Changeset 175f9f4 for libcfa/src/concurrency/ready_queue.cfa
- Timestamp:
- Jan 18, 2022, 1:16:23 PM (2 years ago)
- Branches:
- ADT, ast-experimental, enum, forall-pointer-decay, master, pthread-emulation, qualifiedEnum
- Children:
- 1e8b4b49, adfd125
- Parents:
- 21a5bfb7 (diff), 91a72ef (diff)
Note: this is a merge changeset, the changes displayed below correspond to the merge itself.
Use the(diff)links above to see all the changes relative to each parent. - File:
-
- 1 edited
-
libcfa/src/concurrency/ready_queue.cfa (modified) (24 diffs)
Legend:
- Unmodified
- Added
- Removed
-
libcfa/src/concurrency/ready_queue.cfa
r21a5bfb7 r175f9f4 20 20 21 21 22 #define USE_RELAXED_FIFO22 // #define USE_RELAXED_FIFO 23 23 // #define USE_WORK_STEALING 24 24 // #define USE_CPU_WORK_STEALING 25 #define USE_AWARE_STEALING 25 26 26 27 #include "bits/defs.hfa" … … 29 30 30 31 #include "stdlib.hfa" 32 #include "limits.hfa" 31 33 #include "math.hfa" 32 34 … … 54 56 #endif 55 57 56 #if defined(USE_CPU_WORK_STEALING) 58 #if defined(USE_AWARE_STEALING) 59 #define READYQ_SHARD_FACTOR 2 60 #define SEQUENTIAL_SHARD 2 61 #elif defined(USE_CPU_WORK_STEALING) 57 62 #define READYQ_SHARD_FACTOR 2 58 63 #elif defined(USE_RELAXED_FIFO) … … 138 143 __kernel_rseq_register(); 139 144 140 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p for RW-Lock\n", proc);141 145 bool * handle = (bool *)&kernelTLS().sched_lock; 142 146 … … 174 178 } 175 179 176 __cfadbg_print_safe(ready_queue, "Kernel : Registering proc %p done, id %lu\n", proc, n);177 178 180 // Return new spot. 179 181 /* paranoid */ verify(n < ready); … … 190 192 191 193 __atomic_store_n(cell, 0p, __ATOMIC_RELEASE); 192 193 __cfadbg_print_safe(ready_queue, "Kernel : Unregister proc %p\n", proc);194 194 195 195 __kernel_rseq_unregister(); … … 244 244 245 245 //======================================================================= 246 // caches handling 247 248 struct __attribute__((aligned(128))) __ready_queue_caches_t { 249 // Count States: 250 // - 0 : No one is looking after this cache 251 // - 1 : No one is looking after this cache, BUT it's not empty 252 // - 2+ : At least one processor is looking after this cache 253 volatile unsigned count; 254 }; 255 256 void ?{}(__ready_queue_caches_t & this) { this.count = 0; } 257 void ^?{}(__ready_queue_caches_t & this) {} 258 259 static inline void depart(__ready_queue_caches_t & cache) { 260 /* paranoid */ verify( cache.count > 1); 261 __atomic_fetch_add(&cache.count, -1, __ATOMIC_SEQ_CST); 262 /* paranoid */ verify( cache.count != 0); 263 /* paranoid */ verify( cache.count < 65536 ); // This verify assumes no cluster will have more than 65000 kernel threads mapped to a single cache, which could be correct but is super weird. 264 } 265 266 static inline void arrive(__ready_queue_caches_t & cache) { 267 // for() { 268 // unsigned expected = cache.count; 269 // unsigned desired = 0 == expected ? 2 : expected + 1; 270 // } 271 } 272 273 //======================================================================= 246 274 // Cforall Ready Queue used for scheduling 247 275 //======================================================================= 248 unsigned long long moving_average(unsigned long long nval, unsigned long long oval) { 249 const unsigned long long tw = 16; 250 const unsigned long long nw = 4; 251 const unsigned long long ow = tw - nw; 252 return ((nw * nval) + (ow * oval)) / tw; 276 unsigned long long moving_average(unsigned long long currtsc, unsigned long long instsc, unsigned long long old_avg) { 277 /* paranoid */ verifyf( currtsc < 45000000000000000, "Suspiciously large current time: %'llu (%llx)\n", currtsc, currtsc ); 278 /* paranoid */ verifyf( instsc < 45000000000000000, "Suspiciously large insert time: %'llu (%llx)\n", instsc, instsc ); 279 /* paranoid */ verifyf( old_avg < 15000000000000, "Suspiciously large previous average: %'llu (%llx)\n", old_avg, old_avg ); 280 281 const unsigned long long new_val = currtsc > instsc ? currtsc - instsc : 0; 282 const unsigned long long total_weight = 16; 283 const unsigned long long new_weight = 4; 284 const unsigned long long old_weight = total_weight - new_weight; 285 const unsigned long long ret = ((new_weight * new_val) + (old_weight * old_avg)) / total_weight; 286 return ret; 253 287 } 254 288 … … 271 305 } 272 306 #else 273 lanes.data = 0p; 274 lanes.tscs = 0p; 275 lanes.help = 0p; 276 lanes.count = 0; 307 lanes.data = 0p; 308 lanes.tscs = 0p; 309 lanes.caches = 0p; 310 lanes.help = 0p; 311 lanes.count = 0; 277 312 #endif 278 313 } … … 285 320 free(lanes.data); 286 321 free(lanes.tscs); 322 free(lanes.caches); 287 323 free(lanes.help); 288 324 } 289 325 290 326 //----------------------------------------------------------------------- 327 #if defined(USE_AWARE_STEALING) 328 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) { 329 processor * const proc = kernelTLS().this_processor; 330 const bool external = (!proc) || (cltr != proc->cltr); 331 const bool remote = hint == UNPARK_REMOTE; 332 333 unsigned i; 334 if( external || remote ) { 335 // Figure out where thread was last time and make sure it's valid 336 /* paranoid */ verify(thrd->preferred >= 0); 337 if(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count) { 338 /* paranoid */ verify(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count); 339 unsigned start = thrd->preferred * READYQ_SHARD_FACTOR; 340 do { 341 unsigned r = __tls_rand(); 342 i = start + (r % READYQ_SHARD_FACTOR); 343 /* paranoid */ verify( i < lanes.count ); 344 // If we can't lock it retry 345 } while( !__atomic_try_acquire( &lanes.data[i].lock ) ); 346 } else { 347 do { 348 i = __tls_rand() % lanes.count; 349 } while( !__atomic_try_acquire( &lanes.data[i].lock ) ); 350 } 351 } else { 352 do { 353 unsigned r = proc->rdq.its++; 354 i = proc->rdq.id + (r % READYQ_SHARD_FACTOR); 355 /* paranoid */ verify( i < lanes.count ); 356 // If we can't lock it retry 357 } while( !__atomic_try_acquire( &lanes.data[i].lock ) ); 358 } 359 360 // Actually push it 361 push(lanes.data[i], thrd); 362 363 // Unlock and return 364 __atomic_unlock( &lanes.data[i].lock ); 365 366 #if !defined(__CFA_NO_STATISTICS__) 367 if(unlikely(external || remote)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED); 368 else __tls_stats()->ready.push.local.success++; 369 #endif 370 } 371 372 static inline unsigned long long calc_cutoff(const unsigned long long ctsc, const processor * proc, __ready_queue_t & rdq) { 373 unsigned start = proc->rdq.id; 374 unsigned long long max = 0; 375 for(i; READYQ_SHARD_FACTOR) { 376 unsigned long long ptsc = ts(rdq.lanes.data[start + i]); 377 if(ptsc != -1ull) { 378 /* paranoid */ verify( start + i < rdq.lanes.count ); 379 unsigned long long tsc = moving_average(ctsc, ptsc, rdq.lanes.tscs[start + i].ma); 380 if(tsc > max) max = tsc; 381 } 382 } 383 return (max + 2 * max) / 2; 384 } 385 386 __attribute__((hot)) struct thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) { 387 /* paranoid */ verify( lanes.count > 0 ); 388 /* paranoid */ verify( kernelTLS().this_processor ); 389 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count ); 390 391 processor * const proc = kernelTLS().this_processor; 392 unsigned this = proc->rdq.id; 393 /* paranoid */ verify( this < lanes.count ); 394 __cfadbg_print_safe(ready_queue, "Kernel : pop from %u\n", this); 395 396 // Figure out the current cpu and make sure it is valid 397 const int cpu = __kernel_getcpu(); 398 /* paranoid */ verify(cpu >= 0); 399 /* paranoid */ verify(cpu < cpu_info.hthrd_count); 400 unsigned this_cache = cpu_info.llc_map[cpu].cache; 401 402 // Super important: don't write the same value over and over again 403 // We want to maximise our chances that his particular values stays in cache 404 if(lanes.caches[this / READYQ_SHARD_FACTOR].id != this_cache) 405 __atomic_store_n(&lanes.caches[this / READYQ_SHARD_FACTOR].id, this_cache, __ATOMIC_RELAXED); 406 407 const unsigned long long ctsc = rdtscl(); 408 409 if(proc->rdq.target == MAX) { 410 uint64_t chaos = __tls_rand(); 411 unsigned ext = chaos & 0xff; 412 unsigned other = (chaos >> 8) % (lanes.count); 413 414 if(ext < 3 || __atomic_load_n(&lanes.caches[other / READYQ_SHARD_FACTOR].id, __ATOMIC_RELAXED) == this_cache) { 415 proc->rdq.target = other; 416 } 417 } 418 else { 419 const unsigned target = proc->rdq.target; 420 __cfadbg_print_safe(ready_queue, "Kernel : %u considering helping %u, tcsc %llu\n", this, target, lanes.tscs[target].tv); 421 /* paranoid */ verify( lanes.tscs[target].tv != MAX ); 422 if(target < lanes.count) { 423 const unsigned long long cutoff = calc_cutoff(ctsc, proc, cltr->ready_queue); 424 const unsigned long long age = moving_average(ctsc, lanes.tscs[target].tv, lanes.tscs[target].ma); 425 __cfadbg_print_safe(ready_queue, "Kernel : Help attempt on %u from %u, age %'llu vs cutoff %'llu, %s\n", target, this, age, cutoff, age > cutoff ? "yes" : "no"); 426 if(age > cutoff) { 427 thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help)); 428 if(t) return t; 429 } 430 } 431 proc->rdq.target = MAX; 432 } 433 434 for(READYQ_SHARD_FACTOR) { 435 unsigned i = this + (proc->rdq.itr++ % READYQ_SHARD_FACTOR); 436 if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t; 437 } 438 439 // All lanes where empty return 0p 440 return 0p; 441 442 } 443 __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) { 444 unsigned i = __tls_rand() % lanes.count; 445 return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal)); 446 } 447 __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) { 448 return search(cltr); 449 } 450 #endif 291 451 #if defined(USE_CPU_WORK_STEALING) 292 452 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) { … … 350 510 /* paranoid */ verify( kernelTLS().this_processor ); 351 511 512 processor * const proc = kernelTLS().this_processor; 352 513 const int cpu = __kernel_getcpu(); 353 514 /* paranoid */ verify(cpu >= 0); … … 360 521 /* paranoid */ verifyf((map.start + map.count) * READYQ_SHARD_FACTOR <= lanes.count, "have %zu lanes but map can go up to %u", lanes.count, (map.start + map.count) * READYQ_SHARD_FACTOR); 361 522 362 processor * const proc = kernelTLS().this_processor;363 523 const int start = map.self * READYQ_SHARD_FACTOR; 364 524 const unsigned long long ctsc = rdtscl(); 365 525 366 526 // Did we already have a help target 367 if(proc->rdq.target == -1u) {527 if(proc->rdq.target == MAX) { 368 528 unsigned long long max = 0; 369 529 for(i; READYQ_SHARD_FACTOR) { 370 unsigned long long tsc = moving_average(ctsc -ts(lanes.data[start + i]), lanes.tscs[start + i].ma);530 unsigned long long tsc = moving_average(ctsc, ts(lanes.data[start + i]), lanes.tscs[start + i].ma); 371 531 if(tsc > max) max = tsc; 372 532 } 373 proc->rdq.cutoff = (max + 2 * max) / 2;533 // proc->rdq.cutoff = (max + 2 * max) / 2; 374 534 /* paranoid */ verify(lanes.count < 65536); // The following code assumes max 65536 cores. 375 535 /* paranoid */ verify(map.count < 65536); // The following code assumes max 65536 cores. … … 384 544 } 385 545 386 /* paranoid */ verify(proc->rdq.target != -1u);546 /* paranoid */ verify(proc->rdq.target != MAX); 387 547 } 388 548 else { 389 549 unsigned long long max = 0; 390 550 for(i; READYQ_SHARD_FACTOR) { 391 unsigned long long tsc = moving_average(ctsc -ts(lanes.data[start + i]), lanes.tscs[start + i].ma);551 unsigned long long tsc = moving_average(ctsc, ts(lanes.data[start + i]), lanes.tscs[start + i].ma); 392 552 if(tsc > max) max = tsc; 393 553 } … … 395 555 { 396 556 unsigned target = proc->rdq.target; 397 proc->rdq.target = -1u;557 proc->rdq.target = MAX; 398 558 lanes.help[target / READYQ_SHARD_FACTOR].tri++; 399 if(moving_average(ctsc -lanes.tscs[target].tv, lanes.tscs[target].ma) > cutoff) {559 if(moving_average(ctsc, lanes.tscs[target].tv, lanes.tscs[target].ma) > cutoff) { 400 560 thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help)); 401 561 proc->rdq.last = target; 402 562 if(t) return t; 403 else proc->rdq.target = -1u;404 563 } 405 else proc->rdq.target = -1u;564 proc->rdq.target = MAX; 406 565 } 407 566 408 567 unsigned last = proc->rdq.last; 409 if(last != -1u && lanes.tscs[last].tv < cutoff && ts(lanes.data[last]) <cutoff) {568 if(last != MAX && moving_average(ctsc, lanes.tscs[last].tv, lanes.tscs[last].ma) > cutoff) { 410 569 thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.help)); 411 570 if(t) return t; 412 571 } 413 572 else { 414 proc->rdq.last = -1u;573 proc->rdq.last = MAX; 415 574 } 416 575 } … … 428 587 processor * const proc = kernelTLS().this_processor; 429 588 unsigned last = proc->rdq.last; 430 if(last != -1u) {589 if(last != MAX) { 431 590 struct thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.steal)); 432 591 if(t) return t; 433 proc->rdq.last = -1u;592 proc->rdq.last = MAX; 434 593 } 435 594 … … 560 719 #else 561 720 unsigned preferred = thrd->preferred; 562 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || preferred == -1u|| thrd->curr_cluster != cltr;721 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || preferred == MAX || thrd->curr_cluster != cltr; 563 722 /* paranoid */ verifyf(external || preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count ); 564 723 … … 612 771 processor * proc = kernelTLS().this_processor; 613 772 614 if(proc->rdq.target == -1u) {773 if(proc->rdq.target == MAX) { 615 774 unsigned long long min = ts(lanes.data[proc->rdq.id]); 616 775 for(int i = 0; i < READYQ_SHARD_FACTOR; i++) { … … 623 782 else { 624 783 unsigned target = proc->rdq.target; 625 proc->rdq.target = -1u;784 proc->rdq.target = MAX; 626 785 const unsigned long long bias = 0; //2_500_000_000; 627 786 const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff; … … 658 817 // try to pop from a lane given by index w 659 818 static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) { 819 /* paranoid */ verify( w < lanes.count ); 660 820 __STATS( stats.attempt++; ) 661 821 … … 681 841 // Actually pop the list 682 842 struct thread$ * thrd; 683 unsigned long long tsc_before = ts(lane); 843 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING) 844 unsigned long long tsc_before = ts(lane); 845 #endif 684 846 unsigned long long tsv; 685 847 [thrd, tsv] = pop(lane); … … 695 857 __STATS( stats.success++; ) 696 858 697 #if defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING) 698 unsigned long long now = rdtscl(); 699 lanes.tscs[w].tv = tsv; 700 lanes.tscs[w].ma = moving_average(now > tsc_before ? now - tsc_before : 0, lanes.tscs[w].ma); 859 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING) 860 if (tsv != MAX) { 861 unsigned long long now = rdtscl(); 862 unsigned long long pma = __atomic_load_n(&lanes.tscs[w].ma, __ATOMIC_RELAXED); 863 __atomic_store_n(&lanes.tscs[w].tv, tsv, __ATOMIC_RELAXED); 864 __atomic_store_n(&lanes.tscs[w].ma, moving_average(now, tsc_before, pma), __ATOMIC_RELAXED); 865 } 701 866 #endif 702 867 703 #if defined(USE_ CPU_WORK_STEALING)868 #if defined(USE_AWARE_STEALING) || defined(USE_CPU_WORK_STEALING) 704 869 thrd->preferred = w / READYQ_SHARD_FACTOR; 705 870 #else … … 800 965 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count); 801 966 it->rdq.id = value; 802 it->rdq.target = -1u;967 it->rdq.target = MAX; 803 968 value += READYQ_SHARD_FACTOR; 804 969 it = &(*it)`next; … … 813 978 814 979 static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) { 815 #if defined(USE_ WORK_STEALING)980 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) 816 981 lanes.tscs = alloc(lanes.count, lanes.tscs`realloc); 817 982 for(i; lanes.count) { 818 unsigned long long tsc1 = ts(lanes.data[i]); 819 unsigned long long tsc2 = rdtscl(); 820 lanes.tscs[i].tv = min(tsc1, tsc2); 983 lanes.tscs[i].tv = rdtscl(); 984 lanes.tscs[i].ma = 0; 821 985 } 822 986 #endif … … 864 1028 // Update original 865 1029 lanes.count = ncount; 1030 1031 lanes.caches = alloc( target, lanes.caches`realloc ); 866 1032 } 867 1033 … … 940 1106 fix(lanes.data[idx]); 941 1107 } 1108 1109 lanes.caches = alloc( target, lanes.caches`realloc ); 942 1110 } 943 1111 944 1112 fix_times(cltr); 1113 945 1114 946 1115 reassign_cltr_id(cltr);
Note: See TracChangeset
for help on using the changeset viewer.
