- File:
-
- 1 edited
-
libcfa/src/concurrency/ready_queue.cfa (modified) (26 diffs)
Legend:
- Unmodified
- Added
- Removed
-
libcfa/src/concurrency/ready_queue.cfa
r078fb05 r46bbcaf 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(); … … 201 201 uint_fast32_t ready_mutate_lock( void ) with(*__scheduler_lock) { 202 202 /* paranoid */ verify( ! __preemption_enabled() ); 203 /* paranoid */ verify( ! kernelTLS().sched_lock );204 203 205 204 // Step 1 : lock global lock … … 207 206 // to simply lock their own lock and enter. 208 207 __atomic_acquire( &write_lock ); 208 209 // Make sure we won't deadlock ourself 210 // Checking before acquiring the writer lock isn't safe 211 // because someone else could have locked us. 212 /* paranoid */ verify( ! kernelTLS().sched_lock ); 209 213 210 214 // Step 2 : lock per-proc lock … … 244 248 245 249 //======================================================================= 250 // caches handling 251 252 struct __attribute__((aligned(128))) __ready_queue_caches_t { 253 // Count States: 254 // - 0 : No one is looking after this cache 255 // - 1 : No one is looking after this cache, BUT it's not empty 256 // - 2+ : At least one processor is looking after this cache 257 volatile unsigned count; 258 }; 259 260 void ?{}(__ready_queue_caches_t & this) { this.count = 0; } 261 void ^?{}(__ready_queue_caches_t & this) {} 262 263 static inline void depart(__ready_queue_caches_t & cache) { 264 /* paranoid */ verify( cache.count > 1); 265 __atomic_fetch_add(&cache.count, -1, __ATOMIC_SEQ_CST); 266 /* paranoid */ verify( cache.count != 0); 267 /* 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. 268 } 269 270 static inline void arrive(__ready_queue_caches_t & cache) { 271 // for() { 272 // unsigned expected = cache.count; 273 // unsigned desired = 0 == expected ? 2 : expected + 1; 274 // } 275 } 276 277 //======================================================================= 246 278 // Cforall Ready Queue used for scheduling 247 279 //======================================================================= 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; 280 unsigned long long moving_average(unsigned long long currtsc, unsigned long long instsc, unsigned long long old_avg) { 281 /* paranoid */ verifyf( currtsc < 45000000000000000, "Suspiciously large current time: %'llu (%llx)\n", currtsc, currtsc ); 282 /* paranoid */ verifyf( instsc < 45000000000000000, "Suspiciously large insert time: %'llu (%llx)\n", instsc, instsc ); 283 /* paranoid */ verifyf( old_avg < 15000000000000, "Suspiciously large previous average: %'llu (%llx)\n", old_avg, old_avg ); 284 285 const unsigned long long new_val = currtsc > instsc ? currtsc - instsc : 0; 286 const unsigned long long total_weight = 16; 287 const unsigned long long new_weight = 4; 288 const unsigned long long old_weight = total_weight - new_weight; 289 const unsigned long long ret = ((new_weight * new_val) + (old_weight * old_avg)) / total_weight; 290 return ret; 253 291 } 254 292 … … 271 309 } 272 310 #else 273 lanes.data = 0p; 274 lanes.tscs = 0p; 275 lanes.help = 0p; 276 lanes.count = 0; 311 lanes.data = 0p; 312 lanes.tscs = 0p; 313 lanes.caches = 0p; 314 lanes.help = 0p; 315 lanes.count = 0; 277 316 #endif 278 317 } … … 285 324 free(lanes.data); 286 325 free(lanes.tscs); 326 free(lanes.caches); 287 327 free(lanes.help); 288 328 } 289 329 290 330 //----------------------------------------------------------------------- 331 #if defined(USE_AWARE_STEALING) 332 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) { 333 processor * const proc = kernelTLS().this_processor; 334 const bool external = (!proc) || (cltr != proc->cltr); 335 const bool remote = hint == UNPARK_REMOTE; 336 337 unsigned i; 338 if( external || remote ) { 339 // Figure out where thread was last time and make sure it's valid 340 /* paranoid */ verify(thrd->preferred >= 0); 341 if(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count) { 342 /* paranoid */ verify(thrd->preferred * READYQ_SHARD_FACTOR < lanes.count); 343 unsigned start = thrd->preferred * READYQ_SHARD_FACTOR; 344 do { 345 unsigned r = __tls_rand(); 346 i = start + (r % READYQ_SHARD_FACTOR); 347 /* paranoid */ verify( i < lanes.count ); 348 // If we can't lock it retry 349 } while( !__atomic_try_acquire( &lanes.data[i].lock ) ); 350 } else { 351 do { 352 i = __tls_rand() % lanes.count; 353 } while( !__atomic_try_acquire( &lanes.data[i].lock ) ); 354 } 355 } else { 356 do { 357 unsigned r = proc->rdq.its++; 358 i = proc->rdq.id + (r % READYQ_SHARD_FACTOR); 359 /* paranoid */ verify( i < lanes.count ); 360 // If we can't lock it retry 361 } while( !__atomic_try_acquire( &lanes.data[i].lock ) ); 362 } 363 364 // Actually push it 365 push(lanes.data[i], thrd); 366 367 // Unlock and return 368 __atomic_unlock( &lanes.data[i].lock ); 369 370 #if !defined(__CFA_NO_STATISTICS__) 371 if(unlikely(external || remote)) __atomic_fetch_add(&cltr->stats->ready.push.extrn.success, 1, __ATOMIC_RELAXED); 372 else __tls_stats()->ready.push.local.success++; 373 #endif 374 } 375 376 static inline unsigned long long calc_cutoff(const unsigned long long ctsc, const processor * proc, __ready_queue_t & rdq) { 377 unsigned start = proc->rdq.id; 378 unsigned long long max = 0; 379 for(i; READYQ_SHARD_FACTOR) { 380 unsigned long long ptsc = ts(rdq.lanes.data[start + i]); 381 if(ptsc != -1ull) { 382 /* paranoid */ verify( start + i < rdq.lanes.count ); 383 unsigned long long tsc = moving_average(ctsc, ptsc, rdq.lanes.tscs[start + i].ma); 384 if(tsc > max) max = tsc; 385 } 386 } 387 return (max + 2 * max) / 2; 388 } 389 390 __attribute__((hot)) struct thread$ * pop_fast(struct cluster * cltr) with (cltr->ready_queue) { 391 /* paranoid */ verify( lanes.count > 0 ); 392 /* paranoid */ verify( kernelTLS().this_processor ); 393 /* paranoid */ verify( kernelTLS().this_processor->rdq.id < lanes.count ); 394 395 processor * const proc = kernelTLS().this_processor; 396 unsigned this = proc->rdq.id; 397 /* paranoid */ verify( this < lanes.count ); 398 __cfadbg_print_safe(ready_queue, "Kernel : pop from %u\n", this); 399 400 // Figure out the current cpu and make sure it is valid 401 const int cpu = __kernel_getcpu(); 402 /* paranoid */ verify(cpu >= 0); 403 /* paranoid */ verify(cpu < cpu_info.hthrd_count); 404 unsigned this_cache = cpu_info.llc_map[cpu].cache; 405 406 // Super important: don't write the same value over and over again 407 // We want to maximise our chances that his particular values stays in cache 408 if(lanes.caches[this / READYQ_SHARD_FACTOR].id != this_cache) 409 __atomic_store_n(&lanes.caches[this / READYQ_SHARD_FACTOR].id, this_cache, __ATOMIC_RELAXED); 410 411 const unsigned long long ctsc = rdtscl(); 412 413 if(proc->rdq.target == MAX) { 414 uint64_t chaos = __tls_rand(); 415 unsigned ext = chaos & 0xff; 416 unsigned other = (chaos >> 8) % (lanes.count); 417 418 if(ext < 3 || __atomic_load_n(&lanes.caches[other / READYQ_SHARD_FACTOR].id, __ATOMIC_RELAXED) == this_cache) { 419 proc->rdq.target = other; 420 } 421 } 422 else { 423 const unsigned target = proc->rdq.target; 424 __cfadbg_print_safe(ready_queue, "Kernel : %u considering helping %u, tcsc %llu\n", this, target, lanes.tscs[target].tv); 425 /* paranoid */ verify( lanes.tscs[target].tv != MAX ); 426 if(target < lanes.count) { 427 const unsigned long long cutoff = calc_cutoff(ctsc, proc, cltr->ready_queue); 428 const unsigned long long age = moving_average(ctsc, lanes.tscs[target].tv, lanes.tscs[target].ma); 429 __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"); 430 if(age > cutoff) { 431 thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help)); 432 if(t) return t; 433 } 434 } 435 proc->rdq.target = MAX; 436 } 437 438 for(READYQ_SHARD_FACTOR) { 439 unsigned i = this + (proc->rdq.itr++ % READYQ_SHARD_FACTOR); 440 if(thread$ * t = try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.local))) return t; 441 } 442 443 // All lanes where empty return 0p 444 return 0p; 445 446 } 447 __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr) with (cltr->ready_queue) { 448 unsigned i = __tls_rand() % lanes.count; 449 return try_pop(cltr, i __STATS(, __tls_stats()->ready.pop.steal)); 450 } 451 __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr) { 452 return search(cltr); 453 } 454 #endif 291 455 #if defined(USE_CPU_WORK_STEALING) 292 456 __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint) with (cltr->ready_queue) { … … 350 514 /* paranoid */ verify( kernelTLS().this_processor ); 351 515 516 processor * const proc = kernelTLS().this_processor; 352 517 const int cpu = __kernel_getcpu(); 353 518 /* paranoid */ verify(cpu >= 0); … … 360 525 /* 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); 361 526 362 processor * const proc = kernelTLS().this_processor;363 527 const int start = map.self * READYQ_SHARD_FACTOR; 364 528 const unsigned long long ctsc = rdtscl(); 365 529 366 530 // Did we already have a help target 367 if(proc->rdq.target == -1u) {531 if(proc->rdq.target == MAX) { 368 532 unsigned long long max = 0; 369 533 for(i; READYQ_SHARD_FACTOR) { 370 unsigned long long tsc = moving_average(ctsc -ts(lanes.data[start + i]), lanes.tscs[start + i].ma);534 unsigned long long tsc = moving_average(ctsc, ts(lanes.data[start + i]), lanes.tscs[start + i].ma); 371 535 if(tsc > max) max = tsc; 372 536 } 373 proc->rdq.cutoff = (max + 2 * max) / 2;537 // proc->rdq.cutoff = (max + 2 * max) / 2; 374 538 /* paranoid */ verify(lanes.count < 65536); // The following code assumes max 65536 cores. 375 539 /* paranoid */ verify(map.count < 65536); // The following code assumes max 65536 cores. … … 384 548 } 385 549 386 /* paranoid */ verify(proc->rdq.target != -1u);550 /* paranoid */ verify(proc->rdq.target != MAX); 387 551 } 388 552 else { 389 553 unsigned long long max = 0; 390 554 for(i; READYQ_SHARD_FACTOR) { 391 unsigned long long tsc = moving_average(ctsc -ts(lanes.data[start + i]), lanes.tscs[start + i].ma);555 unsigned long long tsc = moving_average(ctsc, ts(lanes.data[start + i]), lanes.tscs[start + i].ma); 392 556 if(tsc > max) max = tsc; 393 557 } … … 395 559 { 396 560 unsigned target = proc->rdq.target; 397 proc->rdq.target = -1u;561 proc->rdq.target = MAX; 398 562 lanes.help[target / READYQ_SHARD_FACTOR].tri++; 399 if(moving_average(ctsc -lanes.tscs[target].tv, lanes.tscs[target].ma) > cutoff) {563 if(moving_average(ctsc, lanes.tscs[target].tv, lanes.tscs[target].ma) > cutoff) { 400 564 thread$ * t = try_pop(cltr, target __STATS(, __tls_stats()->ready.pop.help)); 401 565 proc->rdq.last = target; 402 566 if(t) return t; 403 else proc->rdq.target = -1u;404 567 } 405 else proc->rdq.target = -1u;568 proc->rdq.target = MAX; 406 569 } 407 570 408 571 unsigned last = proc->rdq.last; 409 if(last != -1u && lanes.tscs[last].tv < cutoff && ts(lanes.data[last]) <cutoff) {572 if(last != MAX && moving_average(ctsc, lanes.tscs[last].tv, lanes.tscs[last].ma) > cutoff) { 410 573 thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.help)); 411 574 if(t) return t; 412 575 } 413 576 else { 414 proc->rdq.last = -1u;577 proc->rdq.last = MAX; 415 578 } 416 579 } … … 428 591 processor * const proc = kernelTLS().this_processor; 429 592 unsigned last = proc->rdq.last; 430 if(last != -1u) {593 if(last != MAX) { 431 594 struct thread$ * t = try_pop(cltr, last __STATS(, __tls_stats()->ready.pop.steal)); 432 595 if(t) return t; 433 proc->rdq.last = -1u;596 proc->rdq.last = MAX; 434 597 } 435 598 … … 560 723 #else 561 724 unsigned preferred = thrd->preferred; 562 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || preferred == -1u|| thrd->curr_cluster != cltr;725 const bool external = (hint != UNPARK_LOCAL) || (!kernelTLS().this_processor) || preferred == MAX || thrd->curr_cluster != cltr; 563 726 /* paranoid */ verifyf(external || preferred < lanes.count, "Invalid preferred queue %u for %u lanes", preferred, lanes.count ); 564 727 … … 612 775 processor * proc = kernelTLS().this_processor; 613 776 614 if(proc->rdq.target == -1u) {777 if(proc->rdq.target == MAX) { 615 778 unsigned long long min = ts(lanes.data[proc->rdq.id]); 616 779 for(int i = 0; i < READYQ_SHARD_FACTOR; i++) { … … 623 786 else { 624 787 unsigned target = proc->rdq.target; 625 proc->rdq.target = -1u;788 proc->rdq.target = MAX; 626 789 const unsigned long long bias = 0; //2_500_000_000; 627 790 const unsigned long long cutoff = proc->rdq.cutoff > bias ? proc->rdq.cutoff - bias : proc->rdq.cutoff; … … 658 821 // try to pop from a lane given by index w 659 822 static inline struct thread$ * try_pop(struct cluster * cltr, unsigned w __STATS(, __stats_readyQ_pop_t & stats)) with (cltr->ready_queue) { 823 /* paranoid */ verify( w < lanes.count ); 660 824 __STATS( stats.attempt++; ) 661 825 … … 681 845 // Actually pop the list 682 846 struct thread$ * thrd; 683 #if defined(USE_ WORK_STEALING) || defined(USE_CPU_WORK_STEALING)847 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING) 684 848 unsigned long long tsc_before = ts(lane); 685 849 #endif … … 697 861 __STATS( stats.success++; ) 698 862 699 #if defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING) 700 unsigned long long now = rdtscl(); 701 lanes.tscs[w].tv = tsv; 702 lanes.tscs[w].ma = moving_average(now > tsc_before ? now - tsc_before : 0, lanes.tscs[w].ma); 863 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) || defined(USE_CPU_WORK_STEALING) 864 if (tsv != MAX) { 865 unsigned long long now = rdtscl(); 866 unsigned long long pma = __atomic_load_n(&lanes.tscs[w].ma, __ATOMIC_RELAXED); 867 __atomic_store_n(&lanes.tscs[w].tv, tsv, __ATOMIC_RELAXED); 868 __atomic_store_n(&lanes.tscs[w].ma, moving_average(now, tsc_before, pma), __ATOMIC_RELAXED); 869 } 703 870 #endif 704 871 705 #if defined(USE_ CPU_WORK_STEALING)872 #if defined(USE_AWARE_STEALING) || defined(USE_CPU_WORK_STEALING) 706 873 thrd->preferred = w / READYQ_SHARD_FACTOR; 707 874 #else … … 802 969 /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count); 803 970 it->rdq.id = value; 804 it->rdq.target = -1u;971 it->rdq.target = MAX; 805 972 value += READYQ_SHARD_FACTOR; 806 973 it = &(*it)`next; … … 815 982 816 983 static void fix_times( struct cluster * cltr ) with( cltr->ready_queue ) { 817 #if defined(USE_ WORK_STEALING)984 #if defined(USE_AWARE_STEALING) || defined(USE_WORK_STEALING) 818 985 lanes.tscs = alloc(lanes.count, lanes.tscs`realloc); 819 986 for(i; lanes.count) { 820 unsigned long long tsc1 = ts(lanes.data[i]); 821 unsigned long long tsc2 = rdtscl(); 822 lanes.tscs[i].tv = min(tsc1, tsc2); 987 lanes.tscs[i].tv = rdtscl(); 988 lanes.tscs[i].ma = 0; 823 989 } 824 990 #endif … … 866 1032 // Update original 867 1033 lanes.count = ncount; 1034 1035 lanes.caches = alloc( target, lanes.caches`realloc ); 868 1036 } 869 1037 … … 942 1110 fix(lanes.data[idx]); 943 1111 } 1112 1113 lanes.caches = alloc( target, lanes.caches`realloc ); 944 1114 } 945 1115 946 1116 fix_times(cltr); 1117 947 1118 948 1119 reassign_cltr_id(cltr);
Note:
See TracChangeset
for help on using the changeset viewer.
