Documentation

Performance Analysis

Using Profiling to Collect Profile Data

  1. On the AMIT process page, click System Profiler to enter the parameter configuration page for performance analysis.
  2. Click New Project on the welcome page or the icon in the upper left corner. The profiling configuration window is displayed, as shown in Figure 1.
    Set Project Name and Project Location under Project Properties. Click Next.
    Figure 1 Configuring project properties
  3. Go to the Executable Properties page. Figure 2 shows detailed configurations.
    Figure 2 Executable Properties page
  4. Go to the Profiling Options page and retain the default settings, as shown in Figure 3.
    Figure 3 Profiling options
  5. After the preceding configurations are complete, click Start in the lower right corner of the window to start profiling.

    After the execution is complete, the profiling result will be automatically displayed in the MindStudio window. See Figure 4.

    Figure 4 Profiling result

Using Profiling to Analyze Profile Data

This section provides only a simple profiling process. The detailed process depends on the actual data. For more profile data analysis examples, see "Profiling" in the MindStudio User Guide.

  1. Check the full statistics on iteration durations: In the timeline view, view the iteration durations under Step Trace and identify a time-consuming iteration and analyze it.
  2. Export the timeline data of the iteration: Click the button and then Yes in the displayed dialog box to export the timeline data of the iteration. See Figure 5.
    Figure 5 Step Trace
  3. Check the operator time consumption within the iteration. If an operator that takes a long time exists, view the operator details to locate the fault. If the communication is time-consuming or the scheduling interval is long, analyze the API time consumption during each call. See Figure 6.
    Figure 6 AI Core task
  4. View the operator statistics table: View the time consumption and details of each AI Core and AI CPU operator within the iteration to find the time-consuming operator. Then, further locate and analyze the operator metrics data and analyze the proportion of operator data transfer and execution pipeline to identify the operator bottleneck. See Figure 7.
    Figure 7 AI Core Metrics
  5. View the time consumption statistics table of component APIs: Check the time consumption of AscendCL APIs and Runtime APIs within the iteration to analyze the impact of API calling on performance. See Figure 8.
    Figure 8 Statistics