[DuckDB] 多核運算元並行的原始碼解析

DuckDB 是近年來頗受關注的OLAP資料庫，號稱是OLAP領域的SQLite，以精巧簡單，效能優異而著稱。筆者前段時間在調研Doris的Pipeline的運算元並行方案，而DuckDB基於論文《Morsel-Driven Parallelism: A NUMA-Aware Query Evaluation Framework for the Many-Core Age》實現SQL運算元的高效並行化的Pipeline執行引擎，所以筆者花了一些時間進行了學習和總結，這裡結合了Mark Raasveldt進行的分享和原始程式碼來一一剖析DuckDB在執行運算元並行上的具體實現。

1. 基礎知識

問題1：並行task的數目由什麼決定？

Pipeline的核心是：Morsel-Driven，資料是拆分成了小部分的資料。所以並行Task的核心是：能夠利用多執行緒來處理資料，每一個資料拆分為小部分，所以拆分並行的數目由Source決定。

DuckDB在GlobalSource上實現了一個虛擬函式MaxThread來決定task數目：

每一個運算元的GlobalSource抽象了自己的並行度：

問題2：並行task的怎麼樣進行多執行緒同步：

多執行緒的競爭只會發生在SinkOperator上，也就是Pipeline的尾端。
parallelism-aware的演演算法需要實現在Sink端
其他的非Sink operators (比如：Hash Join Probe, Projection, Filter等)，不需要感知多執行緒同步的問題

問題3：DuckDB的是如何抽象介面的：

Sink的Opeartor 定義了兩種型別：GlobalState, LocalState

GlobalState: 每個查詢的Operator全域性只有一個GlobalSinkState，記錄全域性部分的資訊

class PhysicalOperator {
public:
	unique_ptr<GlobalSinkState> sink_state;

LocalState: 每個查詢的PipelineExecutor都有一個LocalSinkState，都是區域性私有

//! The Pipeline class represents an execution pipeline
class PipelineExecutor {
private:
	//! The local sink state (if any)
	unique_ptr<LocalSinkState> local_sink_state;

後續會詳細解析不同的sink之間的LocalState和GlobalState如何配合的，核心部分如下：

Sink ：處理LocalState的資料

Combine：合併LocalState到GlobalState之中

2. 核心運算元的並行

這部分進行各個運算元的原始碼剖析，筆者在原始碼的關鍵部分加上了中文註釋，以方便大家的理解

Sort運算元

Sink介面：這裡需要注意的是DuckDB排序是進行了列轉行的工作的，後續讀取時需要行轉列。Sink這部分相當於實現了部分資料的排序工作。

SinkResultType PhysicalOrder::Sink(ExecutionContext &context, GlobalSinkState &gstate_p, LocalSinkState &lstate_p,
                                   DataChunk &input) const {
	auto &lstate = (OrderLocalSinkState &)lstate_p;
        
      // keys 是排序的列block，payload是輸出的排序後資料，這裡呼叫LocalState的SinkChunk，進行資料的轉行，
	local_sort_state.SinkChunk(keys, payload);

	// 資料達到記憶體閾值的時候進行基數排序處理，排序之後的結果存入LocalState的原生的SortedBlock中
	if (local_sort_state.SizeInBytes() >= gstate.memory_per_thread) {
		local_sort_state.Sort(global_sort_state, true);
	}
	return SinkResultType::NEED_MORE_INPUT;
}

Combine介面：加鎖，拷貝sorted block到Global State

void PhysicalOrder::Combine(ExecutionContext &context, GlobalSinkState &gstate_p, LocalSinkState &lstate_p) const {
	auto &gstate = (OrderGlobalSinkState &)gstate_p;
	auto &lstate = (OrderLocalSinkState &)lstate_p;
        // 排序剩餘記憶體中不滿的資料
	local_sort_state.Sort(*this, external || !local_sort_state.sorted_blocks.empty());

	// Append local state sorted data to this global state
	lock_guard<mutex> append_guard(lock);
	for (auto &sb : local_sort_state.sorted_blocks) {
		sorted_blocks.push_back(move(sb));
	}
}

MergeTask：啟動核數相同的task來進行Merge (這裡可以看出DuckDB對於多執行緒的使用是很激進的)，這裡是通過Event的機制實現的

void Schedule() override {
		auto &context = pipeline->GetClientContext();
		idx_t num_threads = ts.NumberOfThreads();

		vector<unique_ptr<Task>> merge_tasks;
		for (idx_t tnum = 0; tnum < num_threads; tnum++) {
			merge_tasks.push_back(make_unique<PhysicalOrderMergeTask>(shared_from_this(), context, gstate));
		}
		SetTasks(move(merge_tasks));
	}

class PhysicalOrderMergeTask : public ExecutorTask {
public:
	TaskExecutionResult ExecuteTask(TaskExecutionMode mode) override {
		// Initialize merge sorted and iterate until done
		auto &global_sort_state = state.global_sort_state;
		MergeSorter merge_sorter(global_sort_state, BufferManager::GetBufferManager(context));
		
        // 加鎖，獲取兩路，不斷進行兩路歸併，最終完成全域性排序。
	while (true) {
		{
			lock_guard<mutex> pair_guard(state.lock);
			if (state.pair_idx == state.num_pairs) {
				break;
			}
			GetNextPartition();
		}
		MergePartition();
	}
		event->FinishTask();
		return TaskExecutionResult::TASK_FINISHED;
	}

聚合運算元(這裡分析的是Prefetch Agg Operator運算元)

Sink介面：和Sort運算元一樣，這裡拆分為Group Chunk和Aggregate Input Chunk，可以理解為代表聚合時的key與value列。注意此時Sink介面上的聚合是在LocalSinkState上完成的。

SinkResultType PhysicalPerfectHashAggregate::Sink(ExecutionContext &context, GlobalSinkState &state,
                                                  LocalSinkState &lstate_p, DataChunk &input) const {
	lstate.ht->AddChunk(group_chunk, aggregate_input_chunk);
}


void PerfectAggregateHashTable::AddChunk(DataChunk &groups, DataChunk &payload) {
	auto address_data = FlatVector::GetData<uintptr_t>(addresses);
	memset(address_data, 0, groups.size() * sizeof(uintptr_t));
	D_ASSERT(groups.ColumnCount() == group_minima.size());

	// 計算group key列對應的entry的位置
	idx_t current_shift = total_required_bits;
	for (idx_t i = 0; i < groups.ColumnCount(); i++) {
		current_shift -= required_bits[i];
		ComputeGroupLocation(groups.data[i], group_minima[i], address_data, current_shift, groups.size());
	}

	// 通過data加上面的entry位置 + tuple的偏移量，計算出對應的記憶體地址，並進行init
	idx_t needs_init = 0;
	for (idx_t i = 0; i < groups.size(); i++) {
		D_ASSERT(address_data[i] < total_groups);
		const auto group = address_data[i];
		address_data[i] = uintptr_t(data) + address_data[i] * tuple_size;
	}
	RowOperations::InitializeStates(layout, addresses, sel, needs_init);

	// after finding the group location we update the aggregates
	idx_t payload_idx = 0;
	auto &aggregates = layout.GetAggregates();
	for (idx_t aggr_idx = 0; aggr_idx < aggregates.size(); aggr_idx++) {
		auto &aggregate = aggregates[aggr_idx];
		auto input_count = (idx_t)aggregate.child_count;
                // 進行聚合的Update操作
		RowOperations::UpdateStates(aggregate, addresses, payload, payload_idx, payload.size());
	}
}

Combine介面：加鎖，merge local hash table 與 global hash table

void PhysicalPerfectHashAggregate::Combine(ExecutionContext &context, GlobalSinkState &gstate_p,
                                           LocalSinkState &lstate_p) const {
	auto &lstate = (PerfectHashAggregateLocalState &)lstate_p;
	auto &gstate = (PerfectHashAggregateGlobalState &)gstate_p;

	lock_guard<mutex> l(gstate.lock);
	gstate.ht->Combine(*lstate.ht);
}

        // local state的地址vector
	Vector source_addresses(LogicalType::POINTER);
       // global state的地址vector
	Vector target_addresses(LogicalType::POINTER);
	auto source_addresses_ptr = FlatVector::GetData<data_ptr_t>(source_addresses);
	auto target_addresses_ptr = FlatVector::GetData<data_ptr_t>(target_addresses);

	// 遍歷所有hash table的表，然後進行合併對應能夠合併的key
	data_ptr_t source_ptr = other.data;
	data_ptr_t target_ptr = data;
	idx_t combine_count = 0;
	idx_t reinit_count = 0;
	const auto &reinit_sel = *FlatVector::IncrementalSelectionVector();
	for (idx_t i = 0; i < total_groups; i++) {
		auto has_entry_source = other.group_is_set[i];
		// we only have any work to do if the source has an entry for this group
		if (has_entry_source) {
			auto has_entry_target = group_is_set[i];
			if (has_entry_target) {
				// both source and target have an entry: need to combine
				source_addresses_ptr[combine_count] = source_ptr;
				target_addresses_ptr[combine_count] = target_ptr;
				combine_count++;
				if (combine_count == STANDARD_VECTOR_SIZE) {
					RowOperations::CombineStates(layout, source_addresses, target_addresses, combine_count);
					combine_count = 0;
				}
			} else {
				group_is_set[i] = true;
				// only source has an entry for this group: we can just memcpy it over
				memcpy(target_ptr, source_ptr, tuple_size);
				// we clear this entry in the other HT as we "consume" the entry here
				other.group_is_set[i] = false;
			}
		}
		source_ptr += tuple_size;
		target_ptr += tuple_size;
	}

        // 做對應的merge操作
	RowOperations::CombineStates(layout, source_addresses, target_addresses, combine_count);

Join運算元

Sink介面：和Sort運算元一樣，注意此時Sink介面上的hash 表是在LocalSinkState上完成的。

SinkResultType PhysicalHashJoin::Sink(ExecutionContext &context, GlobalSinkState &gstate_p, LocalSinkState &lstate_p,
                                      DataChunk &input) const {
	auto &gstate = (HashJoinGlobalSinkState &)gstate_p;
	auto &lstate = (HashJoinLocalSinkState &)lstate_p;

	lstate.join_keys.Reset();
	lstate.build_executor.Execute(input, lstate.join_keys);
	// build the HT
	auto &ht = *lstate.hash_table;
	if (!right_projection_map.empty()) {
		// there is a projection map: fill the build chunk with the projected columns
		lstate.build_chunk.Reset();
		lstate.build_chunk.SetCardinality(input);
		for (idx_t i = 0; i < right_projection_map.size(); i++) {
			lstate.build_chunk.data[i].Reference(input.data[right_projection_map[i]]);
		}
                // 構建local state的hash 表
		ht.Build(lstate.join_keys, lstate.build_chunk)

	return SinkResultType::NEED_MORE_INPUT;
}

Combine介面：加鎖，拷貝local state的hash表到global state

void PhysicalHashJoin::Combine(ExecutionContext &context, GlobalSinkState &gstate_p, LocalSinkState &lstate_p) const {
	auto &gstate = (HashJoinGlobalSinkState &)gstate_p;
	auto &lstate = (HashJoinLocalSinkState &)lstate_p;
	if (lstate.hash_table) {
		lock_guard<mutex> local_ht_lock(gstate.lock);
		gstate.local_hash_tables.push_back(move(lstate.hash_table));
	}
}

MergeTask：啟動核數相同的task來進行Hash table的Merge (這裡可以看出DuckDB對於多執行緒的使用是很激進的)，每個任務merge一部分Block(DuckDB之中的行資料，落盤使用）

void Schedule() override {
		auto &context = pipeline->GetClientContext();

		vector<unique_ptr<Task>> finalize_tasks;
		auto &ht = *sink.hash_table;
		const auto &block_collection = ht.GetBlockCollection();
		const auto &blocks = block_collection.blocks;
		const auto num_blocks = blocks.size();
		if (block_collection.count < PARALLEL_CONSTRUCT_THRESHOLD && !context.config.verify_parallelism) {
			// Single-threaded finalize
			finalize_tasks.push_back(
			    make_unique<HashJoinFinalizeTask>(shared_from_this(), context, sink, 0, num_blocks, false));
		} else {
			// Parallel finalize
			idx_t num_threads = TaskScheduler::GetScheduler(context).NumberOfThreads();
			auto blocks_per_thread = MaxValue<idx_t>((num_blocks + num_threads - 1) / num_threads, 1);

			idx_t block_idx = 0;
			for (idx_t thread_idx = 0; thread_idx < num_threads; thread_idx++) {
				auto block_idx_start = block_idx;
				auto block_idx_end = MinValue<idx_t>(block_idx_start + blocks_per_thread, num_blocks);
				finalize_tasks.push_back(make_unique<HashJoinFinalizeTask>(shared_from_this(), context, sink,
				                                                           block_idx_start, block_idx_end, true));
				block_idx = block_idx_end;
				if (block_idx == num_blocks) {
					break;
				}
			}
		}
		SetTasks(move(finalize_tasks));
	}

template <bool PARALLEL>
static inline void InsertHashesLoop(atomic<data_ptr_t> pointers[], const hash_t indices[], const idx_t count,
                                    const data_ptr_t key_locations[], const idx_t pointer_offset) {
	for (idx_t i = 0; i < count; i++) {
		auto index = indices[i];
		if (PARALLEL) {
			data_ptr_t head;
			do {
				head = pointers[index];
				Store<data_ptr_t>(head, key_locations[i] + pointer_offset);
			} while (!std::atomic_compare_exchange_weak(&pointers[index], &head, key_locations[i]));
		} else {
			// set prev in current key to the value (NOTE: this will be nullptr if there is none)
			Store<data_ptr_t>(pointers[index], key_locations[i] + pointer_offset);

			// set pointer to current tuple
			pointers[index] = key_locations[i];
		}
	}
}

並行掃描hash表，進行outer資料的處理：

void PhysicalHashJoin::GetData(ExecutionContext &context, DataChunk &chunk, GlobalSourceState &gstate_p,
                               LocalSourceState &lstate_p) const {
	auto &sink = (HashJoinGlobalSinkState &)*sink_state;
	auto &gstate = (HashJoinGlobalSourceState &)gstate_p;
	auto &lstate = (HashJoinLocalSourceState &)lstate_p;
	sink.scanned_data = true;

	if (!sink.external) {
		if (IsRightOuterJoin(join_type)) {
			{
				lock_guard<mutex> guard(gstate.lock);
                                // 拆解掃描部分hash表的資料
				lstate.ScanFullOuter(sink, gstate);
			}
                        // 掃描hash表讀取資料
			sink.hash_table->GatherFullOuter(chunk, lstate.addresses, lstate.full_outer_found_entries);
		}
		return;
	}
}


void HashJoinLocalSourceState::ScanFullOuter(HashJoinGlobalSinkState &sink, HashJoinGlobalSourceState &gstate) {
	auto &fo_ss = gstate.full_outer_scan;
	idx_t scan_index_before = fo_ss.scan_index;
	full_outer_found_entries = sink.hash_table->ScanFullOuter(fo_ss, addresses);
	idx_t scanned = fo_ss.scan_index - scan_index_before;
	full_outer_in_progress = scanned;
}

小結

DuckDB在多執行緒同步，核心就是在Combine的時候：加鎖，並行是通過原子變數的方式實現並行寫入hash表的操作
通過local/global 拆分私有記憶體和公共記憶體，並行的基礎是在私有記憶體上進行運算，同步的部分主要在公有記憶體的更新

3. Spill To Disk的實現

DuckDB並沒有如筆者預期的實現非同步IO，所以任意的執行執行緒是有可能Stall在系統的I/O排程上的，我想大概率是DuckDB本身的定位對於高並行場景的支援不是那麼敏感所導致的。這裡他們也作為了後續TODO的計劃之一。

4. 參考資料

DuckDB原始碼

Push-Based Execution in DuckDB

[​DuckDB] 多核運算元並行的原始碼解析