#include <QueryFragmentDescriptor.h>

Public Member Functions
	QueryFragmentDescriptor (const RelAlgExecutionUnit &ra_exe_unit, const std::vector< InputTableInfo > &query_infos, const std::vector< Data_Namespace::MemoryInfo > &gpu_mem_infos, const double gpu_input_mem_limit_percent)

void	buildFragmentKernelMap (const RelAlgExecutionUnit &ra_exe_unit, const std::vector< uint64_t > &frag_offsets, const int device_count, const ExecutorDeviceType &device_type, const bool enable_multifrag_kernels, const bool enable_inner_join_fragment_skipping, Executor *executor)

template<typename DISPATCH_FCN >
void	assignFragsToMultiDispatch (DISPATCH_FCN f) const

template<typename DISPATCH_FCN >
void	assignFragsToKernelDispatch (DISPATCH_FCN f, const RelAlgExecutionUnit &ra_exe_unit) const

bool	shouldCheckWorkUnitWatchdog () const

Static Public Member Functions
static void	computeAllTablesFragments (std::map< int, const TableFragments * > &all_tables_fragments, const RelAlgExecutionUnit &ra_exe_unit, const std::vector< InputTableInfo > &query_infos)

Protected Member Functions
void	buildFragmentPerKernelMap (const RelAlgExecutionUnit &ra_exe_unit, const std::vector< uint64_t > &frag_offsets, const int device_count, const ExecutorDeviceType &device_type, Executor *executor)

void	buildMultifragKernelMap (const RelAlgExecutionUnit &ra_exe_unit, const std::vector< uint64_t > &frag_offsets, const int device_count, const ExecutorDeviceType &device_type, const bool enable_inner_join_fragment_skipping, Executor *executor)

bool	terminateDispatchMaybe (size_t &tuple_count, const RelAlgExecutionUnit &ra_exe_unit, const ExecutionKernel &kernel) const

void	checkDeviceMemoryUsage (const Fragmenter_Namespace::FragmentInfo &fragment, const int device_id, const size_t num_cols)

Protected Attributes
size_t	outer_fragments_size_ = 0

int64_t	rowid_lookup_key_ = -1

std::map< int, const TableFragments * >	selected_tables_fragments_

std::map< int, std::vector < ExecutionKernel > >	execution_kernels_per_device_

double	gpu_input_mem_limit_percent_

std::map< size_t, size_t >	tuple_count_per_device_

std::map< size_t, size_t >	available_gpu_mem_bytes_

Detailed Description

Definition at line 64 of file QueryFragmentDescriptor.h.

Constructor & Destructor Documentation

QueryFragmentDescriptor::QueryFragmentDescriptor	(	const RelAlgExecutionUnit &	ra_exe_unit,
		const std::vector< InputTableInfo > &	query_infos,
		const std::vector< Data_Namespace::MemoryInfo > &	gpu_mem_infos,
		const double	gpu_input_mem_limit_percent
	)

Definition at line 22 of file QueryFragmentDescriptor.cpp.

References available_gpu_mem_bytes_, CHECK_EQ, RelAlgExecutionUnit::input_descs, and selected_tables_fragments_.

     : gpu_input_mem_limit_percent_(gpu_input_mem_limit_percent) {
   const size_t input_desc_count{ra_exe_unit.input_descs.size()};
   CHECK_EQ(query_infos.size(), input_desc_count);
   for (size_t table_idx = 0; table_idx < input_desc_count; ++table_idx) {
     const auto table_id = ra_exe_unit.input_descs[table_idx].getTableId();
     if (!selected_tables_fragments_.count(table_id)) {
       selected_tables_fragments_[ra_exe_unit.input_descs[table_idx].getTableId()] =
           &query_infos[table_idx].info.fragments;
     }
   }
 
   for (size_t device_id = 0; device_id < gpu_mem_infos.size(); device_id++) {
     const auto& gpu_mem_info = gpu_mem_infos[device_id];
     available_gpu_mem_bytes_[device_id] =
         gpu_mem_info.maxNumPages * gpu_mem_info.pageSize;
   }
 }

Member Function Documentation

template<typename DISPATCH_FCN >

void QueryFragmentDescriptor::assignFragsToKernelDispatch	(	DISPATCH_FCN	f,
		const RelAlgExecutionUnit &	ra_exe_unit
	)		const

inline

Dispatch one fragment for each device. Iterate the device map and dispatch one kernel for each device per iteration. This allows balanced dispatch as well as early termination if the number of rows passing the kernel can be computed at dispatch time and the scan limit is reached.

Definition at line 106 of file QueryFragmentDescriptor.h.

References CHECK(), execution_kernels_per_device_, rowid_lookup_key_, and terminateDispatchMaybe().

Referenced by Executor::dispatchFragments().

                                                                                  {
     if (execution_kernels_per_device_.empty()) {
       return;
     }
 
     size_t tuple_count = 0;
 
     std::unordered_map<int, size_t> execution_kernel_index;
     for (const auto& device_itr : execution_kernels_per_device_) {
       CHECK(execution_kernel_index.insert(std::make_pair(device_itr.first, size_t(0)))
                 .second);
     }
 
     bool dispatch_finished = false;
     while (!dispatch_finished) {
       dispatch_finished = true;
       for (const auto& device_itr : execution_kernels_per_device_) {
         auto& kernel_idx = execution_kernel_index[device_itr.first];
         if (kernel_idx < device_itr.second.size()) {
           dispatch_finished = false;
           const auto& execution_kernel = device_itr.second[kernel_idx++];
           f(device_itr.first, execution_kernel.fragments, rowid_lookup_key_);
 
           if (terminateDispatchMaybe(tuple_count, ra_exe_unit, execution_kernel)) {
             return;
           }
         }
       }
     }
   }

Here is the call graph for this function:

Here is the caller graph for this function:

template<typename DISPATCH_FCN >

void QueryFragmentDescriptor::assignFragsToMultiDispatch ( DISPATCH_FCN f ) const

inline

Dispatch multi-fragment kernels. Currently GPU only. Each GPU should have only one kernel, with multiple fragments in its fragments list.

Definition at line 89 of file QueryFragmentDescriptor.h.

References CHECK_EQ, execution_kernels_per_device_, and rowid_lookup_key_.

Referenced by Executor::dispatchFragments().

                                                         {
     for (const auto& device_itr : execution_kernels_per_device_) {
       const auto& execution_kernels = device_itr.second;
       CHECK_EQ(execution_kernels.size(), size_t(1));
 
       const auto& fragments_list = execution_kernels.front().fragments;
       f(device_itr.first, fragments_list, rowid_lookup_key_);
     }
   }

Here is the caller graph for this function:

void QueryFragmentDescriptor::buildFragmentKernelMap	(	const RelAlgExecutionUnit &	ra_exe_unit,
		const std::vector< uint64_t > &	frag_offsets,
		const int	device_count,
		const ExecutorDeviceType &	device_type,
		const bool	enable_multifrag_kernels,
		const bool	enable_inner_join_fragment_skipping,
		Executor *	executor
	)

Definition at line 59 of file QueryFragmentDescriptor.cpp.

References buildFragmentPerKernelMap(), and buildMultifragKernelMap().

Referenced by Executor::dispatchFragments().

                         {
   if (enable_multifrag_kernels) {
     buildMultifragKernelMap(ra_exe_unit,
                             frag_offsets,
                             device_count,
                             device_type,
                             enable_inner_join_fragment_skipping,
                             executor);
   } else {
     buildFragmentPerKernelMap(
         ra_exe_unit, frag_offsets, device_count, device_type, executor);
   }
 }

Here is the call graph for this function:

Here is the caller graph for this function:

void QueryFragmentDescriptor::buildFragmentPerKernelMap	(	const RelAlgExecutionUnit &	ra_exe_unit,
		const std::vector< uint64_t > &	frag_offsets,
		const int	device_count,
		const ExecutorDeviceType &	device_type,
		Executor *	executor
	)

protected

Definition at line 93 of file QueryFragmentDescriptor.cpp.

References CHECK(), CHECK_GT, checkDeviceMemoryUsage(), anonymous_namespace{QueryFragmentDescriptor.cpp}::compute_fragment_tuple_count(), CPU, Data_Namespace::CPU_LEVEL, execution_kernels_per_device_, GPU, Data_Namespace::GPU_LEVEL, RelAlgExecutionUnit::input_descs, outer_fragments_size_, rowid_lookup_key_, selected_tables_fragments_, and RelAlgExecutionUnit::simple_quals.

Referenced by buildFragmentKernelMap().

                         {
   const auto& outer_table_desc = ra_exe_unit.input_descs.front();
   const int outer_table_id = outer_table_desc.getTableId();
   auto it = selected_tables_fragments_.find(outer_table_id);
   CHECK(it != selected_tables_fragments_.end());
   const auto outer_fragments = it->second;
   outer_fragments_size_ = outer_fragments->size();
 
   const auto num_bytes_for_row = executor->getNumBytesForFetchedRow();
 
   const ColumnDescriptor* deleted_cd{nullptr};
   if (outer_table_id > 0) {
     // Temporary tables will not have a table descriptor and will also not have deleted
     // rows.
     const auto& catalog = executor->getCatalog();
     const auto td = catalog->getMetadataForTable(outer_table_id);
     CHECK(td);
     deleted_cd = catalog->getDeletedColumnIfRowsDeleted(td);
   }
 
   for (size_t i = 0; i < outer_fragments->size(); ++i) {
     const auto& fragment = (*outer_fragments)[i];
     const auto skip_frag = executor->skipFragment(
         outer_table_desc, fragment, ra_exe_unit.simple_quals, frag_offsets, i);
     if (skip_frag.first) {
       continue;
     }
     rowid_lookup_key_ = std::max(rowid_lookup_key_, skip_frag.second);
     const int chosen_device_count =
         device_type == ExecutorDeviceType::CPU ? 1 : device_count;
     CHECK_GT(chosen_device_count, 0);
     const auto memory_level = device_type == ExecutorDeviceType::GPU
                                   ? Data_Namespace::GPU_LEVEL
                                   : Data_Namespace::CPU_LEVEL;
     int device_id = (device_type == ExecutorDeviceType::CPU || fragment.shard == -1)
                         ? fragment.deviceIds[static_cast<int>(memory_level)]
                         : fragment.shard % chosen_device_count;
 
     if (device_type == ExecutorDeviceType::GPU) {
       checkDeviceMemoryUsage(fragment, device_id, num_bytes_for_row);
     }
 
     ExecutionKernel execution_kernel{
         device_id, {}, compute_fragment_tuple_count(fragment, deleted_cd)};
     for (size_t j = 0; j < ra_exe_unit.input_descs.size(); ++j) {
       const auto frag_ids =
           executor->getTableFragmentIndices(ra_exe_unit,
                                             device_type,
                                             j,
                                             i,
                                             selected_tables_fragments_,
                                             executor->getInnerTabIdToJoinCond());
       const auto table_id = ra_exe_unit.input_descs[j].getTableId();
       auto table_frags_it = selected_tables_fragments_.find(table_id);
       CHECK(table_frags_it != selected_tables_fragments_.end());
 
       execution_kernel.fragments.emplace_back(FragmentsPerTable{table_id, frag_ids});
     }
 
     if (execution_kernels_per_device_.find(device_id) ==
         execution_kernels_per_device_.end()) {
       CHECK(execution_kernels_per_device_
                 .insert(std::make_pair(device_id,
                                        std::vector<ExecutionKernel>{execution_kernel}))
                 .second);
     } else {
       execution_kernels_per_device_[device_id].emplace_back(execution_kernel);
     }
   }
 }

Here is the call graph for this function:

Here is the caller graph for this function:

void QueryFragmentDescriptor::buildMultifragKernelMap	(	const RelAlgExecutionUnit &	ra_exe_unit,
		const std::vector< uint64_t > &	frag_offsets,
		const int	device_count,
		const ExecutorDeviceType &	device_type,
		const bool	enable_inner_join_fragment_skipping,
		Executor *	executor
	)

protected

Definition at line 169 of file QueryFragmentDescriptor.cpp.

References CHECK(), CHECK_EQ, checkDeviceMemoryUsage(), execution_kernels_per_device_, GPU, Data_Namespace::GPU_LEVEL, RelAlgExecutionUnit::input_descs, outer_fragments_size_, rowid_lookup_key_, selected_tables_fragments_, and RelAlgExecutionUnit::simple_quals.

Referenced by buildFragmentKernelMap().

                         {
   // Allocate all the fragments of the tables involved in the query to available
   // devices. The basic idea: the device is decided by the outer table in the
   // query (the first table in a join) and we need to broadcast the fragments
   // in the inner table to each device. Sharding will change this model.
   const auto& outer_table_desc = ra_exe_unit.input_descs.front();
   const int outer_table_id = outer_table_desc.getTableId();
   auto it = selected_tables_fragments_.find(outer_table_id);
   CHECK(it != selected_tables_fragments_.end());
   const auto outer_fragments = it->second;
   outer_fragments_size_ = outer_fragments->size();
 
   const auto inner_table_id_to_join_condition = executor->getInnerTabIdToJoinCond();
   const auto num_bytes_for_row = executor->getNumBytesForFetchedRow();
 
   for (size_t outer_frag_id = 0; outer_frag_id < outer_fragments->size();
        ++outer_frag_id) {
     const auto& fragment = (*outer_fragments)[outer_frag_id];
     auto skip_frag = executor->skipFragment(outer_table_desc,
                                             fragment,
                                             ra_exe_unit.simple_quals,
                                             frag_offsets,
                                             outer_frag_id);
     if (enable_inner_join_fragment_skipping &&
         (skip_frag == std::pair<bool, int64_t>(false, -1))) {
       skip_frag = executor->skipFragmentInnerJoins(
           outer_table_desc, ra_exe_unit, fragment, frag_offsets, outer_frag_id);
     }
     if (skip_frag.first) {
       continue;
     }
     const int device_id =
         fragment.shard == -1
             ? fragment.deviceIds[static_cast<int>(Data_Namespace::GPU_LEVEL)]
             : fragment.shard % device_count;
     if (device_type == ExecutorDeviceType::GPU) {
       checkDeviceMemoryUsage(fragment, device_id, num_bytes_for_row);
     }
     for (size_t j = 0; j < ra_exe_unit.input_descs.size(); ++j) {
       const auto table_id = ra_exe_unit.input_descs[j].getTableId();
       auto table_frags_it = selected_tables_fragments_.find(table_id);
       CHECK(table_frags_it != selected_tables_fragments_.end());
       const auto frag_ids =
           executor->getTableFragmentIndices(ra_exe_unit,
                                             device_type,
                                             j,
                                             outer_frag_id,
                                             selected_tables_fragments_,
                                             inner_table_id_to_join_condition);
 
       if (execution_kernels_per_device_.find(device_id) ==
           execution_kernels_per_device_.end()) {
         std::vector<ExecutionKernel> kernels{
             ExecutionKernel{device_id, FragmentsList{}, std::nullopt}};
         CHECK(execution_kernels_per_device_.insert(std::make_pair(device_id, kernels))
                   .second);
       }
 
       // Multifrag kernels only have one execution kernel per device. Grab the execution
       // kernel object and push back into its fragments list.
       CHECK_EQ(execution_kernels_per_device_[device_id].size(), size_t(1));
       auto& execution_kernel = execution_kernels_per_device_[device_id].front();
 
       auto& kernel_frag_list = execution_kernel.fragments;
       if (kernel_frag_list.size() < j + 1) {
         kernel_frag_list.emplace_back(FragmentsPerTable{table_id, frag_ids});
       } else {
         CHECK_EQ(kernel_frag_list[j].table_id, table_id);
         auto& curr_frag_ids = kernel_frag_list[j].fragment_ids;
         for (const int frag_id : frag_ids) {
           if (std::find(curr_frag_ids.begin(), curr_frag_ids.end(), frag_id) ==
               curr_frag_ids.end()) {
             curr_frag_ids.push_back(frag_id);
           }
         }
       }
     }
     rowid_lookup_key_ = std::max(rowid_lookup_key_, skip_frag.second);
   }
 }

Here is the call graph for this function:

Here is the caller graph for this function:

void QueryFragmentDescriptor::checkDeviceMemoryUsage	(	const Fragmenter_Namespace::FragmentInfo &	fragment,
		const int	device_id,
		const size_t	num_cols
	)

protected

Definition at line 290 of file QueryFragmentDescriptor.cpp.

References available_gpu_mem_bytes_, CHECK_GE, g_cluster, Fragmenter_Namespace::FragmentInfo::getNumTuples(), gpu_input_mem_limit_percent_, LOG, tuple_count_per_device_, and logger::WARNING.

Referenced by buildFragmentPerKernelMap(), and buildMultifragKernelMap().

                                     {
   if (g_cluster) {
     // Disabled in distributed mode for now
     return;
   }
   CHECK_GE(device_id, 0);
   tuple_count_per_device_[device_id] += fragment.getNumTuples();
   const size_t gpu_bytes_limit =
       available_gpu_mem_bytes_[device_id] * gpu_input_mem_limit_percent_;
   if (tuple_count_per_device_[device_id] * num_bytes_for_row > gpu_bytes_limit) {
     LOG(WARNING) << "Not enough memory on device " << device_id
                  << " for input chunks totaling "
                  << tuple_count_per_device_[device_id] * num_bytes_for_row
                  << " bytes (available device memory: " << gpu_bytes_limit << " bytes)";
     throw QueryMustRunOnCpu();
   }
 }

Here is the call graph for this function:

Here is the caller graph for this function:

void QueryFragmentDescriptor::computeAllTablesFragments	(	std::map< int, const TableFragments * > &	all_tables_fragments,
		const RelAlgExecutionUnit &	ra_exe_unit,
		const std::vector< InputTableInfo > &	query_infos
	)

static

Definition at line 45 of file QueryFragmentDescriptor.cpp.

References CHECK_EQ, and RelAlgExecutionUnit::input_descs.

Referenced by Executor::ExecutionDispatch::runImpl().

                                                   {
   for (size_t tab_idx = 0; tab_idx < ra_exe_unit.input_descs.size(); ++tab_idx) {
     int table_id = ra_exe_unit.input_descs[tab_idx].getTableId();
     CHECK_EQ(query_infos[tab_idx].table_id, table_id);
     const auto& fragments = query_infos[tab_idx].info.fragments;
     if (!all_tables_fragments.count(table_id)) {
       all_tables_fragments.insert(std::make_pair(table_id, &fragments));
     }
   }
 }

Here is the caller graph for this function:

bool QueryFragmentDescriptor::shouldCheckWorkUnitWatchdog ( ) const

inline

Definition at line 138 of file QueryFragmentDescriptor.h.

References execution_kernels_per_device_, and rowid_lookup_key_.

Referenced by Executor::dispatchFragments().

                                            {
     return rowid_lookup_key_ < 0 && !execution_kernels_per_device_.empty();
   }

Here is the caller graph for this function:

bool QueryFragmentDescriptor::terminateDispatchMaybe	(	size_t &	tuple_count,
		const RelAlgExecutionUnit &	ra_exe_unit,
		const ExecutionKernel &	kernel
	)		const

protected

Definition at line 272 of file QueryFragmentDescriptor.cpp.

References anonymous_namespace{QueryFragmentDescriptor.cpp}::is_sample_query(), SortInfo::limit, SortInfo::offset, ExecutionKernel::outer_tuple_count, and RelAlgExecutionUnit::sort_info.

Referenced by assignFragsToKernelDispatch().

                                          {
   const auto sample_query_limit =
       ra_exe_unit.sort_info.limit + ra_exe_unit.sort_info.offset;
   if (!kernel.outer_tuple_count) {
     return false;
   } else {
     tuple_count += *kernel.outer_tuple_count;
     if (is_sample_query(ra_exe_unit) && sample_query_limit > 0 &&
         tuple_count >= sample_query_limit) {
       return true;
     }
   }
   return false;
 }

Here is the call graph for this function:

Here is the caller graph for this function:

Member Data Documentation

std::map<size_t, size_t> QueryFragmentDescriptor::available_gpu_mem_bytes_

protected

Definition at line 152 of file QueryFragmentDescriptor.h.

Referenced by checkDeviceMemoryUsage(), and QueryFragmentDescriptor().

std::map<int, std::vector<ExecutionKernel> > QueryFragmentDescriptor::execution_kernels_per_device_

protected

Definition at line 148 of file QueryFragmentDescriptor.h.

Referenced by assignFragsToKernelDispatch(), assignFragsToMultiDispatch(), buildFragmentPerKernelMap(), buildMultifragKernelMap(), and shouldCheckWorkUnitWatchdog().

double QueryFragmentDescriptor::gpu_input_mem_limit_percent_

protected

Definition at line 150 of file QueryFragmentDescriptor.h.

Referenced by checkDeviceMemoryUsage().

size_t QueryFragmentDescriptor::outer_fragments_size_ = 0

protected

Definition at line 143 of file QueryFragmentDescriptor.h.

Referenced by buildFragmentPerKernelMap(), and buildMultifragKernelMap().

int64_t QueryFragmentDescriptor::rowid_lookup_key_ = -1

protected

Definition at line 144 of file QueryFragmentDescriptor.h.

Referenced by assignFragsToKernelDispatch(), assignFragsToMultiDispatch(), buildFragmentPerKernelMap(), buildMultifragKernelMap(), and shouldCheckWorkUnitWatchdog().

std::map<int, const TableFragments*> QueryFragmentDescriptor::selected_tables_fragments_

protected

Definition at line 146 of file QueryFragmentDescriptor.h.

Referenced by buildFragmentPerKernelMap(), buildMultifragKernelMap(), and QueryFragmentDescriptor().

std::map<size_t, size_t> QueryFragmentDescriptor::tuple_count_per_device_

protected

Definition at line 151 of file QueryFragmentDescriptor.h.

Referenced by checkDeviceMemoryUsage().

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

/home/jenkins-slave/workspace/core-os-doxygen/QueryEngine/Descriptors/QueryFragmentDescriptor.h
/home/jenkins-slave/workspace/core-os-doxygen/QueryEngine/Descriptors/QueryFragmentDescriptor.cpp

Public Member Functions

Static Public Member Functions

Protected Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

Member Function Documentation

Member Data Documentation