Skip to content

Commit

Permalink
DATAS for small HCs (dotnet#100390)
Browse files Browse the repository at this point in the history 
I adjusted the formula for determining a new HC and change how we calculate the gen0 budget based on gen2 size.

changes included -

+ currently we have a very simplistic formula for actually adapting to the size and this basically just makes all the asp.net benchmarks with low surv rate adjust to the min 2.5 mb gen0 budget, while those run ok with such a small budget on a 28 core machine, it doesn't work if we limit the heap count to a small number, eg, 4. what happens is the % time in GC is very high, some benchmarks run with 20% to 40% time in GC. this is obviously not desirable. I reworked this to make it actually adapting to the size. and we'll take the min of this and what we calculated without DATAS.

+ the formula I had previously did not handle small HCs well so I also adjust that.

+ got rid of the adjusting to cache size in gc1 for DATAS, this just makes things unpredictable especially for small workloads.
  • Loading branch information
@Maoni0
Maoni0 committed Apr 3, 2024
1 parent 7c030e8 commit 89bd910
Show file tree
Hide file tree
Showing 3 changed files with 151 additions and 60 deletions.
151 changes: 92 additions & 59 deletions src/coreclr/gc/gc.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22526,6 +22526,15 @@ void gc_heap::gc1()
{
limit = total_generation_count-1;
}

size_t total_max_gen_size = 0;
for (int i = 0; i < gc_heap::n_heaps; i++)
{
gc_heap* hp = gc_heap::g_heaps[i];
dynamic_data* dd = hp->dynamic_data_of (max_generation);
total_max_gen_size += dd_current_size (dd) + dd_desired_allocation (dd);
}

for (int gen = 0; gen <= limit; gen++)
{
size_t total_desired = 0;
Expand Down Expand Up @@ -22554,20 +22563,35 @@ void gc_heap::gc1()
total_already_consumed = temp_total_already_consumed;
}

size_t desired_per_heap = Align (total_desired/gc_heap::n_heaps,
get_alignment_constant (gen <= max_generation));
size_t desired_per_heap = Align (total_desired/gc_heap::n_heaps, get_alignment_constant (gen <= max_generation));

size_t already_consumed_per_heap = total_already_consumed / gc_heap::n_heaps;

if (gen == 0)
{
#if 1 //subsumed by the linear allocation model
#ifdef DYNAMIC_HEAP_COUNT
if (dynamic_adaptation_mode == dynamic_adaptation_to_application_sizes)
{
size_t new_allocation_datas = dynamic_heap_count_data.compute_gen0_new_allocation (total_max_gen_size);
new_allocation_datas = Align (new_allocation_datas, get_alignment_constant (gen <= max_generation));
dprintf (6666, ("gen0 new_alloc %Id (%.3fmb), from datas: %Id (%.3fmb)",
desired_per_heap, ((double)desired_per_heap / 1000.0 / 1000.0),
new_allocation_datas, ((double)new_allocation_datas / 1000.0 / 1000.0)));
desired_per_heap = min (desired_per_heap, new_allocation_datas);
}
#endif //DYNAMIC_HEAP_COUNT

// to avoid spikes in mem usage due to short terms fluctuations in survivorship,
// apply some smoothing.
size_t desired_per_heap_before_smoothing = desired_per_heap;
desired_per_heap = exponential_smoothing (gen, dd_collection_count (dynamic_data_of(gen)), desired_per_heap);
#endif //0
size_t desired_per_heap_after_smoothing = desired_per_heap;

if (!heap_hard_limit)
if (!heap_hard_limit
#ifdef DYNAMIC_HEAP_COUNT
&& (dynamic_adaptation_mode != dynamic_adaptation_to_application_sizes)
#endif //DYNAMIC_HEAP_COUNT
)
{
// if desired_per_heap is close to min_gc_size, trim it
// down to min_gc_size to stay in the cache
Expand All @@ -22584,7 +22608,10 @@ void gc_heap::gc1()
}
#ifdef HOST_64BIT
desired_per_heap = joined_youngest_desired (desired_per_heap);
dprintf (2, ("final gen0 new_alloc: %zd", desired_per_heap));

dprintf (6666, ("final gen0 new_alloc: total desired: %Id (%.3fmb/heap), before smooth %zd -> after smooth %zd -> after joined %zd",
total_desired, ((double)(total_desired / n_heaps)/ 1000.0 / 1000.0),
desired_per_heap_before_smoothing, desired_per_heap_after_smoothing, desired_per_heap));
#endif // HOST_64BIT
gc_data_global.final_youngest_desired = desired_per_heap;
}
Expand Down Expand Up @@ -25347,9 +25374,10 @@ int gc_heap::calculate_new_heap_count ()
// on the way up, we essentially multiply the heap count by 1.5, so we go 1, 2, 3, 5, 8 ...
// we don't go all the way to the number of CPUs, but stay 1 or 2 short
int step_up = (n_heaps + 1) / 2;
int extra_heaps = 1 + (n_max_heaps >= 32);
int extra_heaps = (n_max_heaps >= 16) + (n_max_heaps >= 64);
int actual_n_max_heaps = n_max_heaps - extra_heaps;
int max_growth = max ((n_max_heaps / 4), 2);
int max_growth = max ((n_max_heaps / 4), (1 + (actual_n_max_heaps > 3)));

step_up = min (step_up, (actual_n_max_heaps - n_heaps));

// on the way down, we essentially divide the heap count by 1.5
Expand Down Expand Up @@ -25392,13 +25420,15 @@ int gc_heap::calculate_new_heap_count ()
// target_tcp should be configurable.
float target_tcp = 5.0;
float target_gen2_tcp = 10.0;
float log_base = (float)1.1;
float log_base = (float)1.11;

dynamic_heap_count_data.add_to_recorded_tcp (median_throughput_cost_percent);

// This is the average of whatever is in the recorded tcp buffer.
float avg_recorded_tcp = 0.0;

size_t num_gcs_since_last_change = current_gc_index - dynamic_heap_count_data.last_changed_gc_index;

if (process_eph_samples_p)
{
dynamic_heap_count_data.last_processed_stcp = smoothed_median_throughput_cost_percent;
Expand All @@ -25407,22 +25437,21 @@ int gc_heap::calculate_new_heap_count ()
{
// If median is high but stcp is lower than target, and if this situation continues, stcp will quickly be above target anyway; otherwise
// we treat it as an outlier.
if (smoothed_median_throughput_cost_percent > target_tcp)
if (smoothed_median_throughput_cost_percent >= (target_tcp + 1.0))
{
float step_up_percent = log_with_base ((smoothed_median_throughput_cost_percent - target_tcp + log_base), log_base);
float step_up_float = (float)(step_up_percent / 100.0 * actual_n_max_heaps);
float step_up_float = (float)(1 + actual_n_max_heaps * log_with_base ((smoothed_median_throughput_cost_percent - target_tcp), log_base) / 100.0);
int step_up_int = (int)step_up_float;

dprintf (6666, ("[CHP0] inc %d(%.3f), last inc %d, %Id GCs elapsed, last stcp %.3f",
step_up_int, step_up_float, (int)dynamic_heap_count_data.last_changed_count,
(current_gc_index - dynamic_heap_count_data.last_changed_gc_index), dynamic_heap_count_data.last_changed_stcp));
num_gcs_since_last_change, dynamic_heap_count_data.last_changed_stcp));

// Don't adjust if we just adjusted last time we checked, unless we are in an extreme situation.
if ((smoothed_median_throughput_cost_percent < 20.0f) &&
(avg_throughput_cost_percent < 20.0f) &&
((current_gc_index - dynamic_heap_count_data.last_changed_gc_index) < (2 * dynamic_heap_count_data_t::sample_size)))
(num_gcs_since_last_change < (2 * dynamic_heap_count_data_t::sample_size)))
{
dprintf (6666, ("[CHP0] we just adjusted %Id GCs ago, skipping", (current_gc_index - dynamic_heap_count_data.last_changed_gc_index)));
dprintf (6666, ("[CHP0] we just adjusted %Id GCs ago, skipping", num_gcs_since_last_change));
}
else
{
Expand All @@ -25435,9 +25464,9 @@ int gc_heap::calculate_new_heap_count ()
}

if (((int)dynamic_heap_count_data.last_changed_count > 0) && (dynamic_heap_count_data.last_changed_gc_index > 0.0) &&
((current_gc_index - dynamic_heap_count_data.last_changed_gc_index) <= (3 * dynamic_heap_count_data_t::sample_size)))
(num_gcs_since_last_change <= (3 * dynamic_heap_count_data_t::sample_size)))
{
dprintf (6666, ("[CHP0-0] just grew %d GCs ago, no change", (current_gc_index - dynamic_heap_count_data.last_changed_gc_index)));
dprintf (6666, ("[CHP0-0] just grew %d GCs ago, no change", num_gcs_since_last_change));
step_up_int = 0;
}
else
Expand Down Expand Up @@ -25487,9 +25516,18 @@ int gc_heap::calculate_new_heap_count ()
{
if (((int)dynamic_heap_count_data.last_changed_count > 0) && (dynamic_heap_count_data.last_changed_gc_index > 0.0))
{
(dynamic_heap_count_data.inc_failure_count)++;
dprintf (6666, ("[CHP0-4] just grew %d GCs ago, grow more aggressively from %d -> %d more heaps",
(current_gc_index - dynamic_heap_count_data.last_changed_gc_index), step_up_int, (step_up_int * (dynamic_heap_count_data.inc_failure_count + 1))));
if (num_gcs_since_last_change > (16 * dynamic_heap_count_data_t::sample_size))
{
dynamic_heap_count_data.inc_failure_count = 0;
dprintf (6666, ("[CHP0-4] grew %d GCs ago, too far in the past, set aggressive factor to 0, grow from %d -> %d more heaps",
num_gcs_since_last_change, dynamic_heap_count_data.inc_failure_count, step_up_int, (step_up_int * (dynamic_heap_count_data.inc_failure_count + 1))));
}
else
{
(dynamic_heap_count_data.inc_failure_count)++;
dprintf (6666, ("[CHP0-4] grew %d GCs ago, aggressive factor is %d, grow more aggressively from %d -> %d more heaps",
num_gcs_since_last_change, dynamic_heap_count_data.inc_failure_count, step_up_int, (step_up_int * (dynamic_heap_count_data.inc_failure_count + 1))));
}
step_up_int *= dynamic_heap_count_data.inc_failure_count + 1;
}
}
Expand All @@ -25514,9 +25552,9 @@ int gc_heap::calculate_new_heap_count ()
dynamic_heap_count_data.last_changed_stcp = smoothed_median_throughput_cost_percent;
}

dprintf (6666, ("[CHP0] tcp %.3f, stcp %.3f -> (%d * %.3f%% = %.3f) -> %d + %d = %d -> %d",
dprintf (6666, ("[CHP0] tcp %.3f, stcp %.3f -> (%d -> %.3f) -> %d + %d = %d -> %d",
median_throughput_cost_percent, smoothed_median_throughput_cost_percent,
actual_n_max_heaps, step_up_percent, step_up_float, step_up_int, n_heaps, (n_heaps + step_up_int), new_n_heaps));
actual_n_max_heaps, step_up_float, step_up_int, n_heaps, (n_heaps + step_up_int), new_n_heaps));
}
}
}
Expand All @@ -25533,7 +25571,7 @@ int gc_heap::calculate_new_heap_count ()
}
dprintf (6666, ("[CHP1] last time adjusted %s by %d at GC#%Id (%Id GCs since), stcp was %.3f, now stcp is %.3f",
((dynamic_heap_count_data.last_changed_count > 0.0) ? "up" : "down"), (int)dynamic_heap_count_data.last_changed_count,
dynamic_heap_count_data.last_changed_gc_index, (current_gc_index - dynamic_heap_count_data.last_changed_gc_index),
dynamic_heap_count_data.last_changed_gc_index, num_gcs_since_last_change,
dynamic_heap_count_data.last_changed_stcp, smoothed_median_throughput_cost_percent));

float below_target_diff = target_tcp - median_throughput_cost_percent;
Expand All @@ -25546,10 +25584,16 @@ int gc_heap::calculate_new_heap_count ()
if (dynamic_heap_count_data.below_target_accumulation >= dynamic_heap_count_data.below_target_threshold)
{
int below_target_tcp_count = dynamic_heap_count_data.rearrange_recorded_tcp ();
float below_target_tcp_slope = slope (dynamic_heap_count_data.recorded_tcp, below_target_tcp_count, &avg_recorded_tcp);
float below_target_tcp_slope = slope (dynamic_heap_count_data.recorded_tcp_rearranged, below_target_tcp_count, &avg_recorded_tcp);
float diff_pct = (target_tcp - smoothed_median_throughput_cost_percent) / target_tcp;
int step_down_int = (int)(diff_pct / 2.0 * n_heaps);
dprintf (6666, ("[CHP1] observed %d tcp's <= or ~ target, avg %.3f, slope %.3f, stcp %.3f below target, shrink by %.3f * %d = %d heaps",
if ((step_down_int == 0) && dynamic_heap_count_data.is_tcp_far_below (diff_pct))
{
dprintf (6666, ("[CHP1] we are far below target, reduce by 1 heap"));
step_down_int = 1;
}

dprintf (6666, ("[CHP1] observed %d tcp's <= or ~ target, avg %.3f, slope %.3f, stcp %.3f%% below target, shrink by %.3f%% * %d = %d heaps",
below_target_tcp_count, avg_recorded_tcp, below_target_tcp_slope, (diff_pct * 100.0), (diff_pct * 50.0), n_heaps, step_down_int));

bool shrink_p = false;
Expand Down Expand Up @@ -25629,11 +25673,22 @@ int gc_heap::calculate_new_heap_count ()

if (shrink_p && step_down_int && (new_n_heaps > step_down_int))
{
// TODO - if we see that it wants to shrink by 1 heap too many times, we do want to shrink.
if (step_down_int == 1)
{
step_down_int = 0;
dprintf (6666, ("[CHP1-3] don't shrink if it's just one heap. not worth it"));
if (dynamic_heap_count_data.should_dec_by_one())
{
dprintf (6666, ("[CHP1-3] shrink by one heap"));
}
else
{
step_down_int = 0;
dprintf (6666, ("[CHP1-3] don't shrink just yet if it's just one heap"));
}
}
else
{
dynamic_heap_count_data.reset_dec_by_one();
dprintf (6666, ("[CHP1-3] shrink by %d heap(s), reset dec by one", step_down_int));
}

new_n_heaps -= step_down_int;
Expand Down Expand Up @@ -26265,7 +26320,7 @@ bool gc_heap::change_heap_count (int new_n_heaps)
assert (gen_size >= dd_fragmentation (dd));
dd_current_size (dd) = gen_size - dd_fragmentation (dd);

dprintf (3, ("h%d g%d: budget: %zd, left in budget: %zd, %zd generation_size: %zd fragmentation: %zd current_size: %zd",
dprintf (3, ("h%d g%d: budget: %zd, left in budget: %zd, generation_size: %zd fragmentation: %zd current_size: %zd",
i,
gen_idx,
desired_alloc_per_heap[gen_idx],
Expand Down Expand Up @@ -43608,35 +43663,6 @@ size_t gc_heap::desired_new_allocation (dynamic_data* dd,
new_allocation = min (new_allocation,
max (min_gc_size, (max_size/3)));
}

#ifdef DYNAMIC_HEAP_COUNT
if (dynamic_adaptation_mode == dynamic_adaptation_to_application_sizes)
{
// if this is set, limit gen 0 size to a small multiple of the older generations
float f_older_gen = ((10.0f / conserve_mem_setting) - 1) * 0.5f;

// compute the total size of the older generations
size_t older_size = 0;
for (int gen_index_older = 1; gen_index_older < total_generation_count; gen_index_older++)
{
dynamic_data* dd_older = dynamic_data_of (gen_index_older);
older_size += dd_current_size (dd_older);
}
// derive a new allocation size from it
size_t new_allocation_from_older = (size_t)(older_size*f_older_gen);

// limit the new allocation to this value
new_allocation = min (new_allocation, new_allocation_from_older);

// but make sure it doesn't drop below the minimum size
new_allocation = max (new_allocation, min_gc_size);

dprintf (2, ("f_older_gen: %d%% older_size: %zd new_allocation: %zd",
(int)(f_older_gen*100),
older_size,
new_allocation));
}
#endif //DYNAMIC_HEAP_COUNT
}
}

Expand Down Expand Up @@ -48782,7 +48808,8 @@ HRESULT GCHeap::Initialize()
// start with only 1 heap
gc_heap::smoothed_desired_total[0] /= gc_heap::n_heaps;
int initial_n_heaps = 1;
dprintf (9999, ("gc_heap::n_heaps is %d, initial %d", gc_heap::n_heaps, initial_n_heaps));

dprintf (6666, ("n_heaps is %d, initial n_heaps is %d, %d cores", gc_heap::n_heaps, initial_n_heaps, g_num_processors));

{
if (!gc_heap::prepare_to_change_heap_count (initial_n_heaps))
Expand Down Expand Up @@ -48810,6 +48837,12 @@ HRESULT GCHeap::Initialize()
gc_heap::dynamic_heap_count_data.below_target_threshold = 10.0;
gc_heap::dynamic_heap_count_data.inc_recheck_threshold = 5;
gc_heap::dynamic_heap_count_data.dec_failure_recheck_threshold = 5;
// 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 = min (dd_max_size (gc_heap::g_heaps[0]->dynamic_data_of (0)), (64 * 1024 * 1024));
gc_heap::dynamic_heap_count_data.min_gen0_new_allocation = dd_min_size (gc_heap::g_heaps[0]->dynamic_data_of (0));

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));
}
#endif //DYNAMIC_HEAP_COUNT
GCScan::GcRuntimeStructuresValid (TRUE);
Expand Down
Loading

0 comments on commit 89bd910

Please sign in to comment.