Performance Data in the Profiler Directory

View Source On Gitee

The profiler directory includes some csv, json, and txt files, which contain performance data such as operator execution time, memory usage, and communication during model computation to help users analyze performance bottlenecks. The fields in some of the csv and txt files are explained below. The contents of the files mainly include information about the time consumed by the device-side operators (AI Core operator and AI CPU operator), and the information about the memory occupied at the operator level and the memory occupied at the application level.

aicore_intermediate_*_detail.csv File Descriptions

The aicore_intermediate_*detail.csv file contains statistical AI Core operator information based on the contents of output_timeline_data*.txt and framework_raw_*.csv. Refer to the following table for descriptions of the fields in the file:

Field Names

Descriptions

full_kernel_name

Full name of the device-side execution kernel operator

task_duration

Operator execution time

execution_frequency

Frequency of operator execution

task_type

Task types for the operator

aicore_intermediate_*_type.csv File Descriptions

The aicore_intermediate_*type.csv file includes statistical information about the specific types of AI Core operators based on the contents of output_timeline_data*.txt and framework_raw_*.csv. Refer to the following table for descriptions of the fields in the file:

Field Names

Descriptions

kernel_type

AI Core operator types

task_time

Operator execution time

execution_frequency

Frequency of operator execution

percent

Percentage of time taken of this operator type over the total time taken of all operators

aicpu_intermediate_*.csv File Descriptions

The aicpu_intermediate_*.csv file contains time taken information of the AI CPU operators. Refer to the following table for descriptions of the fields in the file:

Field Names

Descriptions

serial_num

AI CPU operator number

kernel_type

AI CPU operator type

total_time

Operator time taken, which is equal to the sum of the downstream time taken and the execution time taken

dispatch_time

downstream time taken

execution_time

execution time taken

run_start

Start time of operator execution

run_end

End time of operator execution

flops_*.txt File Descriptions

The flops_*.txt file contains information about the number of floating-point computations for the device-side operator, the number of floating-point computations per second, and so on. Refer to the following table for descriptions of the fields in the file:

Field Names

Descriptions

full_kernel_name

Full name of the device-side execution kernel operator

MFLOPs(10^6 cube)

The number of floating-point calculations (10^6 cube)

GFLOPS(10^9 cube)

The number of floating-point calculations per second(10^9 cube)

MFLOPs(10^6 vector)

The number of floating-point calculations(10^6 vector)

GFLOPS(10^9 vector)

The number of floating-point calculations per second(10^9 vector)

output_timeline_data_*.txt File Descriptions

The output_timeline_data_*.txt file contains information about the time taken of the device-side operator. Refer to the following table for descriptions of the fields in the file:

Field Names

Descriptions

kernel_name

Full name of the device-side execution kernel operator

stream_id

Stream ID of the operator

start_time

Start time of operator execution(us)

duration

operator execution time (ms)

cpu_ms_memory_record_*.txt File Descriptions

The cpu_ms_memory_record_*.txt file contains information about application-level memory usage. Refer to the following table for descriptions of the fields in the file:

Field Names

Descriptions

Timestamp

Moment of memory event(ns)

Total Allocated

Total memory allocation(Byte)

Total Reserved

Total memory reservation(Byte)

Total Active

Total memory requested by streams in MindSpore(Byte)

operator_memory_*.csv File Descriptions

The operator_memory_*.csv file contains information about operator-level memory usage. Refer to the following table for descriptions of the fields in the file:

Field Names

Descriptions

Name

Memory Consumption Tensor Name

Size

Size of memory occupied(KB)

Allocation Time

Tensor memory allocation time(us)

Duration

Tensor memory occupation time(us)

Allocation Total Allocated

Total memory allocation at operator memory allocation(MB)

Allocation Total Reserved

Total memory occupation at operator memory allocation(MB)

Release Total Allocated

Total memory allocation at operator memory release(MB)

Release Total Reserved

Total memory occupation at operator memory release(MB)

Device

device type