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-lazy-init.h>
29 #include <mono/utils/mono-logger.h>
30 #include <mono/utils/mono-logger-internals.h>
31 #include <mono/utils/mono-proclib.h>
32 #include <mono/utils/mono-threads.h>
33 #include <mono/utils/mono-time.h>
34 #include <mono/utils/mono-rand.h>
35 #include <mono/utils/refcount.h>
36 #include <mono/io-layer/io-layer.h>
38 #define CPU_USAGE_LOW 80
39 #define CPU_USAGE_HIGH 95
41 #define MONITOR_INTERVAL 500 // ms
42 #define MONITOR_MINIMAL_LIFETIME 60 * 1000 // ms
44 #define WORKER_CREATION_MAX_PER_SEC 10
46 /* The exponent to apply to the gain. 1.0 means to use linear gain,
47 * higher values will enhance large moves and damp small ones.
49 #define HILL_CLIMBING_GAIN_EXPONENT 2.0
51 /* The 'cost' of a thread. 0 means drive for increased throughput regardless
52 * of thread count, higher values bias more against higher thread counts.
54 #define HILL_CLIMBING_BIAS 0.15
56 #define HILL_CLIMBING_WAVE_PERIOD 4
57 #define HILL_CLIMBING_MAX_WAVE_MAGNITUDE 20
58 #define HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER 1.0
59 #define HILL_CLIMBING_WAVE_HISTORY_SIZE 8
60 #define HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO 3.0
61 #define HILL_CLIMBING_MAX_CHANGE_PER_SECOND 4
62 #define HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE 20
63 #define HILL_CLIMBING_SAMPLE_INTERVAL_LOW 10
64 #define HILL_CLIMBING_SAMPLE_INTERVAL_HIGH 200
65 #define HILL_CLIMBING_ERROR_SMOOTHING_FACTOR 0.01
66 #define HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT 0.15
70 TRANSITION_INITIALIZING,
71 TRANSITION_RANDOM_MOVE,
72 TRANSITION_CLIMBING_MOVE,
73 TRANSITION_CHANGE_POINT,
74 TRANSITION_STABILIZING,
75 TRANSITION_STARVATION,
76 TRANSITION_THREAD_TIMED_OUT,
78 } ThreadPoolHeuristicStateTransition;
82 gint32 samples_to_measure;
83 gdouble target_throughput_ratio;
84 gdouble target_signal_to_noise_ratio;
85 gdouble max_change_per_second;
86 gdouble max_change_per_sample;
87 gint32 max_thread_wave_magnitude;
88 gint32 sample_interval_low;
89 gdouble thread_magnitude_multiplier;
90 gint32 sample_interval_high;
91 gdouble throughput_error_smoothing_factor;
92 gdouble gain_exponent;
93 gdouble max_sample_error;
95 gdouble current_control_setting;
97 gint16 last_thread_count;
98 gdouble elapsed_since_last_change;
99 gdouble completions_since_last_change;
101 gdouble average_throughput_noise;
104 gdouble *thread_counts;
106 guint32 current_sample_interval;
107 gpointer random_interval_generator;
109 gint32 accumulated_completion_count;
110 gdouble accumulated_sample_duration;
111 } ThreadPoolHillClimbing;
114 MonoThreadPoolWorkerCallback callback;
116 } ThreadPoolWorkItem;
120 gint16 max_working; /* determined by heuristic */
121 gint16 starting; /* starting, but not yet in worker_thread */
122 gint16 working; /* executing worker_thread */
123 gint16 parked; /* parked */
126 } ThreadPoolWorkerCounter;
128 typedef MonoInternalThread ThreadPoolWorkerThread;
130 struct MonoThreadPoolWorker {
133 ThreadPoolWorkerCounter counters;
135 GPtrArray *threads; // ThreadPoolWorkerThread* []
136 MonoCoopMutex threads_lock; /* protect access to working_threads and parked_threads */
137 gint32 parked_threads_count;
138 MonoCoopCond parked_threads_cond;
139 MonoCoopCond threads_exit_cond;
141 ThreadPoolWorkItem *work_items; // ThreadPoolWorkItem []
142 gint32 work_items_count;
143 gint32 work_items_size;
144 MonoCoopMutex work_items_lock;
146 guint32 worker_creation_current_second;
147 guint32 worker_creation_current_count;
148 MonoCoopMutex worker_creation_lock;
150 gint32 heuristic_completions;
151 gint64 heuristic_sample_start;
152 gint64 heuristic_last_dequeue; // ms
153 gint64 heuristic_last_adjustment; // ms
154 gint64 heuristic_adjustment_interval; // ms
155 ThreadPoolHillClimbing heuristic_hill_climbing;
156 MonoCoopMutex heuristic_lock;
158 gint32 limit_worker_min;
159 gint32 limit_worker_max;
161 MonoCpuUsageState *cpu_usage_state;
164 /* suspended by the debugger */
167 gint32 monitor_status;
171 MONITOR_STATUS_REQUESTED,
172 MONITOR_STATUS_WAITING_FOR_REQUEST,
173 MONITOR_STATUS_NOT_RUNNING,
176 #define COUNTER_CHECK(counter) \
178 g_assert (counter._.max_working > 0); \
179 g_assert (counter._.starting >= 0); \
180 g_assert (counter._.working >= 0); \
183 #define COUNTER_ATOMIC(worker,var,block) \
185 ThreadPoolWorkerCounter __old; \
188 __old = COUNTER_READ (worker); \
191 COUNTER_CHECK (var); \
192 } while (InterlockedCompareExchange64 (&worker->counters.as_gint64, (var).as_gint64, __old.as_gint64) != __old.as_gint64); \
195 static inline ThreadPoolWorkerCounter
196 COUNTER_READ (MonoThreadPoolWorker *worker)
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)
224 MonoThreadPoolWorker *worker;
226 worker = (MonoThreadPoolWorker*) data;
229 // FIXME destroy everything
235 mono_threadpool_worker_init (MonoThreadPoolWorker **worker)
237 MonoThreadPoolWorker *wk;
238 ThreadPoolHillClimbing *hc;
239 const char *threads_per_cpu_env;
240 gint threads_per_cpu;
245 wk = *worker = g_new0 (MonoThreadPoolWorker, 1);
247 mono_refcount_init (wk, destroy);
249 wk->threads = g_ptr_array_new ();
250 mono_coop_mutex_init (&wk->threads_lock);
251 wk->parked_threads_count = 0;
252 mono_coop_cond_init (&wk->parked_threads_cond);
253 mono_coop_cond_init (&wk->threads_exit_cond);
255 /* wk->work_items_size is inited to 0 */
256 mono_coop_mutex_init (&wk->work_items_lock);
258 wk->worker_creation_current_second = -1;
259 mono_coop_mutex_init (&wk->worker_creation_lock);
261 wk->heuristic_adjustment_interval = 10;
262 mono_coop_mutex_init (&wk->heuristic_lock);
266 hc = &wk->heuristic_hill_climbing;
268 hc->wave_period = HILL_CLIMBING_WAVE_PERIOD;
269 hc->max_thread_wave_magnitude = HILL_CLIMBING_MAX_WAVE_MAGNITUDE;
270 hc->thread_magnitude_multiplier = (gdouble) HILL_CLIMBING_WAVE_MAGNITUDE_MULTIPLIER;
271 hc->samples_to_measure = hc->wave_period * HILL_CLIMBING_WAVE_HISTORY_SIZE;
272 hc->target_throughput_ratio = (gdouble) HILL_CLIMBING_BIAS;
273 hc->target_signal_to_noise_ratio = (gdouble) HILL_CLIMBING_TARGET_SIGNAL_TO_NOISE_RATIO;
274 hc->max_change_per_second = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SECOND;
275 hc->max_change_per_sample = (gdouble) HILL_CLIMBING_MAX_CHANGE_PER_SAMPLE;
276 hc->sample_interval_low = HILL_CLIMBING_SAMPLE_INTERVAL_LOW;
277 hc->sample_interval_high = HILL_CLIMBING_SAMPLE_INTERVAL_HIGH;
278 hc->throughput_error_smoothing_factor = (gdouble) HILL_CLIMBING_ERROR_SMOOTHING_FACTOR;
279 hc->gain_exponent = (gdouble) HILL_CLIMBING_GAIN_EXPONENT;
280 hc->max_sample_error = (gdouble) HILL_CLIMBING_MAX_SAMPLE_ERROR_PERCENT;
281 hc->current_control_setting = 0;
282 hc->total_samples = 0;
283 hc->last_thread_count = 0;
284 hc->average_throughput_noise = 0;
285 hc->elapsed_since_last_change = 0;
286 hc->accumulated_completion_count = 0;
287 hc->accumulated_sample_duration = 0;
288 hc->samples = g_new0 (gdouble, hc->samples_to_measure);
289 hc->thread_counts = g_new0 (gdouble, hc->samples_to_measure);
290 hc->random_interval_generator = rand_create ();
291 hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
293 if (!(threads_per_cpu_env = g_getenv ("MONO_THREADS_PER_CPU")))
296 threads_per_cpu = CLAMP (atoi (threads_per_cpu_env), 1, 50);
298 threads_count = mono_cpu_count () * threads_per_cpu;
300 wk->limit_worker_min = threads_count;
302 #if defined (PLATFORM_ANDROID) || defined (HOST_IOS)
303 wk->limit_worker_max = CLAMP (threads_count * 100, MIN (threads_count, 200), MAX (threads_count, 200));
305 wk->limit_worker_max = threads_count * 100;
308 wk->counters._.max_working = wk->limit_worker_min;
310 wk->cpu_usage_state = g_new0 (MonoCpuUsageState, 1);
312 wk->suspended = FALSE;
314 wk->monitor_status = MONITOR_STATUS_NOT_RUNNING;
318 mono_threadpool_worker_cleanup (MonoThreadPoolWorker *worker)
320 MonoInternalThread *current;
322 /* we make the assumption along the code that we are
323 * cleaning up only if the runtime is shutting down */
324 g_assert (mono_runtime_is_shutting_down ());
326 current = mono_thread_internal_current ();
328 while (worker->monitor_status != MONITOR_STATUS_NOT_RUNNING)
329 mono_thread_info_sleep (1, NULL);
331 mono_coop_mutex_lock (&worker->threads_lock);
333 /* unpark all worker->parked_threads */
334 mono_coop_cond_broadcast (&worker->parked_threads_cond);
337 ThreadPoolWorkerCounter counter;
339 counter = COUNTER_READ (worker);
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);
353 mono_coop_mutex_unlock (&worker->threads_lock);
355 mono_refcount_dec (worker);
359 work_item_lock (MonoThreadPoolWorker *worker)
361 mono_coop_mutex_lock (&worker->work_items_lock);
365 work_item_unlock (MonoThreadPoolWorker *worker)
367 mono_coop_mutex_unlock (&worker->work_items_lock);
371 work_item_push (MonoThreadPoolWorker *worker, MonoThreadPoolWorkerCallback callback, gpointer data)
373 ThreadPoolWorkItem work_item;
378 work_item.callback = callback;
379 work_item.data = data;
381 work_item_lock (worker);
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);
397 work_item_unlock (worker);
401 work_item_try_pop (MonoThreadPoolWorker *worker, ThreadPoolWorkItem *work_item)
404 g_assert (work_item);
406 work_item_lock (worker);
408 // printf ("[pop] worker->work_items = %p, worker->work_items_count = %d, worker->work_items_size = %d\n",
409 // worker->work_items, worker->work_items_count, worker->work_items_size);
411 if (worker->work_items_count == 0) {
412 work_item_unlock (worker);
416 *work_item = worker->work_items [-- worker->work_items_count];
418 if (G_UNLIKELY (worker->work_items_count >= 64 * 3 && worker->work_items_count < worker->work_items_size / 2)) {
419 worker->work_items_size -= 64;
420 worker->work_items = g_renew (ThreadPoolWorkItem, worker->work_items, worker->work_items_size);
423 work_item_unlock (worker);
429 work_item_count (MonoThreadPoolWorker *worker)
433 work_item_lock (worker);
434 count = worker->work_items_count;
435 work_item_unlock (worker);
440 static void worker_request (MonoThreadPoolWorker *worker);
443 mono_threadpool_worker_enqueue (MonoThreadPoolWorker *worker, MonoThreadPoolWorkerCallback callback, gpointer data)
445 work_item_push (worker, callback, data);
447 worker_request (worker);
451 worker_wait_interrupt (gpointer data)
453 MonoThreadPoolWorker *worker;
455 worker = (MonoThreadPoolWorker*) data;
458 mono_coop_mutex_lock (&worker->threads_lock);
459 mono_coop_cond_signal (&worker->parked_threads_cond);
460 mono_coop_mutex_unlock (&worker->threads_lock);
462 mono_refcount_dec (worker);
465 /* return TRUE if timeout, FALSE otherwise (worker unpark or interrupt) */
467 worker_park (MonoThreadPoolWorker *worker)
469 gboolean timeout = FALSE;
471 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker parking", mono_native_thread_id_get ());
473 mono_coop_mutex_lock (&worker->threads_lock);
475 if (!mono_runtime_is_shutting_down ()) {
476 static gpointer rand_handle = NULL;
477 MonoInternalThread *thread;
478 gboolean interrupted = FALSE;
479 ThreadPoolWorkerCounter counter;
482 rand_handle = rand_create ();
483 g_assert (rand_handle);
485 thread = mono_thread_internal_current ();
488 COUNTER_ATOMIC (worker, counter, {
489 counter._.working --;
493 worker->parked_threads_count += 1;
495 mono_thread_info_install_interrupt (worker_wait_interrupt, mono_refcount_inc (worker), &interrupted);
497 mono_refcount_dec (worker);
501 if (mono_coop_cond_timedwait (&worker->parked_threads_cond, &worker->threads_lock, rand_next (&rand_handle, 5 * 1000, 60 * 1000)) != 0)
504 mono_thread_info_uninstall_interrupt (&interrupted);
506 mono_refcount_dec (worker);
509 worker->parked_threads_count -= 1;
511 COUNTER_ATOMIC (worker, counter, {
512 counter._.working ++;
517 mono_coop_mutex_unlock (&worker->threads_lock);
519 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker unparking, timeout? %s", mono_native_thread_id_get (), timeout ? "yes" : "no");
525 worker_try_unpark (MonoThreadPoolWorker *worker)
527 gboolean res = FALSE;
529 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker", mono_native_thread_id_get ());
531 mono_coop_mutex_lock (&worker->threads_lock);
532 if (worker->parked_threads_count > 0) {
533 mono_coop_cond_signal (&worker->parked_threads_cond);
536 mono_coop_mutex_unlock (&worker->threads_lock);
538 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try unpark worker, success? %s", mono_native_thread_id_get (), res ? "yes" : "no");
544 worker_thread (gpointer data)
546 MonoThreadPoolWorker *worker;
548 MonoInternalThread *thread;
549 ThreadPoolWorkerCounter counter;
551 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker starting", mono_native_thread_id_get ());
553 worker = (MonoThreadPoolWorker*) data;
556 COUNTER_ATOMIC (worker, counter, {
557 counter._.starting --;
558 counter._.working ++;
561 thread = mono_thread_internal_current ();
564 mono_coop_mutex_lock (&worker->threads_lock);
565 g_ptr_array_add (worker->threads, thread);
566 mono_coop_mutex_unlock (&worker->threads_lock);
568 mono_thread_set_name_internal (thread, mono_string_new (mono_get_root_domain (), "Threadpool worker"), FALSE, &error);
569 mono_error_assert_ok (&error);
571 while (!mono_runtime_is_shutting_down ()) {
572 ThreadPoolWorkItem work_item;
574 if (mono_thread_interruption_checkpoint ())
577 if (!work_item_try_pop (worker, &work_item)) {
580 timeout = worker_park (worker);
587 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] worker executing %p (%p)",
588 mono_native_thread_id_get (), work_item.callback, work_item.data);
590 work_item.callback (work_item.data);
593 mono_coop_mutex_lock (&worker->threads_lock);
595 COUNTER_ATOMIC (worker, counter, {
596 counter._.working --;
599 g_ptr_array_remove (worker->threads, thread);
601 mono_coop_cond_signal (&worker->threads_exit_cond);
603 mono_coop_mutex_unlock (&worker->threads_lock);
605 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_THREADPOOL, "[%p] worker finishing", mono_native_thread_id_get ());
607 mono_refcount_dec (worker);
613 worker_try_create (MonoThreadPoolWorker *worker)
616 MonoInternalThread *thread;
617 gint64 current_ticks;
619 ThreadPoolWorkerCounter counter;
621 if (mono_runtime_is_shutting_down ())
624 mono_coop_mutex_lock (&worker->worker_creation_lock);
626 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker", mono_native_thread_id_get ());
628 current_ticks = mono_100ns_ticks ();
629 if (0 == current_ticks) {
630 g_warning ("failed to get 100ns ticks");
632 now = current_ticks / (10 * 1000 * 1000);
633 if (worker->worker_creation_current_second != now) {
634 worker->worker_creation_current_second = now;
635 worker->worker_creation_current_count = 0;
637 g_assert (worker->worker_creation_current_count <= WORKER_CREATION_MAX_PER_SEC);
638 if (worker->worker_creation_current_count == WORKER_CREATION_MAX_PER_SEC) {
639 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",
640 mono_native_thread_id_get (), worker->worker_creation_current_count);
641 mono_coop_mutex_unlock (&worker->worker_creation_lock);
647 COUNTER_ATOMIC (worker, counter, {
648 if (counter._.working >= counter._.max_working) {
649 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, failed: maximum number of working threads reached",
650 mono_native_thread_id_get ());
651 mono_coop_mutex_unlock (&worker->worker_creation_lock);
654 counter._.starting ++;
657 thread = mono_thread_create_internal (mono_get_root_domain (), worker_thread, mono_refcount_inc (worker), TRUE, 0, &error);
659 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));
660 mono_error_cleanup (&error);
662 COUNTER_ATOMIC (worker, counter, {
663 counter._.starting --;
666 mono_coop_mutex_unlock (&worker->worker_creation_lock);
668 mono_refcount_dec (worker);
673 worker->worker_creation_current_count += 1;
675 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] try create worker, created %p, now = %d count = %d",
676 mono_native_thread_id_get (), (gpointer) thread->tid, now, worker->worker_creation_current_count);
678 mono_coop_mutex_unlock (&worker->worker_creation_lock);
682 static void monitor_ensure_running (MonoThreadPoolWorker *worker);
685 worker_request (MonoThreadPoolWorker *worker)
689 if (worker->suspended)
692 monitor_ensure_running (worker);
694 if (worker_try_unpark (worker)) {
695 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, unparked", mono_native_thread_id_get ());
699 if (worker_try_create (worker)) {
700 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, created", mono_native_thread_id_get ());
704 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] request worker, failed", mono_native_thread_id_get ());
708 monitor_should_keep_running (MonoThreadPoolWorker *worker)
710 static gint64 last_should_keep_running = -1;
712 g_assert (worker->monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || worker->monitor_status == MONITOR_STATUS_REQUESTED);
714 if (InterlockedExchange (&worker->monitor_status, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST) {
715 gboolean should_keep_running = TRUE, force_should_keep_running = FALSE;
717 if (mono_runtime_is_shutting_down ()) {
718 should_keep_running = FALSE;
720 if (work_item_count (worker) == 0)
721 should_keep_running = FALSE;
723 if (!should_keep_running) {
724 if (last_should_keep_running == -1 || mono_100ns_ticks () - last_should_keep_running < MONITOR_MINIMAL_LIFETIME * 1000 * 10) {
725 should_keep_running = force_should_keep_running = TRUE;
730 if (should_keep_running) {
731 if (last_should_keep_running == -1 || !force_should_keep_running)
732 last_should_keep_running = mono_100ns_ticks ();
734 last_should_keep_running = -1;
735 if (InterlockedCompareExchange (&worker->monitor_status, MONITOR_STATUS_NOT_RUNNING, MONITOR_STATUS_WAITING_FOR_REQUEST) == MONITOR_STATUS_WAITING_FOR_REQUEST)
740 g_assert (worker->monitor_status == MONITOR_STATUS_WAITING_FOR_REQUEST || worker->monitor_status == MONITOR_STATUS_REQUESTED);
746 monitor_sufficient_delay_since_last_dequeue (MonoThreadPoolWorker *worker)
752 if (worker->cpu_usage < CPU_USAGE_LOW) {
753 threshold = MONITOR_INTERVAL;
755 ThreadPoolWorkerCounter counter;
756 counter = COUNTER_READ (worker);
757 threshold = counter._.max_working * MONITOR_INTERVAL * 2;
760 return mono_msec_ticks () >= worker->heuristic_last_dequeue + threshold;
763 static void hill_climbing_force_change (MonoThreadPoolWorker *worker, gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition);
766 monitor_thread (gpointer data)
768 MonoThreadPoolWorker *worker;
769 MonoInternalThread *internal;
772 worker = (MonoThreadPoolWorker*) data;
775 internal = mono_thread_internal_current ();
778 mono_cpu_usage (worker->cpu_usage_state);
780 // printf ("monitor_thread: start\n");
782 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, started", mono_native_thread_id_get ());
785 ThreadPoolWorkerCounter counter;
786 gboolean limit_worker_max_reached;
787 gint32 interval_left = MONITOR_INTERVAL;
788 gint32 awake = 0; /* number of spurious awakes we tolerate before doing a round of rebalancing */
790 g_assert (worker->monitor_status != MONITOR_STATUS_NOT_RUNNING);
792 // counter = COUNTER_READ (worker);
793 // printf ("monitor_thread: starting = %d working = %d parked = %d max_working = %d\n",
794 // counter._.starting, counter._.working, counter._.parked, counter._.max_working);
798 gboolean alerted = FALSE;
800 if (mono_runtime_is_shutting_down ())
803 ts = mono_msec_ticks ();
804 if (mono_thread_info_sleep (interval_left, &alerted) == 0)
806 interval_left -= mono_msec_ticks () - ts;
808 g_assert (!(internal->state & ThreadState_StopRequested));
809 mono_thread_interruption_checkpoint ();
810 } while (interval_left > 0 && ++awake < 10);
812 if (mono_runtime_is_shutting_down ())
815 if (worker->suspended)
818 if (work_item_count (worker) == 0)
821 worker->cpu_usage = mono_cpu_usage (worker->cpu_usage_state);
823 if (!monitor_sufficient_delay_since_last_dequeue (worker))
826 limit_worker_max_reached = FALSE;
828 COUNTER_ATOMIC (worker, counter, {
829 if (counter._.max_working >= worker->limit_worker_max) {
830 limit_worker_max_reached = TRUE;
833 counter._.max_working ++;
836 if (limit_worker_max_reached)
839 hill_climbing_force_change (worker, counter._.max_working, TRANSITION_STARVATION);
841 for (i = 0; i < 5; ++i) {
842 if (mono_runtime_is_shutting_down ())
845 if (worker_try_unpark (worker)) {
846 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, unparked", mono_native_thread_id_get ());
850 if (worker_try_create (worker)) {
851 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, created", mono_native_thread_id_get ());
855 } while (monitor_should_keep_running (worker));
857 // printf ("monitor_thread: stop\n");
859 mono_trace (G_LOG_LEVEL_DEBUG, MONO_TRACE_THREADPOOL, "[%p] monitor thread, finished", mono_native_thread_id_get ());
865 monitor_ensure_running (MonoThreadPoolWorker *worker)
869 switch (worker->monitor_status) {
870 case MONITOR_STATUS_REQUESTED:
871 // printf ("monitor_thread: requested\n");
873 case MONITOR_STATUS_WAITING_FOR_REQUEST:
874 // printf ("monitor_thread: waiting for request\n");
875 InterlockedCompareExchange (&worker->monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_WAITING_FOR_REQUEST);
877 case MONITOR_STATUS_NOT_RUNNING:
878 // printf ("monitor_thread: not running\n");
879 if (mono_runtime_is_shutting_down ())
881 if (InterlockedCompareExchange (&worker->monitor_status, MONITOR_STATUS_REQUESTED, MONITOR_STATUS_NOT_RUNNING) == MONITOR_STATUS_NOT_RUNNING) {
882 // printf ("monitor_thread: creating\n");
883 if (!mono_thread_create_internal (mono_get_root_domain (), monitor_thread, worker, TRUE, SMALL_STACK, &error)) {
884 // printf ("monitor_thread: creating failed\n");
885 worker->monitor_status = MONITOR_STATUS_NOT_RUNNING;
886 mono_error_cleanup (&error);
891 default: g_assert_not_reached ();
897 hill_climbing_change_thread_count (MonoThreadPoolWorker *worker, gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
899 ThreadPoolHillClimbing *hc;
903 hc = &worker->heuristic_hill_climbing;
905 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);
907 hc->last_thread_count = new_thread_count;
908 hc->current_sample_interval = rand_next (&hc->random_interval_generator, hc->sample_interval_low, hc->sample_interval_high);
909 hc->elapsed_since_last_change = 0;
910 hc->completions_since_last_change = 0;
914 hill_climbing_force_change (MonoThreadPoolWorker *worker, gint16 new_thread_count, ThreadPoolHeuristicStateTransition transition)
916 ThreadPoolHillClimbing *hc;
920 hc = &worker->heuristic_hill_climbing;
922 if (new_thread_count != hc->last_thread_count) {
923 hc->current_control_setting += new_thread_count - hc->last_thread_count;
924 hill_climbing_change_thread_count (worker, new_thread_count, transition);
928 static double_complex
929 hill_climbing_get_wave_component (MonoThreadPoolWorker *worker, gdouble *samples, guint sample_count, gdouble period)
931 ThreadPoolHillClimbing *hc;
932 gdouble w, cosine, sine, coeff, q0, q1, q2;
936 g_assert (sample_count >= period);
937 g_assert (period >= 2);
939 hc = &worker->heuristic_hill_climbing;
941 w = 2.0 * M_PI / period;
944 coeff = 2.0 * cosine;
947 for (i = 0; i < sample_count; ++i) {
948 q0 = coeff * q1 - q2 + samples [(hc->total_samples - sample_count + i) % hc->samples_to_measure];
953 return mono_double_complex_scalar_div (mono_double_complex_make (q1 - q2 * cosine, (q2 * sine)), ((gdouble)sample_count));
957 hill_climbing_update (MonoThreadPoolWorker *worker, gint16 current_thread_count, guint32 sample_duration, gint32 completions, gint64 *adjustment_interval)
959 ThreadPoolHillClimbing *hc;
960 ThreadPoolHeuristicStateTransition transition;
962 gdouble throughput_error_estimate;
968 gint new_thread_wave_magnitude;
969 gint new_thread_count;
970 double_complex thread_wave_component;
971 double_complex throughput_wave_component;
972 double_complex ratio;
975 g_assert (adjustment_interval);
977 hc = &worker->heuristic_hill_climbing;
979 /* If someone changed the thread count without telling us, update our records accordingly. */
980 if (current_thread_count != hc->last_thread_count)
981 hill_climbing_force_change (worker, current_thread_count, TRANSITION_INITIALIZING);
983 /* Update the cumulative stats for this thread count */
984 hc->elapsed_since_last_change += sample_duration;
985 hc->completions_since_last_change += completions;
987 /* Add in any data we've already collected about this sample */
988 sample_duration += hc->accumulated_sample_duration;
989 completions += hc->accumulated_completion_count;
991 /* We need to make sure we're collecting reasonably accurate data. Since we're just counting the end
992 * of each work item, we are goinng to be missing some data about what really happened during the
993 * sample interval. The count produced by each thread includes an initial work item that may have
994 * started well before the start of the interval, and each thread may have been running some new
995 * work item for some time before the end of the interval, which did not yet get counted. So
996 * our count is going to be off by +/- threadCount workitems.
998 * The exception is that the thread that reported to us last time definitely wasn't running any work
999 * at that time, and the thread that's reporting now definitely isn't running a work item now. So
1000 * we really only need to consider threadCount-1 threads.
1002 * Thus the percent error in our count is +/- (threadCount-1)/numCompletions.
1004 * We cannot rely on the frequency-domain analysis we'll be doing later to filter out this error, because
1005 * of the way it accumulates over time. If this sample is off by, say, 33% in the negative direction,
1006 * then the next one likely will be too. The one after that will include the sum of the completions
1007 * we missed in the previous samples, and so will be 33% positive. So every three samples we'll have
1008 * two "low" samples and one "high" sample. This will appear as periodic variation right in the frequency
1009 * range we're targeting, which will not be filtered by the frequency-domain translation. */
1010 if (hc->total_samples > 0 && ((current_thread_count - 1.0) / completions) >= hc->max_sample_error) {
1011 /* Not accurate enough yet. Let's accumulate the data so
1012 * far, and tell the MonoThreadPoolWorker to collect a little more. */
1013 hc->accumulated_sample_duration = sample_duration;
1014 hc->accumulated_completion_count = completions;
1015 *adjustment_interval = 10;
1016 return current_thread_count;
1019 /* We've got enouugh data for our sample; reset our accumulators for next time. */
1020 hc->accumulated_sample_duration = 0;
1021 hc->accumulated_completion_count = 0;
1023 /* Add the current thread count and throughput sample to our history. */
1024 throughput = ((gdouble) completions) / sample_duration;
1026 sample_index = hc->total_samples % hc->samples_to_measure;
1027 hc->samples [sample_index] = throughput;
1028 hc->thread_counts [sample_index] = current_thread_count;
1029 hc->total_samples ++;
1031 /* Set up defaults for our metrics. */
1032 thread_wave_component = mono_double_complex_make(0, 0);
1033 throughput_wave_component = mono_double_complex_make(0, 0);
1034 throughput_error_estimate = 0;
1035 ratio = mono_double_complex_make(0, 0);
1038 transition = TRANSITION_WARMUP;
1040 /* How many samples will we use? It must be at least the three wave periods we're looking for, and it must also
1041 * be a whole multiple of the primary wave's period; otherwise the frequency we're looking for will fall between
1042 * two frequency bands in the Fourier analysis, and we won't be able to measure it accurately. */
1043 sample_count = ((gint) MIN (hc->total_samples - 1, hc->samples_to_measure) / hc->wave_period) * hc->wave_period;
1045 if (sample_count > hc->wave_period) {
1047 gdouble average_throughput;
1048 gdouble average_thread_count;
1049 gdouble sample_sum = 0;
1050 gdouble thread_sum = 0;
1052 /* Average the throughput and thread count samples, so we can scale the wave magnitudes later. */
1053 for (i = 0; i < sample_count; ++i) {
1054 guint j = (hc->total_samples - sample_count + i) % hc->samples_to_measure;
1055 sample_sum += hc->samples [j];
1056 thread_sum += hc->thread_counts [j];
1059 average_throughput = sample_sum / sample_count;
1060 average_thread_count = thread_sum / sample_count;
1062 if (average_throughput > 0 && average_thread_count > 0) {
1063 gdouble noise_for_confidence, adjacent_period_1, adjacent_period_2;
1065 /* Calculate the periods of the adjacent frequency bands we'll be using to
1066 * measure noise levels. We want the two adjacent Fourier frequency bands. */
1067 adjacent_period_1 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) + 1);
1068 adjacent_period_2 = sample_count / (((gdouble) sample_count) / ((gdouble) hc->wave_period) - 1);
1070 /* Get the the three different frequency components of the throughput (scaled by average
1071 * throughput). Our "error" estimate (the amount of noise that might be present in the
1072 * frequency band we're really interested in) is the average of the adjacent bands. */
1073 throughput_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (worker, hc->samples, sample_count, hc->wave_period), average_throughput);
1074 throughput_error_estimate = cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (worker, hc->samples, sample_count, adjacent_period_1), average_throughput));
1076 if (adjacent_period_2 <= sample_count) {
1077 throughput_error_estimate = MAX (throughput_error_estimate, cabs (mono_double_complex_scalar_div (hill_climbing_get_wave_component (
1078 worker, hc->samples, sample_count, adjacent_period_2), average_throughput)));
1081 /* Do the same for the thread counts, so we have something to compare to. We don't
1082 * measure thread count noise, because there is none; these are exact measurements. */
1083 thread_wave_component = mono_double_complex_scalar_div (hill_climbing_get_wave_component (worker, hc->thread_counts, sample_count, hc->wave_period), average_thread_count);
1085 /* Update our moving average of the throughput noise. We'll use this
1086 * later as feedback to determine the new size of the thread wave. */
1087 if (hc->average_throughput_noise == 0) {
1088 hc->average_throughput_noise = throughput_error_estimate;
1090 hc->average_throughput_noise = (hc->throughput_error_smoothing_factor * throughput_error_estimate)
1091 + ((1.0 + hc->throughput_error_smoothing_factor) * hc->average_throughput_noise);
1094 if (cabs (thread_wave_component) > 0) {
1095 /* Adjust the throughput wave so it's centered around the target wave,
1096 * and then calculate the adjusted throughput/thread ratio. */
1097 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);
1098 transition = TRANSITION_CLIMBING_MOVE;
1100 ratio = mono_double_complex_make (0, 0);
1101 transition = TRANSITION_STABILIZING;
1104 noise_for_confidence = MAX (hc->average_throughput_noise, throughput_error_estimate);
1105 if (noise_for_confidence > 0) {
1106 confidence = cabs (thread_wave_component) / noise_for_confidence / hc->target_signal_to_noise_ratio;
1108 /* there is no noise! */
1114 /* We use just the real part of the complex ratio we just calculated. If the throughput signal
1115 * is exactly in phase with the thread signal, this will be the same as taking the magnitude of
1116 * the complex move and moving that far up. If they're 180 degrees out of phase, we'll move
1117 * backward (because this indicates that our changes are having the opposite of the intended effect).
1118 * If they're 90 degrees out of phase, we won't move at all, because we can't tell wether we're
1119 * having a negative or positive effect on throughput. */
1120 move = creal (ratio);
1121 move = CLAMP (move, -1.0, 1.0);
1123 /* Apply our confidence multiplier. */
1124 move *= CLAMP (confidence, -1.0, 1.0);
1126 /* Now apply non-linear gain, such that values around zero are attenuated, while higher values
1127 * are enhanced. This allows us to move quickly if we're far away from the target, but more slowly
1128 * if we're getting close, giving us rapid ramp-up without wild oscillations around the target. */
1129 gain = hc->max_change_per_second * sample_duration;
1130 move = pow (fabs (move), hc->gain_exponent) * (move >= 0.0 ? 1 : -1) * gain;
1131 move = MIN (move, hc->max_change_per_sample);
1133 /* If the result was positive, and CPU is > 95%, refuse the move. */
1134 if (move > 0.0 && worker->cpu_usage > CPU_USAGE_HIGH)
1137 /* Apply the move to our control setting. */
1138 hc->current_control_setting += move;
1140 /* Calculate the new thread wave magnitude, which is based on the moving average we've been keeping of the
1141 * throughput error. This average starts at zero, so we'll start with a nice safe little wave at first. */
1142 new_thread_wave_magnitude = (gint)(0.5 + (hc->current_control_setting * hc->average_throughput_noise
1143 * hc->target_signal_to_noise_ratio * hc->thread_magnitude_multiplier * 2.0));
1144 new_thread_wave_magnitude = CLAMP (new_thread_wave_magnitude, 1, hc->max_thread_wave_magnitude);
1146 /* Make sure our control setting is within the MonoThreadPoolWorker's limits. */
1147 hc->current_control_setting = CLAMP (hc->current_control_setting, worker->limit_worker_min, worker->limit_worker_max - new_thread_wave_magnitude);
1149 /* Calculate the new thread count (control setting + square wave). */
1150 new_thread_count = (gint)(hc->current_control_setting + new_thread_wave_magnitude * ((hc->total_samples / (hc->wave_period / 2)) % 2));
1152 /* Make sure the new thread count doesn't exceed the MonoThreadPoolWorker's limits. */
1153 new_thread_count = CLAMP (new_thread_count, worker->limit_worker_min, worker->limit_worker_max);
1155 if (new_thread_count != current_thread_count)
1156 hill_climbing_change_thread_count (worker, new_thread_count, transition);
1158 if (creal (ratio) < 0.0 && new_thread_count == worker->limit_worker_min)
1159 *adjustment_interval = (gint)(0.5 + hc->current_sample_interval * (10.0 * MAX (-1.0 * creal (ratio), 1.0)));
1161 *adjustment_interval = hc->current_sample_interval;
1163 return new_thread_count;
1167 heuristic_should_adjust (MonoThreadPoolWorker *worker)
1169 if (worker->heuristic_last_dequeue > worker->heuristic_last_adjustment + worker->heuristic_adjustment_interval) {
1170 ThreadPoolWorkerCounter counter;
1171 counter = COUNTER_READ (worker);
1172 if (counter._.working <= counter._.max_working)
1180 heuristic_adjust (MonoThreadPoolWorker *worker)
1182 if (mono_coop_mutex_trylock (&worker->heuristic_lock) == 0) {
1183 gint32 completions = InterlockedExchange (&worker->heuristic_completions, 0);
1184 gint64 sample_end = mono_msec_ticks ();
1185 gint64 sample_duration = sample_end - worker->heuristic_sample_start;
1187 if (sample_duration >= worker->heuristic_adjustment_interval / 2) {
1188 ThreadPoolWorkerCounter counter;
1189 gint16 new_thread_count;
1191 counter = COUNTER_READ (worker);
1192 new_thread_count = hill_climbing_update (worker, counter._.max_working, sample_duration, completions, &worker->heuristic_adjustment_interval);
1194 COUNTER_ATOMIC (worker, counter, {
1195 counter._.max_working = new_thread_count;
1198 if (new_thread_count > counter._.max_working)
1199 worker_request (worker);
1201 worker->heuristic_sample_start = sample_end;
1202 worker->heuristic_last_adjustment = mono_msec_ticks ();
1205 mono_coop_mutex_unlock (&worker->heuristic_lock);
1210 heuristic_notify_work_completed (MonoThreadPoolWorker *worker)
1214 InterlockedIncrement (&worker->heuristic_completions);
1215 worker->heuristic_last_dequeue = mono_msec_ticks ();
1217 if (heuristic_should_adjust (worker))
1218 heuristic_adjust (worker);
1222 mono_threadpool_worker_notify_completed (MonoThreadPoolWorker *worker)
1224 ThreadPoolWorkerCounter counter;
1226 heuristic_notify_work_completed (worker);
1228 counter = COUNTER_READ (worker);
1229 return counter._.working <= counter._.max_working;
1233 mono_threadpool_worker_get_min (MonoThreadPoolWorker *worker)
1235 return worker->limit_worker_min;
1239 mono_threadpool_worker_set_min (MonoThreadPoolWorker *worker, gint32 value)
1241 if (value <= 0 || value > worker->limit_worker_max)
1244 worker->limit_worker_min = value;
1249 mono_threadpool_worker_get_max (MonoThreadPoolWorker *worker)
1251 return worker->limit_worker_max;
1255 mono_threadpool_worker_set_max (MonoThreadPoolWorker *worker, gint32 value)
1257 gint32 cpu_count = mono_cpu_count ();
1259 if (value < worker->limit_worker_min || value < cpu_count)
1262 worker->limit_worker_max = value;
1267 mono_threadpool_worker_set_suspended (MonoThreadPoolWorker *worker, gboolean suspended)
1269 worker->suspended = suspended;
1271 worker_request (worker);