mindspore-mindinsight/mindinsight/profiler/README.md

• MindInsight Profiler Introduction
• • Add profiling code to MindSpore script
  • Operator Performance Analysis
  • Limitations

MindInsight Profiler Introduction

MindInsight Profiler is a performance analysis tool for MindSpore. It can help to analyse and optimize the performance of the neural networks.

The Profiler enables users to:

• Start/finish profiling the neural networks by adding two simple Profiler apis to the script.
• Analyse the performance of the operators in the neural network.

Add profiling code to MindSpore script

To enable profiling on MindSpore, the MindInsight Profiler apis should be added to the script:

1. Import MindInsight Profiler

   from mindinsight.profiler import Profiler
   

2. Initialize the Profiler after set context, and before the network initialization.

   Example:

    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", device_id=int(os.environ["DEVICE_ID"]))
    profiler = Profiler(output_path="./data", is_detail=True, is_show_op_path=False, subgraph='All')
    net = Net()
   

   Parameters of Profiler including：

     subgraph (str): Defines which subgraph to monitor and analyse, can be 'all', 'Default', 'Gradients'.
     is_detail (bool): Whether to show profiling data for op_instance level, only show optype level if False.
     is_show_op_path (bool): Whether to save the full path for each op instance.
     output_path (str): Output data path.
     optypes_to_deal (list): Op type names, the data of which optype should be collected and analysed,
                            will deal with all op if null.
     optypes_not_deal (list): Op type names, the data of which optype will not be collected and analysed.
   

3. Call Profiler.analyse() at the end of the program

   Profiler.analyse() will collect profiling data and generate the analysis results.

After training, we can open MindInsight UI to analyse the performance.

Operator Performance Analysis

The operator performance analysis component is used to display the execution time of the operators during MindSpore run.

Figure 17: Statistics for Operator Types

Figure 17 displays the statistics for the operator types, including:

• Choose pie or bar graph to show the proportion time occupied by each operator type. The time of one operator type is calculated by accumulating the execution time of operators belong to this type.
• Display top 20 operator types with longest execution time, show the proportion and execution time (ms) of each operator type.

Figure 18: Statistics for Operators

Figure 18 displays the statistics table for the operators, including:

• Choose All: Display statistics for the operators, including operator name、type、excution time、full scope time、information etc. The table will be sorted by execution time by default.
• Choose Type: Display statistics for the operator types, including operator type name、execution time、execution frequency and proportion of total time. Users can click on each line, querying for all the operators belong to this type.
• Search: There is a search box on the right, which can support fuzzy search for operators/operator types.

Limitations

The Profiler has the following limitations now:

• Only programs running on Ascend chip is supported.
• To limit the data size generated by the Profiler, MindInsight suggests that for large neural network, the profiled steps should better below 10.