+ introduce a few configs for DATAS fine tuning

+ if the GCGen0MaxBudget config is specified, always honor it including for DATAS

Author: Maoni0

src/coreclr/gc/gc.cpp

@@ -44216,38 +44216,42 @@ void gc_heap::init_static_data()
 {
     size_t gen0_min_size = get_gen0_min_size();
 
-    size_t gen0_max_size =
-#ifdef MULTIPLE_HEAPS
-        max ((size_t)6*1024*1024, min ( Align(soh_segment_size/2), (size_t)200*1024*1024));
-#else //MULTIPLE_HEAPS
-        (
-#ifdef BACKGROUND_GC
-            gc_can_use_concurrent ?
-            6*1024*1024 :
-#endif //BACKGROUND_GC
-            max ((size_t)6*1024*1024,  min ( Align(soh_segment_size/2), (size_t)200*1024*1024))
-        );
-#endif //MULTIPLE_HEAPS
-
-    gen0_max_size = max (gen0_min_size, gen0_max_size);
-
-    if (heap_hard_limit)
-    {
-        size_t gen0_max_size_seg = soh_segment_size / 4;
-        dprintf (GTC_LOG, ("limit gen0 max %zd->%zd", gen0_max_size, gen0_max_size_seg));
-        gen0_max_size = min (gen0_max_size, gen0_max_size_seg);
-    }
+    size_t gen0_max_size = 0;
 
     size_t gen0_max_size_config = (size_t)GCConfig::GetGCGen0MaxBudget();
 
     if (gen0_max_size_config)
     {
-        gen0_max_size = min (gen0_max_size, gen0_max_size_config);
+        gen0_max_size = gen0_max_size_config;
 
 #ifdef FEATURE_EVENT_TRACE
         gen0_max_budget_from_config = gen0_max_size;
 #endif //FEATURE_EVENT_TRACE
     }
+    else
+    {
+        gen0_max_size =
+#ifdef MULTIPLE_HEAPS
+            max ((size_t)6 * 1024 * 1024, min (Align(soh_segment_size / 2), (size_t)200 * 1024 * 1024));
+#else //MULTIPLE_HEAPS
+            (
+#ifdef BACKGROUND_GC
+                gc_can_use_concurrent ?
+                6 * 1024 * 1024 :
+#endif //BACKGROUND_GC
+                max ((size_t)6 * 1024 * 1024, min (Align(soh_segment_size / 2), (size_t)200 * 1024 * 1024))
+                );
+#endif //MULTIPLE_HEAPS
+
+        gen0_max_size = max (gen0_min_size, gen0_max_size);
+
+        if (heap_hard_limit)
+        {
+            size_t gen0_max_size_seg = soh_segment_size / 4;
+            dprintf (GTC_LOG, ("limit gen0 max %zd->%zd", gen0_max_size, gen0_max_size_seg));
+            gen0_max_size = min (gen0_max_size, gen0_max_size_seg);
+        }
+    }
 
     gen0_max_size = Align (gen0_max_size);
     gen0_min_size = min (gen0_min_size, gen0_max_size);
@@ -44284,6 +44288,17 @@ void gc_heap::init_static_data()
         static_data_table[i][0].max_size = gen0_max_size;
         static_data_table[i][1].max_size = gen1_max_size;
     }
+
+#ifdef DYNAMIC_HEAP_COUNT
+    if (gc_heap::dynamic_adaptation_mode == dynamic_adaptation_to_application_sizes)
+    {
+        gc_heap::dynamic_heap_count_data.min_gen0_new_allocation = gen0_min_size;
+        if (gen0_max_size_config)
+        {
+            gc_heap::dynamic_heap_count_data.max_gen0_new_allocation = gen0_max_size;
+        }
+    }
+#endif //DYNAMIC_HEAP_COUNT
 }
 
 bool gc_heap::init_dynamic_data()
@@ -49735,13 +49750,43 @@ HRESULT GCHeap::Initialize()
             }
             // This should be adjusted based on the target tcp. See comments in gcpriv.h
             gc_heap::dynamic_heap_count_data.around_target_threshold = 10.0;
-            // This should really be set as part of computing static data and should take conserve_mem_setting into consideration.
-            gc_heap::dynamic_heap_count_data.max_gen0_new_allocation = Align (min (dd_max_size (gc_heap::g_heaps[0]->dynamic_data_of (0)), (size_t)(64 * 1024 * 1024)), get_alignment_constant (TRUE));
-            gc_heap::dynamic_heap_count_data.min_gen0_new_allocation = Align (dd_min_size (gc_heap::g_heaps[0]->dynamic_data_of (0)), get_alignment_constant (TRUE));
 
-            dprintf (6666, ("datas max gen0 budget %Id, min %Id",
-                gc_heap::dynamic_heap_count_data.max_gen0_new_allocation, gc_heap::dynamic_heap_count_data.min_gen0_new_allocation));
+            int gen0_growth_soh_ratio_percent = (int)GCConfig::GetGCDGen0GrowthPercent();
+            if (gen0_growth_soh_ratio_percent)
+            {
+                gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_percent = (int)GCConfig::GetGCDGen0GrowthPercent() * 0.01f;
+            }
+            // You can specify what sizes you want to allow DATAS to stay within wrt the SOH stable size.
+            // By default DATAS allows 10x this size for gen0 budget when the size is small, and 0.1x when the size is large.
+            int gen0_growth_min_permil = (int)GCConfig::GetGCDGen0GrowthMinFactor();
+            int gen0_growth_max_permil = (int)GCConfig::GetGCDGen0GrowthMaxFactor();
+            if (gen0_growth_min_permil)
+            {
+                gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_min = gen0_growth_min_permil * 0.001f;
+            }
+            if (gen0_growth_max_permil)
+            {
+                gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_max = gen0_growth_max_permil * 0.001f;
+            }
+
+            if (gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_min > gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_max)
+            {
+                log_init_error_to_host ("DATAS min permil for gen0 growth %d is greater than max %d, it needs to be lower",
+                    gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_min, gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_max);
+                return E_FAIL;
+            }
+
+            GCConfig::SetGCDTargetTCP ((int)gc_heap::dynamic_heap_count_data.target_tcp);
+            GCConfig::SetGCDGen0GrowthPercent ((int)(gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_percent * 100.0f));
+            GCConfig::SetGCDGen0GrowthMinFactor ((int)(gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_min * 1000.0f));
+            GCConfig::SetGCDGen0GrowthMaxFactor ((int)(gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_max * 1000.0f));
+            dprintf (6666, ("DATAS gen0 growth multiplier will be adjusted by %d%%, cap %.3f-%.3f, min budget %Id, max %Id",
+                (int)GCConfig::GetGCDGen0GrowthPercent(),
+                gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_min, gc_heap::dynamic_heap_count_data.gen0_growth_soh_ratio_max,
+                gc_heap::dynamic_heap_count_data.min_gen0_new_allocation, gc_heap::dynamic_heap_count_data.max_gen0_new_allocation));
         }
+
+        GCConfig::SetGCDynamicAdaptationMode (gc_heap::dynamic_adaptation_mode);
 #endif //DYNAMIC_HEAP_COUNT
         GCScan::GcRuntimeStructuresValid (TRUE);
 
@@ -52280,13 +52325,11 @@ size_t gc_heap::get_gen0_min_size()
         gen0size = gen0size / 8 * 5;
     }
 
-#ifdef USE_REGIONS
 #ifdef STRESS_REGIONS
     // This is just so we can test allocation using more than one region on machines with very
     // small caches.
     gen0size = ((size_t)1 << min_segment_size_shr) * 3;
 #endif //STRESS_REGIONS
-#endif //USE_REGIONS
 
     gen0size = Align (gen0size);
 
@@ -53883,6 +53926,8 @@ bool gc_heap::compute_memory_settings(bool is_initialization, uint32_t& nhp, uin
     m_high_memory_load_th = min ((high_memory_load_th + 5), v_high_memory_load_th);
     almost_high_memory_load_th = (high_memory_load_th > 5) ? (high_memory_load_th - 5) : 1; // avoid underflow of high_memory_load_th - 5
 
+    GCConfig::SetGCHighMemPercent (high_memory_load_th);
+
     return true;
 }
 

src/coreclr/gc/gcconfig.h

@@ -98,7 +98,7 @@ class GCConfigStringHolder
                                                                                                                                           "prefixed by the CPU group number. Example: Unix - 1,3,5,7-9,12, Windows - 0:1,1:7-9")    \
     INT_CONFIG   (GCHighMemPercent,          "GCHighMemPercent",          "System.GC.HighMemoryPercent",       0,                  "The percent for GC to consider as high memory")                                           \
     INT_CONFIG   (GCProvModeStress,          "GCProvModeStress",          NULL,                                0,                  "Stress the provisional modes")                                                           \
-    INT_CONFIG   (GCGen0MaxBudget,           "GCGen0MaxBudget",           NULL,                                0,                  "Specifies the largest gen0 allocation budget")                                           \
+    INT_CONFIG   (GCGen0MaxBudget,           "GCGen0MaxBudget",           "System.GC.Gen0MaxBudget",         0,                  "Specifies the largest gen0 allocation budget")                                           \
     INT_CONFIG   (GCGen1MaxBudget,           "GCGen1MaxBudget",           NULL,                                0,                  "Specifies the largest gen1 allocation budget")                                           \
     INT_CONFIG   (GCLowSkipRatio,            "GCLowSkipRatio",            NULL,                                30,                 "Specifies the low generation skip ratio")                                                \
     INT_CONFIG   (GCHeapHardLimit,           "GCHeapHardLimit",           "System.GC.HeapHardLimit",           0,                  "Specifies a hard limit for the GC heap")                                                 \
@@ -144,6 +144,9 @@ class GCConfigStringHolder
     INT_CONFIG   (GCDynamicAdaptationMode,   "GCDynamicAdaptationMode",   "System.GC.DynamicAdaptationMode",   1,                  "Enable the GC to dynamically adapt to application sizes.")                               \
     INT_CONFIG   (GCDTargetTCP,              "GCDTargetTCP",              "System.GC.DTargetTCP",              0,                  "Specifies the target tcp for DATAS")                                                     \
     INT_CONFIG   (GCDBGCRatio,               "GCDBGCRatio",               NULL,                                0,                  "Specifies the ratio of BGC to NGC2 for HC change")                                       \
+    INT_CONFIG   (GCDGen0GrowthPercent,      "GCDGen0GrowthPercent",      "System.GC.DGen0GrowthPercent",      0,                  "Specifies the percentage of the default growth factor")                                  \
+    INT_CONFIG   (GCDGen0GrowthMinFactor,    "GCDGen0GrowthMinFactor",    "System.GC.DGen0GrowthMinFactor",    0,                  "Specifies the minimum growth factor in permil")                                          \
+    INT_CONFIG   (GCDGen0GrowthMaxFactor,    "GCDGen0GrowthMaxFactor",    "System.GC.DGen0GrowthMaxFactor",    0,                  "Specifies the maximum growth factor in permil")                                          \
     BOOL_CONFIG  (GCCacheSizeFromSysConf,    "GCCacheSizeFromSysConf",    NULL,                                false,              "Specifies using sysconf to retrieve the last level cache size for Unix.")
 
 // This class is responsible for retreiving configuration information

src/coreclr/gc/gcpriv.h

@@ -4388,6 +4388,18 @@ class gc_heap
         float target_tcp = 2.0;
         float target_gen2_tcp = 10.0;
 
+        // The following 3 constants are used in the computation for the total gen0 budget relative to the stable soh size.
+        // 
+        // By default DATAS computes a multiplier (gen0_growth_soh_ratio) that scales the size. This multiplier follows
+        // a power decay curve where the multiplier decreases rapidly as the size increases. We cap it at 10x at the low
+        // end and 10% at the high end.
+        //
+        // You can choose to modify these by specifying gen0_growth_factor_percent to reduce or increase this multiplier
+        // and the min/max multipliers.
+        float gen0_growth_soh_ratio_percent = 1.0;
+        float gen0_growth_soh_ratio_min = 0.1f;
+        float gen0_growth_soh_ratio_max = 10.0;
+
         static const int recorded_adjustment_size = 4;
         static const int sample_size = 3;
         static const int recorded_tcp_array_size = 64;
@@ -5069,23 +5081,23 @@ class gc_heap
         // time in msl).
         //
 
-        size_t          max_gen0_new_allocation;
-        size_t          min_gen0_new_allocation;
+        size_t max_gen0_new_allocation = 64 * 1024 * 1024;
+        size_t min_gen0_new_allocation = 0;
 
         size_t compute_total_gen0_budget (size_t total_soh_stable_size)
         {
             assert (total_soh_stable_size > 0);
 
-            float factor = (float)(20 - conserve_mem_setting);
-            double old_gen_growth_factor = factor / sqrt ((double)total_soh_stable_size / 1000.0 / 1000.0);
-            double saved_old_gen_growth_factor = old_gen_growth_factor;
-            old_gen_growth_factor = min (10.0, old_gen_growth_factor);
-            old_gen_growth_factor = max (0.1, old_gen_growth_factor);
+            float factor = (float)(20 - conserve_mem_setting) * gen0_growth_soh_ratio_percent;
+            double gen0_growth_soh_ratio = factor / sqrt ((double)total_soh_stable_size / 1000.0 / 1000.0);
+            double saved_gen0_growth_soh_ratio = gen0_growth_soh_ratio;
+            gen0_growth_soh_ratio = min ((double)gen0_growth_soh_ratio_max, gen0_growth_soh_ratio);
+            gen0_growth_soh_ratio = max ((double)gen0_growth_soh_ratio_min, gen0_growth_soh_ratio);
 
-            size_t total_new_allocation_old_gen = (size_t)(old_gen_growth_factor * (double)total_soh_stable_size);
+            size_t total_new_allocation_old_gen = (size_t)(gen0_growth_soh_ratio * (double)total_soh_stable_size);
             dprintf (6666, ("stable soh %Id (%.3fmb), factor %.3f=>%.3f -> total gen0 new_alloc %Id (%.3fmb)",
                 total_soh_stable_size, ((double)total_soh_stable_size / 1000.0 / 1000.0),
-                saved_old_gen_growth_factor, old_gen_growth_factor, total_new_allocation_old_gen,
+                saved_gen0_growth_soh_ratio, gen0_growth_soh_ratio, total_new_allocation_old_gen,
                 ((double)total_new_allocation_old_gen  / 1000.0 / 1000.0)));
             return total_new_allocation_old_gen;
         }