source: libcfa/src/concurrency/actor.hfa@ 3c82bf6

Last change on this file since 3c82bf6 was 9235192c, checked in by caparsons <caparson@…>, 2 years ago

added support for missed gulps benchmark
  • Property mode set to 100644
File size: 25.0 KB
Line 
1#pragma once
2
3#include <locks.hfa>
4#include <limits.hfa>
5#include <kernel.hfa>
6#include <iofwd.hfa>
7#include <virtual_dtor.hfa>
8
9#ifdef __CFA_DEBUG__
10#define CFA_DEBUG( stmt ) stmt
11#else
12#define CFA_DEBUG( stmt )
13#endif // CFA_DEBUG
14
15#define DEBUG_ABORT( cond, string ) CFA_DEBUG( if ( cond ) abort( string ) )
16
17// Define the default number of processors created in the executor. Must be greater than 0.
18#define __DEFAULT_EXECUTOR_PROCESSORS__ 2
19
20// Define the default number of threads created in the executor. Must be greater than 0.
21#define __DEFAULT_EXECUTOR_WORKERS__ 2
22
23// Define the default number of executor request-queues (mailboxes) written to by actors and serviced by the
24// actor-executor threads. Must be greater than 0.
25#define __DEFAULT_EXECUTOR_RQUEUES__ 4
26
27// Define if executor is created in a separate cluster
28#define __DEFAULT_EXECUTOR_SEPCLUS__ false
29
30#define __DEFAULT_EXECUTOR_BUFSIZE__ 10
31
32#define __STEAL 1 // workstealing toggle. Disjoint from toggles above
33
34// workstealing heuristic selection (only set one to be 1)
35// #define RAND 0
36#define SEARCH 1
37
38// show stats
39// #define ACTOR_STATS
40
41// used to run and only track missed queue gulps
42#ifdef ACTOR_STATS
43#define ACTOR_STATS_QUEUE_MISSED
44#endif
45
46// forward decls
47struct actor;
48struct message;
49struct executor;
50
51enum allocation { Nodelete, Delete, Destroy, Finished }; // allocation status
52
53typedef allocation (*__receive_fn)(actor &, message &, actor **, message **);
54struct request {
55 actor * receiver;
56 message * msg;
57 __receive_fn fn;
58};
59
60struct a_msg {
61 int m;
62};
63static inline void ?{}( request & this ) {}
64static inline void ?{}( request & this, actor * receiver, message * msg, __receive_fn fn ) {
65 this.receiver = receiver;
66 this.msg = msg;
67 this.fn = fn;
68}
69static inline void ?{}( request & this, request & copy ) {
70 this.receiver = copy.receiver;
71 this.msg = copy.msg;
72 this.fn = copy.fn;
73}
74
75// Vector-like data structure that supports O(1) queue operations with no bound on size
76// assumes gulping behaviour (once a remove occurs, removes happen until empty beforw next insert)
77struct copy_queue {
78 request * buffer;
79 size_t count, buffer_size, index, utilized, last_size;
80};
81static inline void ?{}( copy_queue & this ) {}
82static inline void ?{}( copy_queue & this, size_t buf_size ) with(this) {
83 buffer_size = buf_size;
84 buffer = aalloc( buffer_size );
85 count = 0;
86 utilized = 0;
87 index = 0;
88 last_size = 0;
89}
90static inline void ^?{}( copy_queue & this ) with(this) {
91 DEBUG_ABORT( count != 0, "Actor system terminated with messages sent but not received\n" );
92 adelete(buffer);
93}
94
95static inline void insert( copy_queue & this, request & elem ) with(this) {
96 if ( count >= buffer_size ) { // increase arr size
97 last_size = buffer_size;
98 buffer_size = 2 * buffer_size;
99 buffer = realloc( buffer, sizeof( request ) * buffer_size );
100 /* paranoid */ verify( buffer );
101 }
102 memcpy( &buffer[count], &elem, sizeof(request) );
103 count++;
104}
105
106// once you start removing you need to remove all elements
107// it is not supported to call insert() before the array is fully empty
108static inline request & remove( copy_queue & this ) with(this) {
109 if ( count > 0 ) {
110 count--;
111 size_t old_idx = index;
112 index = count == 0 ? 0 : index + 1;
113 return buffer[old_idx];
114 }
115 request * ret = 0p;
116 return *0p;
117}
118
119// try to reclaim some memory if less than half of buffer is utilized
120static inline void reclaim( copy_queue & this ) with(this) {
121 if ( utilized >= last_size || buffer_size <= 4 ) { utilized = 0; return; }
122 utilized = 0;
123 buffer_size--;
124 buffer = realloc( buffer, sizeof( request ) * buffer_size ); // try to reclaim some memory
125}
126
127static inline bool is_empty( copy_queue & this ) with(this) { return count == 0; }
128
129struct work_queue {
130 __spinlock_t mutex_lock;
131 copy_queue * owned_queue; // copy queue allocated and cleaned up by this work_queue
132 copy_queue * c_queue; // current queue
133 volatile bool being_processed; // flag to prevent concurrent processing
134 #ifdef ACTOR_STATS
135 unsigned int id;
136 #endif
137 #ifdef ACTOR_STATS_QUEUE_MISSED
138 size_t missed; // transfers skipped due to being_processed flag being up
139 #endif
140}; // work_queue
141static inline void ?{}( work_queue & this, size_t buf_size, unsigned int i ) with(this) {
142 owned_queue = alloc(); // allocated separately to avoid false sharing
143 (*owned_queue){ buf_size };
144 c_queue = owned_queue;
145 being_processed = false;
146 #ifdef ACTOR_STATS
147 id = i;
148 missed = 0;
149 #endif
150}
151
152// clean up copy_queue owned by this work_queue
153static inline void ^?{}( work_queue & this ) with(this) { delete( owned_queue ); }
154
155static inline void insert( work_queue & this, request & elem ) with(this) {
156 lock( mutex_lock __cfaabi_dbg_ctx2 );
157 insert( *c_queue, elem );
158 unlock( mutex_lock );
159} // insert
160
161static inline void transfer( work_queue & this, copy_queue ** transfer_to ) with(this) {
162 lock( mutex_lock __cfaabi_dbg_ctx2 );
163 #ifdef __STEAL
164
165 // check if queue is being processed elsewhere
166 if ( unlikely( being_processed ) ) {
167 #ifdef ACTOR_STATS
168 missed++;
169 #endif
170 unlock( mutex_lock );
171 return;
172 }
173
174 being_processed = c_queue->count != 0;
175 #endif // __STEAL
176
177 c_queue->utilized = c_queue->count;
178
179 // swap copy queue ptrs
180 copy_queue * temp = *transfer_to;
181 *transfer_to = c_queue;
182 c_queue = temp;
183 unlock( mutex_lock );
184} // transfer
185
186// needed since some info needs to persist past worker lifetimes
187struct worker_info {
188 volatile unsigned long long stamp;
189 #ifdef ACTOR_STATS
190 size_t stolen_from, try_steal, stolen, empty_stolen, failed_swaps, msgs_stolen;
191 unsigned long long processed;
192 size_t gulps;
193 #endif
194};
195static inline void ?{}( worker_info & this ) {
196 #ifdef ACTOR_STATS
197 this.stolen_from = 0;
198 this.try_steal = 0; // attempts to steal
199 this.stolen = 0; // successful steals
200 this.processed = 0; // requests processed
201 this.gulps = 0; // number of gulps
202 this.failed_swaps = 0; // steal swap failures
203 this.empty_stolen = 0; // queues empty after steal
204 this.msgs_stolen = 0; // number of messages stolen
205 #endif
206 this.stamp = rdtscl();
207}
208
209// #ifdef ACTOR_STATS
210// unsigned int * stolen_arr;
211// unsigned int * replaced_queue;
212// #endif
213thread worker {
214 work_queue ** request_queues;
215 copy_queue * current_queue;
216 executor * executor_;
217 unsigned int start, range;
218 int id;
219};
220
221#ifdef ACTOR_STATS
222// aggregate counters for statistics
223size_t __total_tries = 0, __total_stolen = 0, __total_workers, __all_gulps = 0, __total_empty_stolen = 0,
224 __total_failed_swaps = 0, __all_processed = 0, __num_actors_stats = 0, __all_msgs_stolen = 0;
225#endif
226static inline void ?{}( worker & this, cluster & clu, work_queue ** request_queues, copy_queue * current_queue, executor * executor_,
227 unsigned int start, unsigned int range, int id ) {
228 ((thread &)this){ clu };
229 this.request_queues = request_queues; // array of all queues
230 this.current_queue = current_queue; // currently gulped queue (start with empty queue to use in swap later)
231 this.executor_ = executor_; // pointer to current executor
232 this.start = start; // start of worker's subrange of request_queues
233 this.range = range; // size of worker's subrange of request_queues
234 this.id = id; // worker's id and index in array of workers
235}
236
237static bool no_steal = false;
238struct executor {
239 cluster * cluster; // if workers execute on separate cluster
240 processor ** processors; // array of virtual processors adding parallelism for workers
241 work_queue * request_queues; // master array of work request queues
242 copy_queue * local_queues; // array of all worker local queues to avoid deletion race
243 work_queue ** worker_req_queues; // secondary array of work queues to allow for swapping
244 worker ** workers; // array of workers executing work requests
245 worker_info * w_infos; // array of info about each worker
246 unsigned int nprocessors, nworkers, nrqueues; // number of processors/threads/request queues
247 bool seperate_clus; // use same or separate cluster for executor
248 volatile bool is_shutdown; // flag to communicate shutdown to worker threads
249}; // executor
250
251// #ifdef ACTOR_STATS
252// __spinlock_t out_lock;
253// #endif
254static inline void ^?{}( worker & mutex this ) with(this) {
255 #ifdef ACTOR_STATS
256 __atomic_add_fetch(&__all_gulps, executor_->w_infos[id].gulps,__ATOMIC_SEQ_CST);
257 __atomic_add_fetch(&__all_processed, executor_->w_infos[id].processed,__ATOMIC_SEQ_CST);
258 __atomic_add_fetch(&__all_msgs_stolen, executor_->w_infos[id].msgs_stolen,__ATOMIC_SEQ_CST);
259 __atomic_add_fetch(&__total_tries, executor_->w_infos[id].try_steal, __ATOMIC_SEQ_CST);
260 __atomic_add_fetch(&__total_stolen, executor_->w_infos[id].stolen, __ATOMIC_SEQ_CST);
261 __atomic_add_fetch(&__total_failed_swaps, executor_->w_infos[id].failed_swaps, __ATOMIC_SEQ_CST);
262 __atomic_add_fetch(&__total_empty_stolen, executor_->w_infos[id].empty_stolen, __ATOMIC_SEQ_CST);
263
264 // per worker steal stats (uncomment alongside the lock above this routine to print)
265 // lock( out_lock __cfaabi_dbg_ctx2 );
266 // printf("Worker id: %d, processed: %llu messages, attempted %lu, stole: %lu, stolen from: %lu\n", id, processed, try_steal, stolen, __atomic_add_fetch(&executor_->w_infos[id].stolen_from, 0, __ATOMIC_SEQ_CST) );
267 // int count = 0;
268 // int count2 = 0;
269 // for ( i; range ) {
270 // if ( replaced_queue[start + i] > 0 ){
271 // count++;
272 // // printf("%d: %u, ",i, replaced_queue[i]);
273 // }
274 // if (__atomic_add_fetch(&stolen_arr[start + i],0,__ATOMIC_SEQ_CST) > 0)
275 // count2++;
276 // }
277 // printf("swapped with: %d of %u indices\n", count, executor_->nrqueues / executor_->nworkers );
278 // printf("%d of %u indices were stolen\n", count2, executor_->nrqueues / executor_->nworkers );
279 // unlock( out_lock );
280 #endif
281}
282
283static inline void ?{}( executor & this, unsigned int nprocessors, unsigned int nworkers, unsigned int nrqueues, bool seperate_clus, size_t buf_size ) with(this) {
284 if ( nrqueues < nworkers ) abort( "nrqueues needs to be >= nworkers\n" );
285 this.nprocessors = nprocessors;
286 this.nworkers = nworkers;
287 this.nrqueues = nrqueues;
288 this.seperate_clus = seperate_clus;
289 this.is_shutdown = false;
290
291 if ( nworkers == nrqueues )
292 no_steal = true;
293
294 #ifdef ACTOR_STATS
295 // stolen_arr = aalloc( nrqueues );
296 // replaced_queue = aalloc( nrqueues );
297 __total_workers = nworkers;
298 #endif
299
300 if ( seperate_clus ) {
301 cluster = alloc();
302 (*cluster){};
303 } else cluster = active_cluster();
304
305 request_queues = aalloc( nrqueues );
306 worker_req_queues = aalloc( nrqueues );
307 for ( i; nrqueues ) {
308 request_queues[i]{ buf_size, i };
309 worker_req_queues[i] = &request_queues[i];
310 }
311
312 processors = aalloc( nprocessors );
313 for ( i; nprocessors )
314 (*(processors[i] = alloc())){ *cluster };
315
316 local_queues = aalloc( nworkers );
317 workers = aalloc( nworkers );
318 w_infos = aalloc( nworkers );
319 unsigned int reqPerWorker = nrqueues / nworkers, extras = nrqueues % nworkers;
320
321 for ( i; nworkers ) {
322 w_infos[i]{};
323 local_queues[i]{ buf_size };
324 }
325
326 for ( unsigned int i = 0, start = 0, range; i < nworkers; i += 1, start += range ) {
327 range = reqPerWorker + ( i < extras ? 1 : 0 );
328 (*(workers[i] = alloc())){ *cluster, worker_req_queues, &local_queues[i], &this, start, range, i };
329 } // for
330}
331static inline void ?{}( executor & this, unsigned int nprocessors, unsigned int nworkers, unsigned int nrqueues, bool seperate_clus ) { this{ nprocessors, nworkers, nrqueues, seperate_clus, __DEFAULT_EXECUTOR_BUFSIZE__ }; }
332static inline void ?{}( executor & this, unsigned int nprocessors, unsigned int nworkers, unsigned int nrqueues ) { this{ nprocessors, nworkers, nrqueues, __DEFAULT_EXECUTOR_SEPCLUS__ }; }
333static inline void ?{}( executor & this, unsigned int nprocessors, unsigned int nworkers ) { this{ nprocessors, nworkers, __DEFAULT_EXECUTOR_RQUEUES__ }; }
334static inline void ?{}( executor & this, unsigned int nprocessors ) { this{ nprocessors, __DEFAULT_EXECUTOR_WORKERS__ }; }
335static inline void ?{}( executor & this ) { this{ __DEFAULT_EXECUTOR_PROCESSORS__ }; }
336
337static inline void ^?{}( executor & this ) with(this) {
338 is_shutdown = true;
339
340 for ( i; nworkers )
341 delete( workers[i] );
342
343 for ( i; nprocessors ) {
344 delete( processors[i] );
345 } // for
346
347 #ifdef ACTOR_STATS_QUEUE_MISSED
348 size_t misses = 0;
349 for ( i; nrqueues ) {
350 misses += worker_req_queues[i]->missed;
351 }
352 // adelete( stolen_arr );
353 // adelete( replaced_queue );
354 #endif
355
356 adelete( workers );
357 adelete( w_infos );
358 adelete( local_queues );
359 adelete( request_queues );
360 adelete( worker_req_queues );
361 adelete( processors );
362 if ( seperate_clus ) delete( cluster );
363
364 #ifdef ACTOR_STATS // print formatted stats
365 printf(" Actor System Stats:\n");
366 printf("\tActors Created:\t\t\t\t%lu\n\tMessages Sent:\t\t\t\t%lu\n", __num_actors_stats, __all_processed);
367 size_t avg_gulps = __all_gulps == 0 ? 0 : __all_processed / __all_gulps;
368 printf("\tGulps:\t\t\t\t\t%lu\n\tAverage Gulp Size:\t\t\t%lu\n\tMissed gulps:\t\t\t\t%lu\n", __all_gulps, avg_gulps, misses);
369 printf("\tSteal attempts:\t\t\t\t%lu\n\tSteals:\t\t\t\t\t%lu\n\tSteal failures (no candidates):\t\t%lu\n\tSteal failures (failed swaps):\t\t%lu\t Empty steals:\t\t%lu\n",
370 __total_tries, __total_stolen, __total_tries - __total_stolen - __total_failed_swaps, __total_failed_swaps, __total_empty_stolen);
371 size_t avg_steal = __total_stolen == 0 ? 0 : __all_msgs_stolen / __total_stolen;
372 printf("\tMessages stolen:\t\t\t%lu\n\tAverage steal size:\t\t\t%lu\n", __all_msgs_stolen, avg_steal);
373 #endif
374
375 #ifndef ACTOR_STATS
376 #ifdef ACTOR_STATS_QUEUE_MISSED
377 printf("\t%lu", misses);
378 #endif
379 #endif
380
381}
382
383// this is a static field of executor but have to forward decl for get_next_ticket
384static size_t __next_ticket = 0;
385
386static inline size_t __get_next_ticket( executor & this ) with(this) {
387 #ifdef __CFA_DEBUG__
388 size_t temp = __atomic_fetch_add( &__next_ticket, 1, __ATOMIC_SEQ_CST) % nrqueues;
389
390 // reserve MAX for dead actors
391 if ( unlikely( temp == MAX ) ) temp = __atomic_fetch_add( &__next_ticket, 1, __ATOMIC_SEQ_CST) % nrqueues;
392 return temp;
393 #else
394 return __atomic_fetch_add( &__next_ticket, 1, __ATOMIC_RELAXED) % nrqueues;
395 #endif
396} // tickets
397
398// TODO: update globals in this file to be static fields once the static fields project is done
399static executor * __actor_executor_ = 0p;
400static bool __actor_executor_passed = false; // was an executor passed to start_actor_system
401static size_t __num_actors_ = 0; // number of actor objects in system
402static struct thread$ * __actor_executor_thd = 0p; // used to wake executor after actors finish
403struct actor {
404 size_t ticket; // executor-queue handle
405 allocation alloc; // allocation action
406 inline virtual_dtor;
407};
408
409static inline void ?{}( actor & this ) with(this) {
410 // Once an actor is allocated it must be sent a message or the actor system cannot stop. Hence, its receive
411 // member must be called to end it
412 DEBUG_ABORT( __actor_executor_ == 0p, "Creating actor before calling start_actor_system() can cause undefined behaviour.\n" );
413 alloc = Nodelete;
414 ticket = __get_next_ticket( *__actor_executor_ );
415 __atomic_fetch_add( &__num_actors_, 1, __ATOMIC_RELAXED );
416 #ifdef ACTOR_STATS
417 __atomic_fetch_add( &__num_actors_stats, 1, __ATOMIC_SEQ_CST );
418 #endif
419}
420
421static inline void check_actor( actor & this ) {
422 if ( this.alloc != Nodelete ) {
423 switch( this.alloc ) {
424 case Delete: delete( &this ); break;
425 case Destroy:
426 CFA_DEBUG( this.ticket = MAX; ); // mark as terminated
427 ^?{}(this);
428 break;
429 case Finished:
430 CFA_DEBUG( this.ticket = MAX; ); // mark as terminated
431 break;
432 default: ; // stop warning
433 }
434
435 if ( unlikely( __atomic_add_fetch( &__num_actors_, -1, __ATOMIC_RELAXED ) == 0 ) ) { // all actors have terminated
436 unpark( __actor_executor_thd );
437 }
438 }
439}
440
441struct message {
442 allocation alloc; // allocation action
443 inline virtual_dtor;
444};
445
446static inline void ?{}( message & this ) {
447 this.alloc = Nodelete;
448}
449static inline void ?{}( message & this, allocation alloc ) {
450 memcpy( &this.alloc, &alloc, sizeof(allocation) ); // optimization to elide ctor
451 CFA_DEBUG( if ( this.alloc == Finished ) this.alloc = Nodelete; );
452}
453static inline void ^?{}( message & this ) with(this) {
454 CFA_DEBUG(
455 if ( alloc == Nodelete ) {
456 printf( "CFA warning (UNIX pid:%ld) : program terminating with message %p allocated but never sent.\n",
457 (long int)getpid(), &this );
458 }
459 )
460}
461
462static inline void check_message( message & this ) {
463 switch ( this.alloc ) { // analyze message status
464 case Nodelete: CFA_DEBUG( this.alloc = Finished ); break;
465 case Delete: delete( &this ); break;
466 case Destroy: ^?{}( this ); break;
467 case Finished: break;
468 } // switch
469}
470static inline allocation set_allocation( message & this, allocation state ) {
471 CFA_DEBUG( if ( state == Nodelete ) state = Finished; );
472 allocation prev = this.alloc;
473 this.alloc = state;
474 return prev;
475}
476static inline allocation get_allocation( message & this ) {
477 return this.alloc;
478}
479
480static inline void deliver_request( request & this ) {
481 DEBUG_ABORT( this.receiver->ticket == (unsigned long int)MAX, "Attempted to send message to deleted/dead actor\n" );
482 actor * base_actor;
483 message * base_msg;
484 allocation temp = this.fn( *this.receiver, *this.msg, &base_actor, &base_msg );
485 memcpy( &base_actor->alloc, &temp, sizeof(allocation) ); // optimization to elide ctor
486 check_message( *base_msg );
487 check_actor( *base_actor );
488}
489
490// tries to atomically swap two queues and returns 0p if the swap failed
491// returns ptr to newly owned queue if swap succeeds
492static inline work_queue * try_swap_queues( worker & this, unsigned int victim_idx, unsigned int my_idx ) with(this) {
493 work_queue * my_queue = request_queues[my_idx];
494 work_queue * other_queue = request_queues[victim_idx];
495
496 // if either queue is 0p then they are in the process of being stolen
497 if ( other_queue == 0p ) return 0p;
498
499 // try to set our queue ptr to be 0p. If it fails someone moved our queue so return false
500 if ( !__atomic_compare_exchange_n( &request_queues[my_idx], &my_queue, 0p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) )
501 return 0p;
502
503 // try to set other queue ptr to be our queue ptr. If it fails someone moved the other queue so fix up then return false
504 if ( !__atomic_compare_exchange_n( &request_queues[victim_idx], &other_queue, my_queue, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST ) ) {
505 /* paranoid */ verify( request_queues[my_idx] == 0p );
506 request_queues[my_idx] = my_queue; // reset my queue ptr back to appropriate val
507 return 0p;
508 }
509
510 // we have successfully swapped and since our queue is 0p no one will touch it so write back new queue ptr non atomically
511 request_queues[my_idx] = other_queue; // last write does not need to be atomic
512 return other_queue;
513}
514
515// once a worker to steal from has been chosen, choose queue to steal from
516static inline void choose_queue( worker & this, unsigned int victim_id, unsigned int swap_idx ) with(this) {
517 // have to calculate victim start and range since victim may be deleted before us in shutdown
518 const unsigned int queues_per_worker = executor_->nrqueues / executor_->nworkers;
519 const unsigned int extras = executor_->nrqueues % executor_->nworkers;
520 unsigned int vic_start, vic_range;
521 if ( extras > victim_id ) {
522 vic_range = queues_per_worker + 1;
523 vic_start = vic_range * victim_id;
524 } else {
525 vic_start = extras + victim_id * queues_per_worker;
526 vic_range = queues_per_worker;
527 }
528 unsigned int start_idx = prng( vic_range );
529
530 unsigned int tries = 0;
531 work_queue * curr_steal_queue;
532
533 for ( unsigned int i = start_idx; tries < vic_range; i = (i + 1) % vic_range ) {
534 tries++;
535 curr_steal_queue = request_queues[ i + vic_start ];
536 // avoid empty queues and queues that are being operated on
537 if ( curr_steal_queue == 0p || curr_steal_queue->being_processed || is_empty( *curr_steal_queue->c_queue ) )
538 continue;
539
540 #ifdef ACTOR_STATS
541 curr_steal_queue = try_swap_queues( this, i + vic_start, swap_idx );
542 if ( curr_steal_queue ) {
543 executor_->w_infos[id].msgs_stolen += curr_steal_queue->c_queue->count;
544 executor_->w_infos[id].stolen++;
545 if ( is_empty( *curr_steal_queue->c_queue ) ) executor_->w_infos[id].empty_stolen++;
546 // __atomic_add_fetch(&executor_->w_infos[victim_id].stolen_from, 1, __ATOMIC_RELAXED);
547 // replaced_queue[swap_idx]++;
548 // __atomic_add_fetch(&stolen_arr[ i + vic_start ], 1, __ATOMIC_RELAXED);
549 } else {
550 executor_->w_infos[id].failed_swaps++;
551 }
552 #else
553 curr_steal_queue = try_swap_queues( this, i + vic_start, swap_idx );
554 #endif // ACTOR_STATS
555
556 return;
557 }
558
559 return;
560}
561
562// choose a worker to steal from
563static inline void steal_work( worker & this, unsigned int swap_idx ) with(this) {
564 #if RAND
565 unsigned int victim = prng( executor_->nworkers );
566 if ( victim == id ) victim = ( victim + 1 ) % executor_->nworkers;
567 choose_queue( this, victim, swap_idx );
568 #elif SEARCH
569 unsigned long long min = MAX; // smaller timestamp means longer since service
570 int min_id = 0; // use ints not uints to avoid integer underflow without hacky math
571 int n_workers = executor_->nworkers;
572 unsigned long long curr_stamp;
573 int scount = 1;
574 for ( int i = (id + 1) % n_workers; scount < n_workers; i = (i + 1) % n_workers, scount++ ) {
575 curr_stamp = executor_->w_infos[i].stamp;
576 if ( curr_stamp < min ) {
577 min = curr_stamp;
578 min_id = i;
579 }
580 }
581 choose_queue( this, min_id, swap_idx );
582 #endif
583}
584
585#define CHECK_TERMINATION if ( unlikely( executor_->is_shutdown ) ) break Exit
586void main( worker & this ) with(this) {
587 // #ifdef ACTOR_STATS
588 // for ( i; executor_->nrqueues ) {
589 // replaced_queue[i] = 0;
590 // __atomic_store_n( &stolen_arr[i], 0, __ATOMIC_SEQ_CST );
591 // }
592 // #endif
593
594 // threshold of empty queues we see before we go stealing
595 const unsigned int steal_threshold = 2 * range;
596
597 // Store variable data here instead of worker struct to avoid any potential false sharing
598 unsigned int empty_count = 0;
599 request & req;
600 work_queue * curr_work_queue;
601
602 Exit:
603 for ( unsigned int i = 0;; i = (i + 1) % range ) { // cycle through set of request buffers
604 curr_work_queue = request_queues[i + start];
605
606 #ifndef __STEAL
607 CHECK_TERMINATION;
608 #endif
609
610 // check if queue is empty before trying to gulp it
611 if ( is_empty( *curr_work_queue->c_queue ) ) {
612 #ifdef __STEAL
613 empty_count++;
614 if ( empty_count < steal_threshold ) continue;
615 #else
616 continue;
617 #endif
618 }
619 transfer( *curr_work_queue, &current_queue );
620 #ifdef ACTOR_STATS
621 executor_->w_infos[id].gulps++;
622 #endif // ACTOR_STATS
623 #ifdef __STEAL
624 if ( is_empty( *current_queue ) ) {
625 if ( unlikely( no_steal ) ) { CHECK_TERMINATION; continue; }
626 empty_count++;
627 if ( empty_count < steal_threshold ) continue;
628 empty_count = 0;
629
630 CHECK_TERMINATION; // check for termination
631
632 __atomic_store_n( &executor_->w_infos[id].stamp, rdtscl(), __ATOMIC_RELAXED );
633
634 #ifdef ACTOR_STATS
635 executor_->w_infos[id].try_steal++;
636 #endif // ACTOR_STATS
637
638 steal_work( this, start + prng( range ) );
639 continue;
640 }
641 #endif // __STEAL
642 while ( ! is_empty( *current_queue ) ) {
643 #ifdef ACTOR_STATS
644 executor_->w_infos[id].processed++;
645 #endif
646 &req = &remove( *current_queue );
647 if ( !&req ) continue;
648 deliver_request( req );
649 }
650 #ifdef __STEAL
651 curr_work_queue->being_processed = false; // set done processing
652 empty_count = 0; // we found work so reset empty counter
653 #endif
654
655 CHECK_TERMINATION;
656
657 // potentially reclaim some of the current queue's vector space if it is unused
658 reclaim( *current_queue );
659 } // for
660}
661
662static inline void send( executor & this, request & req, unsigned long int ticket ) with(this) {
663 insert( request_queues[ticket], req);
664}
665
666static inline void send( actor & this, request & req ) {
667 DEBUG_ABORT( this.ticket == (unsigned long int)MAX, "Attempted to send message to deleted/dead actor\n" );
668 send( *__actor_executor_, req, this.ticket );
669}
670
671static inline void __reset_stats() {
672 #ifdef ACTOR_STATS
673 __total_tries = 0;
674 __total_stolen = 0;
675 __all_gulps = 0;
676 __total_failed_swaps = 0;
677 __total_empty_stolen = 0;
678 __all_processed = 0;
679 __num_actors_stats = 0;
680 __all_msgs_stolen = 0;
681 #endif
682}
683
684static inline void start_actor_system( size_t num_thds ) {
685 __reset_stats();
686 __actor_executor_thd = active_thread();
687 __actor_executor_ = alloc();
688 (*__actor_executor_){ 0, num_thds, num_thds == 1 ? 1 : num_thds * 16 };
689}
690
691static inline void start_actor_system() { start_actor_system( get_proc_count( *active_cluster() ) ); }
692
693static inline void start_actor_system( executor & this ) {
694 __reset_stats();
695 __actor_executor_thd = active_thread();
696 __actor_executor_ = &this;
697 __actor_executor_passed = true;
698}
699
700static inline void stop_actor_system() {
701 park(); // unparked when actor system is finished
702
703 if ( !__actor_executor_passed ) delete( __actor_executor_ );
704 __actor_executor_ = 0p;
705 __actor_executor_thd = 0p;
706 __next_ticket = 0;
707 __actor_executor_passed = false;
708}
709
710// Default messages to send to any actor to change status
711// assigned at creation to __base_msg_finished to avoid unused message warning
712message __base_msg_finished @= { .alloc : Finished };
713struct delete_msg_t { inline message; } delete_msg = __base_msg_finished;
714struct destroy_msg_t { inline message; } destroy_msg = __base_msg_finished;
715struct finished_msg_t { inline message; } finished_msg = __base_msg_finished;
716
717allocation receive( actor & this, delete_msg_t & msg ) { return Delete; }
718allocation receive( actor & this, destroy_msg_t & msg ) { return Destroy; }
719allocation receive( actor & this, finished_msg_t & msg ) { return Finished; }
Note: See TracBrowser for help on using the repository browser.