//extract the high bits [high, 32] of a uint32_t
#define highbits(wrd, bits) ((wrd) & ~((1 << (bits))-1))
+// Things we need to manually initialize:
+// gen0 min_size - based on cache
+// gen0/1 max_size - based on segment size
+static static_data static_data_table[latency_level_last - latency_level_first + 1][NUMBERGENERATIONS] =
+{
+ // latency_level_memory_footprint
+ {
+ // gen0
+ {0, 0, 40000, 0.5f, 9.0f, 20.0f, 1000, 1},
+ // gen1
+ {163840, 0, 80000, 0.5f, 2.0f, 7.0f, 10000, 10},
+ // gen2
+ {256*1024, SSIZE_T_MAX, 200000, 0.25f, 1.2f, 1.8f, 100000, 100},
+ // gen3
+ {3*1024*1024, SSIZE_T_MAX, 0, 0.0f, 1.25f, 4.5f, 0, 0}
+ },
+
+ // latency_level_balanced
+ {
+ // gen0
+ {0, 0, 40000, 0.5f,
+#ifdef MULTIPLE_HEAPS
+ 20.0f, 40.0f,
+#else
+ 9.0f, 20.0f,
+#endif //MULTIPLE_HEAPS
+ 1000, 1},
+ // gen1
+ {9*32*1024, 0, 80000, 0.5f, 2.0f, 7.0f, 10000, 10},
+ // gen2
+ {256*1024, SSIZE_T_MAX, 200000, 0.25f, 1.2f, 1.8f, 100000, 100},
+ // gen3
+ {3*1024*1024, SSIZE_T_MAX, 0, 0.0f, 1.25f, 4.5f, 0, 0}
+ },
+};
class mark;
class generation;
static gc_reason gc_trigger_reason = reason_empty;
#endif //DACCESS_COMPILE
+gc_latency_level gc_heap::latency_level = latency_level_default;
+
gc_mechanisms gc_heap::settings;
gc_history_global gc_heap::gc_data_global;
dprintf (3, ("evaluating allocation rate"));
dynamic_data* dd0 = dynamic_data_of (0);
if ((allocation_running_amount - dd_new_allocation (dd0)) >
- dd_min_gc_size (dd0))
+ dd_min_size (dd0))
{
uint32_t ctime = GCToOSInterface::GetLowPrecisionTimeStamp();
if ((ctime - allocation_running_time) > 1000)
settings.first_init();
+ int latency_level_from_config = static_cast<int>(GCConfig::GetLatencyLevel());
+ if (latency_level_from_config >= latency_level_first && latency_level_from_config <= latency_level_last)
+ {
+ gc_heap::latency_level = static_cast<gc_latency_level>(latency_level_from_config);
+ }
+
+ init_static_data();
+
g_gc_card_table = make_card_table (g_gc_lowest_address, g_gc_highest_address);
if (!g_gc_card_table)
//needs to be done after the dynamic data has been initialized
#ifndef MULTIPLE_HEAPS
- allocation_running_amount = dd_min_gc_size (dynamic_data_of (0));
+ allocation_running_amount = dd_min_size (dynamic_data_of (0));
#endif //!MULTIPLE_HEAPS
- fgn_last_alloc = dd_min_gc_size (dynamic_data_of (0));
+ fgn_last_alloc = dd_min_size (dynamic_data_of (0));
mark* arr = new (nothrow) (mark [MARK_STACK_INITIAL_LENGTH]);
if (!arr)
background_soh_alloc_count = 0;
background_loh_alloc_count = 0;
bgc_overflow_count = 0;
- end_loh_size = dd_min_gc_size (dynamic_data_of (max_generation + 1));
+ end_loh_size = dd_min_size (dynamic_data_of (max_generation + 1));
#endif //BACKGROUND_GC
#ifdef GC_CONFIG_DRIVEN
if (reason == oom_budget)
{
- alloc_size = dd_min_gc_size (dynamic_data_of (0)) / 2;
+ alloc_size = dd_min_size (dynamic_data_of (0)) / 2;
}
if ((reason == oom_budget) && ((!fgm_result.loh_p) && (fgm_result.fgm != fgm_no_failure)))
BOOL gc_heap::bgc_loh_should_allocate()
{
- size_t min_gc_size = dd_min_gc_size(dynamic_data_of (max_generation + 1));
+ size_t min_gc_size = dd_min_size(dynamic_data_of (max_generation + 1));
if ((bgc_begin_loh_size + bgc_loh_size_increased) < (min_gc_size * 10))
{
for (i = (temp_gen+1); i <= n_time_max; i++)
{
dynamic_data* dd = dynamic_data_of (i);
- if ((now > dd_time_clock(dd) + power (10, i)*1000) &&
- (dd_gc_clock (dd0) > (dd_gc_clock (dd) + (power (10, i)))) &&
+ if ((now > dd_time_clock(dd) + dd_time_clock_interval(dd)) &&
+ (dd_gc_clock (dd0) > (dd_gc_clock (dd) + dd_gc_clock_interval(dd))) &&
((n < max_generation) || ((dd_current_size (dd) < dd_max_size (dd0)))))
{
n = min (i, n_time_max);
// down to min_gc_size to stay in the cache
gc_heap* hp = gc_heap::g_heaps[0];
dynamic_data* dd = hp->dynamic_data_of (gen);
- size_t min_gc_size = dd_min_gc_size(dd);
+ size_t min_gc_size = dd_min_size(dd);
// if min GC size larger than true on die cache, then don't bother
// limiting the desired size
if ((min_gc_size <= GCToOSInterface::GetCacheSizePerLogicalCpu(TRUE)) &&
size_t m = 0;
for (int n = 0; n <= condemned_gen_number;n++)
{
- m += (size_t)(dd_min_gc_size (dynamic_data_of (n))*(n+1)*0.1);
+ m += (size_t)(dd_min_size (dynamic_data_of (n))*(n+1)*0.1);
}
for (int i = 0; i < n_heaps; i++)
size_t m = 0;
for (int n = 0; n <= condemned_gen_number;n++)
{
- m += (size_t)(dd_min_gc_size (dynamic_data_of (n))*(n+1)*0.06);
+ m += (size_t)(dd_min_size (dynamic_data_of (n))*(n+1)*0.06);
}
dynamic_data* dd = dynamic_data_of (min (condemned_gen_number +1,
max_generation));
return consing_gen;
}
+void gc_heap::set_static_data()
+{
+ static_data* pause_mode_sdata = static_data_table[latency_level];
+ for (int i = 0; i < NUMBERGENERATIONS; i++)
+ {
+ dynamic_data* dd = dynamic_data_of (i);
+ static_data* sdata = &pause_mode_sdata[i];
+
+ dd->sdata = sdata;
+ dd->min_size = sdata->min_size;
+
+ dprintf (GTC_LOG, ("PM: %d - min: %Id, max: %Id, fr_l: %Id, fr_b: %d%%",
+ settings.pause_mode,
+ dd->min_size, dd_max_size,
+ dd->fragmentation_limit, (int)(dd->fragmentation_burden_limit * 100)));
+ }
+}
+
+// Initialize the values that are not const.
+void gc_heap::init_static_data()
+{
+ size_t gen0size = GCHeap::GetValidGen0MaxSize(get_valid_segment_size());
+ size_t gen0_min_size = Align(gen0size / 8 * 5);
+
+ size_t gen0_max_size =
+#ifdef MULTIPLE_HEAPS
+ max (6*1024*1024, min ( Align(soh_segment_size/2), 200*1024*1024));
+#else //MULTIPLE_HEAPS
+ (gc_can_use_concurrent ?
+ 6*1024*1024 :
+ max (6*1024*1024, min ( Align(soh_segment_size/2), 200*1024*1024)));
+#endif //MULTIPLE_HEAPS
+
+ // TODO: gen0_max_size has a 200mb cap; gen1_max_size should also have a cap.
+ size_t gen1_max_size =
+#ifdef MULTIPLE_HEAPS
+ max (6*1024*1024, Align(soh_segment_size/2));
+#else //MULTIPLE_HEAPS
+ (gc_can_use_concurrent ?
+ 6*1024*1024 :
+ max (6*1024*1024, Align(soh_segment_size/2)));
+#endif //MULTIPLE_HEAPS
+
+ dprintf (GTC_LOG, ("gen0size: %Id, gen0 min: %Id, max: %Id, gen1 max: %Id",
+ gen0size, gen0_min_size, gen0_max_size, gen1_max_size));
+
+ for (int i = latency_level_first; i <= latency_level_last; i++)
+ {
+ static_data_table[i][0].min_size = gen0_min_size;
+ static_data_table[i][0].max_size = gen0_max_size;
+ static_data_table[i][1].max_size = gen1_max_size;
+ }
+}
+
bool gc_heap::init_dynamic_data()
{
qpf = GCToOSInterface::QueryPerformanceFrequency();
uint32_t now = (uint32_t)GetHighPrecisionTimeStamp();
- //clear some fields
- for (int i = 0; i < max_generation+1; i++)
+ set_static_data();
+
+ for (int i = 0; i <= max_generation+1; i++)
{
dynamic_data* dd = dynamic_data_of (i);
dd->gc_clock = 0;
dd->time_clock = now;
+ dd->current_size = 0;
+ dd->promoted_size = 0;
+ dd->collection_count = 0;
+ dd->new_allocation = dd->min_size;
+ dd->gc_new_allocation = dd->new_allocation;
+ dd->desired_allocation = dd->new_allocation;
+ dd->fragmentation = 0;
}
#ifdef GC_CONFIG_DRIVEN
time_init = now;
#endif //GC_CONFIG_DRIVEN
- // get the registry setting for generation 0 size
- size_t gen0size = GCHeap::GetValidGen0MaxSize(soh_segment_size);
-
- dprintf (2, ("gen 0 size: %Id", gen0size));
-
- dynamic_data* dd = dynamic_data_of (0);
- dd->current_size = 0;
- dd->promoted_size = 0;
- dd->collection_count = 0;
-// dd->limit = 3.0f;
-#ifdef MULTIPLE_HEAPS
- dd->limit = 20.0f; // be more aggressive on server gc
- dd->max_limit = 40.0f;
-#else
- dd->limit = 9.0f;
-// dd->max_limit = 15.0f; //10.0f;
- dd->max_limit = 20.0f;
-#endif //MULTIPLE_HEAPS
- dd->min_gc_size = Align(gen0size / 8 * 5);
- dd->min_size = dd->min_gc_size;
- //dd->max_size = Align (gen0size);
-
-#ifdef BACKGROUND_GC
- //gc_can_use_concurrent is not necessarily 0 for server builds
- bool can_use_concurrent = gc_can_use_concurrent;
-#else // !BACKGROUND_GC
- bool can_use_concurrent = false;
-#endif // BACKGROUND_GC
-
-#ifdef MULTIPLE_HEAPS
- dd->max_size = max (6*1024*1024, min ( Align(soh_segment_size/2), 200*1024*1024));
-#else //MULTIPLE_HEAPS
- dd->max_size = (can_use_concurrent ?
- 6*1024*1024 :
- max (6*1024*1024, min ( Align(soh_segment_size/2), 200*1024*1024)));
-#endif //MULTIPLE_HEAPS
- dd->new_allocation = dd->min_gc_size;
- dd->gc_new_allocation = dd->new_allocation;
- dd->desired_allocation = dd->new_allocation;
- dd->default_new_allocation = dd->min_gc_size;
- dd->fragmentation = 0;
- dd->fragmentation_limit = 40000;
- dd->fragmentation_burden_limit = 0.5f;
-
- dd = dynamic_data_of (1);
- dd->current_size = 0;
- dd->promoted_size = 0;
- dd->collection_count = 0;
- dd->limit = 2.0f;
-// dd->max_limit = 15.0f;
- dd->max_limit = 7.0f;
- dd->min_gc_size = 9*32*1024;
- dd->min_size = dd->min_gc_size;
-// dd->max_size = 2397152;
-#ifdef MULTIPLE_HEAPS
- dd->max_size = max (6*1024*1024, Align(soh_segment_size/2));
-#else //MULTIPLE_HEAPS
- dd->max_size = (can_use_concurrent ?
- 6*1024*1024 :
- max (6*1024*1024, Align(soh_segment_size/2)));
-#endif //MULTIPLE_HEAPS
- dd->new_allocation = dd->min_gc_size;
- dd->gc_new_allocation = dd->new_allocation;
- dd->desired_allocation = dd->new_allocation;
- dd->default_new_allocation = dd->min_gc_size;
- dd->fragmentation = 0;
- dd->fragmentation_limit = 80000;
- dd->fragmentation_burden_limit = 0.5f;
-
- dd = dynamic_data_of (2);
- dd->current_size = 0;
- dd->promoted_size = 0;
- dd->collection_count = 0;
- dd->limit = 1.2f;
- dd->max_limit = 1.8f;
- dd->min_gc_size = 256*1024;
- dd->min_size = dd->min_gc_size;
- dd->max_size = SSIZE_T_MAX;
- dd->new_allocation = dd->min_gc_size;
- dd->gc_new_allocation = dd->new_allocation;
- dd->desired_allocation = dd->new_allocation;
- dd->default_new_allocation = dd->min_gc_size;
- dd->fragmentation = 0;
- dd->fragmentation_limit = 200000;
- dd->fragmentation_burden_limit = 0.25f;
-
- //dynamic data for large objects
- dd = dynamic_data_of (3);
- dd->current_size = 0;
- dd->promoted_size = 0;
- dd->collection_count = 0;
- dd->limit = 1.25f;
- dd->max_limit = 4.5f;
- dd->min_gc_size = 3*1024*1024;
- dd->min_size = dd->min_gc_size;
- dd->max_size = SSIZE_T_MAX;
- dd->new_allocation = dd->min_gc_size;
- dd->gc_new_allocation = dd->new_allocation;
- dd->desired_allocation = dd->new_allocation;
- dd->default_new_allocation = dd->min_gc_size;
- dd->fragmentation = 0;
- dd->fragmentation_limit = 0;
- dd->fragmentation_burden_limit = 0.0f;
-
return true;
}
if (dd_begin_data_size (dd) == 0)
{
- size_t new_allocation = dd_default_new_allocation (dd);
+ size_t new_allocation = dd_min_size (dd);
current_gc_data_per_heap->gen_data[gen_number].new_allocation = new_allocation;
return new_allocation;
}
size_t current_size = dd_current_size (dd);
float max_limit = dd_max_limit (dd);
float limit = dd_limit (dd);
- size_t min_gc_size = dd_min_gc_size (dd);
+ size_t min_gc_size = dd_min_size (dd);
float f = 0;
size_t max_size = dd_max_size (dd);
size_t new_allocation = 0;
dynamic_data* dd = dynamic_data_of (0);
size_t current = dd_desired_allocation (dd);
- size_t candidate = max (Align ((committed_mem / 10), get_alignment_constant(FALSE)), dd_min_gc_size (dd));
+ size_t candidate = max (Align ((committed_mem / 10), get_alignment_constant(FALSE)), dd_min_size (dd));
dd_desired_allocation (dd) = min (current, candidate);
}
// a reasonable amount of allocation requests.
size_t gc_heap::end_space_after_gc()
{
- return max ((dd_min_gc_size (dynamic_data_of (0))/2), (END_SPACE_AFTER_GC + Align (min_obj_size)));
+ return max ((dd_min_size (dynamic_data_of (0))/2), (END_SPACE_AFTER_GC + Align (min_obj_size)));
}
BOOL gc_heap::ephemeral_gen_fit_p (gc_tuning_point tp)
gc_heap::g_low_memory_status = (reason == reason_lowmemory) ||
(reason == reason_lowmemory_blocking) ||
- g_bLowMemoryFromHost;
-
- if (g_bLowMemoryFromHost)
- reason = reason_lowmemory_host;
+ (gc_heap::latency_level == latency_level_memory_footprint);
gc_trigger_reason = reason;
set_pause_mode_no_gc = 1 // NoGCRegion is in progress, can't change pause mode.
};
+/*
+ Latency modes required user to have specific GC knowledge (eg, budget, full blocking GC).
+ We are trying to move away from them as it makes a lot more sense for users to tell
+ us what's the most important out of the perf aspects that make sense to them.
+
+ In general there are 3 such aspects:
+
+ + memory footprint
+ + throughput
+ + pause predictibility
+
+ Currently the following levels are supported. We may (and will likely) add more
+ in the future.
+
+ +----------+--------------------+---------------------------------------+
+ | Level | Optimization Goals | Latency Charactaristics |
+ +==========+====================+=======================================+
+ | 0 | memory footprint | pauses can be long and more frequent |
+ +----------+--------------------+---------------------------------------+
+ | 1 | balanced | pauses are more predictable and more |
+ | | | frequent. the longest pauses are |
+ | | | shorter than 1. |
+ +----------+--------------------+---------------------------------------+
+*/
+enum gc_latency_level
+{
+ latency_level_first = 0,
+ latency_level_memory_footprint = latency_level_first,
+ latency_level_balanced = 1,
+ latency_level_last = latency_level_balanced,
+ latency_level_default = latency_level_balanced
+};
+
enum gc_tuning_point
{
tuning_deciding_condemned_gen,
static_assert(offsetof(dac_generation, start_segment) == offsetof(generation, start_segment), "DAC generation offset mismatch");
static_assert(offsetof(dac_generation, allocation_start) == offsetof(generation, allocation_start), "DAC generation offset mismatch");
+// static data remains the same after it's initialized.
+// It's per generation.
+// TODO: for gen_time_tuning, we should put the multipliers in static data.
+struct static_data
+{
+ size_t min_size;
+ size_t max_size;
+ size_t fragmentation_limit;
+ float fragmentation_burden_limit;
+ float limit;
+ float max_limit;
+ size_t time_clock; // time after which to collect generation, in performance counts (see QueryPerformanceCounter)
+ size_t gc_clock; // nubmer of gcs after which to collect generation
+};
// The dynamic data fields are grouped into 3 categories:
//
// calculated logical data (like desired_allocation)
// physical data (like fragmentation)
-// const data (like min_gc_size), initialized at the beginning
+// const data (sdata), initialized at the beginning
class dynamic_data
{
public:
size_t gc_elapsed_time; // Time it took for the gc to complete
float gc_speed; // speed in bytes/msec for the gc to complete
- // min_size is always the same as min_gc_size..
- size_t min_gc_size;
- size_t max_size;
size_t min_size;
- size_t default_new_allocation;
- size_t fragmentation_limit;
- float fragmentation_burden_limit;
- float limit;
- float max_limit;
+
+ static_data* sdata;
};
#define ro_in_entry 0x1
PER_HEAP
void save_ephemeral_generation_starts();
+ PER_HEAP
+ void set_static_data();
+
+ PER_HEAP_ISOLATED
+ void init_static_data();
+
PER_HEAP
bool init_dynamic_data ();
PER_HEAP
#endif //MULTIPLE_HEAPS
+ PER_HEAP_ISOLATED
+ gc_latency_level latency_level;
+
PER_HEAP_ISOLATED
gc_mechanisms settings;
inline
float& dd_limit (dynamic_data* inst)
{
- return inst->limit;
+ return inst->sdata->limit;
}
inline
float& dd_max_limit (dynamic_data* inst)
{
- return inst->max_limit;
-}
-inline
-size_t& dd_min_gc_size (dynamic_data* inst)
-{
- return inst->min_gc_size;
+ return inst->sdata->max_limit;
}
inline
size_t& dd_max_size (dynamic_data* inst)
{
- return inst->max_size;
+ return inst->sdata->max_size;
}
inline
size_t& dd_min_size (dynamic_data* inst)
return inst->gc_new_allocation;
}
inline
-size_t& dd_default_new_allocation (dynamic_data* inst)
-{
- return inst->default_new_allocation;
-}
-inline
size_t& dd_fragmentation_limit (dynamic_data* inst)
{
- return inst->fragmentation_limit;
+ return inst->sdata->fragmentation_limit;
}
inline
float& dd_fragmentation_burden_limit (dynamic_data* inst)
{
- return inst->fragmentation_burden_limit;
+ return inst->sdata->fragmentation_burden_limit;
}
inline
float dd_v_fragmentation_burden_limit (dynamic_data* inst)
return inst->time_clock;
}
+inline
+size_t& dd_gc_clock_interval (dynamic_data* inst)
+{
+ return inst->sdata->gc_clock;
+}
+inline
+size_t& dd_time_clock_interval (dynamic_data* inst)
+{
+ return inst->sdata->time_clock;
+}
+
inline
size_t& dd_gc_elapsed_time (dynamic_data* inst)
{
projects / platform / upstream / coreclr.git / commitdiff