reduce contention on page metadata lists during the sweeping phase (J…

…uliaLang#52943) **EDIT**: fixes JuliaLang#52937 by decreasing the contention on the page lists and only waking GC threads up if we have a sufficiently large number of pages. Seems to address the regression from the MWE of JuliaLang#52937: - master: ``` ../julia-master/julia --project=. run_benchmarks.jl serial obj_arrays issue-52937 -n5 --gcthreads=1 bench = "issue-52937.jl" ┌─────────┬────────────┬─────────┬───────────┬────────────┬──────────────┬───────────────────┬──────────┬────────────┐ │ │ total time │ gc time │ mark time │ sweep time │ max GC pause │ time to safepoint │ max heap │ percent gc │ │ │ ms │ ms │ ms │ ms │ ms │ us │ MB │ % │ ├─────────┼────────────┼─────────┼───────────┼────────────┼──────────────┼───────────────────┼──────────┼────────────┤ │ minimum │ 24841 │ 818 │ 78 │ 740 │ 44 │ 10088 │ 96 │ 3 │ │ median │ 24881 │ 834 │ 83 │ 751 │ 45 │ 10738 │ 97 │ 3 │ │ maximum │ 25002 │ 891 │ 87 │ 803 │ 48 │ 11074 │ 112 │ 4 │ │ stdev │ 78 │ 29 │ 4 │ 26 │ 1 │ 393 │ 7 │ 0 │ └─────────┴────────────┴─────────┴───────────┴────────────┴──────────────┴───────────────────┴──────────┴────────────┘ ../julia-master/julia --project=. run_benchmarks.jl serial obj_arrays issue-52937 -n5 --gcthreads=8 bench = "issue-52937.jl" ┌─────────┬────────────┬─────────┬───────────┬────────────┬──────────────┬───────────────────┬──────────┬────────────┐ │ │ total time │ gc time │ mark time │ sweep time │ max GC pause │ time to safepoint │ max heap │ percent gc │ │ │ ms │ ms │ ms │ ms │ ms │ us │ MB │ % │ ├─────────┼────────────┼─────────┼───────────┼────────────┼──────────────┼───────────────────┼──────────┼────────────┤ │ minimum │ 29113 │ 5200 │ 68 │ 5130 │ 12 │ 9724 │ 95 │ 18 │ │ median │ 29354 │ 5274 │ 69 │ 5204 │ 12 │ 10456 │ 96 │ 18 │ │ maximum │ 29472 │ 5333 │ 70 │ 5264 │ 14 │ 11913 │ 97 │ 18 │ │ stdev │ 138 │ 54 │ 1 │ 55 │ 1 │ 937 │ 1 │ 0 │ └─────────┴────────────┴─────────┴───────────┴────────────┴──────────────┴───────────────────┴──────────┴────────────┘ ``` - PR: ``` ../julia-master/julia --project=. run_benchmarks.jl serial obj_arrays issue-52937 -n5 --gcthreads=1 bench = "issue-52937.jl" ┌─────────┬────────────┬─────────┬───────────┬────────────┬──────────────┬───────────────────┬──────────┬────────────┐ │ │ total time │ gc time │ mark time │ sweep time │ max GC pause │ time to safepoint │ max heap │ percent gc │ │ │ ms │ ms │ ms │ ms │ ms │ us │ MB │ % │ ├─────────┼────────────┼─────────┼───────────┼────────────┼──────────────┼───────────────────┼──────────┼────────────┤ │ minimum │ 24475 │ 761 │ 77 │ 681 │ 40 │ 9499 │ 94 │ 3 │ │ median │ 24845 │ 775 │ 80 │ 698 │ 43 │ 10793 │ 97 │ 3 │ │ maximum │ 25128 │ 811 │ 85 │ 726 │ 47 │ 12820 │ 113 │ 3 │ │ stdev │ 240 │ 22 │ 3 │ 21 │ 3 │ 1236 │ 8 │ 0 │ └─────────┴────────────┴─────────┴───────────┴────────────┴──────────────┴───────────────────┴──────────┴────────────┘ ../julia-master/julia --project=. run_benchmarks.jl serial obj_arrays issue-52937 -n5 --gcthreads=8 bench = "issue-52937.jl" ┌─────────┬────────────┬─────────┬───────────┬────────────┬──────────────┬───────────────────┬──────────┬────────────┐ │ │ total time │ gc time │ mark time │ sweep time │ max GC pause │ time to safepoint │ max heap │ percent gc │ │ │ ms │ ms │ ms │ ms │ ms │ us │ MB │ % │ ├─────────┼────────────┼─────────┼───────────┼────────────┼──────────────┼───────────────────┼──────────┼────────────┤ │ minimum │ 24709 │ 679 │ 70 │ 609 │ 11 │ 9981 │ 95 │ 3 │ │ median │ 24869 │ 702 │ 70 │ 631 │ 12 │ 10705 │ 96 │ 3 │ │ maximum │ 24911 │ 708 │ 72 │ 638 │ 13 │ 10820 │ 98 │ 3 │ │ stdev │ 79 │ 12 │ 1 │ 12 │ 1 │ 401 │ 1 │ 0 │ └─────────┴────────────┴─────────┴───────────┴────────────┴──────────────┴───────────────────┴──────────┴────────────┘ ``` Also, performance on `objarray.jl` (an example of benchmark in which sweeping parallelizes well with the current implementation) seems fine: - master: ``` ../julia-master/julia --project=. run_benchmarks.jl multithreaded bigarrays -n5 --gcthreads=1 bench = "objarray.jl" ┌─────────┬────────────┬─────────┬───────────┬────────────┬──────────────┬───────────────────┬──────────┬────────────┐ │ │ total time │ gc time │ mark time │ sweep time │ max GC pause │ time to safepoint │ max heap │ percent gc │ │ │ ms │ ms │ ms │ ms │ ms │ us │ MB │ % │ ├─────────┼────────────┼─────────┼───────────┼────────────┼──────────────┼───────────────────┼──────────┼────────────┤ │ minimum │ 19301 │ 10792 │ 7485 │ 3307 │ 1651 │ 196 │ 4519 │ 56 │ │ median │ 21415 │ 12646 │ 9094 │ 3551 │ 1985 │ 241 │ 6576 │ 59 │ │ maximum │ 21873 │ 13118 │ 9353 │ 3765 │ 2781 │ 330 │ 8793 │ 60 │ │ stdev │ 1009 │ 932 │ 757 │ 190 │ 449 │ 50 │ 1537 │ 2 │ └─────────┴────────────┴─────────┴───────────┴────────────┴──────────────┴───────────────────┴──────────┴────────────┘ ../julia-master/julia --project=. run_benchmarks.jl multithreaded bigarrays -n5 --gcthreads=8 bench = "objarray.jl" ┌─────────┬────────────┬─────────┬───────────┬────────────┬──────────────┬───────────────────┬──────────┬────────────┐ │ │ total time │ gc time │ mark time │ sweep time │ max GC pause │ time to safepoint │ max heap │ percent gc │ │ │ ms │ ms │ ms │ ms │ ms │ us │ MB │ % │ ├─────────┼────────────┼─────────┼───────────┼────────────┼──────────────┼───────────────────┼──────────┼────────────┤ │ minimum │ 13135 │ 4377 │ 3350 │ 1007 │ 491 │ 231 │ 6062 │ 33 │ │ median │ 13164 │ 4540 │ 3370 │ 1177 │ 669 │ 256 │ 6383 │ 35 │ │ maximum │ 13525 │ 4859 │ 3675 │ 1184 │ 748 │ 320 │ 7528 │ 36 │ │ stdev │ 183 │ 189 │ 146 │ 77 │ 129 │ 42 │ 584 │ 1 │ └─────────┴────────────┴─────────┴───────────┴────────────┴──────────────┴───────────────────┴──────────┴────────────┘ ``` - PR: ``` ../julia-master/julia --project=. run_benchmarks.jl multithreaded bigarrays -n5 --gcthreads=1 bench = "objarray.jl" ┌─────────┬────────────┬─────────┬───────────┬────────────┬──────────────┬───────────────────┬──────────┬────────────┐ │ │ total time │ gc time │ mark time │ sweep time │ max GC pause │ time to safepoint │ max heap │ percent gc │ │ │ ms │ ms │ ms │ ms │ ms │ us │ MB │ % │ ├─────────┼────────────┼─────────┼───────────┼────────────┼──────────────┼───────────────────┼──────────┼────────────┤ │ minimum │ 19642 │ 10931 │ 7566 │ 3365 │ 1653 │ 204 │ 5688 │ 56 │ │ median │ 21441 │ 12717 │ 8948 │ 3770 │ 1796 │ 217 │ 6972 │ 59 │ │ maximum │ 23494 │ 14643 │ 10576 │ 4067 │ 2513 │ 248 │ 8229 │ 62 │ │ stdev │ 1408 │ 1339 │ 1079 │ 267 │ 393 │ 19 │ 965 │ 2 │ └─────────┴────────────┴─────────┴───────────┴────────────┴──────────────┴───────────────────┴──────────┴────────────┘ ../julia-master/julia --project=. run_benchmarks.jl multithreaded bigarrays -n5 --gcthreads=8 bench = "objarray.jl" ┌─────────┬────────────┬─────────┬───────────┬────────────┬──────────────┬───────────────────┬──────────┬────────────┐ │ │ total time │ gc time │ mark time │ sweep time │ max GC pause │ time to safepoint │ max heap │ percent gc │ │ │ ms │ ms │ ms │ ms │ ms │ us │ MB │ % │ ├─────────┼────────────┼─────────┼───────────┼────────────┼──────────────┼───────────────────┼──────────┼────────────┤ │ minimum │ 13365 │ 4544 │ 3389 │ 1104 │ 516 │ 255 │ 6349 │ 34 │ │ median │ 13445 │ 4624 │ 3404 │ 1233 │ 578 │ 275 │ 6385 │ 34 │ │ maximum │ 14413 │ 5278 │ 3837 │ 1441 │ 753 │ 300 │ 7547 │ 37 │ │ stdev │ 442 │ 303 │ 194 │ 121 │ 89 │ 18 │ 522 │ 1 │ └─────────┴────────────┴─────────┴───────────┴────────────┴──────────────┴───────────────────┴──────────┴────────────┘ ```
RelationalAI · Jan 25, 2024 · 8acb66c · 8acb66c
1 parent 575a553
commit 8acb66c
Show file tree

Hide file tree

Showing 3 changed files with 154 additions and 18 deletions.
diff --git a/src/gc.c b/src/gc.c
@@ -21,8 +21,8 @@ int jl_n_sweepthreads;
 _Atomic(int) gc_n_threads_marking;
 // Number of threads sweeping
 _Atomic(int) gc_n_threads_sweeping;
-// Temporary for the `ptls->page_metadata_allocd` used during parallel sweeping
-_Atomic(jl_gc_page_stack_t *) gc_allocd_scratch;
+// Temporary for the `ptls->page_metadata_allocd` used during parallel sweeping (padded to avoid false sharing)
+_Atomic(jl_gc_padded_page_stack_t *) gc_allocd_scratch;
 // `tid` of mutator thread that triggered GC
 _Atomic(int) gc_master_tid;
 // `tid` of first GC thread
@@ -1586,8 +1586,72 @@ static void gc_pool_sync_nfree(jl_gc_pagemeta_t *pg, jl_taggedvalue_t *last) JL_
     pg->nfree = nfree;
 }
 
-void gc_sweep_wake_all(void)
+// pre-scan pages to check whether there are enough pages so that's worth parallelizing
+// also sweeps pages that don't need to be linearly scanned
+int gc_sweep_prescan(jl_ptls_t ptls, jl_gc_padded_page_stack_t *new_gc_allocd_scratch)
 {
+    // 4MB worth of pages is worth parallelizing
+    const int n_pages_worth_parallel_sweep = (int)(4 * (1 << 20) / GC_PAGE_SZ);
+    int n_pages_to_scan = 0;
+    gc_page_profiler_serializer_t serializer = gc_page_serializer_create();
+    for (int t_i = 0; t_i < gc_n_threads; t_i++) {
+        jl_ptls_t ptls2 = gc_all_tls_states[t_i];
+        if (ptls2 == NULL) {
+            continue;
+        }
+        jl_gc_page_stack_t *dest = &new_gc_allocd_scratch[ptls2->tid].stack;
+        jl_gc_page_stack_t tmp;
+        jl_gc_pagemeta_t *tail = NULL;
+        memset(&tmp, 0, sizeof(tmp));
+        while (1) {
+            jl_gc_pagemeta_t *pg = pop_lf_back_nosync(&ptls2->page_metadata_allocd);
+            if (pg == NULL) {
+                break;
+            }
+            int should_scan = 1;
+            if (!pg->has_marked) {
+                should_scan = 0;
+            }
+            if (!current_sweep_full && !pg->has_young) {
+                assert(!prev_sweep_full || pg->prev_nold >= pg->nold);
+                if (!prev_sweep_full || pg->prev_nold == pg->nold) {
+                    should_scan = 0;
+                }
+            }
+            if (should_scan) {
+                if (tail == NULL) {
+                    tail = pg;
+                }
+                n_pages_to_scan++;
+                push_lf_back_nosync(&tmp, pg);
+            }
+            else {
+                gc_sweep_pool_page(&serializer, dest, &ptls2->page_metadata_buffered, pg);
+            }
+            if (n_pages_to_scan >= n_pages_worth_parallel_sweep) {
+                break;
+            }
+        }
+        if (tail != NULL) {
+            tail->next = jl_atomic_load_relaxed(&ptls2->page_metadata_allocd.bottom);
+        }
+        ptls2->page_metadata_allocd = tmp;
+        if (n_pages_to_scan >= n_pages_worth_parallel_sweep) {
+            break;
+        }
+    }
+    gc_page_serializer_destroy(&serializer);
+    return n_pages_to_scan >= n_pages_worth_parallel_sweep;
+}
+
+// wake up all threads to sweep the pages
+void gc_sweep_wake_all(jl_ptls_t ptls, jl_gc_padded_page_stack_t *new_gc_allocd_scratch)
+{
+    int parallel_sweep_worthwhile = gc_sweep_prescan(ptls, new_gc_allocd_scratch);
+    jl_atomic_store(&gc_allocd_scratch, new_gc_allocd_scratch);
+    if (!parallel_sweep_worthwhile) {
+        return;
+    }
     uv_mutex_lock(&gc_threads_lock);
     for (int i = gc_first_tid; i < gc_first_tid + jl_n_markthreads; i++) {
         jl_ptls_t ptls2 = gc_all_tls_states[i];
@@ -1597,6 +1661,7 @@ void gc_sweep_wake_all(void)
     uv_mutex_unlock(&gc_threads_lock);
 }
 
+// wait for all threads to finish sweeping
 void gc_sweep_wait_for_all(void)
 {
     jl_atomic_store(&gc_allocd_scratch, NULL);
@@ -1605,36 +1670,58 @@ void gc_sweep_wait_for_all(void)
     }
 }
 
-void gc_sweep_pool_parallel(void)
+// sweep all pools
+void gc_sweep_pool_parallel(jl_ptls_t ptls)
 {
     jl_atomic_fetch_add(&gc_n_threads_sweeping, 1);
-    jl_gc_page_stack_t *allocd_scratch = jl_atomic_load(&gc_allocd_scratch);
+    jl_gc_padded_page_stack_t *allocd_scratch = jl_atomic_load(&gc_allocd_scratch);
     if (allocd_scratch != NULL) {
         gc_page_profiler_serializer_t serializer = gc_page_serializer_create();
         while (1) {
             int found_pg = 0;
+            // sequentially walk the threads and sweep the pages
             for (int t_i = 0; t_i < gc_n_threads; t_i++) {
                 jl_ptls_t ptls2 = gc_all_tls_states[t_i];
+                // skip foreign threads that already exited
                 if (ptls2 == NULL) {
                     continue;
                 }
-                jl_gc_page_stack_t *allocd = &allocd_scratch[t_i];
-                jl_gc_pagemeta_t *pg = pop_lf_back(&ptls2->page_metadata_allocd);
+                jl_gc_page_stack_t *dest = &allocd_scratch[ptls2->tid].stack;
+                jl_gc_pagemeta_t *pg = try_pop_lf_back(&ptls2->page_metadata_allocd);
+                // failed steal attempt
                 if (pg == NULL) {
                     continue;
                 }
-                gc_sweep_pool_page(&serializer, allocd, &ptls2->page_metadata_buffered, pg);
+                gc_sweep_pool_page(&serializer, dest, &ptls2->page_metadata_buffered, pg);
                 found_pg = 1;
             }
             if (!found_pg) {
-                break;
+                // check for termination
+                int no_more_work = 1;
+                for (int t_i = 0; t_i < gc_n_threads; t_i++) {
+                    jl_ptls_t ptls2 = gc_all_tls_states[t_i];
+                    // skip foreign threads that already exited
+                    if (ptls2 == NULL) {
+                        continue;
+                    }
+                    jl_gc_pagemeta_t *pg = jl_atomic_load_relaxed(&ptls2->page_metadata_allocd.bottom);
+                    if (pg != NULL) {
+                        no_more_work = 0;
+                        break;
+                    }
+                }
+                if (no_more_work) {
+                    break;
+                }
             }
+            jl_cpu_pause();
         }
         gc_page_serializer_destroy(&serializer);
     }
     jl_atomic_fetch_add(&gc_n_threads_sweeping, -1);
 }
 
+// free all pages (i.e. through `madvise` on Linux) that were lazily freed
 void gc_free_pages(void)
 {
     while (1) {
@@ -1659,7 +1746,7 @@ static void gc_sweep_pool(void)
 
     // allocate enough space to hold the end of the free list chain
     // for every thread and pool size
-    jl_taggedvalue_t ***pfl = (jl_taggedvalue_t ***) alloca(n_threads * JL_GC_N_POOLS * sizeof(jl_taggedvalue_t**));
+    jl_taggedvalue_t ***pfl = (jl_taggedvalue_t ***) malloc_s(n_threads * JL_GC_N_POOLS * sizeof(jl_taggedvalue_t**));
 
     // update metadata of pages that were pointed to by freelist or newpages from a pool
     // i.e. pages being the current allocation target
@@ -1701,17 +1788,18 @@ static void gc_sweep_pool(void)
     }
 
     // the actual sweeping
-    jl_gc_page_stack_t *tmp = (jl_gc_page_stack_t *)alloca(n_threads * sizeof(jl_gc_page_stack_t));
-    memset(tmp, 0, n_threads * sizeof(jl_gc_page_stack_t));
-    jl_atomic_store(&gc_allocd_scratch, tmp);
-    gc_sweep_wake_all();
-    gc_sweep_pool_parallel();
+    jl_gc_padded_page_stack_t *new_gc_allocd_scratch = (jl_gc_padded_page_stack_t *) malloc_s(n_threads * sizeof(jl_gc_padded_page_stack_t));
+    memset(new_gc_allocd_scratch, 0, n_threads * sizeof(jl_gc_padded_page_stack_t));
+    jl_ptls_t ptls = jl_current_task->ptls;
+    gc_sweep_wake_all(ptls, new_gc_allocd_scratch);
+    gc_sweep_pool_parallel(ptls);
     gc_sweep_wait_for_all();
 
+    // reset half-pages pointers
     for (int t_i = 0; t_i < n_threads; t_i++) {
         jl_ptls_t ptls2 = gc_all_tls_states[t_i];
         if (ptls2 != NULL) {
-            ptls2->page_metadata_allocd = tmp[t_i];
+            ptls2->page_metadata_allocd = new_gc_allocd_scratch[t_i].stack;
             for (int i = 0; i < JL_GC_N_POOLS; i++) {
                 jl_gc_pool_t *p = &ptls2->heap.norm_pools[i];
                 p->newpages = NULL;
@@ -1749,6 +1837,10 @@ static void gc_sweep_pool(void)
         }
     }
 
+    // cleanup
+    free(pfl);
+    free(new_gc_allocd_scratch);
+
 #ifdef _P64 // only enable concurrent sweeping on 64bit
     // wake thread up to sweep concurrently
     if (jl_n_sweepthreads > 0) {

diff --git a/src/gc.h b/src/gc.h
@@ -195,6 +195,23 @@ extern jl_gc_page_stack_t global_page_pool_freed;
 // in the sweeping phase, which also doesn't push a node into the
 // same stack after it's popped
 
+STATIC_INLINE void push_lf_back_nosync(jl_gc_page_stack_t *pool, jl_gc_pagemeta_t *elt) JL_NOTSAFEPOINT
+{
+    jl_gc_pagemeta_t *old_back = jl_atomic_load_relaxed(&pool->bottom);
+    elt->next = old_back;
+    jl_atomic_store_relaxed(&pool->bottom, elt);
+}
+
+STATIC_INLINE jl_gc_pagemeta_t *pop_lf_back_nosync(jl_gc_page_stack_t *pool) JL_NOTSAFEPOINT
+{
+    jl_gc_pagemeta_t *old_back = jl_atomic_load_relaxed(&pool->bottom);
+    if (old_back == NULL) {
+        return NULL;
+    }
+    jl_atomic_store_relaxed(&pool->bottom, old_back->next);
+    return old_back;
+}
+
 STATIC_INLINE void push_lf_back(jl_gc_page_stack_t *pool, jl_gc_pagemeta_t *elt) JL_NOTSAFEPOINT
 {
     while (1) {
@@ -207,6 +224,23 @@ STATIC_INLINE void push_lf_back(jl_gc_page_stack_t *pool, jl_gc_pagemeta_t *elt)
     }
 }
 
+#define MAX_POP_ATTEMPTS (1 << 10)
+
+STATIC_INLINE jl_gc_pagemeta_t *try_pop_lf_back(jl_gc_page_stack_t *pool) JL_NOTSAFEPOINT
+{
+    for (int i = 0; i < MAX_POP_ATTEMPTS; i++) {
+        jl_gc_pagemeta_t *old_back = jl_atomic_load_relaxed(&pool->bottom);
+        if (old_back == NULL) {
+            return NULL;
+        }
+        if (jl_atomic_cmpswap(&pool->bottom, &old_back, old_back->next)) {
+            return old_back;
+        }
+        jl_cpu_pause();
+    }
+    return NULL;
+}
+
 STATIC_INLINE jl_gc_pagemeta_t *pop_lf_back(jl_gc_page_stack_t *pool) JL_NOTSAFEPOINT
 {
     while (1) {
@@ -220,6 +254,16 @@ STATIC_INLINE jl_gc_pagemeta_t *pop_lf_back(jl_gc_page_stack_t *pool) JL_NOTSAFE
         jl_cpu_pause();
     }
 }
+typedef struct {
+    jl_gc_page_stack_t stack;
+    // pad to 128 bytes to avoid false-sharing
+#ifdef _P64
+    void *_pad[15];
+#else
+    void *_pad[31];
+#endif
+} jl_gc_padded_page_stack_t;
+static_assert(sizeof(jl_gc_padded_page_stack_t) == 128, "jl_gc_padded_page_stack_t is not 128 bytes");
 
 typedef struct {
     _Atomic(size_t) n_freed_objs;
@@ -461,7 +505,7 @@ void gc_mark_finlist(jl_gc_markqueue_t *mq, arraylist_t *list, size_t start) JL_
 void gc_mark_loop_serial_(jl_ptls_t ptls, jl_gc_markqueue_t *mq);
 void gc_mark_loop_serial(jl_ptls_t ptls);
 void gc_mark_loop_parallel(jl_ptls_t ptls, int master);
-void gc_sweep_pool_parallel(void);
+void gc_sweep_pool_parallel(jl_ptls_t ptls);
 void gc_free_pages(void);
 void sweep_stack_pools(void);
 void jl_gc_debug_init(void);

diff --git a/src/partr.c b/src/partr.c
@@ -143,7 +143,7 @@ void jl_parallel_gc_threadfun(void *arg)
             gc_mark_loop_parallel(ptls, 0);
         }
         if (may_sweep(ptls)) { // not an else!
-            gc_sweep_pool_parallel();
+            gc_sweep_pool_parallel(ptls);
             jl_atomic_fetch_add(&ptls->gc_sweeps_requested, -1);
         }
     }