Using Ascend Graph APIs for Data Profiling

Ascend Graph APIs are used to collect profile data during graph construction.

Availability

Atlas Inference Series Product

Atlas Training Series Product

Atlas A2 Training Series Product/Atlas 800I A2 Inference Product

Atlas A3 Training Series Product

Enabling Modes

**Table 1** Profiling methods
Method	API
Pass options through the GEInitialize API.	ge.exec.profilingMode ge.exec.profilingOptions You can pass ge.exec.profilingOptions through the GEInitialize API to profile iteration track data. The passed fields include training_trace, bp_point, and fp_point. This method saves profile data to the path set in the output parameter of ge.exec.profilingOptions.
aclgrph APIs	aclgrphProfInit aclgrphProfFinalize aclgrphProfCreateConfig aclgrphProfDestroyConfig aclgrphProfStart aclgrphProfStop This method saves profile data to the path set in the profiler_path parameter of aclgrphProfInit.
Note: For details about the APIs, see "GEInitialize" and "aclgrph API".

Raw Performance Data Profiling (Passing Options Through the GEInitialize API)

The following is an example of using GEInitialize to pass required option arguments:

// 0. System init
std::map<AscendString, AscendString> config = {{"ge.exec.deviceId", "0"},
                        {"ge.graphRunMode", "1"},
                        {"ge.exec.precision_mode", "allow_fp32_to_fp16"},
                        {"ge.exec.profilingMode", "1"},
                        {"ge.exec.profilingOptions",  R"({"output":"/tmp/profiling","training_trace":"on","fp_point":"","bp_point":""})"}};
    Status ret = ge::GEInitialize(config);
    if (ret != SUCCESS) {
        return FAILED;
    }

Raw Performance Data Profiling (aclgrph APIs)

This following is an example of calling APIs to collect profile data:

  // Construct a graph. The details are omitted here.
  // ......

  // init ge
  std::map<std::string, std::string> ge_options = {{"ge.socVersion", "Ascendxxx"}, {"ge.graphRunMode", "1"}};
  ge::GEInitialize(ge_options);

  std::string profilerResultPath = "/home/test/prof";       // Create the path in advance.
  uint32_t length = strlen("/home/test/prof");
  ret = ge::aclgrphProfInit(profilerResultPath.c_str(), length);     

  std::map<string, string> options = {{"a", "b"}, {"c", "d"}};
  uint32_t graphId = 0;

  ge::Session *session = new Session(options);
  ret = session->AddGraph(graphId, graph);

  uint32_t deviceid_list[1] = {0};
  uint32_t device_nums = 1;
  uint64_t data_type_config = ProfDataTypeConfig::kProfTaskTime | ProfDataTypeConfig::kProfAiCoreMetrics | ProfDataTypeConfig::kProfAicpu | ProfDataTypeConfig::kProfTrainingTrace;
  ProfAicoreEvents *aicore_events = NULL;
  ProfilingAicoreMetrics aicore_metrics = ProfilingAicoreMetrics::kAicoreArithmeticUtilization;  
  ge::aclgrphProfConfig *pro_config = ge::aclgrphProfCreateConfig(deviceid_list, device_nums, aicore_metrics, aicore_events, data_type_config);

  ge::aclgrphProfStart(pro_config);

  session->RunGraph(graphId, inputs_r, outputs_r);

  ge::aclgrphProfStop(pro_config);

  ge::aclgrphProfDestroyConfig(pro_config);

  ge::aclgrphProfFinalize();

  delete session;
  ge::GEFinalize();

Profiling Description

After the Ascend Graph API method is configured for profiling, parse the raw data and export the parsing results as visualized timeline and summary files. For details, see Offline Parsing. The following table shows the generated profile data.

**Table 2** Introduction to the profile data files (options are passed by GEInitialize)
Argument	Profile Data File
Automatically generated by default	msprof (Timeline Report) The Ascend Hardware layer in msprof_.json step_trace (iteration trace data) fusion_op_.csv
task_time, task_trace	The CANN layer in msprof_.json and the api_statistic_.csv file task_time_.csv The Communication layer in msprof_.json and the communication_statistic_.csv file op_summary_.csv op_statistic_*.csv
runtime_api	The CANN_Runtime layer in msprof_.json and the api_statistic_.csv file
hccl	The Communication layer in msprof_.json and the communication_statistic_.csv file api_statistic_*.csv
aicpu	aicpu_*.csv
host_sys_usage	System CPU usage on the host CPU usage of processes on the host System memory usage on the host Memory usage of processes on the host

**Table 3** Introduction to the profile data files (aclgrph)
Argument	Profile Data File
kProfTaskTime	The CANN layer in msprof_.json and the api_statistic_.csv file The Ascend Hardware layer in msprof_.json and the task_time_.csv file The Communication layer in msprof_.json and the communication_statistic_.csv file op_summary_.csv op_statistic_.csv step_trace (iteration trace data) fusion_op_*.csv
kProfHccl	The Communication layer in msprof_.json and the communication_statistic_.csv file api_statistic_*.csv
kProfAicpu	aicpu_*.csv
kProfL2cache	l2_cache_*.csv
Automatically generated by default	System CPU usage on the host CPU usage of processes on the host System memory usage on the host Memory usage of processes on the host

Parent topic: Other Profiling Methods