Profile Data Collection
Description
Profile the performance.
Applicability
Product |
Supported |
|---|---|
Atlas 350 Accelerator Card |
√ |
√ |
|
√ |
|
☓ |
|
☓ |
|
☓ |
Command
asys profiling -r=aicore -p=time -d=deviceId --output=./ --aic_metrics=PipeUtilization
Parameters
- r : specifies the collection type. This parameter is mandatory. The value is an enumerated string. Multiple enumerated values can be entered and separated by commas (,).
- dvpp: collects DVPP profile data, such as the execution time and usage.
- aicore: collects AI Core profile data, such as the time consumption and proportion of cube and vector instructions, and the time consumption proportion of compute units and MTEs.
- os: collects system memory data, AI CPU usage, and Ctrl CPU usage.
- memory: collects the memory read rate and bandwidth data, including the on-chip memory and level-3 cache.
- link: collects bandwidth data, such as collective communication bandwidth and PCIe bandwidth.
- power: collects low-power data.
- p: specifies the collection interval, in seconds. This parameter is mandatory. The minimum value is 1 and the maximum value is 30 x 24 x 3600.
- d: specifies the ID of the device to be operated. This parameter is optional. Only one device ID can be entered. The default value is 0.
- output: optional parameter. Its value is used as the prefix of the result output directory of the asys tool. That is, the final output directory is {output}/asys_profiling_result_ timestamp. If the command does not contain the output parameter, the output is stored in the command execution directory. If the value of output is empty or invalid, the specified directory does not have the write permission, or the directory fails to be created, the asys tool exits and an error is reported.
For details about the result file, see Profile Data File References in Performance Tuning Tool.
- aic_metrics: This parameter is optional. It specifies the AI Core performance monitor unit (PMU) type. This parameter is valid only when the collection type contains AI Cores.
Value range:
- PipeUtilization: percentages of time taken by compute units and MTEs. This is the default value.
- ArithmeticUtilization: time consumptions and percentages of Cube and Vector instructions.
- Memory: memory read/write bandwidth rate.
- MemoryL0: L0 read/write bandwidth rate.
- MemoryUB: UB read/write bandwidth rate.
- ResourceConflictRatio: resource conflict ratio
- L2Cache: L2 cache hit ratio.
- MemoryAccess: memory access bandwidth of the operator on the AI Core.
Usage Example and Output Description
# Collect AI Core profile data. asys profiling -r=aicore -p=10 -d=0 --output=./ --aic_metrics=PipeUtilization
After the command is executed successfully, the following information is displayed and the collection result file is generated in the {output}/asys_profiling_result_ timestamp directory:
2025-11-27 20:15:45,141 [ASYS] [INFO]: asys start. 2025-11-27 20:15:45,141 [ASYS] [INFO]: Start run: msprof --output=./ --sys-period=10 --sys-devices=0 --ai-core=on --aic-mode=sample-based --aic-metrics=PipeUtilization, please wait about 10 seconds. 2025-11-27 20:16:04,335 [ASYS] [INFO]: Succeeded in running aicore profiling, [INFO] Start profiling.... [INFO] Start export data in PROF_000001_20251127201545157_03062849EPFNHDPB. ...... [INFO] Query all data in PROF_000001_20251127201545157_03062849EPFNHDPB done. [INFO] Profiling finished. [INFO] Process profiling data complete. Data is saved in /xxx/ascend_system_advisor/asys/asys_profiling_result_20251127201545110/PROF_000001_20251127201545157_03062849EPFNHDPB 2025-11-27 20:16:04,336 [ASYS] [INFO]: profiling task execute finish. 2025-11-27 20:16:04,336 [ASYS] [INFO]: asys finish.
Parent topic: asys Tool Usage Guide