Displaying System Functions

Statistics

MindStudio Insight allows you to view operator statistics and details about a single operator.

You can select some operators on a single level-3 unit or select some operators across multiple units at a single-card level. After selecting some operators, the operator statistics are displayed on the Slice List tab page, as shown in Figure 1. Table 1 describes the fields.

You can move the cursor to the Slice List tab page and click

in the upper right corner of the table to copy the content displayed in the Slice List tab page and paste the content to an Excel file for analysis.

When operators are selected across multiple units under a single card, the selected parts of the HBM, LLC, NPU_MEM, Stars Soc Info, and acc_pmu histogram units are not counted on the Slice List tab page.

Click an operator in the Slice List column. All operators with the same name as the operator are displayed in the More list on the right. Click a row in the More list to locate the operator in the timeline view, and go to the Slice Detail page, where you can view the details about the operator.

Figure 1 Slice List

**Table 1** Slice List fields
Field	Description
Name	Operator name.
Wall Duration	Total duration of operator execution.
Self Time	Operator execution time (excluding the time of the called sub-operator).
Average Wall Duration	Average operator execution time.
Max Wall Duration	Maximum operator execution duration.
Min Wall Duration	Minimum operator execution duration.
Occurrences	Number of operator calls.
Index	Sequence number.
Start Time	Timestamp in the graphical pane.
Duration(ms)	Execution duration.

If you select a single operator, you can view the operator details in the Slice Detail area, as shown in Figure 2. Table 2 describes the fields.

Select a single operator and press M to select the Timeline area to which the operator belongs. Press M again to cancel the selection.

Figure 2 Slice Detail

**Table 2** Slice Detail fields
Field	Description
Title	Name.
Start	Start time.
Start(Raw Timestamp)	Original start time of data collection.
Wall Duration	Total duration.
Self Time	Total time (excluding sub-classes).
Input Shapes	Input dimension of the operator. When task-time is set to l0 during data collection, this field is not collected and is displayed as N/A. This field is available only for operators collected on the NPU accelerator core.
Input Data Types	Input data type of the operator. When task-time is set to l0 during data collection, this field is not collected and is displayed as N/A. This field is available only for operators collected on the NPU accelerator core.
Input Formats	Input format of the operator. When task-time is set to l0 during data collection, this field is not collected and is displayed as N/A. This field is available only for operators collected on the NPU accelerator core.
Output Shapes	Output shape of the operator. When task-time is set to l0 during data collection, this field is not collected and is displayed as N/A. This field is available only for operators collected on the NPU accelerator core.
Output Data Types	Output data type of the operator. When task-time is set to l0 during data collection, this field is not collected and is displayed as N/A. This field is available only for operators collected on the NPU accelerator core.
Output Formats	Output format of the operator. When task-time is set to l0 during data collection, this field is not collected and is displayed as N/A. This field is available only for operators collected on the NPU accelerator core.
Attr Info	Operator attributes. When task-time is set to l0 or l1 during data collection, this field is not collected and is displayed as N/A. This field is available only when aclnn is enabled and task-time is set to l2.
Args	Operator parameters.

Statistics System View

On the System View tab page, when you select Stats System View, the tab page contains Rank ID, Overall Metrics, summary statistics of five types of operators, and Kernel Details (details about the operators on the NPU). You can select the card to be viewed from the Rank ID selection box. In the DB scenario, select Host Name and Rank ID in sequence.

The Overall Metrics area displays the overall information about all operators, as shown in Figure 3. Table 3 describes the fields. When you select a sub-level in the Computing Time column, you can click any operator in the More area, and the specific position of the operator in the timeline view is displayed.

Figure 3 Overall Metrics

**Table 3** Field description of Overall Metrics
Field	Description
Category	Category. Multi-level information can be displayed. Parent layers: Computing Time, Communication(Not Overlapped) Time, Free Time, and E2E Time. Child layers: Computing Time is further divided into the disassembling results of computing stream operators such as Flash Attention, Conv, Matmul, Cube, and Vector. Forward and Backward are used to distinguish the forward and backward propagation. The child layers of Communication(Not Overlapped) Time are the grouping and disassembling results of each communication group. The waiting time and transmission time are the intersection results of the not-overlapped communication.
Total Time(μs)	Total time of the category.
Time Ratio	Duration percentage of the category.
Number	Number of operators of the category.
Avg(μs)	Average time of the category.
Min(μs)	Minimum time of the category.
Max(μs)	Maximum time of the category.
More	When you select a child layer from the Computing Time list, this area displays details about all operators at the selected layer. You can click any operator to go to the specific location of the operator in the timeline view.

The five operator types are Python API Summary, CANN API Summary, Ascend HardWare Task Summary, Communication Summary, and Overlap Analysis. Figure 4 shows the operator information. Table 4 describes the fields.

Figure 4 Operator Summary tab page

**Table 4** Stats System View fields
Field	Description
Name	Name.
Time(%)	Total time ratio = Total time of the category/Total time of all categories When the statistical type is Overlap Analysis, Time (%) = Total Time (us)/(Total Communication(Not Overlapped) + Total Computing + Total Free)
Total Time(us)	Total time of the category.
Num Calls	Number of calls.
Avg(us)	Average time of the category.
Min(us)	Minimum time of the category.
Max(us)	Maximum time of the category.

The Kernel Details area displays details about the operator on the NPU, as shown in Figure 5. Table 5 describes the fields. Click Click in the Click to Timeline column to go to the specific location of the operator in the timeline view. Area 4 (data pane) displays the selected details, that is, the specific information about the operator. You can click next to a field name in the operator details table to perform fuzzy search on the related field.

Figure 5 Operator details

**Table 5** Operator detail fields
Field	Description
Name	Operator name.
Type	Operator type.
Accelerator Core	Computing core type.
Start Time	Start time of the task.
Duration(μs)	Duration of the task.
Wait Time(μs)	Interval between the end time of the previous task and the start time of the current task, in microseconds.
Task ID	Task ID.
Block Dim	Number of task running splits, which corresponds to the number of cores during task running.
Input Shapes	Input shape of the operator.
Input Data Types	Input data type of the operator.
Input Formats	Input format of the operator.
Output Shapes	Output shape of the operator.
Output Data Types	Output data type of the operator.
Output Formats	Output format of the operator.
Click To Timeline	Click Click to go to the specific location of the operator in the timeline view. The details about the operator are displayed in area 4 (data pane).

Expert System View

On the System View tab page, when you select Expert System View, the Rank ID selection box, Expert Analysis, and six advice system tabs are displayed. You can select the card to be viewed from the Rank ID selection box. In the DB scenario, select Host Name and Rank ID in sequence.

The Expert Analysis tab page displays the abnormal metrics in the unit.

The six advice systems include Affinity API, Affinity Optimizer, AICPU Operators, ACLNN Operators, Operators Fusion, and Operators Dispatch as shown in Figure 6. Table 6 describes the fields.

If you select any advice system, the details about the advice system are displayed in the right pane. Click Click in the Click to Timeline column to go to the specific location of the operator in the timeline view. The Slice Detail tab page in area 4 (data pane) displays the details about the operator.

Figure 6 Expert System View

**Table 6** Expert System View fields
Field	Description
Name	Operator name. This parameter is unavailable when the advice system is Affinity Optimizer.
Origin API	Used to fuse the API sequence. This parameter is available only when the advice system is Affinity API.
Replacement API	Equivalent to Affinity API. This parameter is available only when the advice system is Affinity API.
Origin Optimizer	Used to fuse the optimizer. This parameter is available only when the advice system is Affinity Optimizer.
Replacement Optimizer	Optimizer that can be replaced. This parameter is available only when the advice system is Affinity Optimizer.
Origin Operators	Operators that can be fused. This parameter is available only when the advice system is Operators Fusion.
Fused Operator	An operator that has been fused at the CANN layer. This parameter is available only when the advice system is Operators Fusion.
Start Time	Start time of the task.
Duration(μs)	Duration of the task.
Process Id	Process ID.
Thread Id	Thread ID.
Notes	Prompt information. This parameter is unavailable when the advice system is Affinity Optimizer.
Click To Timeline	Click Click to go to the specific location of the operator in the timeline view. The details about the operator are displayed in area 4 (data pane).

Event View

On the Timeline tab page, operator information can be displayed in the event view.

On the Timeline tab page, right-click the required unit and choose Show in events view from the shortcut menu. The System View tab page is displayed. By default, Events View is selected in the left pane, and details about all operators in the unit are displayed in the right pane, as shown in Figure 7. Table 7 describes the fields.

Figure 7 Event View

**Table 7** Events View fields
Field	Description
Name	Operator name.
Start	Operator execution start time.
Duration(ns)	Total time of running the operator.
TID	Thread ID. This parameter is available when you select the Python and CANN units and their sub-units.
PID	Process ID This parameter is available when you select the Python and CANN units and their sub-units.
Stream Name	Name of the stream task flow in the Ascend Hardware unit. This parameter is available only when you select the Ascend Hardware unit and its sub-units.
Group Name	Name of the communication operator cluster. This parameter is available only when you select the Communication unit and its Group sub-units.
Analysis Type	Analysis operator type. This parameter is available only when you select the Overlap Analysis unit or its sub-units.
Rank ID	Rank ID of the operator. This parameter is available when you select the Ascend Hardware, Communication, and Overlap Analysis units and their sub-units.
Click To Timeline	Click Click. The specific location of the operator is displayed in the timeline view. The operator details are displayed on the Slice Detail tab page.

This function is not supported at the card level.
This function is not supported in the HBM, LLC, NPU_MEM, Stars Soc Info, and acc_pmu histogram units.
This function is not supported in the Plane sub-unit of the Communication unit.

Parent topic: Instructions