OmniSciDB  16c4e035a1
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Groups Pages
ExecutionKernel Class Reference

#include <ExecutionKernel.h>

+ Collaboration diagram for ExecutionKernel:

Public Member Functions

 ExecutionKernel (const RelAlgExecutionUnit &ra_exe_unit, const ExecutorDeviceType chosen_device_type, int chosen_device_id, const ExecutionOptions &eo, const ColumnFetcher &column_fetcher, const QueryCompilationDescriptor &query_comp_desc, const QueryMemoryDescriptor &query_mem_desc, const FragmentsList &frag_list, const ExecutorDispatchMode kernel_dispatch_mode, RenderInfo *render_info, const int64_t rowid_lookup_key)
 
void run (Executor *executor, const size_t thread_idx, SharedKernelContext &shared_context)
 

Public Attributes

const RelAlgExecutionUnitra_exe_unit_
 

Private Member Functions

void runImpl (Executor *executor, const size_t thread_idx, SharedKernelContext &shared_context)
 

Private Attributes

const ExecutorDeviceType chosen_device_type
 
int chosen_device_id
 
const ExecutionOptionseo
 
const ColumnFetchercolumn_fetcher
 
const QueryCompilationDescriptorquery_comp_desc
 
const QueryMemoryDescriptorquery_mem_desc
 
const FragmentsList frag_list
 
const ExecutorDispatchMode kernel_dispatch_mode
 
RenderInforender_info_
 
const int64_t rowid_lookup_key
 
ResultSetPtr device_results_
 

Friends

class KernelSubtask
 

Detailed Description

Definition at line 72 of file ExecutionKernel.h.

Constructor & Destructor Documentation

ExecutionKernel::ExecutionKernel ( const RelAlgExecutionUnit ra_exe_unit,
const ExecutorDeviceType  chosen_device_type,
int  chosen_device_id,
const ExecutionOptions eo,
const ColumnFetcher column_fetcher,
const QueryCompilationDescriptor query_comp_desc,
const QueryMemoryDescriptor query_mem_desc,
const FragmentsList frag_list,
const ExecutorDispatchMode  kernel_dispatch_mode,
RenderInfo render_info,
const int64_t  rowid_lookup_key 
)
inline

Definition at line 74 of file ExecutionKernel.h.

85  : ra_exe_unit_(ra_exe_unit)
88  , eo(eo)
89  , column_fetcher(column_fetcher)
90  , query_comp_desc(query_comp_desc)
91  , query_mem_desc(query_mem_desc)
94  , render_info_(render_info)
const ExecutionOptions & eo
const ExecutorDispatchMode kernel_dispatch_mode
const RelAlgExecutionUnit & ra_exe_unit_
const int64_t rowid_lookup_key
const ExecutorDeviceType chosen_device_type
RenderInfo * render_info_
const QueryMemoryDescriptor & query_mem_desc
const QueryCompilationDescriptor & query_comp_desc
const FragmentsList frag_list
const ColumnFetcher & column_fetcher

Member Function Documentation

void ExecutionKernel::run ( Executor executor,
const size_t  thread_idx,
SharedKernelContext shared_context 
)

Definition at line 122 of file ExecutionKernel.cpp.

References DEBUG_TIMER, Executor::ERR_COLUMNAR_CONVERSION_NOT_SUPPORTED, Executor::ERR_OUT_OF_CPU_MEM, Executor::ERR_OUT_OF_GPU_MEM, Executor::ERR_OUT_OF_RENDER_MEM, Executor::ERR_STRING_CONST_IN_RESULTSET, Executor::ERR_TOO_MANY_LITERALS, QueryMemoryDescriptor::getQueryDescriptionType(), INJECT_TIMER, kernel_dispatch_mode, MultifragmentKernel, query_mem_desc, RelAlgExecutionUnit::query_state, ra_exe_unit_, runImpl(), and OutOfHostMemory::what().

Referenced by Executor::executeUpdate(), and Executor::executeWorkUnitPerFragment().

124  {
125  DEBUG_TIMER("ExecutionKernel::run");
126  INJECT_TIMER(kernel_run);
127  std::optional<logger::QidScopeGuard> qid_scope_guard;
129  qid_scope_guard.emplace(ra_exe_unit_.query_state->setThreadLocalQueryId());
130  }
131  try {
132  runImpl(executor, thread_idx, shared_context);
133  } catch (const OutOfHostMemory& e) {
135  } catch (const std::bad_alloc& e) {
137  } catch (const OutOfRenderMemory& e) {
139  } catch (const OutOfMemory& e) {
140  throw QueryExecutionError(
142  e.what(),
146  } catch (const ColumnarConversionNotSupported& e) {
148  } catch (const TooManyLiterals& e) {
150  } catch (const StringConstInResultSet& e) {
152  } catch (const QueryExecutionError& e) {
153  throw e;
154  }
155 }
const ExecutorDispatchMode kernel_dispatch_mode
const RelAlgExecutionUnit & ra_exe_unit_
static const int32_t ERR_TOO_MANY_LITERALS
Definition: Execute.h:1174
static const int32_t ERR_STRING_CONST_IN_RESULTSET
Definition: Execute.h:1175
static const int32_t ERR_COLUMNAR_CONVERSION_NOT_SUPPORTED
Definition: Execute.h:1173
#define INJECT_TIMER(DESC)
Definition: measure.h:93
static const int32_t ERR_OUT_OF_RENDER_MEM
Definition: Execute.h:1168
const QueryMemoryDescriptor & query_mem_desc
static const int32_t ERR_OUT_OF_GPU_MEM
Definition: Execute.h:1165
QueryDescriptionType getQueryDescriptionType() const
void runImpl(Executor *executor, const size_t thread_idx, SharedKernelContext &shared_context)
#define DEBUG_TIMER(name)
Definition: Logger.h:358
const char * what() const noexceptfinal
Definition: checked_alloc.h:39
std::shared_ptr< const query_state::QueryState > query_state
static const int32_t ERR_OUT_OF_CPU_MEM
Definition: Execute.h:1169

+ Here is the call graph for this function:

+ Here is the caller graph for this function:

void ExecutionKernel::runImpl ( Executor executor,
const size_t  thread_idx,
SharedKernelContext shared_context 
)
private

Definition at line 157 of file ExecutionKernel.cpp.

References gpu_enabled::accumulate(), SharedKernelContext::addDeviceResults(), ExecutionOptions::allow_runtime_query_interrupt, CHECK, CHECK_EQ, CHECK_GE, CHECK_GT, CHECK_LT, chosen_device_id, chosen_device_type, column_fetcher, QueryFragmentDescriptor::computeAllTablesFragments(), CPU, Data_Namespace::CPU_LEVEL, device_results_, dynamic_watchdog_init(), SharedKernelContext::dynamic_watchdog_set, ExecutionOptions::dynamic_watchdog_time_limit, eo, Executor::ERR_OUT_OF_CPU_MEM, Executor::ERR_OUT_OF_GPU_MEM, RelAlgExecutionUnit::estimator, ExecutionOptions::executor_type, Extern, frag_list, g_cpu_sub_task_size, QueryCompilationDescriptor::getCompilationResult(), SharedKernelContext::getFragOffsets(), QueryMemoryDescriptor::getQueryDescriptionType(), QueryMemoryDescriptor::getQueryExecutionContext(), SharedKernelContext::getQueryInfos(), GPU, Data_Namespace::GPU_LEVEL, RelAlgExecutionUnit::groupby_exprs, QueryCompilationDescriptor::hoistLiterals(), logger::INFO, RelAlgExecutionUnit::input_descs, RenderInfo::isPotentialInSituRender(), kernel_dispatch_mode, KernelPerFragment, LOG, Executor::max_gpu_count, MultifragmentKernel, Native, anonymous_namespace{ExecutionKernel.cpp}::need_to_hold_chunk(), CompilationResult::output_columnar, Projection, query_comp_desc, anonymous_namespace{ExecutionKernel.cpp}::query_has_inner_join(), query_mem_desc, ra_exe_unit_, render_info_, rowid_lookup_key, run_query_external(), RelAlgExecutionUnit::scan_limit, serialize_to_sql(), QueryMemoryDescriptor::sortOnGpu(), RelAlgExecutionUnit::target_exprs, target_exprs_to_infos(), to_string(), RelAlgExecutionUnit::union_all, VLOG, and ExecutionOptions::with_dynamic_watchdog.

Referenced by run().

159  {
160  CHECK(executor);
161  const auto memory_level = chosen_device_type == ExecutorDeviceType::GPU
164  CHECK_GE(frag_list.size(), size_t(1));
165  // frag_list[0].table_id is how we tell which query we are running for UNION ALL.
166  const int outer_table_id = ra_exe_unit_.union_all
167  ? frag_list[0].table_id
168  : ra_exe_unit_.input_descs[0].getTableId();
169  CHECK_EQ(frag_list[0].table_id, outer_table_id);
170  const auto& outer_tab_frag_ids = frag_list[0].fragment_ids;
171 
174 
175  auto catalog = executor->getCatalog();
176  CHECK(catalog);
177 
178  auto data_mgr = executor->getDataMgr();
179 
180  // need to own them while query executes
181  auto chunk_iterators_ptr = std::make_shared<std::list<ChunkIter>>();
182  std::list<std::shared_ptr<Chunk_NS::Chunk>> chunks;
183  std::unique_ptr<std::lock_guard<std::mutex>> gpu_lock;
184  std::unique_ptr<CudaAllocator> device_allocator;
186  gpu_lock.reset(
187  new std::lock_guard<std::mutex>(executor->gpu_exec_mutex_[chosen_device_id]));
188  device_allocator = std::make_unique<CudaAllocator>(data_mgr, chosen_device_id);
189  }
190  std::shared_ptr<FetchResult> fetch_result(new FetchResult);
191  try {
192  std::map<int, const TableFragments*> all_tables_fragments;
194  all_tables_fragments, ra_exe_unit_, shared_context.getQueryInfos());
195 
196  *fetch_result = ra_exe_unit_.union_all
197  ? executor->fetchUnionChunks(column_fetcher,
198  ra_exe_unit_,
200  memory_level,
201  all_tables_fragments,
202  frag_list,
203  *catalog,
204  *chunk_iterators_ptr,
205  chunks,
206  device_allocator.get(),
207  thread_idx,
209  : executor->fetchChunks(column_fetcher,
210  ra_exe_unit_,
212  memory_level,
213  all_tables_fragments,
214  frag_list,
215  *catalog,
216  *chunk_iterators_ptr,
217  chunks,
218  device_allocator.get(),
219  thread_idx,
221  if (fetch_result->num_rows.empty()) {
222  return;
223  }
225  !shared_context.dynamic_watchdog_set.test_and_set(std::memory_order_acquire)) {
228  LOG(INFO) << "Dynamic Watchdog budget: CPU: "
230  << std::to_string(cycle_budget) << " cycles";
231  }
232  } catch (const OutOfMemory&) {
233  throw QueryExecutionError(
239  return;
240  }
241 
243  if (ra_exe_unit_.input_descs.size() > 1) {
244  throw std::runtime_error("Joins not supported through external execution");
245  }
246  const auto query = serialize_to_sql(&ra_exe_unit_, catalog);
247  GroupByAndAggregate group_by_and_aggregate(executor,
249  ra_exe_unit_,
250  shared_context.getQueryInfos(),
251  executor->row_set_mem_owner_,
252  std::nullopt);
253  const auto query_mem_desc =
254  group_by_and_aggregate.initQueryMemoryDescriptor(false, 0, 8, nullptr, false);
256  query,
257  *fetch_result,
258  executor->plan_state_.get(),
262  executor});
263  shared_context.addDeviceResults(std::move(device_results_), outer_tab_frag_ids);
264  return;
265  }
266  const CompilationResult& compilation_result = query_comp_desc.getCompilationResult();
267  std::unique_ptr<QueryExecutionContext> query_exe_context_owned;
268  const bool do_render = render_info_ && render_info_->isPotentialInSituRender();
269 
270  int64_t total_num_input_rows{-1};
272  query_mem_desc.getQueryDescriptionType() == QueryDescriptionType::Projection) {
273  total_num_input_rows = 0;
274  std::for_each(fetch_result->num_rows.begin(),
275  fetch_result->num_rows.end(),
276  [&total_num_input_rows](const std::vector<int64_t>& frag_row_count) {
277  total_num_input_rows = std::accumulate(frag_row_count.begin(),
278  frag_row_count.end(),
279  total_num_input_rows);
280  });
281  VLOG(2) << "total_num_input_rows=" << total_num_input_rows;
282  // TODO(adb): we may want to take this early out for all queries, but we are most
283  // likely to see this query pattern on the kernel per fragment path (e.g. with HAVING
284  // 0=1)
285  if (total_num_input_rows == 0) {
286  return;
287  }
288 
290  total_num_input_rows *= ra_exe_unit_.input_descs.size();
291  }
292  }
293 
294  uint32_t start_rowid{0};
295  if (rowid_lookup_key >= 0) {
296  if (!frag_list.empty()) {
297  const auto& all_frag_row_offsets = shared_context.getFragOffsets();
298  start_rowid = rowid_lookup_key -
299  all_frag_row_offsets[frag_list.begin()->fragment_ids.front()];
300  }
301  }
302 
303 #ifdef HAVE_TBB
304  bool can_run_subkernels = shared_context.getThreadPool() != nullptr;
305 
306  // Sub-tasks are supported for groupby queries and estimators only for now.
307  bool is_groupby =
308  (ra_exe_unit_.groupby_exprs.size() > 1) ||
309  (ra_exe_unit_.groupby_exprs.size() == 1 && ra_exe_unit_.groupby_exprs.front());
310  can_run_subkernels = can_run_subkernels && (is_groupby || ra_exe_unit_.estimator);
311 
312  // In case some column is lazily fetched, we cannot mix different fragments in a single
313  // ResultSet.
314  can_run_subkernels =
315  can_run_subkernels && !executor->hasLazyFetchColumns(ra_exe_unit_.target_exprs);
316 
317  // TODO: Use another structure to hold chunks. Currently, ResultSet holds them, but with
318  // sub-tasks chunk can be referenced by many ResultSets. So, some outer structure to
319  // hold all ResultSets and all chunks is required.
320  can_run_subkernels =
321  can_run_subkernels &&
323  chunks, ra_exe_unit_, std::vector<ColumnLazyFetchInfo>(), chosen_device_type);
324 
325  // TODO: check for literals? We serialize literals before execution and hold them in
326  // result sets. Can we simply do it once and holdin an outer structure?
327  if (can_run_subkernels) {
328  size_t total_rows = fetch_result->num_rows[0][0];
329  size_t sub_size = g_cpu_sub_task_size;
330 
331  for (size_t sub_start = start_rowid; sub_start < total_rows; sub_start += sub_size) {
332  sub_size = (sub_start + sub_size > total_rows) ? total_rows - sub_start : sub_size;
333  auto subtask = std::make_shared<KernelSubtask>(*this,
334  shared_context,
335  fetch_result,
336  chunk_iterators_ptr,
337  total_num_input_rows,
338  sub_start,
339  sub_size,
340  thread_idx);
341  shared_context.getThreadPool()->run(
342  [subtask, executor] { subtask->run(executor); });
343  }
344 
345  return;
346  }
347 #endif // HAVE_TBB
348 
350  try {
351  // std::unique_ptr<QueryExecutionContext> query_exe_context_owned
352  // has std::unique_ptr<QueryMemoryInitializer> query_buffers_
353  // has std::vector<std::unique_ptr<ResultSet>> result_sets_
354  // has std::unique_ptr<ResultSetStorage> storage_
355  // which are initialized and possibly allocated here.
356  query_exe_context_owned =
357  query_mem_desc.getQueryExecutionContext(ra_exe_unit_,
358  executor,
362  outer_table_id,
363  total_num_input_rows,
364  fetch_result->col_buffers,
365  fetch_result->frag_offsets,
366  executor->getRowSetMemoryOwner(),
367  compilation_result.output_columnar,
368  query_mem_desc.sortOnGpu(),
369  thread_idx,
370  do_render ? render_info_ : nullptr);
371  } catch (const OutOfHostMemory& e) {
372  throw QueryExecutionError(Executor::ERR_OUT_OF_CPU_MEM);
373  }
374  }
375  QueryExecutionContext* query_exe_context{query_exe_context_owned.get()};
376  CHECK(query_exe_context);
377  int32_t err{0};
378 
379  if (ra_exe_unit_.groupby_exprs.empty()) {
380  err = executor->executePlanWithoutGroupBy(ra_exe_unit_,
381  compilation_result,
386  fetch_result->col_buffers,
387  query_exe_context,
388  fetch_result->num_rows,
389  fetch_result->frag_offsets,
390  data_mgr,
392  start_rowid,
393  ra_exe_unit_.input_descs.size(),
395  do_render ? render_info_ : nullptr);
396  } else {
397  if (ra_exe_unit_.union_all) {
398  VLOG(1) << "outer_table_id=" << outer_table_id
399  << " ra_exe_unit_.scan_limit=" << ra_exe_unit_.scan_limit;
400  }
401  err = executor->executePlanWithGroupBy(ra_exe_unit_,
402  compilation_result,
406  fetch_result->col_buffers,
407  outer_tab_frag_ids,
408  query_exe_context,
409  fetch_result->num_rows,
410  fetch_result->frag_offsets,
411  data_mgr,
413  outer_table_id,
415  start_rowid,
416  ra_exe_unit_.input_descs.size(),
418  do_render ? render_info_ : nullptr);
419  }
420  if (device_results_) {
421  std::list<std::shared_ptr<Chunk_NS::Chunk>> chunks_to_hold;
422  for (const auto& chunk : chunks) {
423  if (need_to_hold_chunk(chunk.get(),
424  ra_exe_unit_,
425  device_results_->getLazyFetchInfo(),
427  chunks_to_hold.push_back(chunk);
428  }
429  }
430  device_results_->holdChunks(chunks_to_hold);
431  device_results_->holdChunkIterators(chunk_iterators_ptr);
432  } else {
433  VLOG(1) << "null device_results.";
434  }
435  if (err) {
436  throw QueryExecutionError(err);
437  }
438  shared_context.addDeviceResults(std::move(device_results_), outer_tab_frag_ids);
439 }
bool need_to_hold_chunk(const Chunk_NS::Chunk *chunk, const RelAlgExecutionUnit &ra_exe_unit, const std::vector< ColumnLazyFetchInfo > &lazy_fetch_info, const ExecutorDeviceType device_type)
std::vector< Analyzer::Expr * > target_exprs
#define CHECK_EQ(x, y)
Definition: Logger.h:219
std::atomic_flag dynamic_watchdog_set
const ExecutionOptions & eo
size_t g_cpu_sub_task_size
Definition: Execute.cpp:79
const std::vector< uint64_t > & getFragOffsets()
static const int max_gpu_count
Definition: Execute.h:1080
const std::optional< bool > union_all
#define LOG(tag)
Definition: Logger.h:205
const ExecutorDispatchMode kernel_dispatch_mode
const RelAlgExecutionUnit & ra_exe_unit_
const int64_t rowid_lookup_key
void addDeviceResults(ResultSetPtr &&device_results, std::vector< size_t > outer_table_fragment_ids)
std::vector< InputDescriptor > input_descs
const ExecutorDeviceType chosen_device_type
#define CHECK_GE(x, y)
Definition: Logger.h:224
const std::list< std::shared_ptr< Analyzer::Expr > > groupby_exprs
std::unique_ptr< ResultSet > run_query_external(const ExecutionUnitSql &sql, const FetchResult &fetch_result, const PlanState *plan_state, const ExternalQueryOutputSpec &output_spec)
RenderInfo * render_info_
#define CHECK_GT(x, y)
Definition: Logger.h:223
std::string to_string(char const *&&v)
ExecutorType executor_type
const QueryMemoryDescriptor & query_mem_desc
DEVICE auto accumulate(ARGS &&...args)
Definition: gpu_enabled.h:42
const QueryCompilationDescriptor & query_comp_desc
const std::shared_ptr< Analyzer::Estimator > estimator
static const int32_t ERR_OUT_OF_GPU_MEM
Definition: Execute.h:1165
QueryDescriptionType getQueryDescriptionType() const
RUNTIME_EXPORT uint64_t dynamic_watchdog_init(unsigned ms_budget)
#define CHECK_LT(x, y)
Definition: Logger.h:221
const FragmentsList frag_list
bool query_has_inner_join(const RelAlgExecutionUnit &ra_exe_unit)
const std::vector< InputTableInfo > & getQueryInfos() const
ResultSetPtr device_results_
bool isPotentialInSituRender() const
Definition: RenderInfo.cpp:63
#define CHECK(condition)
Definition: Logger.h:211
std::vector< TargetInfo > target_exprs_to_infos(const std::vector< Analyzer::Expr * > &targets, const QueryMemoryDescriptor &query_mem_desc)
unsigned dynamic_watchdog_time_limit
ExecutionUnitSql serialize_to_sql(const RelAlgExecutionUnit *ra_exe_unit, const Catalog_Namespace::Catalog *catalog)
static const int32_t ERR_OUT_OF_CPU_MEM
Definition: Execute.h:1169
static void computeAllTablesFragments(std::map< int, const TableFragments * > &all_tables_fragments, const RelAlgExecutionUnit &ra_exe_unit, const std::vector< InputTableInfo > &query_infos)
#define VLOG(n)
Definition: Logger.h:305
const ColumnFetcher & column_fetcher

+ Here is the call graph for this function:

+ Here is the caller graph for this function:

Friends And Related Function Documentation

friend class KernelSubtask
friend

Definition at line 121 of file ExecutionKernel.h.

Member Data Documentation

int ExecutionKernel::chosen_device_id
private

Definition at line 105 of file ExecutionKernel.h.

Referenced by runImpl().

const ExecutorDeviceType ExecutionKernel::chosen_device_type
private

Definition at line 104 of file ExecutionKernel.h.

Referenced by runImpl().

const ColumnFetcher& ExecutionKernel::column_fetcher
private

Definition at line 107 of file ExecutionKernel.h.

Referenced by runImpl().

ResultSetPtr ExecutionKernel::device_results_
private

Definition at line 115 of file ExecutionKernel.h.

Referenced by runImpl().

const ExecutionOptions& ExecutionKernel::eo
private

Definition at line 106 of file ExecutionKernel.h.

Referenced by runImpl().

const FragmentsList ExecutionKernel::frag_list
private

Definition at line 110 of file ExecutionKernel.h.

Referenced by runImpl().

const ExecutorDispatchMode ExecutionKernel::kernel_dispatch_mode
private

Definition at line 111 of file ExecutionKernel.h.

Referenced by run(), and runImpl().

const QueryCompilationDescriptor& ExecutionKernel::query_comp_desc
private

Definition at line 108 of file ExecutionKernel.h.

Referenced by runImpl().

const QueryMemoryDescriptor& ExecutionKernel::query_mem_desc
private

Definition at line 109 of file ExecutionKernel.h.

Referenced by run(), and runImpl().

const RelAlgExecutionUnit& ExecutionKernel::ra_exe_unit_

Definition at line 101 of file ExecutionKernel.h.

Referenced by run(), and runImpl().

RenderInfo* ExecutionKernel::render_info_
private

Definition at line 112 of file ExecutionKernel.h.

Referenced by runImpl().

const int64_t ExecutionKernel::rowid_lookup_key
private

Definition at line 113 of file ExecutionKernel.h.

Referenced by runImpl().


The documentation for this class was generated from the following files: