GB/T 47514-2026Distributed computing - Computing power measurement method (English PDF)

6.2.1 Initial Resource Information for Data Processing

6.2.1.1 Overview Initial resource information for data processing includes computing configuration, peak computing power, average computing power, caching capacity, and memory performance.

6.2.1.2 Calculation Configuration The computing configuration includes the number of computing units, the number of computing unit cores, and the computing unit frequency. The computing configuration obtains computing unit architecture information from the operating system's output data by executing operating system commands, and outputs computing... Unit state parameter values.

6.2.1.3 Peak computing power Peak computing power includes.

a) Peak number of integer operations per second per computing unit (e.g., peak number of 32-bit integer operations per second per computing unit, peak number of integer operations per second per computing unit) (Peak number of 64-bit integer operations, etc.)

b) Peak number of floating-point operations per second for the computing unit (e.g., peak number of 32-bit floating-point operations per second for the computing unit, peak number of floating-point operations per second for the computing unit) (Peak number of 64-bit floating-point operations, etc.). Peak computing power can be obtained through product specifications or random documentation, or by performing commonly used industry benchmarking tools.

6.2.1.4 Average Computational Ability The average computing power includes.

a) Average number of integer operations per second per computing unit (e.g., average number of 32-bit integer operations per second per computing unit, average number of integer operations per second per computing unit) (average number of 64-bit integer operations, etc.)

b) Average number of floating-point operations per second per computing unit (e.g., average number of 32-bit floating-point operations per second per computing unit, average number of floating-point operations per second per computing unit) (average number of 64-bit floating-point operations, etc.)

c) Average number of hash operations per second per computing unit. The average computing power was obtained by executing commonly used industry benchmarking tools.

6.2.1.5 Caching Capability Caching capabilities include.

a) Computational unit cache capacity (e.g., CPU L1 cache capacity, CPU L2 cache capacity, CPU L3 cache capacity) (Cache capacity, GPU cache capacity, etc.)

b) Computational unit cache bandwidth (e.g., CPU L1 cache bandwidth, CPU L2 cache bandwidth, CPU L3 cache bandwidth) (Cache bandwidth, GPU cache bandwidth, etc.)

c) Cache access latency of computing units (e.g., CPUL1 cache access latency, CPUL2 cache access latency, CPU L3 cache access latency, GPU cache access latency, etc. Caching capabilities are obtained through product specifications or random documentation, or by executing operating system commands.

6.2.1.6 Memory Performance Memory performance includes.

6.3 Data Storage

6.3.1 Initial Resource Information for Data Storage Initial resource information for data storage includes.

a) Storage unit capacity;

b) Read/write bandwidth;

c) Theoretical throughput of the storage unit;

d) Theoretical speed values for sequential read/write;

e) Theoretical random read/write speed values. Initial resource information for data storage is obtained through operating system commands or product specifications.

6.3.2 Data storage resource occupancy status Data storage resource occupancy status includes.

a) Storage space utilization rate;

b) IOPS of the storage unit;

c) Storage unit read/write bandwidth utilization;

d) Current sequential read/write speed of the storage unit;

e) Current random read/write speed of the storage unit. The status of data storage resource occupancy can be obtained from the device interface.

6.4 Data Transmission

6.4.1 Initial Resource Information for Data Transmission Initial resource information for data transmission includes.

a) Network bandwidth;

b) Network latency;

c) Virtual network speed;

d) Data received and transmitted per second;

e) Send queue length. Initial resource information for data transmission is obtained through original manufacturer configuration information or operating system commands, or by executing industry-standard testing tools. Acquisition is possible.

6.4.2 Data Transmission Resource Occupancy Status Data transmission resource occupancy status includes.

a) Current bandwidth utilization;

b) Current data transmission delay. Data transmission resource usage status is obtained by executing commonly used industry testing tools.

6.5.1 Initial Resource Information for Collaborative Computing

6.5.1.1 Collaborative Computing Performance Metrics The Collaborative Computing Performance Measure (COPS) is calculated according to the method given in Formula (1). COPS= OPi Tc (1) In the formula. OPi

---The theoretical number of operations performed by the task in device i from the start to the end of execution. See the appendix for the calculation of the theoretical number of operations. Record A of A.1; Tc

--- The maximum time taken for a distributed computing task to run from start to finish. Example. Embed an operation counting program in the device's task execution code to count the number of various computational operations during task execution, thereby obtaining the appropriate... The number of operations (OPi) is considered; the time span from task initiation to completion is obtained through the built-in task scheduling and monitoring module of the distributed computing system, thus yielding the task... Execution time Tc.

6.5.1.2 Equipment Collaborative Computing Measurement Configuration Requirements The configuration requirements for device collaborative computing metrics are as follows:

a) The device being measured shall meet all the requirements in 6.6.2;

b) The device being measured should be able to execute at least one distributed coordination and synchronization algorithm. Example. In distributed machine learning training, the FedAvg algorithm in federated learning can be used as a benchmark program.

6.5.1.3 Collaborative Computing Metrics The collaborative computing metric method is as follows, and the testing procedure is shown in Appendix B, section B.2.

a) Calculate the theoretical number of operations for the distributed benchmark program in the measured device. The calculation of the theoretical number of operations is shown in A.1.

b) Execute the distributed benchmark program and record the execution time of the device being measured;

c) Calculate the collaborative computing performance metric of the device being measured using the method in 6.5.1.1;

d) Return the collaborative computing performance metric of the measured device.

6.5.2 Collaborative Computing Resource Usage Status Collaborative computing resource occupancy status includes.

a) Online/offline status; Note

1.Call the distributed sensing program to return the online or offline status of devices within the system.

6.6 Dynamic Calculation

6.6.1 Dynamically calculate performance metrics Dynamic computing refers to evaluating the overall performance of distributed computing devices by running benchmark programs in typical application scenarios in a real-world environment. Performance is measured by dynamically calculating the performance metric OPS according to the method given in formula (2).

6.6.2 Equipment Dynamic Calculation Measurement Configuration Requirements The configuration requirements for dynamic calculation metrics of the equipment are as follows:

a) The device being measured shall comply with the relevant equipment requirements in GB/T 47502-2026.

b) All devices being measured should have the same version of the operating system installed and use the same version of the software stack (e.g., programming language, compiler). (Drivers, etc.)

c) The device being measured should have at least one dedicated calculation function. 1) Encryption/decryption; 2) Ray tracing; 3) Matrix floating-point multiplication; 4) Graphics rendering; 5) Intelligent computing tasks.

Note. See A.2 for the calculation of operands for intelligent computing tasks.

d) All measured devices should select the same dedicated computational task as the benchmark program for execution.

e) The device being measured should have at least one of the following data accuracy execution benchmark procedures. 1) 4-bit integer (INT4); 2) 8-bit integer (INT8); 3) 16-bit integer (INT16); 4) 32-bit integer (INT32); 5) 64-bit integer (INT64); 6) 8-bit floating-point type (FP8); 7) 16-bit floating-point type (FP16); 8) 32-bit floating-point type (FP32); 9) 64-bit floating-point type (FP64); 10) 16-bit brain floating-point type (BF16); 11) 32-bit tensor floating-point type (TF32); 12) Mixing accuracy.

f) All measured devices should be subjected to the same data precision when performing the benchmark procedure.

6.6.3 Dynamic Calculation Measurement Method The dynamic calculation method for metrics is as follows; the test procedure is shown in B.1.

a) Calculate the theoretical number of operations for the baseline procedure. See A.1 for the calculation of the theoretical number of operations.

b) Execute the benchmark program and record the execution time;

c) Calculate the dynamic performance metrics of the equipment using the method in 6.6.1;

d) Returns the device's dynamically calculated performance metrics.

7.1 Initial value preparation for the device being measured

7.1.1 Environmental Conditions When measuring the equipment, the temperature and humidity environmental conditions should meet the requirements of

5.6.1 in GB/T 2887-2011.

7.1.2 Requirements for Initial Value Preparation of the Measuring Device The requirements for obtaining the initial resource information of the device being measured are as follows:

a) Collect initial resource information according to the methods given in 6.2.1, 6.3.1, 6.4.1, 6.5.1, and 6.6, with the metric values being the initial values.

b) The selection of models and datasets shall be determined by the user and the party being measured in consultation according to the application scenario.

c) Process the model under test within the device (e.g., model quantization, model conversion, etc.).

d) The device being measured provides the necessary interface implementations for the measurement program to call (e.g., remote read/write/execute primitive commands, initialization, etc.). (e.g., model loading, execution, unloading, throughput measurement, resource usage measurement, utilization measurement, etc.)

e) The device being measured must meet at least one of the following requirements. 1) If the device being measured is a SoC, the measurement program is deployed on the SoC; 2) If the device being measured is a board, mount the board to the host and deploy the measurement program on the host; 3) If the device being measured is a server, the measurement program is deployed on the server host; 4) If the device being measured is a cloud service, the measurement program is deployed on a cloud host.

7.2 Computational Power Methods for Distributed Computing Devices

7.2.1 Static Calculation Method In a distributed computing environment, when a single device has no running programs other than the operating system, its computing power can be utilized by analyzing its initial resources. The source information parameters are measured by weighting the entropy values and then summing them. The specific measurement method is as follows:

a) Initial resources of primary parameters under the data processing characteristics, data storage characteristics, data transmission characteristics, and dynamic computing characteristics of the equipment. Information v is used to determine parameter weights wv using AHP (see Appendix C).

b) Initial resources for primary parameters under the data processing characteristics, data storage characteristics, data transmission characteristics, and dynamic computing characteristics of the equipment. Information v collects initial resource information and secondary parameters of each device, and collects the dynamic calculation characteristics of the devices according to the method in 6.6.3, and according to the formula... Equation (3) gives the method for forming the second-level metric parameter vector Xv.

7.2.2 Dynamic Calculation Method In a distributed computing environment, when a single device is executing a program, the device's computing power can be measured by analyzing various dynamic computing power parameters of the device. Entropy is used as the basis for weighting, and then a weighted sum is applied for measurement. The specific measurement method is as follows:

a) Collect data processing characteristics, data storage characteristics, and data transmission characteristics of the equipment according to the methods given in 6.2.2, 6.3.2, and 6.4.2. Based on the source occupancy status, calculate the remaining amount of equipment resources according to the method given in Table 2.