IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix coredump with numactl (#21393)

Browse Source 
* fix coredump with numactl

* record original socketid and numaid

* update comment

* update comments

---------

Co-authored-by: Chen Peter <peter.chen@intel.com>
This commit is contained in:
Fang Xu
2023-12-08 14:20:35 +08:00
committed by GitHub
parent 2086041552
commit 488abd0ad2
6 changed files with 82 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
src/inference/dev_api/openvino/runtime/system_conf.hpp
View File

@@ -260,6 +260,24 @@ OPENVINO_RUNTIME_API void set_cpu_used(const std::vector<int>& cpu_ids, const in
*/
OPENVINO_RUNTIME_API int get_socket_by_numa_node(int numa_node_id);
/**
* @brief Get original socket id by current socket id, the input socket id is recalculated after filtering (like
* numactl), while the original socket id is the original id before filtering
* @ingroup ie_dev_api_system_conf
* @param[in] socket_id socket id
* @return socket id
*/
OPENVINO_RUNTIME_API int get_org_socket_id(int socket_id);
/**
* @brief Get original numa node id by current numa node id, the input numa node id is recalculated after filtering
* (like numactl), while the original numa node id is the original id before filtering
* @ingroup ie_dev_api_system_conf
* @param[in] numa_node_id numa node id
* @return numa node id
*/
OPENVINO_RUNTIME_API int get_org_numa_id(int numa_node_id);
/**
* @enum ColumnOfCPUMappingTable
* @brief This enum contains definition of each columns in CPU mapping table which use processor id as index.

2
src/inference/src/dev/threading/cpu_streams_executor.cpp
View File

@@ -144,7 +144,7 @@ struct CPUStreamsExecutor::Impl {
.set_max_threads_per_core(max_threads_per_core)});
} else if (stream_type == STREAM_WITH_NUMA_ID) {
_taskArena.reset(new custom::task_arena{custom::task_arena::constraints{}
.set_numa_id(_numaNodeId)
.set_numa_id(get_org_numa_id(_numaNodeId))
.set_max_concurrency(concurrency)
.set_max_threads_per_core(max_threads_per_core)});
} else if (stream_type == STREAM_WITH_CORE_TYPE) {

4
src/inference/src/dev/threading/istreams_executor.cpp
View File

@@ -563,6 +563,10 @@ void IStreamsExecutor::Config::update_executor_config(int stream_nums,
return;
}
if (proc_type_table.size() > 1) {
core_type = ov::threading::IStreamsExecutor::Config::ANY;
}
// IStreamsExecutor::Config config = initial;
const auto total_num_cores = proc_type_table[0][ALL_PROC];
const auto total_num_big_cores = proc_type_table[0][MAIN_CORE_PROC] + proc_type_table[0][HYPER_THREADING_PROC];

2
src/inference/src/os/cpu_map_info.hpp
View File

@@ -27,6 +27,8 @@ public:
std::vector<std::vector<int>> _org_proc_type_table;
std::vector<std::vector<int>> _proc_type_table;
std::vector<std::vector<int>> _cpu_mapping_table;
std::map<int, int> _socketid_mapping_table;
std::map<int, int> _numaid_mapping_table;
std::mutex _cpu_mutex;
int _socket_idx = 0;
};

17
src/inference/src/os/lin/lin_system_conf.cpp
View File

@@ -169,10 +169,16 @@ CPU::CPU() {
}
}
for (size_t i = 0; i < valid_cpu_mapping_table.size(); i++) {
valid_cpu_mapping_table[i][CPU_MAP_NUMA_NODE_ID] =
numa_node_map.at(valid_cpu_mapping_table[i][CPU_MAP_NUMA_NODE_ID]);
valid_cpu_mapping_table[i][CPU_MAP_SOCKET_ID] =
sockets_map.at(valid_cpu_mapping_table[i][CPU_MAP_SOCKET_ID]);
auto new_numa_id = numa_node_map.at(valid_cpu_mapping_table[i][CPU_MAP_NUMA_NODE_ID]);
auto new_socket_id = sockets_map.at(valid_cpu_mapping_table[i][CPU_MAP_SOCKET_ID]);
if (_numaid_mapping_table.find(new_numa_id) == _numaid_mapping_table.end()) {
_numaid_mapping_table.insert({new_numa_id, valid_cpu_mapping_table[i][CPU_MAP_NUMA_NODE_ID]});
}
if (_socketid_mapping_table.find(new_socket_id) == _socketid_mapping_table.end()) {
_socketid_mapping_table.insert({new_socket_id, valid_cpu_mapping_table[i][CPU_MAP_SOCKET_ID]});
}
valid_cpu_mapping_table[i][CPU_MAP_NUMA_NODE_ID] = new_numa_id;
valid_cpu_mapping_table[i][CPU_MAP_SOCKET_ID] = new_socket_id;
}
}
@@ -265,12 +271,13 @@ CPU::CPU() {
_cores);
}
}
_org_proc_type_table = _proc_type_table;
std::vector<std::vector<std::string>>().swap(system_info_table);
if (check_valid_cpu() < 0) {
OPENVINO_THROW("CPU affinity check failed. No CPU is eligible to run inference.");
};
_org_proc_type_table = _proc_type_table;
}
void parse_node_info_linux(const std::vector<std::string> node_info_table,

45
src/inference/src/system_conf.cpp
View File

@@ -215,6 +215,14 @@ int get_socket_by_numa_node(int numa_node_id) {
return -1;
};
int get_org_socket_id(int socket_id) {
return -1;
}
int get_org_numa_id(int numa_node_id) {
return -1;
}
#elif defined(__APPLE__)
// for Linux and Windows the getNumberOfCPUCores (that accounts only for physical cores) implementation is OS-specific
// (see cpp files in corresponding folders), for __APPLE__ it is default :
@@ -267,6 +275,24 @@ int get_socket_by_numa_node(int numa_node_id) {
return -1;
};
int get_org_socket_id(int socket_id) {
CPU& cpu = cpu_info();
auto iter = cpu._socketid_mapping_table.find(socket_id);
if (iter != cpu._socketid_mapping_table.end()) {
return iter->second;
}
return -1;
}
int get_org_numa_id(int numa_node_id) {
CPU& cpu = cpu_info();
auto iter = cpu._numaid_mapping_table.find(numa_node_id);
if (iter != cpu._numaid_mapping_table.end()) {
return iter->second;
}
return -1;
}
#else
# ifndef _WIN32
@@ -417,6 +443,25 @@ int get_number_of_logical_cpu_cores(bool bigCoresOnly) {
# endif
return logical_cores;
}
int get_org_socket_id(int socket_id) {
CPU& cpu = cpu_info();
auto iter = cpu._socketid_mapping_table.find(socket_id);
if (iter != cpu._socketid_mapping_table.end()) {
return iter->second;
}
return -1;
}
int get_org_numa_id(int numa_node_id) {
CPU& cpu = cpu_info();
auto iter = cpu._numaid_mapping_table.find(numa_node_id);
if (iter != cpu._numaid_mapping_table.end()) {
return iter->second;
}
return -1;
}
#endif
#if ((OV_THREAD == OV_THREAD_TBB) || (OV_THREAD == OV_THREAD_TBB_AUTO))