Overview

MindStudio Insight can display full profile data in a visualized manner, helping you further analyze and confirm issues.

Figure 1 shows the process of analyzing an issue by using MindStudio Insight.

Figure 1 MindStudio Insight analysis flowchart
  1. Use the cluster analysis function for preliminary demarcation.
    1. Go to the Summary page and preliminarily determine the issue category through multi-card computing, communication, and scheduling comparison. For details, see Summary.
    2. Go to the Communication page and further locate the slow card or slow link issue by communication domain. After confirming the abnormal card or link, you can directly go to the Timeline based on the communication operator to locate the fault. For details, see Communication.
      • If the number of cards is small, you can directly import the raw performance data to automatically generate the cluster analysis result (the visualization tool invokes the msprof-analyze CLI).
      • If there are too many cards and the full performance data is too heavy, you are advised to manually invoke the msprof-analyze tool in CLI mode to open the cluster_analysis_output deliverables with the visualization tool MindStudio Insight for faster and more convenient analysis.
  2. After preliminary demarcation, select the required card and perform further analysis from the single-card dimension.
    • Timeline page: intuitively displays the running details of both the host and devices during training or inference. It shows the API execution duration on the host side and the task execution duration on the device side. For details, see Timeline.
    • Memory page: shows the overall memory usage trend in a line chart. You can also select and zoom in on the peak area in the line chart to precisely identify the processes or operators with high memory consumption. For details, see Memory.
    • Operator page: displays the duration statistics of compute operators and communication operators by type and shape. In addition, the comparison between two cards is supported, allowing you to view operator details more intuitively. For details, see Operator.