2 * threadpool-worker.c: native threadpool worker
5 * Ludovic Henry (ludovic.henry@xamarin.com)
7 * Licensed under the MIT license. See LICENSE file in the project root for full license information.
11 #define _USE_MATH_DEFINES // needed by MSVC to define math constants
16 #include <mono/metadata/class-internals.h>
17 #include <mono/metadata/exception.h>
18 #include <mono/metadata/gc-internals.h>
19 #include <mono/metadata/object.h>
20 #include <mono/metadata/object-internals.h>
21 #include <mono/metadata/threadpool.h>
22 #include <mono/metadata/threadpool-worker.h>
23 #include <mono/metadata/threadpool-io.h>
24 #include <mono/metadata/w32event.h>
25 #include <mono/utils/atomic.h>
26 #include <mono/utils/mono-compiler.h>
27 #include <mono/utils/mono-complex.h>
28 #include <mono/utils/mono-logger.h>
29 #include <mono/utils/mono-logger-internals.h>
30 #include <mono/utils/mono-proclib.h>
31 #include <mono/utils/mono-threads.h>
32 #include <mono/utils/mono-time.h>
33 #include <mono/utils/mono-rand.h>
34 #include <mono/utils/refcount.h>
35 #include <mono/utils/w32api.h>
37 #define CPU_USAGE_LOW 80
38 #define CPU_USAGE_HIGH 95
40 #define MONITOR_INTERVAL 500 // ms
41 #define MONITOR_MINIMAL_LIFETIME 60 * 1000 // ms
43 #define WORKER_CREATION_MAX_PER_SEC 10
45 /* The exponent to apply to the gain. 1.0 means to use linear gain,
46 * higher values will enhance large moves and damp small ones.
48 #define HILL_CLIMBING_GAIN_EXPONENT 2.0
50 /* The 'cost' of a thread. 0 means drive for increased throughput regardless
51 * of thread count, higher values bias more against higher thread counts.
53 #define HILL_CLIMBING_BIAS 0.15
55 #define HILL_CLIMBING_WAVE_PERIOD 4
56 #define HILL_CLIMBING_MAX_WAVE_MAGNITUDE 20
57 #define HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER 1.0
58 #define HILL_CLIMBING_WAVE_HISTORY_SIZE 8
59 #define HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO 3.0
60 #define HILL_CLIMBING_MAX_CHANGE_PER_SECOND 4
61 #define HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE 20
62 #define HILL_CLIMBING_SAMPLE_INTERVAL_LOW 10
63 #define HILL_CLIMBING_SAMPLE_INTERVAL_HIGH 200
64 #define HILL_CLIMBING_ERROR_SMOOTHING_FACTOR 0.01
65 #define HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT 0.15
69 TRANSITION_INITIALIZING,
70 TRANSITION_RANDOM_MOVE,
71 TRANSITION_CLIMBING_MOVE,
72 TRANSITION_CHANGE_POINT,
73 TRANSITION_STABILIZING,
74 TRANSITION_STARVATION,
75 TRANSITION_THREAD_TIMED_OUT,
77 } ThreadPoolHeuristicStateTransition;
81 gint32 samples_to_measure;
82 gdouble target_throughput_ratio;
83 gdouble target_signal_to_noise_ratio;
84 gdouble max_change_per_second;
85 gdouble max_change_per_sample;
86 gint32 max_thread_wave_magnitude;
87 gint32 sample_interval_low;
88 gdouble thread_magnitude_multiplier;
89 gint32 sample_interval_high;
90 gdouble throughput_error_smoothing_factor;
91 gdouble gain_exponent;
92 gdouble max_sample_error;
94 gdouble current_control_setting;
96 gint16 last_thread_count;
97 gdouble elapsed_since_last_change;
98 gdouble completions_since_last_change;
100 gdouble average_throughput_noise;
103 gdouble *thread_counts;
105 guint32 current_sample_interval;
106 gpointer random_interval_generator;
108 gint32 accumulated_completion_count;
109 gdouble accumulated_sample_duration;
110 } ThreadPoolHillClimbing;
113 MonoThreadPoolWorkerCallback callback;
115 } ThreadPoolWorkItem;
119 gint16 max_working; /* determined by heuristic */
120 gint16 starting; /* starting, but not yet in worker_thread */
121 gint16 working; /* executing worker_thread */
122 gint16 parked; /* parked */
125 } ThreadPoolWorkerCounter;
127 typedef MonoInternalThread ThreadPoolWorkerThread;
132 ThreadPoolWorkerCounter counters;
134 GPtrArray *threads; // ThreadPoolWorkerThread* []
135 MonoCoopMutex threads_lock; /* protect access to working_threads and parked_threads */
136 gint32 parked_threads_count;
137 MonoCoopCond parked_threads_cond;
138 MonoCoopCond threads_exit_cond;
140 ThreadPoolWorkItem *work_items; // ThreadPoolWorkItem []
141 gint32 work_items_count;
142 gint32 work_items_size;
143 MonoCoopMutex work_items_lock;
145 guint32 worker_creation_current_second;
146 guint32 worker_creation_current_count;
147 MonoCoopMutex worker_creation_lock;
149 gint32 heuristic_completions;
150 gint64 heuristic_sample_start;
151 gint64 heuristic_last_dequeue; // ms
152 gint64 heuristic_last_adjustment; // ms
153 gint64 heuristic_adjustment_interval; // ms
154 ThreadPoolHillClimbing heuristic_hill_climbing;
155 MonoCoopMutex heuristic_lock;
157 gint32 limit_worker_min;
158 gint32 limit_worker_max;
160 MonoCpuUsageState *cpu_usage_state;
163 /* suspended by the debugger */
166 gint32 monitor_status;
170 MONITOR_STATUS_REQUESTED,
171 MONITOR_STATUS_WAITING_FOR_REQUEST,
172 MONITOR_STATUS_NOT_RUNNING,
175 static ThreadPoolWorker worker;
177 #define COUNTER_CHECK(counter) \
179 g_assert (counter._.max_working > 0); \
180 g_assert (counter._.starting >= 0); \
181 g_assert (counter._.working >= 0); \
184 #define COUNTER_ATOMIC(var,block) \
186 ThreadPoolWorkerCounter __old; \
188 __old = COUNTER_READ (); \
191 COUNTER_CHECK (var); \
192 } while (InterlockedCompareExchange64 (&worker.counters.as_gint64, (var).as_gint64, __old.as_gint64) != __old.as_gint64); \
195 static inline ThreadPoolWorkerCounter
198 ThreadPoolWorkerCounter counter;
199 counter.as_gint64 = InterlockedRead64 (&worker.counters.as_gint64);
207 return mono_rand_init (NULL, 0);
211 rand_next (gpointer *handle, guint32 min, guint32 max)
215 mono_rand_try_get_uint32 (handle, &val, min, max, &error);
216 // FIXME handle error
217 mono_error_assert_ok (&error);
222 destroy (gpointer data)
225 mono_coop_mutex_destroy (&worker.threads_lock);
226 mono_coop_cond_destroy (&worker.parked_threads_cond);
228 mono_coop_mutex_destroy (&worker.work_items_lock);
230 mono_coop_mutex_destroy (&worker.worker_creation_lock);
232 mono_coop_mutex_destroy (&worker.heuristic_lock);
234 g_free (worker.cpu_usage_state);
239 mono_threadpool_worker_init (void)
241 ThreadPoolHillClimbing *hc;
242 const char *threads_per_cpu_env;
243 gint threads_per_cpu;
246 mono_refcount_init (&worker, destroy);
248 worker.threads = g_ptr_array_new ();
249 mono_coop_mutex_init (&worker.threads_lock);
250 worker.parked_threads_count = 0;
251 mono_coop_cond_init (&worker.parked_threads_cond);
252 mono_coop_cond_init (&worker.threads_exit_cond);
254 /* worker.work_items_size is inited to 0 */
255 mono_coop_mutex_init (&worker.work_items_lock);
257 worker.worker_creation_current_second = -1;
258 mono_coop_mutex_init (&worker.worker_creation_lock);
260 worker.heuristic_adjustment_interval = 10;
261 mono_coop_mutex_init (&worker.heuristic_lock);
265 hc = &worker.heuristic_hill_climbing;
267 hc->wave_period = HILL_CLIMBING_WAVE_PERIOD;
268 hc->max_thread_wave_magnitude = HILL_CLIMBING_MAX_WAVE_MAGNITUDE;
269 hc->thread_magnitude_multiplier = (gdouble) HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER;
270 hc->samples_to_measure = hc->wave_period * HILL_CLIMBING_WAVE_HISTORY_SIZE;
271 hc->target_throughput_ratio = (gdouble) HILL_CLIMBING_BIAS;
272 hc->target_signal_to_noise_ratio = (gdouble) HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO;
273 hc->max_change_per_second = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SECOND;
274 hc->max_change_per_sample = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE;
275 hc->sample_interval_low = HILL_CLIMBING_SAMPLE_INTERVAL_LOW;
276 hc->sample_interval_high = HILL_CLIMBING_SAMPLE_INTERVAL_HIGH;
277 hc->throughput_error_smoothing_factor = (gdouble) HILL_CLIMBING_ERROR_SMOOTHING_FACTOR;
278 hc->gain_exponent = (gdouble) HILL_CLIMBING_GAIN_EXPONENT;
279 hc->max_sample_error = (gdouble) HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT;
280 hc->current_control_setting = 0;
281 hc->total_samples = 0;
282 hc->last_thread_count = 0;
283 hc->average_throughput_noise = 0;
284 hc->elapsed_since_last_change = 0;
285 hc->accumulated_completion_count = 0;
286 hc->accumulated_sample_duration = 0;
287 hc->samples = g_new0 (gdouble, hc->samples_to_measure);
288 hc->thread_counts = g_new0 (gdouble, hc->samples_to_measure);
289 hc->random_interval_generator = rand_create ();
290 hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
292 if (!(threads_per_cpu_env = g_getenv ("MONO_THREADS_PER_CPU")))
295 threads_per_cpu = CLAMP (atoi (threads_per_cpu_env), 1, 50);
297 threads_count = mono_cpu_count () * threads_per_cpu;
299 worker.limit_worker_min = threads_count;
301 #if defined (PLATFORM_ANDROID) || defined (HOST_IOS)
302 worker.limit_worker_max = CLAMP (threads_count * 100, MIN (threads_count, 200), MAX (threads_count, 200));
304 worker.limit_worker_max = threads_count * 100;
307 worker.counters._.max_working = worker.limit_worker_min;
309 worker.cpu_usage_state = g_new0 (MonoCpuUsageState, 1);
311 worker.suspended = FALSE;
313 worker.monitor_status = MONITOR_STATUS_NOT_RUNNING;
317 mono_threadpool_worker_cleanup (void)
319 MonoInternalThread *current;
321 /* we make the assumption along the code that we are
322 * cleaning up only if the runtime is shutting down */
323 g_assert (mono_runtime_is_shutting_down ());
325 current = mono_thread_internal_current ();
327 while (worker.monitor_status != MONITOR_STATUS_NOT_RUNNING)
328 mono_thread_info_sleep (1, NULL);
330 mono_coop_mutex_lock (&worker.threads_lock);
332 /* unpark all worker.parked_threads */
333 mono_coop_cond_broadcast (&worker.parked_threads_cond);
337 ThreadPoolWorkerCounter counter;
339 counter = COUNTER_READ ();
340 if (counter._.starting + counter._.working + counter._.parked == 0)
343 if (counter._.starting + counter._.working + counter._.parked == 1) {
344 if (worker.threads->len == 1 && g_ptr_array_index (worker.threads, 0) == current) {
345 /* We are waiting on ourselves */
350 mono_coop_cond_wait (&worker.threads_exit_cond, &worker.threads_lock);
354 mono_coop_mutex_unlock (&worker.threads_lock);
356 mono_refcount_dec (&worker);
360 work_item_lock (void)
362 mono_coop_mutex_lock (&worker.work_items_lock);
366 work_item_unlock (void)
368 mono_coop_mutex_unlock (&worker.work_items_lock);
372 work_item_push (MonoThreadPoolWorkerCallback callback, gpointer data)
374 ThreadPoolWorkItem work_item;
378 work_item.callback = callback;
379 work_item.data = data;
383 g_assert (worker.work_items_count <= worker.work_items_size);
385 if (G_UNLIKELY (worker.work_items_count == worker.work_items_size)) {
386 worker.work_items_size += 64;
387 worker.work_items = g_renew (ThreadPoolWorkItem, worker.work_items, worker.work_items_size);
390 g_assert (worker.work_items);
392 worker.work_items [worker.work_items_count ++] = work_item;
394 // printf ("[push] worker.work_items = %p, worker.work_items_count = %d, worker.work_items_size = %d\n",
395 // worker.work_items, worker.work_items_count, worker.work_items_size);
401 work_item_try_pop (ThreadPoolWorkItem *work_item)
403 g_assert (work_item);
407 // printf ("[pop] worker.work_items = %p, worker.work_items_count = %d, worker.work_items_size = %d\n",
408 // worker.work_items, worker.work_items_count, worker.work_items_size);
410 if (worker.work_items_count == 0) {
415 *work_item = worker.work_items [-- worker.work_items_count];
417 if (G_UNLIKELY (worker.work_items_count >= 64 * 3 && worker.work_items_count < worker.work_items_size / 2)) {
418 worker.work_items_size -= 64;
419 worker.work_items = g_renew (ThreadPoolWorkItem, worker.work_items, worker.work_items_size);
428 work_item_count (void)
433 count = worker.work_items_count;
439 static void worker_request (void);
442 mono_threadpool_worker_enqueue (MonoThreadPoolWorkerCallback callback, gpointer data)
444 work_item_push (callback, data);
450 worker_wait_interrupt (gpointer unused)
452 mono_coop_mutex_lock (&worker.threads_lock);
453 mono_coop_cond_signal (&worker.parked_threads_cond);
454 mono_coop_mutex_unlock (&worker.threads_lock);
456 mono_refcount_dec (&worker);
459 /* return TRUE if timeout, FALSE otherwise (worker unpark or interrupt) */
463 gboolean timeout = FALSE;
465 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker parking", mono_native_thread_id_get ());
467 mono_coop_mutex_lock (&worker.threads_lock);
469 if (!mono_runtime_is_shutting_down ()) {
470 static gpointer rand_handle = NULL;
471 MonoInternalThread *thread;
472 gboolean interrupted = FALSE;
473 ThreadPoolWorkerCounter counter;
476 rand_handle = rand_create ();
477 g_assert (rand_handle);
479 thread = mono_thread_internal_current ();
482 COUNTER_ATOMIC (counter, {
483 counter._.working --;
487 worker.parked_threads_count += 1;
489 mono_refcount_inc (&worker);
490 mono_thread_info_install_interrupt (worker_wait_interrupt, NULL, &interrupted);
492 mono_refcount_dec (&worker);
496 if (mono_coop_cond_timedwait (&worker.parked_threads_cond, &worker.threads_lock, rand_next (&rand_handle, 5 * 1000, 60 * 1000)) != 0)
499 mono_thread_info_uninstall_interrupt (&interrupted);
501 mono_refcount_dec (&worker);
504 worker.parked_threads_count -= 1;
506 COUNTER_ATOMIC (counter, {
507 counter._.working ++;
512 mono_coop_mutex_unlock (&worker.threads_lock);
514 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker unparking, timeout? %s", mono_native_thread_id_get (), timeout ? "yes" : "no");
520 worker_try_unpark (void)
522 gboolean res = FALSE;
524 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker", mono_native_thread_id_get ());
526 mono_coop_mutex_lock (&worker.threads_lock);
527 if (worker.parked_threads_count > 0) {
528 mono_coop_cond_signal (&worker.parked_threads_cond);
531 mono_coop_mutex_unlock (&worker.threads_lock);
533 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker, success? %s", mono_native_thread_id_get (), res ? "yes" : "no");
539 worker_thread (gpointer unused)
541 MonoInternalThread *thread;
542 ThreadPoolWorkerCounter counter;
544 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker starting", mono_native_thread_id_get ());
546 COUNTER_ATOMIC (counter, {
547 counter._.starting --;
548 counter._.working ++;
551 thread = mono_thread_internal_current ();
554 mono_coop_mutex_lock (&worker.threads_lock);
555 g_ptr_array_add (worker.threads, thread);
556 mono_coop_mutex_unlock (&worker.threads_lock);
558 while (!mono_runtime_is_shutting_down ()) {
559 ThreadPoolWorkItem work_item;
561 if (mono_thread_interruption_checkpoint ())
564 if (!work_item_try_pop (&work_item)) {
567 timeout = worker_park ();
574 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker executing %p (%p)",
575 mono_native_thread_id_get (), work_item.callback, work_item.data);
577 work_item.callback (work_item.data);
580 mono_coop_mutex_lock (&worker.threads_lock);
582 COUNTER_ATOMIC (counter, {
583 counter._.working --;
586 g_ptr_array_remove (worker.threads, thread);
588 mono_coop_cond_signal (&worker.threads_exit_cond);
590 mono_coop_mutex_unlock (&worker.threads_lock);
592 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker finishing", mono_native_thread_id_get ());
594 mono_refcount_dec (&worker);
600 worker_try_create (void)
603 MonoInternalThread *thread;
604 gint64 current_ticks;
606 ThreadPoolWorkerCounter counter;
608 if (mono_runtime_is_shutting_down ())
611 mono_coop_mutex_lock (&worker.worker_creation_lock);
613 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker", mono_native_thread_id_get ());
615 current_ticks = mono_100ns_ticks ();
616 if (0 == current_ticks) {
617 g_warning ("failed to get 100ns ticks");
619 now = current_ticks / (10 * 1000 * 1000);
620 if (worker.worker_creation_current_second != now) {
621 worker.worker_creation_current_second = now;
622 worker.worker_creation_current_count = 0;
624 g_assert (worker.worker_creation_current_count <= WORKER_CREATION_MAX_PER_SEC);
625 if (worker.worker_creation_current_count == WORKER_CREATION_MAX_PER_SEC) {
626 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of worker created per second reached, current count = %d",
627 mono_native_thread_id_get (), worker.worker_creation_current_count);
628 mono_coop_mutex_unlock (&worker.worker_creation_lock);
634 COUNTER_ATOMIC (counter, {
635 if (counter._.working >= counter._.max_working) {
636 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of working threads reached",
637 mono_native_thread_id_get ());
638 mono_coop_mutex_unlock (&worker.worker_creation_lock);
641 counter._.starting ++;
644 mono_refcount_inc (&worker);
645 thread = mono_thread_create_internal (mono_get_root_domain (), worker_thread, NULL, TRUE, 0, &error);
647 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: could not create thread due to %s", mono_native_thread_id_get (), mono_error_get_message (&error));
648 mono_error_cleanup (&error);
650 COUNTER_ATOMIC (counter, {
651 counter._.starting --;
654 mono_coop_mutex_unlock (&worker.worker_creation_lock);
656 mono_refcount_dec (&worker);
661 worker.worker_creation_current_count += 1;
663 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, created %p, now = %d count = %d",
664 mono_native_thread_id_get (), (gpointer) thread->tid, now, worker.worker_creation_current_count);
666 mono_coop_mutex_unlock (&worker.worker_creation_lock);
670 static void monitor_ensure_running (void);
673 worker_request (void)
675 if (worker.suspended)
678 monitor_ensure_running ();
680 if (worker_try_unpark ()) {
681 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, unparked", mono_native_thread_id_get ());
685 if (worker_try_create ()) {
686 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, created", mono_native_thread_id_get ());
690 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, failed", mono_native_thread_id_get ());
694 monitor_should_keep_running (void)
696 static gint64 last_should_keep_running = -1;
698 g_assert (worker.monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || worker.monitor_status == MONITOR_STATUS_REQUESTED);
700 if (InterlockedExchange (&worker.monitor_status, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST) {
701 gboolean should_keep_running = TRUE, force_should_keep_running = FALSE;
703 if (mono_runtime_is_shutting_down ()) {
704 should_keep_running = FALSE;
706 if (work_item_count () == 0)
707 should_keep_running = FALSE;
709 if (!should_keep_running) {
710 if (last_should_keep_running == -1 || mono_100ns_ticks () - last_should_keep_running < MONITOR_MINIMAL_LIFETIME * 1000 * 10) {
711 should_keep_running = force_should_keep_running = TRUE;
716 if (should_keep_running) {
717 if (last_should_keep_running == -1 || !force_should_keep_running)
718 last_should_keep_running = mono_100ns_ticks ();
720 last_should_keep_running = -1;
721 if (InterlockedCompareExchange (&worker.monitor_status, MONITOR_STATUS_NOT_RUNNING, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST)
726 g_assert (worker.monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || worker.monitor_status == MONITOR_STATUS_REQUESTED);
732 monitor_sufficient_delay_since_last_dequeue (void)
736 if (worker.cpu_usage < CPU_USAGE_LOW) {
737 threshold = MONITOR_INTERVAL;
739 ThreadPoolWorkerCounter counter;
740 counter = COUNTER_READ ();
741 threshold = counter._.max_working * MONITOR_INTERVAL * 2;
744 return mono_msec_ticks () >= worker.heuristic_last_dequeue + threshold;
747 static void hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition);
750 monitor_thread (gpointer unused)
752 MonoInternalThread *internal;
755 internal = mono_thread_internal_current ();
758 mono_cpu_usage (worker.cpu_usage_state);
760 // printf ("monitor_thread: start\n");
762 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, started", mono_native_thread_id_get ());
765 ThreadPoolWorkerCounter counter;
766 gboolean limit_worker_max_reached;
767 gint32 interval_left = MONITOR_INTERVAL;
768 gint32 awake = 0; /* number of spurious awakes we tolerate before doing a round of rebalancing */
770 g_assert (worker.monitor_status != MONITOR_STATUS_NOT_RUNNING);
772 // counter = COUNTER_READ ();
773 // printf ("monitor_thread: starting = %d working = %d parked = %d max_working = %d\n",
774 // counter._.starting, counter._.working, counter._.parked, counter._.max_working);
778 gboolean alerted = FALSE;
780 if (mono_runtime_is_shutting_down ())
783 ts = mono_msec_ticks ();
784 if (mono_thread_info_sleep (interval_left, &alerted) == 0)
786 interval_left -= mono_msec_ticks () - ts;
788 g_assert (!(internal->state & ThreadState_StopRequested));
789 mono_thread_interruption_checkpoint ();
790 } while (interval_left > 0 && ++awake < 10);
792 if (mono_runtime_is_shutting_down ())
795 if (worker.suspended)
798 if (work_item_count () == 0)
801 worker.cpu_usage = mono_cpu_usage (worker.cpu_usage_state);
803 if (!monitor_sufficient_delay_since_last_dequeue ())
806 limit_worker_max_reached = FALSE;
808 COUNTER_ATOMIC (counter, {
809 if (counter._.max_working >= worker.limit_worker_max) {
810 limit_worker_max_reached = TRUE;
813 counter._.max_working ++;
816 if (limit_worker_max_reached)
819 hill_climbing_force_change (counter._.max_working, TRANSITION_STARVATION);
821 for (i = 0; i < 5; ++i) {
822 if (mono_runtime_is_shutting_down ())
825 if (worker_try_unpark ()) {
826 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, unparked", mono_native_thread_id_get ());
830 if (worker_try_create ()) {
831 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, created", mono_native_thread_id_get ());
835 } while (monitor_should_keep_running ());
837 // printf ("monitor_thread: stop\n");
839 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, finished", mono_native_thread_id_get ());
845 monitor_ensure_running (void)
849 switch (worker.monitor_status) {
850 case MONITOR_STATUS_REQUESTED:
851 // printf ("monitor_thread: requested\n");
853 case MONITOR_STATUS_WAITING_FOR_REQUEST:
854 // printf ("monitor_thread: waiting for request\n");
855 InterlockedCompareExchange (&worker.monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_WAITING_FOR_REQUEST);
857 case MONITOR_STATUS_NOT_RUNNING:
858 // printf ("monitor_thread: not running\n");
859 if (mono_runtime_is_shutting_down ())
861 if (InterlockedCompareExchange (&worker.monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_NOT_RUNNING) == MONITOR_STATUS_NOT_RUNNING) {
862 // printf ("monitor_thread: creating\n");
863 if (!mono_thread_create_internal (mono_get_root_domain (), monitor_thread, NULL, TRUE, SMALL_STACK, &error)) {
864 // printf ("monitor_thread: creating failed\n");
865 worker.monitor_status = MONITOR_STATUS_NOT_RUNNING;
866 mono_error_cleanup (&error);
871 default: g_assert_not_reached ();
877 hill_climbing_change_thread_count (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
879 ThreadPoolHillClimbing *hc;
881 hc = &worker.heuristic_hill_climbing;
883 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] hill climbing, change max number of threads %d", mono_native_thread_id_get (), new_thread_count);
885 hc->last_thread_count = new_thread_count;
886 hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
887 hc->elapsed_since_last_change = 0;
888 hc->completions_since_last_change = 0;
892 hill_climbing_force_change (gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
894 ThreadPoolHillClimbing *hc;
896 hc = &worker.heuristic_hill_climbing;
898 if (new_thread_count != hc->last_thread_count) {
899 hc->current_control_setting += new_thread_count - hc->last_thread_count;
900 hill_climbing_change_thread_count (new_thread_count, transition);
904 static double_complex
905 hill_climbing_get_wave_component (gdouble *samples, guint sample_count, gdouble period)
907 ThreadPoolHillClimbing *hc;
908 gdouble w, cosine, sine, coeff, q0, q1, q2;
911 g_assert (sample_count >= period);
912 g_assert (period >= 2);
914 hc = &worker.heuristic_hill_climbing;
916 w = 2.0 * M_PI / period;
919 coeff = 2.0 * cosine;
922 for (i = 0; i < sample_count; ++i) {
923 q0 = coeff * q1 - q2 + samples [(hc->total_samples - sample_count + i) % hc->samples_to_measure];
928 return mono_double_complex_scalar_div (mono_double_complex_make (q1 - q2 * cosine, (q2 * sine)), ((gdouble)sample_count));
932 hill_climbing_update (gint16 current_thread_count, guint32 sample_duration, gint32 completions, gint64 *adjustment_interval)
934 ThreadPoolHillClimbing *hc;
935 ThreadPoolHeuristicStateTransition transition;
937 gdouble throughput_error_estimate;
943 gint new_thread_wave_magnitude;
944 gint new_thread_count;
945 double_complex thread_wave_component;
946 double_complex throughput_wave_component;
947 double_complex ratio;
949 g_assert (adjustment_interval);
951 hc = &worker.heuristic_hill_climbing;
953 /* If someone changed the thread count without telling us, update our records accordingly. */
954 if (current_thread_count != hc->last_thread_count)
955 hill_climbing_force_change (current_thread_count, TRANSITION_INITIALIZING);
957 /* Update the cumulative stats for this thread count */
958 hc->elapsed_since_last_change += sample_duration;
959 hc->completions_since_last_change += completions;
961 /* Add in any data we've already collected about this sample */
962 sample_duration += hc->accumulated_sample_duration;
963 completions += hc->accumulated_completion_count;
965 /* We need to make sure we're collecting reasonably accurate data. Since we're just counting the end
966 * of each work item, we are goinng to be missing some data about what really happened during the
967 * sample interval. The count produced by each thread includes an initial work item that may have
968 * started well before the start of the interval, and each thread may have been running some new
969 * work item for some time before the end of the interval, which did not yet get counted. So
970 * our count is going to be off by +/- threadCount workitems.
972 * The exception is that the thread that reported to us last time definitely wasn't running any work
973 * at that time, and the thread that's reporting now definitely isn't running a work item now. So
974 * we really only need to consider threadCount-1 threads.
976 * Thus the percent error in our count is +/- (threadCount-1)/numCompletions.
978 * We cannot rely on the frequency-domain analysis we'll be doing later to filter out this error, because
979 * of the way it accumulates over time. If this sample is off by, say, 33% in the negative direction,
980 * then the next one likely will be too. The one after that will include the sum of the completions
981 * we missed in the previous samples, and so will be 33% positive. So every three samples we'll have
982 * two "low" samples and one "high" sample. This will appear as periodic variation right in the frequency
983 * range we're targeting, which will not be filtered by the frequency-domain translation. */
984 if (hc->total_samples > 0 && ((current_thread_count - 1.0) / completions) >= hc->max_sample_error) {
985 /* Not accurate enough yet. Let's accumulate the data so
986 * far, and tell the ThreadPoolWorker to collect a little more. */
987 hc->accumulated_sample_duration = sample_duration;
988 hc->accumulated_completion_count = completions;
989 *adjustment_interval = 10;
990 return current_thread_count;
993 /* We've got enouugh data for our sample; reset our accumulators for next time. */
994 hc->accumulated_sample_duration = 0;
995 hc->accumulated_completion_count = 0;
997 /* Add the current thread count and throughput sample to our history. */
998 throughput = ((gdouble) completions) / sample_duration;
1000 sample_index = hc->total_samples % hc->samples_to_measure;
1001 hc->samples [sample_index] = throughput;
1002 hc->thread_counts [sample_index] = current_thread_count;
1003 hc->total_samples ++;
1005 /* Set up defaults for our metrics. */
1006 thread_wave_component = mono_double_complex_make(0, 0);
1007 throughput_wave_component = mono_double_complex_make(0, 0);
1008 throughput_error_estimate = 0;
1009 ratio = mono_double_complex_make(0, 0);
1012 transition = TRANSITION_WARMUP;
1014 /* How many samples will we use? It must be at least the three wave periods we're looking for, and it must also
1015 * be a whole multiple of the primary wave's period; otherwise the frequency we're looking for will fall between
1016 * two frequency bands in the Fourier analysis, and we won't be able to measure it accurately. */
1017 sample_count = ((gint) MIN (hc->total_samples - 1, hc->samples_to_measure) / hc->wave_period) * hc->wave_period;
1019 if (sample_count > hc->wave_period) {
1021 gdouble average_throughput;
1022 gdouble average_thread_count;
1023 gdouble sample_sum = 0;
1024 gdouble thread_sum = 0;
1026 /* Average the throughput and thread count samples, so we can scale the wave magnitudes later. */
1027 for (i = 0; i < sample_count; ++i) {
1028 guint j = (hc->total_samples - sample_count + i) % hc->samples_to_measure;
1029 sample_sum += hc->samples [j];
1030 thread_sum += hc->thread_counts [j];
1033 average_throughput = sample_sum / sample_count;
1034 average_thread_count = thread_sum / sample_count;
1036 if (average_throughput > 0 && average_thread_count > 0) {
1037 gdouble noise_for_confidence, adjacent_period_1, adjacent_period_2;
1039 /* Calculate the periods of the adjacent frequency bands we'll be using to
1040 * measure noise levels. We want the two adjacent Fourier frequency bands. */
1041 adjacent_period_1 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) + 1);
1042 adjacent_period_2 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) - 1);
1044 /* Get the the three different frequency components of the throughput (scaled by average
1045 * throughput). Our "error" estimate (the amount of noise that might be present in the
1046 * frequency band we're really interested in) is the average of the adjacent bands. */
1047 throughput_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, hc->wave_period), average_throughput);
1048 throughput_error_estimate = cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->samples, sample_count, adjacent_period_1), average_throughput));
1050 if (adjacent_period_2 <= sample_count) {
1051 throughput_error_estimate = MAX (throughput_error_estimate, cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (
1052 hc->samples, sample_count, adjacent_period_2), average_throughput)));
1055 /* Do the same for the thread counts, so we have something to compare to. We don't
1056 * measure thread count noise, because there is none; these are exact measurements. */
1057 thread_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (hc->thread_counts, sample_count, hc->wave_period), average_thread_count);
1059 /* Update our moving average of the throughput noise. We'll use this
1060 * later as feedback to determine the new size of the thread wave. */
1061 if (hc->average_throughput_noise == 0) {
1062 hc->average_throughput_noise = throughput_error_estimate;
1064 hc->average_throughput_noise = (hc->throughput_error_smoothing_factor * throughput_error_estimate)
1065 + ((1.0 + hc->throughput_error_smoothing_factor) * hc->average_throughput_noise);
1068 if (cabs (thread_wave_component) > 0) {
1069 /* Adjust the throughput wave so it's centered around the target wave,
1070 * and then calculate the adjusted throughput/thread ratio. */
1071 ratio = mono_double_complex_div (mono_double_complex_sub (throughput_wave_component, mono_double_complex_scalar_mul(thread_wave_component, hc->target_throughput_ratio)), thread_wave_component);
1072 transition = TRANSITION_CLIMBING_MOVE;
1074 ratio = mono_double_complex_make (0, 0);
1075 transition = TRANSITION_STABILIZING;
1078 noise_for_confidence = MAX (hc->average_throughput_noise, throughput_error_estimate);
1079 if (noise_for_confidence > 0) {
1080 confidence = cabs (thread_wave_component) / noise_for_confidence / hc->target_signal_to_noise_ratio;
1082 /* there is no noise! */
1088 /* We use just the real part of the complex ratio we just calculated. If the throughput signal
1089 * is exactly in phase with the thread signal, this will be the same as taking the magnitude of
1090 * the complex move and moving that far up. If they're 180 degrees out of phase, we'll move
1091 * backward (because this indicates that our changes are having the opposite of the intended effect).
1092 * If they're 90 degrees out of phase, we won't move at all, because we can't tell wether we're
1093 * having a negative or positive effect on throughput. */
1094 move = creal (ratio);
1095 move = CLAMP (move, -1.0, 1.0);
1097 /* Apply our confidence multiplier. */
1098 move *= CLAMP (confidence, -1.0, 1.0);
1100 /* Now apply non-linear gain, such that values around zero are attenuated, while higher values
1101 * are enhanced. This allows us to move quickly if we're far away from the target, but more slowly
1102 * if we're getting close, giving us rapid ramp-up without wild oscillations around the target. */
1103 gain = hc->max_change_per_second * sample_duration;
1104 move = pow (fabs (move), hc->gain_exponent) * (move >= 0.0 ? 1 : -1) * gain;
1105 move = MIN (move, hc->max_change_per_sample);
1107 /* If the result was positive, and CPU is > 95%, refuse the move. */
1108 if (move > 0.0 && worker.cpu_usage > CPU_USAGE_HIGH)
1111 /* Apply the move to our control setting. */
1112 hc->current_control_setting += move;
1114 /* Calculate the new thread wave magnitude, which is based on the moving average we've been keeping of the
1115 * throughput error. This average starts at zero, so we'll start with a nice safe little wave at first. */
1116 new_thread_wave_magnitude = (gint)(0.5 + (hc->current_control_setting * hc->average_throughput_noise
1117 * hc->target_signal_to_noise_ratio * hc->thread_magnitude_multiplier * 2.0));
1118 new_thread_wave_magnitude = CLAMP (new_thread_wave_magnitude, 1, hc->max_thread_wave_magnitude);
1120 /* Make sure our control setting is within the ThreadPoolWorker's limits. */
1121 hc->current_control_setting = CLAMP (hc->current_control_setting, worker.limit_worker_min, worker.limit_worker_max - new_thread_wave_magnitude);
1123 /* Calculate the new thread count (control setting + square wave). */
1124 new_thread_count = (gint)(hc->current_control_setting + new_thread_wave_magnitude * ((hc->total_samples / (hc->wave_period / 2)) % 2));
1126 /* Make sure the new thread count doesn't exceed the ThreadPoolWorker's limits. */
1127 new_thread_count = CLAMP (new_thread_count, worker.limit_worker_min, worker.limit_worker_max);
1129 if (new_thread_count != current_thread_count)
1130 hill_climbing_change_thread_count (new_thread_count, transition);
1132 if (creal (ratio) < 0.0 && new_thread_count == worker.limit_worker_min)
1133 *adjustment_interval = (gint)(0.5 + hc->current_sample_interval * (10.0 * MAX (-1.0 * creal (ratio), 1.0)));
1135 *adjustment_interval = hc->current_sample_interval;
1137 return new_thread_count;
1141 heuristic_should_adjust (void)
1143 if (worker.heuristic_last_dequeue > worker.heuristic_last_adjustment + worker.heuristic_adjustment_interval) {
1144 ThreadPoolWorkerCounter counter;
1145 counter = COUNTER_READ ();
1146 if (counter._.working <= counter._.max_working)
1154 heuristic_adjust (void)
1156 if (mono_coop_mutex_trylock (&worker.heuristic_lock) == 0) {
1157 gint32 completions = InterlockedExchange (&worker.heuristic_completions, 0);
1158 gint64 sample_end = mono_msec_ticks ();
1159 gint64 sample_duration = sample_end - worker.heuristic_sample_start;
1161 if (sample_duration >= worker.heuristic_adjustment_interval / 2) {
1162 ThreadPoolWorkerCounter counter;
1163 gint16 new_thread_count;
1165 counter = COUNTER_READ ();
1166 new_thread_count = hill_climbing_update (counter._.max_working, sample_duration, completions, &worker.heuristic_adjustment_interval);
1168 COUNTER_ATOMIC (counter, {
1169 counter._.max_working = new_thread_count;
1172 if (new_thread_count > counter._.max_working)
1175 worker.heuristic_sample_start = sample_end;
1176 worker.heuristic_last_adjustment = mono_msec_ticks ();
1179 mono_coop_mutex_unlock (&worker.heuristic_lock);
1184 heuristic_notify_work_completed (void)
1186 InterlockedIncrement (&worker.heuristic_completions);
1187 worker.heuristic_last_dequeue = mono_msec_ticks ();
1189 if (heuristic_should_adjust ())
1190 heuristic_adjust ();
1194 mono_threadpool_worker_notify_completed (void)
1196 ThreadPoolWorkerCounter counter;
1198 heuristic_notify_work_completed ();
1200 counter = COUNTER_READ ();
1201 return counter._.working <= counter._.max_working;
1205 mono_threadpool_worker_get_min (void)
1207 return worker.limit_worker_min;
1211 mono_threadpool_worker_set_min (gint32 value)
1213 if (value <= 0 || value > worker.limit_worker_max)
1216 worker.limit_worker_min = value;
1221 mono_threadpool_worker_get_max (void)
1223 return worker.limit_worker_max;
1227 mono_threadpool_worker_set_max (gint32 value)
1231 cpu_count = mono_cpu_count ();
1232 if (value < worker.limit_worker_min || value < cpu_count)
1235 if (value < worker.limit_worker_min || value < cpu_count)
1238 worker.limit_worker_max = value;
1243 mono_threadpool_worker_set_suspended (gboolean suspended)
1245 worker.suspended = suspended;