#include <AvailabilityMacros.h>
#include <mach/thread_policy.h>
#include <mach/mach.h>
#include <mach/mach_error.h>
#include <mach/mach_time.h>
#include <pthread.h>
#include <sys/queue.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <err.h>
#include <errno.h>
#include <sys/sysctl.h>
/*
* Sets is a multithreaded test/benchmarking program to evaluate
* affinity set placement in Leopard.
*
* The picture here, for each set, is:
*
* free work
* -> queue --> producer --> queue --> consumer --
* | |
* -----------------------------------------------
*
* <------ "stage" -----> <------ "stage" ----->
*
* We spin off sets of production line threads (2 sets by default).
* All threads of each line sets the same affinity tag (unless disabled).
* By default there are 2 stage (worker) threads per production line.
* A worker thread removes a buffer from an input queue, processses it and
* queues it on an output queue. By default the initial stage (producer)
* writes every byte in a buffer and the other (consumer) stages read every
* byte. By default the buffers are 1MB (256 pages) in size but this can be
* overidden. By default there are 2 buffers per set (again overridable).
* Worker threads process (iterate over) 10000 buffers by default.
*
* With affinity enabled, each producer and consumer thread sets its affinity
* to the set number, 1 .. N. So the threads of each set share an L2 cache.
*
* Buffer management uses pthread mutex/condition variables. A thread blocks
* when no buffer is available on a queue and it is signaled when a buffer
* is placed on an empty queue. Queues are tailq'a a la <sys/queue.h>.
* The queue management is centralized in a single routine: what queues to
* use as input and output and what function to call for processing is
* data-driven.
*/
pthread_mutex_t funnel;
pthread_cond_t barrier;
uint64_t timer;
int threads;
int threads_ready = 0;
int iterations = 10000;
boolean_t affinity = FALSE;
boolean_t halting = FALSE;
boolean_t cache_config = FALSE;
int verbosity = 1;
typedef struct work {
TAILQ_ENTRY(work) link;
int *data;
} work_t;
/*
* A work queue, complete with pthread objects for its management
*/
typedef struct work_queue {
pthread_mutex_t mtx;
pthread_cond_t cnd;
TAILQ_HEAD(, work) queue;
boolean_t waiters;
} work_queue_t;
/* Worker functions take a integer array and size */
typedef void (worker_fn_t)(int *, int);
/* This struct controls the function of a thread */
typedef struct {
int stagenum;
char *name;
worker_fn_t *fn;
work_queue_t *input;
work_queue_t *output;
struct line_info *set;
pthread_t thread;
work_queue_t bufq;
} stage_info_t;
/* This defines a thread set */
#define WORKERS_MAX 10
typedef struct line_info {
int setnum;
int *data;
int isize;
stage_info_t *stage[WORKERS_MAX];
} line_info_t;
#define DBG(x...) do { \
if (verbosity > 1) { \
pthread_mutex_lock(&funnel); \
printf(x); \
pthread_mutex_unlock(&funnel); \
} \
} while (0)
#define mutter(x...) do { \
if (verbosity > 0) { \
printf(x); \
} \
} while (0)
#define s_if_plural(x) (((x) > 1) ? "s" : "")
static void
usage()
{
fprintf(stderr,
"usage: sets [-a] Turn affinity on (off)\n"
" [-b B] Number of buffers per set/line (2)\n"
" [-c] Configure for max cache performance\n"
" [-h] Print this\n"
" [-i I] Number of items/buffers to process (1000)\n"
" [-s S] Number of stages per set/line (2)\n"
" [-t] Halt for keyboard input to start\n"
" [-p P] Number of pages per buffer (256=1MB)]\n"
" [-w] Consumer writes data\n"
" [-v V] Level of verbosity 0..2 (1)\n"
" [N] Number of sets/lines (2)\n"
);
exit(1);
}
/* Trivial producer: write to each byte */
void
writer_fn(int *data, int isize)
{
int i;
for (i = 0; i < isize; i++) {
data[i] = i;
}
}
/* Trivial consumer: read each byte */
void
reader_fn(int *data, int isize)
{
int i;
int datum;
for (i = 0; i < isize; i++) {
datum = data[i];
}
}
/* Consumer reading and writing the buffer */
void
reader_writer_fn(int *data, int isize)
{
int i;
for (i = 0; i < isize; i++) {
data[i] += 1;
}
}
/*
* This is the central function for every thread.
* For each invocation, its role is ets by (a pointer to) a stage_info_t.
*/
void *
manager_fn(void *arg)
{
stage_info_t *sp = (stage_info_t *) arg;
line_info_t *lp = sp->set;
kern_return_t ret;
long iteration = 0;
/*
* If we're using affinity sets (we are by default)
* set our tag to by our thread set number.
*/
thread_extended_policy_data_t epolicy;
thread_affinity_policy_data_t policy;
epolicy.timeshare = FALSE;
ret = thread_policy_set(
mach_thread_self(), THREAD_EXTENDED_POLICY,
(thread_policy_t) &epolicy,
THREAD_EXTENDED_POLICY_COUNT);
if (ret != KERN_SUCCESS) {
printf("thread_policy_set(THREAD_EXTENDED_POLICY) returned %d\n", ret);
}
if (affinity) {
policy.affinity_tag = lp->setnum;
ret = thread_policy_set(
mach_thread_self(), THREAD_AFFINITY_POLICY,
(thread_policy_t) &policy,
THREAD_AFFINITY_POLICY_COUNT);
if (ret != KERN_SUCCESS) {
printf("thread_policy_set(THREAD_AFFINITY_POLICY) returned %d\n", ret);
}
}
DBG("Starting %s set: %d stage: %d\n", sp->name, lp->setnum, sp->stagenum);
/*
* Start barrier.
* The tets thread to get here releases everyone and starts the timer.
*/
pthread_mutex_lock(&funnel);
threads_ready++;
if (threads_ready == threads) {
pthread_mutex_unlock(&funnel);
if (halting) {
printf(" all threads ready for process %d, "
"hit any key to start", getpid());
fflush(stdout);
(void) getchar();
}
pthread_cond_broadcast(&barrier);
timer = mach_absolute_time();
} else {
pthread_cond_wait(&barrier, &funnel);
pthread_mutex_unlock(&funnel);
}
do {
int i;
work_t *workp;
/*
* Get a buffer from the input queue.
* Block if none.
*/
pthread_mutex_lock(&sp->input->mtx);
while (1) {
workp = TAILQ_FIRST(&(sp->input->queue));
if (workp != NULL) {
break;
}
DBG(" %s[%d,%d] iteration %d waiting for buffer\n",
sp->name, lp->setnum, sp->stagenum, iteration);
sp->input->waiters = TRUE;
pthread_cond_wait(&sp->input->cnd, &sp->input->mtx);
sp->input->waiters = FALSE;
}
TAILQ_REMOVE(&(sp->input->queue), workp, link);
pthread_mutex_unlock(&sp->input->mtx);
DBG(" %s[%d,%d] iteration %d work %p data %p\n",
sp->name, lp->setnum, sp->stagenum, iteration, workp, workp->data);
/* Do our stuff with the buffer */
(void) sp->fn(workp->data, lp->isize);
/*
* Place the buffer on the input queue.
* Signal waiters if required.
*/
pthread_mutex_lock(&sp->output->mtx);
TAILQ_INSERT_TAIL(&(sp->output->queue), workp, link);
if (sp->output->waiters) {
DBG(" %s[%d,%d] iteration %d signaling work\n",
sp->name, lp->setnum, sp->stagenum, iteration);
pthread_cond_signal(&sp->output->cnd);
}
pthread_mutex_unlock(&sp->output->mtx);
} while (++iteration < iterations);
DBG("Ending %s[%d,%d]\n", sp->name, lp->setnum, sp->stagenum);
return (void *) iteration;
}
#define MAX_CACHE_DEPTH 10
static void
auto_config(int npages, int *nbufs, int *nsets)
{
size_t len;
int ncpu;
int llc;
int64_t cacheconfig[MAX_CACHE_DEPTH];
int64_t cachesize[MAX_CACHE_DEPTH];
mutter("Autoconfiguring...\n");
len = sizeof(cacheconfig);
if (sysctlbyname("hw.cacheconfig",
&cacheconfig[0], &len, NULL, 0) != 0) {
printf("Unable to get hw.cacheconfig, %d\n", errno);
exit(1);
}
len = sizeof(cachesize);
if (sysctlbyname("hw.cachesize",
&cachesize[0], &len, NULL, 0) != 0) {
printf("Unable to get hw.cachesize, %d\n", errno);
exit(1);
}
/*
* Find LLC
*/
for (llc = MAX_CACHE_DEPTH - 1; llc > 0; llc--) {
if (cacheconfig[llc] != 0) {
break;
}
}
/*
* Calculate number of buffers of size pages*4096 bytes
* fit into 90% of an L2 cache.
*/
*nbufs = cachesize[llc] * 9 / (npages * 4096 * 10);
mutter(" L%d (LLC) cache %qd bytes: "
"using %d buffers of size %d bytes\n",
llc, cachesize[llc], *nbufs, (npages * 4096));
/*
* Calcalute how many sets:
*/
*nsets = cacheconfig[0] / cacheconfig[llc];
mutter(" %qd cpus; %qd cpus per L%d cache: using %d sets\n",
cacheconfig[0], cacheconfig[llc], llc, *nsets);
}
void (*producer_fnp)(int *data, int isize) = &writer_fn;
void (*consumer_fnp)(int *data, int isize) = &reader_fn;
int
main(int argc, char *argv[])
{
int i;
int j;
int pages = 256; /* 1MB */
int buffers = 2;
int sets = 2;
int stages = 2;
int *status;
line_info_t *line_info;
line_info_t *lp;
stage_info_t *stage_info;
stage_info_t *sp;
kern_return_t ret;
int c;
/* Do switch parsing: */
while ((c = getopt(argc, argv, "ab:chi:p:s:twv:")) != -1) {
switch (c) {
case 'a':
affinity = !affinity;
break;
case 'b':
buffers = atoi(optarg);
break;
case 'c':
cache_config = TRUE;
break;
case 'i':
iterations = atoi(optarg);
break;
case 'p':
pages = atoi(optarg);
break;
case 's':
stages = atoi(optarg);
if (stages >= WORKERS_MAX) {
usage();
}
break;
case 't':
halting = TRUE;
break;
case 'w':
consumer_fnp = &reader_writer_fn;
break;
case 'v':
verbosity = atoi(optarg);
break;
case '?':
case 'h':
default:
usage();
}
}
argc -= optind; argv += optind;
if (argc > 0) {
sets = atoi(*argv);
}
if (cache_config) {
auto_config(pages, &buffers, &sets);
}
pthread_mutex_init(&funnel, NULL);
pthread_cond_init(&barrier, NULL);
/*
* Fire up the worker threads.
*/
threads = sets * stages;
mutter("Launching %d set%s of %d threads with %saffinity, "
"consumer reads%s data\n",
sets, s_if_plural(sets), stages, affinity? "": "no ",
(consumer_fnp == &reader_writer_fn)? " and writes" : "");
if (pages < 256) {
mutter(" %dkB bytes per buffer, ", pages * 4);
} else {
mutter(" %dMB bytes per buffer, ", pages / 256);
}
mutter("%d buffer%s per set ",
buffers, s_if_plural(buffers));
if (buffers * pages < 256) {
mutter("(total %dkB)\n", buffers * pages * 4);
} else {
mutter("(total %dMB)\n", buffers * pages / 256);
}
mutter(" processing %d buffer%s...\n",
iterations, s_if_plural(iterations));
line_info = (line_info_t *) malloc(sets * sizeof(line_info_t));
stage_info = (stage_info_t *) malloc(sets * stages * sizeof(stage_info_t));
for (i = 0; i < sets; i++) {
work_t *work_array;
lp = &line_info[i];
lp->setnum = i + 1;
lp->isize = pages * 4096 / sizeof(int);
lp->data = (int *) malloc(buffers * pages * 4096);
/* Set up the queue for the workers of this thread set: */
for (j = 0; j < stages; j++) {
sp = &stage_info[(i * stages) + j];
sp->stagenum = j;
sp->set = lp;
lp->stage[j] = sp;
pthread_mutex_init(&sp->bufq.mtx, NULL);
pthread_cond_init(&sp->bufq.cnd, NULL);
TAILQ_INIT(&sp->bufq.queue);
sp->bufq.waiters = FALSE;
}
/*
* Take a second pass through the stages
* to define what the workers are and to interconnect their input/outputs
*/
for (j = 0; j < stages; j++) {
sp = lp->stage[j];
if (j == 0) {
sp->fn = producer_fnp;
sp->name = "producer";
} else {
sp->fn = consumer_fnp;
sp->name = "consumer";
}
sp->input = &lp->stage[j]->bufq;
sp->output = &lp->stage[(j + 1) % stages]->bufq;
}
/* Set up the buffers on the first worker of the set. */
work_array = (work_t *) malloc(buffers * sizeof(work_t));
for (j = 0; j < buffers; j++) {
work_array[j].data = lp->data + (lp->isize * j);
TAILQ_INSERT_TAIL(&lp->stage[0]->bufq.queue, &work_array[j], link);
DBG(" empty work item %p for set %d data %p\n",
&work_array[j], i, work_array[j].data);
}
/* Create this set of threads */
for (j = 0; j < stages; j++) {
if (ret = pthread_create(&lp->stage[j]->thread, NULL,
&manager_fn,
(void *) lp->stage[j])) {
err(1, "pthread_create %d,%d", i, j);
}
}
}
/*
* We sit back anf wait for the slave to finish.
*/
for (i = 0; i < sets; i++) {
lp = &line_info[i];
for (j = 0; j < stages; j++) {
if (ret = pthread_join(lp->stage[j]->thread, (void **)&status)) {
err(1, "pthread_join %d,%d", i, j);
}
DBG("Thread %d,%d status %d\n", i, j, status);
}
}
/*
* See how long the work took.
*/
timer = mach_absolute_time() - timer;
timer = timer / 1000000ULL;
printf("%d.%03d seconds elapsed.\n",
(int) (timer / 1000ULL), (int) (timer % 1000ULL));
return 0;
}