[improvement](agg) iterate aggregation data in memory written order (#12704)

Following the iteration order of the hash table will result in out-of-order access to aggregate states, which is very inefficient.
Traversing aggregate states in memory write order can significantly improve memory read efficiency.

Test
hash table items count: 3.35M

Before this optimization: insert keys into column takes 500ms
With this optimization only takes 80ms
This commit is contained in:
Jerry Hu
2022-09-21 14:58:50 +08:00
committed by GitHub
parent 27f7ae258d
commit 8f4bb0f804
12 changed files with 523 additions and 51 deletions

View File

@ -257,6 +257,37 @@ struct HashMethodSingleLowNullableColumn : public SingleColumnMethod {
return EmplaceResult(inserted);
}
template <typename Data, typename Func, typename CreatorForNull>
ALWAYS_INLINE typename std::enable_if_t<has_mapped, Mapped>& lazy_emplace_key(
Data& data, size_t row, Arena& pool, Func&& f, CreatorForNull&& null_creator) {
if (key_columns[0]->is_null_at(row)) {
bool has_null_key = data.has_null_key_data();
data.has_null_key_data() = true;
if (!has_null_key) std::forward<CreatorForNull>(null_creator)(data.get_null_key_data());
return data.get_null_key_data();
}
auto key_holder = Base::get_key_holder(row, pool);
typename Data::LookupResult it;
data.lazy_emplace(key_holder, it, std::forward<Func>(f));
return *lookup_result_get_mapped(it);
}
template <typename Data, typename Func, typename CreatorForNull>
ALWAYS_INLINE typename std::enable_if_t<has_mapped, Mapped>& lazy_emplace_key(
Data& data, size_t row, Arena& pool, size_t hash_value, Func&& f,
CreatorForNull&& null_creator) {
if (key_columns[0]->is_null_at(row)) {
bool has_null_key = data.has_null_key_data();
data.has_null_key_data() = true;
if (!has_null_key) std::forward<CreatorForNull>(null_creator)(data.get_null_key_data());
return data.get_null_key_data();
}
auto key_holder = Base::get_key_holder(row, pool);
typename Data::LookupResult it;
data.lazy_emplace(key_holder, it, hash_value, std::forward<Func>(f));
return *lookup_result_get_mapped(it);
}
template <typename Data>
ALWAYS_INLINE FindResult find_key(Data& data, size_t row, Arena& pool) {
if (key_columns[0]->is_null_at(row)) {
@ -276,5 +307,16 @@ struct HashMethodSingleLowNullableColumn : public SingleColumnMethod {
}
};
template <typename HashMethod>
struct IsSingleNullableColumnMethod {
static constexpr bool value = false;
};
template <typename SingleColumnMethod, typename Mapped, bool use_cache>
struct IsSingleNullableColumnMethod<
HashMethodSingleLowNullableColumn<SingleColumnMethod, Mapped, use_cache>> {
static constexpr bool value = true;
};
} // namespace ColumnsHashing
} // namespace doris::vectorized

View File

@ -140,6 +140,22 @@ public:
return emplaceImpl(key_holder, hash_value, data);
}
template <typename Data, typename Func>
ALWAYS_INLINE typename std::enable_if_t<has_mapped, Mapped>& lazy_emplace_key(Data& data,
size_t row,
Arena& pool,
Func&& f) {
auto key_holder = static_cast<Derived&>(*this).get_key_holder(row, pool);
return lazy_emplace_impl(key_holder, data, std::forward<Func>(f));
}
template <typename Data, typename Func>
ALWAYS_INLINE typename std::enable_if_t<has_mapped, Mapped>& lazy_emplace_key(
Data& data, size_t hash_value, size_t row, Arena& pool, Func&& f) {
auto key_holder = static_cast<Derived&>(*this).get_key_holder(row, pool);
return lazy_emplace_impl(key_holder, hash_value, data, std::forward<Func>(f));
}
template <typename Data>
ALWAYS_INLINE FindResult find_key(Data& data, size_t row, Arena& pool) {
auto key_holder = static_cast<Derived&>(*this).get_key_holder(row, pool);
@ -264,6 +280,23 @@ protected:
return EmplaceResult(inserted);
}
template <typename Data, typename KeyHolder, typename Func>
ALWAYS_INLINE typename std::enable_if_t<has_mapped, Mapped>& lazy_emplace_impl(
KeyHolder& key_holder, Data& data, Func&& f) {
typename Data::LookupResult it;
data.lazy_emplace(key_holder, it, std::forward<Func>(f));
return *lookup_result_get_mapped(it);
}
template <typename Data, typename KeyHolder, typename Func>
ALWAYS_INLINE typename std::enable_if_t<has_mapped, Mapped>& lazy_emplace_impl(
KeyHolder& key_holder, size_t hash_value, Data& data, Func&& f) {
typename Data::LookupResult it;
data.lazy_emplace(key_holder, it, hash_value, std::forward<Func>(f));
return *lookup_result_get_mapped(it);
}
template <typename Data, typename Key>
ALWAYS_INLINE FindResult find_key_impl(Key key, Data& data) {
if constexpr (Cache::consecutive_keys_optimization) {

View File

@ -78,6 +78,7 @@ struct FixedHashMapImplicitZeroCell {
FixedHashMapImplicitZeroCell() {}
FixedHashMapImplicitZeroCell(const Key&, const State&) {}
FixedHashMapImplicitZeroCell(const Key&, const Mapped& mapped_) : mapped(mapped_) {}
FixedHashMapImplicitZeroCell(const value_type& value_, const State&) : mapped(value_.second) {}
const VoidKey get_first() const { return {}; }

View File

@ -269,6 +269,31 @@ public:
this->increase_size();
}
class Constructor {
public:
friend class FixedHashTable;
template <typename... Args>
void operator()(Args&&... args) const {
new (_cell) Cell(std::forward<Args>(args)...);
}
private:
Constructor(Cell* cell) : _cell(cell) {}
Cell* _cell;
};
template <typename Func>
void ALWAYS_INLINE lazy_emplace(const Key& x, LookupResult& it, Func&& f) {
it = &buf[x];
if (!buf[x].is_zero(*this)) {
return;
}
f(Constructor(&buf[x]), x);
this->increase_size();
}
std::pair<LookupResult, bool> ALWAYS_INLINE insert(const value_type& x) {
std::pair<LookupResult, bool> res;
emplace(Cell::get_key(x), res.first, res.second);

View File

@ -59,6 +59,7 @@ struct HashMapCell {
HashMapCell() {}
HashMapCell(const Key& key_, const State&) : value(key_, NoInitTag()) {}
HashMapCell(const Key& key_, const Mapped& mapped_) : value(key_, mapped_) {}
HashMapCell(const value_type& value_, const State&) : value(value_) {}
const Key& get_first() const { return value.first; }

View File

@ -765,6 +765,27 @@ protected:
return false;
}
template <typename Func>
bool ALWAYS_INLINE lazy_emplace_if_zero(const Key& x, LookupResult& it, size_t hash_value,
Func&& f) {
/// If it is claimed that the zero key can not be inserted into the table.
if (!Cell::need_zero_value_storage) return false;
if (Cell::is_zero(x, *this)) {
it = this->zero_value();
if (!this->get_has_zero()) {
++m_size;
this->set_get_has_zero();
std::forward<Func>(f)(Constructor(it), x);
this->zero_value()->set_hash(hash_value);
}
return true;
}
return false;
}
template <typename KeyHolder>
void ALWAYS_INLINE emplace_non_zero_impl(size_t place_value, KeyHolder&& key_holder,
LookupResult& it, bool& inserted, size_t hash_value) {
@ -804,6 +825,43 @@ protected:
}
}
template <typename KeyHolder, typename Func>
void ALWAYS_INLINE lazy_emplace_non_zero_impl(size_t place_value, KeyHolder&& key_holder,
LookupResult& it, size_t hash_value, Func&& f) {
it = &buf[place_value];
if (!buf[place_value].is_zero(*this)) {
key_holder_discard_key(key_holder);
return;
}
key_holder_persist_key(key_holder);
const auto& key = key_holder_get_key(key_holder);
f(Constructor(&buf[place_value]), key);
buf[place_value].set_hash(hash_value);
++m_size;
if (UNLIKELY(grower.overflow(m_size))) {
try {
resize();
} catch (...) {
/** If we have not resized successfully, then there will be problems.
* There remains a key, but uninitialized mapped-value,
* which, perhaps, can not even be called a destructor.
*/
--m_size;
buf[place_value].set_zero();
throw;
}
// The hash table was rehashed, so we have to re-find the key.
size_t new_place = find_cell(key, hash_value, grower.place(hash_value));
assert(!buf[new_place].is_zero(*this));
it = &buf[new_place];
}
}
/// Only for non-zero keys. Find the right place, insert the key there, if it does not already exist. Set iterator to the cell in output parameter.
template <typename KeyHolder>
void ALWAYS_INLINE emplace_non_zero(KeyHolder&& key_holder, LookupResult& it, bool& inserted,
@ -813,6 +871,14 @@ protected:
emplace_non_zero_impl(place_value, key_holder, it, inserted, hash_value);
}
template <typename KeyHolder, typename Func>
void ALWAYS_INLINE lazy_emplace_non_zero(KeyHolder&& key_holder, LookupResult& it,
size_t hash_value, Func&& f) {
const auto& key = key_holder_get_key(key_holder);
size_t place_value = find_cell(key, hash_value, grower.place(hash_value));
lazy_emplace_non_zero_impl(place_value, key_holder, it, hash_value, std::forward<Func>(f));
}
public:
void expanse_for_add_elem(size_t num_elem) {
if (add_elem_size_overflow(num_elem)) {
@ -847,6 +913,19 @@ public:
reinsert(*it.get_ptr(), hash_value);
}
class Constructor {
public:
friend class HashTable;
template <typename... Args>
void operator()(Args&&... args) const {
new (_cell) Cell(std::forward<Args>(args)...);
}
private:
Constructor(Cell* cell) : _cell(cell) {}
Cell* _cell;
};
/** Insert the key.
* Return values:
* 'it' -- a LookupResult pointing to the corresponding key/mapped pair.
@ -877,6 +956,20 @@ public:
emplace_non_zero(key_holder, it, inserted, hash_value);
}
template <typename KeyHolder, typename Func>
void ALWAYS_INLINE lazy_emplace(KeyHolder&& key_holder, LookupResult& it, Func&& f) {
const auto& key = key_holder_get_key(key_holder);
lazy_emplace(key_holder, it, hash(key), std::forward<Func>(f));
}
template <typename KeyHolder, typename Func>
void ALWAYS_INLINE lazy_emplace(KeyHolder&& key_holder, LookupResult& it, size_t hash_value,
Func&& f) {
const auto& key = key_holder_get_key(key_holder);
if (!lazy_emplace_if_zero(key, it, hash_value, std::forward<Func>(f)))
lazy_emplace_non_zero(key_holder, it, hash_value, std::forward<Func>(f));
}
/// Copy the cell from another hash table. It is assumed that the cell is not zero, and also that there was no such key in the table yet.
void ALWAYS_INLINE insert_unique_non_zero(const Cell* cell, size_t hash_value) {
size_t place_value = find_empty_cell(grower.place(hash_value));

View File

@ -22,4 +22,5 @@ template <typename T>
struct HashTableTraits {
static constexpr bool is_phmap = false;
static constexpr bool is_parallel_phmap = false;
static constexpr bool is_string_hash_table = false;
};

View File

@ -118,6 +118,16 @@ public:
it = &*it_;
}
template <typename KeyHolder, typename Func>
void ALWAYS_INLINE lazy_emplace(KeyHolder&& key_holder, LookupResult& it, Func&& f) {
const auto& key = key_holder_get_key(key_holder);
auto it_ = _hash_map.lazy_emplace(key, [&](const auto& ctor) {
key_holder_persist_key(key_holder);
f(ctor, key);
});
it = &*it_;
}
template <typename KeyHolder>
void ALWAYS_INLINE emplace(KeyHolder&& key_holder, LookupResult& it, size_t hash_value,
bool& inserted) {
@ -140,6 +150,25 @@ public:
}
}
template <typename KeyHolder, typename Func>
void ALWAYS_INLINE lazy_emplace(KeyHolder&& key_holder, LookupResult& it, size_t hash_value,
Func&& f) {
const auto& key = key_holder_get_key(key_holder);
if constexpr (use_parallel) {
auto it_ = _hash_map.lazy_emplace_with_hash(hash_value, key, [&](const auto& ctor) {
key_holder_persist_key(key_holder);
f(ctor, key);
});
it = &*it_;
} else {
auto it_ = _hash_map.lazy_emplace_with_hash(key, hash_value, [&](const auto& ctor) {
key_holder_persist_key(key_holder);
f(ctor, key);
});
it = &*it_;
}
}
template <typename KeyHolder>
LookupResult ALWAYS_INLINE find(KeyHolder&& key_holder) {
const auto& key = key_holder_get_key(key_holder);
@ -197,4 +226,5 @@ template <typename Key, typename Mapped, typename Hash, bool use_parallel>
struct HashTableTraits<PHHashMap<Key, Mapped, Hash, use_parallel>> {
static constexpr bool is_phmap = true;
static constexpr bool is_parallel_phmap = use_parallel;
static constexpr bool is_string_hash_table = false;
};

View File

@ -209,3 +209,17 @@ public:
char* get_null_key_data() { return nullptr; }
bool has_null_key_data() const { return false; }
};
template <typename TMapped, typename Allocator>
struct HashTableTraits<StringHashMap<TMapped, Allocator>> {
static constexpr bool is_phmap = false;
static constexpr bool is_parallel_phmap = false;
static constexpr bool is_string_hash_table = true;
};
template <template <typename> class Derived, typename TMapped, typename Allocator>
struct HashTableTraits<Derived<StringHashMap<TMapped, Allocator>>> {
static constexpr bool is_phmap = false;
static constexpr bool is_parallel_phmap = false;
static constexpr bool is_string_hash_table = true;
};

View File

@ -130,6 +130,30 @@ public:
it = zero_value();
}
class Constructor {
public:
friend struct StringHashTableEmpty;
template <typename... Args>
void operator()(Args&&... args) const {
new (_cell) Cell(std::forward<Args>(args)...);
}
private:
Constructor(Cell* cell) : _cell(cell) {}
Cell* _cell;
};
template <typename KeyHolder, typename Func>
void ALWAYS_INLINE lazy_emplace(KeyHolder&& key_holder, LookupResult& it, size_t hash_value,
Func&& f) {
if (!has_zero()) {
const auto& key = key_holder_get_key(key_holder);
set_has_zero(key);
std::forward<Func>(f)(Constructor(zero_value()), key);
}
it = zero_value();
}
template <typename Key>
LookupResult ALWAYS_INLINE find(const Key&, size_t = 0) {
return has_zero() ? zero_value() : nullptr;
@ -179,7 +203,7 @@ struct StringHashTableLookupResult {
};
template <typename Mapped>
ALWAYS_INLINE inline auto lookup_result_get_mapped(StringHashTableLookupResult<Mapped*> cell) {
ALWAYS_INLINE inline auto lookup_result_get_mapped(StringHashTableLookupResult<Mapped> cell) {
return &cell.get_mapped();
}
@ -548,6 +572,27 @@ public:
this->dispatch(*this, key_holder, EmplaceCallable(it, inserted));
}
template <typename Func>
struct LazyEmplaceCallable {
LookupResult& mapped;
Func&& f;
LazyEmplaceCallable(LookupResult& mapped_, Func&& f_)
: mapped(mapped_), f(std::forward<Func>(f_)) {}
template <typename Map, typename KeyHolder>
void ALWAYS_INLINE operator()(Map& map, KeyHolder&& key_holder, size_t hash) {
typename Map::LookupResult result;
map.lazy_emplace(key_holder, result, hash, std::forward<Func>(f));
mapped = &result->get_second();
}
};
template <typename KeyHolder, typename Func>
void ALWAYS_INLINE lazy_emplace(KeyHolder&& key_holder, LookupResult& it, Func&& f) {
this->dispatch(*this, key_holder, LazyEmplaceCallable<Func>(it, std::forward<Func>(f)));
}
struct FindCallable {
// find() doesn't need any key memory management, so we don't work with
// any key holders here, only with normal keys. The key type is still
@ -613,3 +658,10 @@ public:
size_t get_collisions() const { return 0; }
#endif
};
template <typename SubMaps>
struct HashTableTraits<StringHashTable<SubMaps>> {
static constexpr bool is_phmap = false;
static constexpr bool is_parallel_phmap = false;
static constexpr bool is_string_hash_table = true;
};

View File

@ -97,6 +97,8 @@ AggregationNode::AggregationNode(ObjectPool* pool, const TPlanNode& tnode,
_serialize_result_timer(nullptr),
_deserialize_data_timer(nullptr),
_hash_table_compute_timer(nullptr),
_hash_table_iterate_timer(nullptr),
_insert_keys_to_column_timer(nullptr),
_streaming_agg_timer(nullptr),
_hash_table_size_counter(nullptr),
_hash_table_input_counter(nullptr) {
@ -295,6 +297,8 @@ Status AggregationNode::prepare(RuntimeState* state) {
_serialize_result_timer = ADD_TIMER(runtime_profile(), "SerializeResultTime");
_deserialize_data_timer = ADD_TIMER(runtime_profile(), "DeserializeDataTime");
_hash_table_compute_timer = ADD_TIMER(runtime_profile(), "HashTableComputeTime");
_hash_table_iterate_timer = ADD_TIMER(runtime_profile(), "HashTableIterateTime");
_insert_keys_to_column_timer = ADD_TIMER(runtime_profile(), "InsertKeysToColumnTime");
_streaming_agg_timer = ADD_TIMER(runtime_profile(), "StreamingAggTime");
_hash_table_size_counter = ADD_COUNTER(runtime_profile(), "HashTableSize", TUnit::UNIT);
_hash_table_input_counter = ADD_COUNTER(runtime_profile(), "HashTableInputCount", TUnit::UNIT);
@ -384,6 +388,20 @@ Status AggregationNode::prepare(RuntimeState* state) {
_executor.close = std::bind<void>(&AggregationNode::_close_without_key, this);
} else {
_init_hash_method(_probe_expr_ctxs);
std::visit(
[&](auto&& agg_method) {
using HashTableType = std::decay_t<decltype(agg_method.data)>;
using KeyType = typename HashTableType::key_type;
/// some aggregate functions (like AVG for decimal) have align issues.
_aggregate_data_container.reset(new AggregateDataContainer(
sizeof(KeyType),
((_total_size_of_aggregate_states + _align_aggregate_states - 1) /
_align_aggregate_states) *
_align_aggregate_states));
},
_agg_data._aggregated_method_variant);
if (_is_merge) {
_executor.execute = std::bind<Status>(&AggregationNode::_merge_with_serialized_key,
this, std::placeholders::_1);
@ -787,43 +805,64 @@ void AggregationNode::_emplace_into_hash_table(AggregateDataPtr* places, ColumnR
}
}
auto creator = [this](const auto& ctor, const auto& key) {
using KeyType = std::decay_t<decltype(key)>;
if constexpr (HashTableTraits<HashTableType>::is_string_hash_table &&
!std::is_same_v<StringRef, KeyType>) {
StringRef string_ref = to_string_ref(key);
ArenaKeyHolder key_holder {string_ref, _agg_arena_pool};
key_holder_persist_key(key_holder);
auto mapped = _aggregate_data_container->append_data(key_holder.key);
_create_agg_status(mapped);
ctor(key, mapped);
} else {
auto mapped = _aggregate_data_container->append_data(key);
_create_agg_status(mapped);
ctor(key, mapped);
}
};
auto creator_for_null_key = [this](auto& mapped) {
mapped = _agg_arena_pool.aligned_alloc(_total_size_of_aggregate_states,
_align_aggregate_states);
_create_agg_status(mapped);
};
/// For all rows.
COUNTER_UPDATE(_hash_table_input_counter, num_rows);
for (size_t i = 0; i < num_rows; ++i) {
AggregateDataPtr aggregate_data = nullptr;
auto emplace_result = [&]() {
if constexpr (HashTableTraits<HashTableType>::is_phmap) {
if (LIKELY(i + HASH_MAP_PREFETCH_DIST < num_rows)) {
if constexpr (HashTableTraits<HashTableType>::is_parallel_phmap) {
agg_method.data.prefetch_by_key(state.get_key_holder(
i + HASH_MAP_PREFETCH_DIST, _agg_arena_pool));
} else
agg_method.data.prefetch_by_hash(
_hash_values[i + HASH_MAP_PREFETCH_DIST]);
}
return state.emplace_key(agg_method.data, _hash_values[i], i,
_agg_arena_pool);
} else {
return state.emplace_key(agg_method.data, i, _agg_arena_pool);
AggregateDataPtr mapped = nullptr;
if constexpr (HashTableTraits<HashTableType>::is_phmap) {
if (LIKELY(i + HASH_MAP_PREFETCH_DIST < num_rows)) {
if constexpr (HashTableTraits<HashTableType>::is_parallel_phmap) {
agg_method.data.prefetch_by_key(state.get_key_holder(
i + HASH_MAP_PREFETCH_DIST, _agg_arena_pool));
} else
agg_method.data.prefetch_by_hash(
_hash_values[i + HASH_MAP_PREFETCH_DIST]);
}
}();
/// If a new key is inserted, initialize the states of the aggregate functions, and possibly something related to the key.
if (emplace_result.is_inserted()) {
/// exception-safety - if you can not allocate memory or create states, then destructors will not be called.
emplace_result.set_mapped(nullptr);
if constexpr (ColumnsHashing::IsSingleNullableColumnMethod<
AggState>::value) {
mapped = state.lazy_emplace_key(agg_method.data, _hash_values[i], i,
_agg_arena_pool, creator,
creator_for_null_key);
} else {
mapped = state.lazy_emplace_key(agg_method.data, _hash_values[i], i,
_agg_arena_pool, creator);
}
} else {
if constexpr (ColumnsHashing::IsSingleNullableColumnMethod<
AggState>::value) {
mapped = state.lazy_emplace_key(agg_method.data, i, _agg_arena_pool,
creator, creator_for_null_key);
} else {
mapped = state.lazy_emplace_key(agg_method.data, i, _agg_arena_pool,
creator);
}
}
aggregate_data = _agg_arena_pool.aligned_alloc(
_total_size_of_aggregate_states, _align_aggregate_states);
_create_agg_status(aggregate_data);
emplace_result.set_mapped(aggregate_data);
} else
aggregate_data = emplace_result.get_mapped();
places[i] = aggregate_data;
places[i] = mapped;
assert(places[i] != nullptr);
}
},
@ -1051,24 +1090,33 @@ Status AggregationNode::_get_with_serialized_key_result(RuntimeState* state, Blo
std::visit(
[&](auto&& agg_method) -> void {
auto& data = agg_method.data;
auto& iter = agg_method.iterator;
agg_method.init_once();
const auto size = std::min(data.size(), size_t(state->batch_size()));
using KeyType = std::decay_t<decltype(iter->get_first())>;
using KeyType = std::decay_t<decltype(agg_method.iterator->get_first())>;
std::vector<KeyType> keys(size);
if (_values.size() < size) {
_values.resize(size);
}
size_t num_rows = 0;
while (iter != data.end() && num_rows < state->batch_size()) {
keys[num_rows] = iter->get_first();
_values[num_rows] = iter->get_second();
++iter;
++num_rows;
_aggregate_data_container->init_once();
auto& iter = _aggregate_data_container->iterator;
{
SCOPED_TIMER(_hash_table_iterate_timer);
while (iter != _aggregate_data_container->end() &&
num_rows < state->batch_size()) {
keys[num_rows] = iter.get_key<KeyType>();
_values[num_rows] = iter.get_aggregate_data();
++iter;
++num_rows;
}
}
agg_method.insert_keys_into_columns(keys, key_columns, num_rows, _probe_key_sz);
{
SCOPED_TIMER(_insert_keys_to_column_timer);
agg_method.insert_keys_into_columns(keys, key_columns, num_rows, _probe_key_sz);
}
for (size_t i = 0; i < _aggregate_evaluators.size(); ++i) {
_aggregate_evaluators[i]->insert_result_info_vec(
@ -1076,7 +1124,7 @@ Status AggregationNode::_get_with_serialized_key_result(RuntimeState* state, Blo
num_rows);
}
if (iter == data.end()) {
if (iter == _aggregate_data_container->end()) {
if (agg_method.data.has_null_key_data()) {
// only one key of group by support wrap null key
// here need additional processing logic on the null key / value
@ -1137,27 +1185,37 @@ Status AggregationNode::_serialize_with_serialized_key_result(RuntimeState* stat
[&](auto&& agg_method) -> void {
agg_method.init_once();
auto& data = agg_method.data;
auto& iter = agg_method.iterator;
const auto size = std::min(data.size(), size_t(state->batch_size()));
using KeyType = std::decay_t<decltype(iter->get_first())>;
using KeyType = std::decay_t<decltype(agg_method.iterator->get_first())>;
std::vector<KeyType> keys(size);
if (_values.size() < size + 1) {
_values.resize(size + 1);
}
size_t num_rows = 0;
while (iter != data.end() && num_rows < state->batch_size()) {
keys[num_rows] = iter->get_first();
_values[num_rows] = iter->get_second();
++iter;
++num_rows;
_aggregate_data_container->init_once();
auto& iter = _aggregate_data_container->iterator;
{
SCOPED_TIMER(_hash_table_iterate_timer);
while (iter != _aggregate_data_container->end() &&
num_rows < state->batch_size()) {
keys[num_rows] = iter.get_key<KeyType>();
_values[num_rows] = iter.get_aggregate_data();
++iter;
++num_rows;
}
}
agg_method.insert_keys_into_columns(keys, key_columns, num_rows, _probe_key_sz);
{
SCOPED_TIMER(_insert_keys_to_column_timer);
agg_method.insert_keys_into_columns(keys, key_columns, num_rows, _probe_key_sz);
}
if (iter == data.end()) {
if (iter == _aggregate_data_container->end()) {
if (agg_method.data.has_null_key_data()) {
// only one key of group by support wrap null key
// here need additional processing logic on the null key / value
DCHECK(key_columns.size() == 1);
DCHECK(key_columns[0]->is_nullable());
if (agg_method.data.has_null_key_data()) {

View File

@ -616,6 +616,125 @@ struct AggregatedDataVariants {
using AggregatedDataVariantsPtr = std::shared_ptr<AggregatedDataVariants>;
struct AggregateDataContainer {
public:
AggregateDataContainer(size_t size_of_key, size_t size_of_aggregate_states)
: _size_of_key(size_of_key), _size_of_aggregate_states(size_of_aggregate_states) {
_expand();
}
template <typename KeyType>
AggregateDataPtr append_data(const KeyType& key) {
assert(sizeof(KeyType) == _size_of_key);
if (UNLIKELY(_index_in_sub_container == SUB_CONTAINER_CAPACITY)) {
_expand();
}
*reinterpret_cast<KeyType*>(_current_keys) = key;
auto aggregate_data = _current_agg_data;
++_total_count;
++_index_in_sub_container;
_current_agg_data += _size_of_aggregate_states;
_current_keys += _size_of_key;
return aggregate_data;
}
template <typename Derived, bool IsConst>
class IteratorBase {
using Container =
std::conditional_t<IsConst, const AggregateDataContainer, AggregateDataContainer>;
Container* container;
uint32_t index;
uint32_t sub_container_index;
uint32_t index_in_sub_container;
friend class HashTable;
public:
IteratorBase() {}
IteratorBase(Container* container_, uint32_t index_)
: container(container_), index(index_) {
sub_container_index = index / SUB_CONTAINER_CAPACITY;
index_in_sub_container = index % SUB_CONTAINER_CAPACITY;
}
bool operator==(const IteratorBase& rhs) const { return index == rhs.index; }
bool operator!=(const IteratorBase& rhs) const { return index != rhs.index; }
Derived& operator++() {
index++;
sub_container_index = index / SUB_CONTAINER_CAPACITY;
index_in_sub_container = index % SUB_CONTAINER_CAPACITY;
return static_cast<Derived&>(*this);
}
template <typename KeyType>
KeyType get_key() {
assert(sizeof(KeyType) == container->_size_of_key);
return ((KeyType*)(container->_key_containers[sub_container_index]))
[index_in_sub_container];
}
AggregateDataPtr get_aggregate_data() {
return &(container->_value_containers[sub_container_index]
[container->_size_of_aggregate_states *
index_in_sub_container]);
}
};
class Iterator : public IteratorBase<Iterator, false> {
public:
using IteratorBase<Iterator, false>::IteratorBase;
};
class ConstIterator : public IteratorBase<ConstIterator, true> {
public:
using IteratorBase<ConstIterator, true>::IteratorBase;
};
ConstIterator begin() const { return ConstIterator(this, 0); }
ConstIterator cbegin() const { return begin(); }
Iterator begin() { return Iterator(this, 0); }
ConstIterator end() const { return ConstIterator(this, _total_count); }
ConstIterator cend() const { return end(); }
Iterator end() { return Iterator(this, _total_count); }
void init_once() {
if (_inited) return;
_inited = true;
iterator = begin();
}
Iterator iterator;
private:
void _expand() {
_index_in_sub_container = 0;
_current_keys = _arena_pool.alloc(_size_of_key * SUB_CONTAINER_CAPACITY);
_key_containers.emplace_back(_current_keys);
_current_agg_data = (AggregateDataPtr)_arena_pool.alloc(_size_of_aggregate_states *
SUB_CONTAINER_CAPACITY);
_value_containers.emplace_back(_current_agg_data);
}
private:
static constexpr uint32_t SUB_CONTAINER_CAPACITY = 8192;
Arena _arena_pool;
std::vector<char*> _key_containers;
std::vector<AggregateDataPtr> _value_containers;
AggregateDataPtr _current_agg_data;
char* _current_keys;
size_t _size_of_key {};
size_t _size_of_aggregate_states {};
uint32_t _index_in_sub_container {};
uint32_t _total_count {};
bool _inited = false;
};
// not support spill
class AggregationNode : public ::doris::ExecNode {
public:
@ -675,6 +794,8 @@ private:
RuntimeProfile::Counter* _serialize_result_timer;
RuntimeProfile::Counter* _deserialize_data_timer;
RuntimeProfile::Counter* _hash_table_compute_timer;
RuntimeProfile::Counter* _hash_table_iterate_timer;
RuntimeProfile::Counter* _insert_keys_to_column_timer;
RuntimeProfile::Counter* _streaming_agg_timer;
RuntimeProfile::Counter* _hash_table_size_counter;
RuntimeProfile::Counter* _hash_table_input_counter;
@ -690,6 +811,7 @@ private:
std::vector<char> _deserialize_buffer;
std::vector<size_t> _hash_values;
std::vector<AggregateDataPtr> _values;
std::unique_ptr<AggregateDataContainer> _aggregate_data_container;
private:
/// Return true if we should keep expanding hash tables in the preagg. If false,