergence,", "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV },
+   { " 8x4-convergence-NOTHP,",
+			  "mem",  "-p",  "8", "-t",  "4", "-P",  "512", OPT_CONV_NOTHP },
+   { " 3x1-convergence,", "mem",  "-p",  "3", "-t",  "1", "-P",  "512", OPT_CONV },
+   { " 4x1-convergence,", "mem",  "-p",  "4", "-t",  "1", "-P",  "512", OPT_CONV },
+   { " 8x1-convergence,", "mem",  "-p",  "8", "-t",  "1", "-P",  "512", OPT_CONV },
+   { "16x1-convergence,", "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_CONV },
+   { "32x1-convergence,", "mem",  "-p", "32", "-t",  "1", "-P",  "128", OPT_CONV },
+
+   /* Various NUMA process/thread layout bandwidth measurements: */
+   { " 2x1-bw-process,",  "mem",  "-p",  "2", "-t",  "1", "-P", "1024", OPT_BW },
+   { " 3x1-bw-process,",  "mem",  "-p",  "3", "-t",  "1", "-P", "1024", OPT_BW },
+   { " 4x1-bw-process,",  "mem",  "-p",  "4", "-t",  "1", "-P", "1024", OPT_BW },
+   { " 8x1-bw-process,",  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW },
+   { " 8x1-bw-process-NOTHP,",
+			  "mem",  "-p",  "8", "-t",  "1", "-P", " 512", OPT_BW_NOTHP },
+   { "16x1-bw-process,",  "mem",  "-p", "16", "-t",  "1", "-P",  "256", OPT_BW },
+
+   { " 4x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "4", "-T",  "256", OPT_BW },
+   { " 8x1-bw-thread,",	  "mem",  "-p",  "1", "-t",  "8", "-T",  "256", OPT_BW },
+   { "16x1-bw-thread,",   "mem",  "-p",  "1", "-t", "16", "-T",  "128", OPT_BW },
+   { "32x1-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-T",   "64", OPT_BW },
+
+   { " 2x3-bw-thread,",	  "mem",  "-p",  "2", "-t",  "3", "-P",  "512", OPT_BW },
+   { " 4x4-bw-thread,",	  "mem",  "-p",  "4", "-t",  "4", "-P",  "512", OPT_BW },
+   { " 4x6-bw-thread,",	  "mem",  "-p",  "4", "-t",  "6", "-P",  "512", OPT_BW },
+   { " 4x8-bw-thread,",	  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW },
+   { " 4x8-bw-thread-NOTHP,",
+			  "mem",  "-p",  "4", "-t",  "8", "-P",  "512", OPT_BW_NOTHP },
+   { " 3x3-bw-thread,",	  "mem",  "-p",  "3", "-t",  "3", "-P",  "512", OPT_BW },
+   { " 5x5-bw-thread,",	  "mem",  "-p",  "5", "-t",  "5", "-P",  "512", OPT_BW },
+
+   { "2x16-bw-thread,",   "mem",  "-p",  "2", "-t", "16", "-P",  "512", OPT_BW },
+   { "1x32-bw-thread,",   "mem",  "-p",  "1", "-t", "32", "-P", "2048", OPT_BW },
+
+   { "numa02-bw,",	  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW },
+   { "numa02-bw-NOTHP,",  "mem",  "-p",  "1", "-t", "32", "-T",   "32", OPT_BW_NOTHP },
+   { "numa01-bw-thread,", "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW },
+   { "numa01-bw-thread-NOTHP,",
+			  "mem",  "-p",  "2", "-t", "16", "-T",  "192", OPT_BW_NOTHP },
+};
+
+static int bench_all(void)
+{
+	int nr = ARRAY_SIZE(tests);
+	int ret;
+	int i;
+
+	ret = system("echo ' #'; echo ' # Running test on: '$(uname -a); echo ' #'");
+	BUG_ON(ret < 0);
+
+	for (i = 0; i < nr; i++) {
+		if (run_bench_numa(tests[i][0], tests[i] + 1))
+			return -1;
+	}
+
+	printf("\n");
+
+	return 0;
+}
+
+int bench_numa(int argc, const char **argv, const char *prefix __maybe_unused)
+{
+	init_params(&p0, "main,", argc, argv);
+	argc = parse_options(argc, argv, options, bench_numa_usage, 0);
+	if (argc)
+		goto err;
+
+	if (p0.run_all)
+		return bench_all();
+
+	if (__bench_numa(NULL))
+		goto err;
+
+	return 0;
+
+err:
+	usage_with_options(numa_usage, options);
+	return -1;
+}
diff --git a/tools/perf/builtin-bench.c b/tools/perf/builtin-bench.c
index afd1255a632f..e5d514bf5365 100644
--- a/tools/perf/builtin-bench.c
+++ b/tools/perf/builtin-bench.c
@@ -35,6 +35,16 @@ struct bench_suite {
 /* sentinel: easy for help */
 #define suite_all { "all", "Test all benchmark suites", NULL }
 
+static struct bench_suite numa_suites[] = {
+	{ "mem",
+	  "Benchmark for NUMA workloads",
+	  bench_numa },
+	suite_all,
+	{ NULL,
+	  NULL,
+	  NULL                  }
+};
+
 static struct bench_suite sched_suites[] = {
 	{ "messaging",
 	  "Benchmark for scheduler and IPC mechanisms",
@@ -68,6 +78,9 @@ struct bench_subsys {
 };
 
 static struct bench_subsys subsystems[] = {
+	{ "numa",
+	  "NUMA scheduling and MM behavior",
+	  numa_suites },
 	{ "sched",
 	  "scheduler and IPC mechanism",
 	  sched_suites },