add some instrumentation to measure page utilization per size class (#…

…52164)

One of the limitations is that it's only accurate right after the GC.
Still might be helpful for observability purposes.

src/gc.c

@@ -1583,6 +1583,38 @@ int jl_gc_classify_pools(size_t sz, int *osize)
 
 // sweep phase
 
+gc_fragmentation_stat_t gc_page_fragmentation_stats[JL_GC_N_POOLS];
+
+extern gc_fragmentation_stat_t gc_page_fragmentation_stats[JL_GC_N_POOLS];
+
+STATIC_INLINE void gc_update_page_fragmentation_data(jl_gc_pagemeta_t *pg) JL_NOTSAFEPOINT
+{
+#ifdef GC_MEASURE_PAGE_FRAGMENTATION
+    gc_fragmentation_stat_t *stats = &gc_page_fragmentation_stats[pg->pool_n];
+    jl_atomic_fetch_add(&stats->n_freed_objs, pg->nfree);
+    jl_atomic_fetch_add(&stats->n_pages_allocd, 1);
+#endif
+}
+
+STATIC_INLINE void gc_dump_page_utilization_data(void) JL_NOTSAFEPOINT
+{
+#ifdef GC_MEASURE_PAGE_FRAGMENTATION
+    for (int i = 0; i < JL_GC_N_POOLS; i++) {
+        gc_fragmentation_stat_t *stats = &gc_page_fragmentation_stats[i];
+        double utilization = 1.0;
+        size_t n_freed_objs = jl_atomic_load_relaxed(&stats->n_freed_objs);
+        size_t n_pages_allocd = jl_atomic_load_relaxed(&stats->n_pages_allocd);
+        if (n_pages_allocd != 0) {
+            utilization -= ((double)n_freed_objs * (double)jl_gc_sizeclasses[i]) / (double)n_pages_allocd / (double)GC_PAGE_SZ;
+        }
+        jl_safe_printf("Size class %d: %.2f%% utilization\n", jl_gc_sizeclasses[i], utilization * 100.0);
+        jl_atomic_store_relaxed(&stats->n_freed_objs, 0);
+        jl_atomic_store_relaxed(&stats->n_pages_allocd, 0);
+    }
+    jl_safe_printf("-----------------------------------------\n");
+#endif
+}
+
 int64_t lazy_freed_pages = 0;
 
 // Returns pointer to terminal pointer of list rooted at *pfl.
@@ -1697,6 +1729,7 @@ static jl_taggedvalue_t **sweep_page(jl_gc_pool_t *p, jl_gc_pagemeta_t *pg, jl_t
     nfree = pg->nfree;
 
 done:
+    gc_update_page_fragmentation_data(pg);
     gc_time_count_page(freedall, pg_skpd);
     jl_ptls_t ptls = gc_all_tls_states[pg->thread_n];
     jl_atomic_fetch_add(&ptls->gc_num.pool_live_bytes, GC_PAGE_SZ - GC_PAGE_OFFSET - nfree * osize);
@@ -1876,6 +1909,7 @@ static void gc_sweep_pool(int sweep_full)
         }
     }
 
+    gc_dump_page_utilization_data();
     gc_time_pool_end(sweep_full);
 }
 

src/gc.h

@@ -331,9 +331,14 @@ typedef struct {
 //  structure of this representation allows us to partially unroll and optimize
 //  various conditions at each level.
 
-//  The following constants define the branching factors at each level.
-//  The constants and GC_PAGE_LG2 must therefore sum to sizeof(void*).
-//  They should all be multiples of 32 (sizeof(uint32_t)) except that REGION2_PG_COUNT may also be 1.
+// data structures for tracking fragmentation in the pool allocator
+// #define GC_MEASURE_PAGE_FRAGMENTATION
+
+typedef struct {
+    _Atomic(size_t) n_freed_objs;
+    _Atomic(size_t) n_pages_allocd;
+} gc_fragmentation_stat_t;
+
 #ifdef _P64
 #define REGION0_PG_COUNT (1 << 16)
 #define REGION1_PG_COUNT (1 << 16)