CN-116916393-B - Time delay constraint scheduling method for heterogeneous computing resources in computing network

CN116916393BCN 116916393 BCN116916393 BCN 116916393BCN-116916393-B

Abstract

The invention discloses a time delay constraint scheduling method for heterogeneous computing power resources in a computing power network, and belongs to the field of communication. The method comprises the following steps of obtaining the number of calculation tasks, the data volume, the calculation volume and the time delay constraint, obtaining the performance parameters of the heterogeneous calculation force server, and obtaining the transmission bandwidth and the transmission power of a communication channel. The method comprises the following steps of following the basic idea of carrying out joint scheduling and optimizing total energy consumption on communication and calculation, carrying out alternate iterative optimization on a scheduling strategy and server performance parameters, wherein the scheduling strategy is fixed, the server performance parameter configuration is optimized by utilizing a dynamic voltage frequency scaling technology, and the server performance parameters are fixed, and the updated scheduling strategy is optimized by utilizing a projection gradient method. The scheduling method can provide energy-optimized heterogeneous computing power resource scheduling strategy and server performance parameter configuration. The method and the device can be used for scheduling optimization of heterogeneous computing power resources in the computing power network so as to obtain the beneficial effect of reducing energy consumption for completing computing tasks under time delay constraint.

Inventors

ZHANG WENKAI
YE CHAOYANG
ZHANG SHICONG
GU CHENHUI
WANG WEI
ZHANG CHAOYANG

Assignees

浙江省新型互联网交换中心有限责任公司
浙江大学

Dates

Publication Date: 20260505
Application Date: 20230628

Claims (4)

1. A time delay constraint scheduling method for heterogeneous computing power resources in a computing power network is characterized by comprising the following steps: 1) The multi-core CPU server in the power computing network is obtained Number of (3) Core frequency CPU-GPU heterogeneous server Number of (3) GPU core voltage GPU core frequency And memory frequency Wherein M+1 is not less than j is not less than M+K, and task to be scheduled Number of (3) Data volume Calculated amount Acceleration factor And delay constraints Tasks in a computing power network And a server Transmission bandwidth between And transmission power ; 2) Initializing the performance parameter configuration of the server, namely initializing the core frequency of the multi-core CPU server And heterogeneous server memory frequency The following are provided: ; Wherein, the And Respectively representing the upper limit and the lower limit of the dynamic scaling range of the corresponding parameters; Initializing heterogeneous server GPU core frequencies The heterogeneous server GPU core voltage is as follows Is arranged as Obtaining ; Wherein, the The server hardware correlation coefficient is obtained through fitting of measurement results; 3) Initializing a scheduling policy matrix using projection gradient method Wherein The value of (2) represents a task Is dispatched to a server The proportion calculated above satisfies The initial scheduling policy matrix is noted as ; 4) In the first place In the multiple iterations, the scheduling policy is fixed Optimizing the calculation energy consumption of the server according to a dynamic voltage frequency scaling technology, and updating the performance parameters of the server: ; 5) Calculate the first Total energy consumption at the next iteration Regarding scheduling policy The gradient of (2) is noted as Fixing server performance parameters and updating scheduling strategy ; 6) Repeating steps 4) and 5) until convergence conditions are satisfied Obtaining an optimal allocation strategy and server performance parameters, wherein Optimizing scheduling policy for alternation A convergence threshold at that time.
2. The method according to claim 1, wherein the initializing of the scheduling strategy by the projection gradient method in step 3) comprises the following steps: 1) Fixing server performance parameters, and calculating to obtain gradient of total energy consumption on scheduling strategy , ; Updating scheduling policy using projection gradient method ; Wherein, the Representing an initialized scheduling policy Is used for the number of iterations of (a), Is a constant coefficient depending on hardware characteristics; representing multicore CPU server power versus frequency Sensitivity of scaling; Respectively representing CPU-GPU heterogeneous server power versus memory frequency Scaling and GPU core voltage Frequency/frequency Scaled sensitivity; representing the power consumption of a CPU (Central processing Unit) in the heterogeneous server, which is used for controlling the GPU to operate; is an acceleration factor related to the application characteristics; 2) Repeating the step 3.1) until the convergence condition is satisfied Obtaining an initialization scheduling strategy with optimal total energy consumption Wherein Is the convergence threshold when initializing the scheduling policy.
3. The method according to claim 1, wherein the method for updating the server performance parameters in step 4) comprises the following steps: 1) Updating , ; 2) Updating ; Order the In the dichotomy Solution of search equation, noted as , ; Wherein, the Calculating energy consumption for the heterogeneous server; if no solution exists in the search range, then ; If it is Then ; If it is Then ; According to Handle Is arranged as ; 3) Updating , Let equation Solution to (1) ; Then ; 。
4. The method according to claim 1, wherein the updating scheduling policy method of step 3.1) and step 5) comprises the following steps: 1) Will be Dimension scheduling policy Column vectorization, noted as Constructing equivalent vector constraints, constraining the equations Conversion to Wherein Corresponding to a coefficient vector constrained by an equation; Classifying and converting the inequality constraint into a vector constraint form ; At a known position Calculating whether the inequality constraint takes the equal sign or not, classifying the inequality constraint based on the equal sign, and converting the inequality constraint taking the equal sign into Wherein Corresponding to a coefficient of the inequality constraint taking the equal sign, converting the rest constraint into Wherein Corresponding to a coefficient of an inequality constraint that does not take an equal sign; 2) Calculate the first Total energy consumption at the next iteration Regarding scheduling policy Gradient of (2) , ; Order the Projection matrix Gradient after projection If (if) Jump to step 5.4), if Step 5.3) is carried out; 3) Order the ; Wherein the method comprises the steps of Dimension and matrix of (a) The number of rows of (a) is the same, Dimension and matrix of (a) The number of lines is the same if If all components in the set are not negative, stopping calculation to make I.e. For the K-T point, if If negative components exist, the smallest component is selected, and the matrix is removed Returning to step 5.2); 4) Will be Generating a matrix by column Updating scheduling policies , For the step of gradient decrease at the kth iteration, by Internally linearly searching for an optimal value such that updated The corresponding total energy consumption is minimum, wherein Is the upper limit of the gradient descent step size.

Description

Time delay constraint scheduling method for heterogeneous computing resources in computing network Technical Field The invention relates to the field of wireless communication, in particular to a time delay constraint scheduling method for heterogeneous computing power resources in a computing power network. Background The computing power network is a novel service concept, integrates heterogeneous resources and distributed computing power, completes uniform scheduling, and innovates a service mode. Key technologies of the computing power network include resource identification, heterogeneous resource scheduling, distributed computing power scheduling and the like. In a computing power network, heterogeneous computing power resources include CPU, GPU, NPU, FPGA, ASIC and the like. Among these, CPU is the most common computational power resource, while GPU is the most common heterogeneous computational power accelerator. For massive heterogeneous computing forces, different computing forces are needed for cooperative processing in different scenes. The heterogeneous computing power is cooperated with the nano-tube scheduling through the computing power network, and a heterogeneous computing power cooperated network integrating cloud, edge and end is constructed, so that computing power requirements of various scenes including high-performance computing, the Internet of things, edge computing, artificial intelligence and the like are better met. The dynamic voltage frequency scaling (Dynamic Voltage and Frequency Scaling, DVFS) is a low power consumption technique, which aims to set the operating voltage and clock frequency according to the actual power consumption requirement of the chip at the time, so that the provided power can be ensured to meet the requirement and not to have excessive performance, and the power consumption can be reduced. The DVFS technology utilizes the characteristics of the CMOS chip, namely, dynamically adjusts the running frequency and voltage of the chip according to different requirements of the application program running on the chip on the computing power under the condition of not changing the chip structure, thereby achieving the purpose of energy saving. The performance modeling of the heterogeneous computing power server based on the DVFS technology can better realize energy consumption optimization. Scheduling optimization of heterogeneous computational resources presents the following two challenges. Firstly, the performance modeling and the same scheduling of different types of computing force servers, and the characteristics of the computing task determine that the computing task will have different performance performances on the different types of computing force servers. Secondly, the interaction of scheduling between nodes in a multi-tasking multi-server computing network environment increases the complexity of the optimization. Disclosure of Invention In order to overcome the defects of the prior art, the invention aims to provide a scheduling optimization algorithm for heterogeneous computing power resources in a computing power network. And (3) alternately and iteratively optimizing the scheduling strategy and the server performance parameters according to the basic idea of comprehensive optimization of the communication model and the calculation model. On one hand, the scheduling strategy is fixed, the server performance parameter configuration is optimized by utilizing a dynamic voltage frequency scaling technology, and on the other hand, the server performance parameter is fixed, and the scheduling strategy is optimized and updated by utilizing a projection gradient method. And (5) alternately iterating until the total energy consumption optimization converges to obtain an optimal scheduling strategy of heterogeneous computing power resources and optimal server performance parameter configuration. A time delay constraint scheduling method for heterogeneous computing power resources in a computing power network comprises the following steps: 1. the number M of multi-core CPU servers { S 1,S2,...,SM } in the computing network and the core frequency are obtained The number K of CPU-GPU heterogeneous servers { S M+1,SM+2,...,SM+K }, the core voltage frequency and the memory frequencyThe number N of tasks to be scheduled { Γ 1,Γ2,...,ΓN }, the data amount J i, the calculated amount L i, the acceleration factor theta i and the time delay constraint D, the transmission bandwidth R i,j and the transmission power between the task Γ i and the server S j in the power calculation network 2. Initializing the performance parameter configuration of the server, namely initializing the core frequency of the multi-core CPU serverAnd heterogeneous server memory frequencyThe following are provided: Wherein, [. Cndot. max ] and [. Cndot. min represent the upper and lower limits of the dynamic scaling range of the respective parameters, respectively. Initializing heterogeneous ser