Timeline View

Timeline View consists of the navigation pane on the left, graphical pane on the right, and data pane at the bottom, as shown in the following figure.

Figure 1 Timeline View

Figure 2 Timeline View (PyTorch E2E Profiling scenario)

On the navigation pane, you can view each timeline's label and their dependencies.
On the graphical pane, events are displayed in timelines.
The data pane displays the collected profile data in tables, including Event View, Statistics, and AI Core Metrics.

The timeline (not timestamps) involved in the profile data indicates the system monotonic time, which is related only to the system and is not the real time.
Start Time and End Time indicate the time range of the collected profile data.
Current Time indicates the start time of the time segment where the cursor is located.

Move the cursor to a sampling point to view the detailed analysis data.
To see the sequence of a timeline in the Event View window, right-click a timeline label in the left navigation pane and choose Show in Event View from the shortcut menu.
In TimeLine View, you can check APIs and operations used.
If multiple OS Runtime APIs within a thread are executed simultaneously, they are displayed in different lines.
If multiple AI Core tasks within a stream are executed simultaneously, they are displayed in different lines.
To zoom in or out the timeline of interest, select a time point, and hold down Ctrl as scrolling the mouse wheel upwards or downwards. You can also click or in the upper right corner of the window. To restore the timeline view, click in the upper right corner of the window.
When selecting a time point in a timeline, you can drag the cursor leftward or rightward to select the duration. With Current Time(us) as the boundary point, drag the mouse leftwards or rightwards to display the selected time range.
In cluster or multi-device scenarios, profile data of the first iteration of the model (ID) with the largest number of iterations on the rank or device with the minimum ID is exported by default.
Under the PyTorch framework, profile data of the first iteration on the rank or device with the minimum ID is exported by default, as shown in Figure 2.

After the profiling project is executed, the inference/training process is displayed in a timeline view based on the scheduling procedure. The actual display depends on the options selected during profiling and the device in use.

The following table describes the fields in the display sequence.

**Table 1** Hardware information
Field	Description
CPU	CPU
Memory	Memory
Disk	Disk
Network	Network bandwidth
Start Time	Process start time of the CPU, memory, disk, and network bandwidth, in μs.
End Time	Process end time of the CPU, memory, disk, and network bandwidth, in μs.
Duration	Running duration of the CPU, memory, disk, and network bandwidth, in μs.
Usage	Usage of the CPU, memory, disk, and network bandwidth.
Note: When you move the cursor over a timeline, Start Time, End Time, Duration, and Usage are displayed.

**Table 2** Time spent by components
Field	Description
Process {ID}	Process ID
Thread {ID}	Thread ID
MsprofTX	MsprofTX profile data
Os Runtime	Timeline of each thread to call OS Runtime APIs.
PTA	Timeline of the task queue asynchronously delivered by operators at the PyTorch layer.
AscendCL API	Time taken by models, operators, and Runtime APIs. If a certain metric is not collected, the profiling result will be not available.
ACL_RTS	AscendCL API of the RTS type.
ACL_MODEL	AscendCL API of the MODEL type.
Runtime API	Timeline of each thread to call Runtime APIs.
GE	Time taken by the model to input, infer, and output data.
Start Time	Time when an API starts to run, in μs.
End Time	Time when an API stops running, in μs.
Duration	Time taken by the calls to the current API, in μs.
Name	API name.
Note: When you move the cursor over a timeline, Start Time, End Time, Duration, and Name are displayed.

**Table 3** System data of the Ascend AI Processor
Field	Description
NPU {ID}	ID of the Ascend AI Processor.
Step Trace	Iteration trace data, which records the time required for each iteration.
model {ID}	Model IDs, which are displayed in sequence under Step Trace. You can use either of the following methods to export and display data of an iteration of a model: Click of an iteration of a model. In the dialog box displayed, click Yes. Specify the Device ID, Model ID, and Iteration ID in the upper left corner of the page, and click Export.
Name	API name
Iteration ID	Iteration ID
FP Start	FP start time, in μs.
Iteration End	End time of each iteration, in μs.
Iteration Time	Iteration duration, in μs.
Stream {ID}	Stream ID.
AI Core task	Timeline of AI Core tasks within each stream.
AI CPU task	Timeline of AI CPU tasks within each stream.
Other task	Timeline of other tasks within each stream.
Start Time	Time when the AI Core task, AI CPU task, or other task starts to run, in μs.
End Time	Time when the AI Core task, AI CPU task, or other task stops running, in μs.
Duration	Time consumed by the AI Core task, AI CPU task, or other task, in μs.
Status	Running status of the AI Core task, AI CPU task, or other task.
Task Type	Type of the AI Core task, AI CPU task, or other task.
Stream ID	Stream ID of the AI Core task, AI CPU task, or other task.
Op Name	Operator name.
Task ID	ID of the AI Core task, AI CPU task, or other task.
Note: When you move the cursor over a timeline, Name, Iteration ID, FP Start, Iteration End, Iteration Time, Start Time, End Time, Duration, Status, Task Type, Stream ID, Op Name, and Task ID are displayed.

Parent topic: Timeline Views