Hardware chip calculation force measurement system based on calculation force node

Abstract

The invention discloses a computing power node-based hardware chip computing power quantity system, and particularly relates to the technical field of computing power quantity, which is characterized in that computing power chips and computing power service scenes are classified based on an existing computing power background, and a service scene computing power evaluation module is arranged to evaluate the computing power of the classified computing power service scenes; the range of a computing power service scene is expanded, a computing power node data acquisition module is arranged to acquire static configuration data, dynamic resource data and historical service data of different computing power nodes, the acquisition range of original parameters is expanded, and the accuracy of computing power magnitude is improved; and setting a computing power node recommendation module to screen computing power nodes based on the identified computing power demand scene, sorting the screened computing power nodes based on computing power service requirements and node residual computing power service quality indexes, and generating a recommendation scheme, thereby being beneficial to realizing accurate demand matching of a computing power demand side.

Description

Technical Field

[0001] The present invention relates to the field of computing power measurement technology, and more specifically, to a hardware chip computing power measurement system based on computing power nodes. Background Art

[0002] With the continuous development of new network services and the increasing demand for computing power, computing network technology has gradually entered people's field of vision and continued to grow and develop. Computing power measurement, as a method of measuring the computing and storage capabilities of various computing power platforms, plays a key role in the perception of computing network services and the efficient scheduling of computing power resources.

[0003] The existing computing power measurement system uses the CPU processor performance indicator TOPS / W, the number of instructions executed per second (MIPS), and the storage performance indicator RAM as static indicators, and uses the CPU idle rate, GPU idle rate, storage remaining amount, and throughput rate as dynamic indicators to measure computing power resources. It takes into account the basic performance of the computing power node and the changes in its dynamic working state, and also conducts comprehensive considerations in the selection of static and dynamic measurement indicators, which can effectively improve the utilization rate of computing power resources and the accuracy of computing power resource matching.

[0004] However, the above system still has some problems: the selection range of static and dynamic indicators is relatively narrow, the measurement of computing power is not accurate enough, only some indicators of CPU and GPU processors are collected, and the supported computing power service scenarios are relatively single. The accuracy of computing power measurement should be improved, and the comprehensiveness of computing power service scenarios should be improved. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a hardware chip computing power measurement system based on computing power nodes to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a hardware chip computing power measurement system based on computing power nodes, comprising:

[0007] Intelligent classification module: classifies computing chips and computing service scenarios based on the existing computing power background;

[0008] Service scenario computing capacity evaluation module: evaluates the computing capacity of classified computing service scenarios and automatically backs up specific evaluation algorithms;

[0009] Computing power node data collection module: collects static configuration data, dynamic resource data, and historical service data of different computing power nodes;

[0010] Computing power node data processing module: processes the collected computing power node data and calculates the node static configuration performance excellence coefficient, node remaining service capacity coefficient, and node computing power service stability control coefficient;

[0011] Computing power node comprehensive evaluation module: Calculates the node's remaining computing power service quality index based on the node's static configuration performance excellence coefficient, the node's remaining service capacity coefficient, and the node's computing power service stability control coefficient, and automatically backs up the calculation results;

[0012] Computing power demand information identification module: This module identifies the computing power demand information of computing power demanders in the order in which they apply for computing power demand orders, confirms the computing power demand scenarios and the computing power demanders' requirements for computing power services;

[0013] Computing power node recommendation module: This module selects computing power nodes based on the identified computing power demand scenarios, ranks the selected computing power nodes based on the computing power service requirements and the node remaining computing power service quality index, and generates a recommendation plan;

[0014] Database: used to store data of all modules in the system.

[0015] Preferably, the intelligent classification module includes a computing power chip classification unit and a computing power service scenario classification unit. The computing power chip classification unit divides the types of computing power chips into four categories: central processing unit (CPU), graphics processing unit (GPU), embedded neural network processor (NPU), and tensor processor (TPU); the computing power service scenario classification unit divides computing power service scenarios into three categories: general computing power, artificial intelligence computing power, and supercomputing computing power. The general computing power is mainly carried by the central processing unit (CPU), and is mainly aimed at general software applications and executing logical operations. The artificial intelligence computing power is mainly carried by the embedded neural network processor (NPU), tensor processor (TPU), and graphics processing unit (GPU), and is mainly aimed at AI applications, with the characteristics of simple logic, intensive calculations, and high concurrency. The supercomputing computing power is mainly carried by the central processing unit (CPU) and graphics processing unit (GPU), and is mainly aimed at scientific computing and industrial simulation scenarios, with complex calculations and high requirements for calculation accuracy.

[0016] Preferably, the service scenario computing capacity evaluation module includes a general computing capacity evaluation unit, an artificial intelligence computing capacity evaluation unit, a super computing capacity evaluation unit, and an automatic backup unit. The general computing capacity evaluation unit is based on the number n of CPUs of the i-th model. cai 、CPU core number m cai , single core main frequency h cai , and the number of floating-point operations per CPU cycle f cai Calculate the general computing power coefficient X TS The specific formula is: nxca The number of CPU models used in general computing power evaluation; the number of NPUs of the i-th model n based on the artificial intelligence computing power evaluation unit ai , the precision d of the jth data type processed sij , the working frequency h when processing the jth data type precision gij , the number of GPUs of the kth model n gak , the number of instructions executed in the clock cycle f gak 、GPU core number m gak , GPU core operating frequency h gak , the number of TPUs of the first model n tl , and the amount of computation f processed by the TPU in a single clock cycle tl Calculate the AI ​​computing power coefficient X RZ The specific formula is:

[0017] Among them, a1, a2, and a3 are weight coefficients, a1>0, a2>0, a3>0, n xa 、n xdi 、n xga 、n xt They are the number of NPU models, the number of data type precision types of the i-th NPU model, the number of GPU models when evaluating artificial intelligence computing power, and the number of TPU models; the supercomputing computing power evaluation unit is based on the number n of i-th CPU models cbi 、CPU core number m cbi , single core main frequency h cbi 、Number of floating-point operations per CPU cycle f cbi , the number of GPUs of type j n gbj , the number of instructions executed in the clock cycle f gbj 、GPU core number m gbj , and the GPU core operating frequency h gbj Calculate the supercomputing power coefficient X CS The specific formula is: where n xcb 、n xgb The number of CPU models and the number of GPU models used in supercomputing computing power evaluation are b1 and b2, respectively, and b1>0 and b2>0. The automatic backup unit is used to automatically back up the calculation formulas for computing power in different service scenarios.

[0018] Preferably, the computing power node data collection module includes a node static configuration data collection unit, a node dynamic resource data collection unit, a node historical service data collection unit, and a data output unit. The node static data collection unit is used to collect the model, quantity, performance parameters, node storage capacity, storage bandwidth, number of read and write operations per second of the node's processor, node memory capacity, node memory bandwidth, and node network connection bandwidth; the node dynamic data collection unit is used to collect the node's CPU idle rate, GPU idle rate, NPU idle rate, TPU idle rate, node remaining storage capacity, node remaining memory capacity, data transmission delay, number of node service connections, and throughput; the node historical service data collection unit is used to collect the total computing power service time, fault handling time, and abnormality duration of each historical computing power service; the data output unit is used to send the collected node computing power data to the computing power node data processing module.

[0019] Preferably, the computing power node data processing module includes a data receiving unit, a static configuration performance evaluation unit, a dynamic remaining service support capability evaluation unit, a node computing power service stability evaluation unit, and a data output unit. The data receiving unit is used to receive the collected node static configuration data, dynamic resource data, and historical service data; the static configuration performance evaluation unit calculates the node static configuration performance excellence coefficient X based on the collected static configuration data of the i-th computing power node. PZi The dynamic remaining service support capability evaluation unit calculates the node remaining service capability coefficient Y based on the collected dynamic data of the i-th computing power node αi The node computing power service stability evaluation unit calculates the node computing power service stability control coefficient X based on the collected historical computing power service data of the i-th computing power node wdi ; The data output unit is used to send the calculated node static configuration performance excellence coefficient, node remaining service capacity coefficient and node computing power service stability control coefficient to the computing power node comprehensive evaluation module.

[0020] Preferably, the specific data processing process in the computing power node data processing module is as follows:

[0021] A1. Based on the model, quantity, and performance parameters of the processors contained in the i-th computing power node, calculate the general computing power coefficient X that the node can provide. TSi , artificial intelligence computing power coefficient X RZi And the supercomputing power coefficient X CSi ;

[0022] A2, based on the node storage capacity r of the collected i-th computing power node ci , storage bandwidth v ciAnd the number of read and write operations per second of the off-node memory m cdi Compute node storage capacity coefficient X CCi , the specific formula is: Based on the node memory capacity r of the collected i-th computing power node hi and node memory bandwidth v hi Compute node cache capacity coefficient X HCi , the specific formula is: Based on the node network connection bandwidth v of the collected i-th computing power node wi Calculate the node communication capability coefficient X Ti , the specific formula is: X Ti =v wi 2 +v wi +1;

[0023] A3. Compute node static configuration performance excellence coefficient X PZi , the specific formula is:

[0024] A4. Based on the collected CPU idle rate θ of the i-th computing power node ci , GPU idle rate θ gi , NPU idle rate θ ai , TPU idle rate θ ti , node remaining storage capacity r yci , node remaining memory capacity r yhi Calculate the computing power resource surplus rate α syi , the specific formula is: e is a natural constant, based on the data transmission delay T of the collected i-th computing power node ai Number of connections to node services m li Calculate the communication quality control capability coefficient α txi , the specific formula is: Based on the collected throughput θ of the i-th computing power node ei Calculate the business processing capacity coefficient α ywi , the specific formula is: ywi =∫(θ ei ) 2 d(θ ei );

[0025] A5. Calculate the node's remaining service capacity coefficient Y αi , the specific formula is: αi =ln(α syi +e)+ln(α txi +e)+ln(α ywi +e), e is a natural constant;

[0026] A6. The total computing service time T based on the collected historical computing service of the i-th computing node zij , Troubleshooting time gij And the abnormal duration T yij Computing node computing power service stability control coefficient X wdi , the specific formula is: n fwi is the number of historical computing power services collected for the i-th computing power node, and e is a natural constant.

[0027] Preferably, the computing power node comprehensive evaluation module includes a data receiving unit, a node remaining computing power service quality index calculation unit and an automatic backup unit, wherein the data receiving unit is used to receive the calculated node static configuration performance excellence coefficient, node remaining service capacity coefficient and node computing power service stability control coefficient of different computing power nodes; the node remaining computing power service quality index calculation unit is based on the received node static configuration performance excellence coefficient X of the i-th computing power node. PZi , node remaining service capacity coefficient Y αi And the node computing power service stability control coefficient X wdi Compute node remaining computing power service quality index Y sei The specific formula is: e is a natural constant; the automatic backup unit is used to automatically back up the calculated node residual computing power service quality index of each computing power node.

[0028] Preferably, the computing power node recommendation module includes a computing power node condition screening unit, a computing power service requirement judgment unit, a computing power service requirement keyword extraction unit, a keyword weight setting unit, a node recommendation index calculation unit, a node recommendation scheme determination unit, and a node recommendation scheme output unit. The computing power node condition screening unit screens the existing computing power nodes based on the identified computing power demand scenario of the computing power demander, retains the computing power nodes containing the same computing power service scenario, and screens out the computing power nodes that do not contain the same computing power service scenario; the computing power service requirement judgment unit is used to judge whether the computing power demander has computing power service requirements, and if so, sends a computing power service requirement prompt to the computing power service requirement keyword extraction unit; if not, sends a no computing power service requirement prompt to the node recommendation scheme determination unit; the computing power service requirement keyword extraction unit is used to extract the computing power service requirement keywords of the computing power demander; the keyword weight setting unit is used to set The weights of the extracted computing power service requirement keywords are set to the same weight when the computing power demander does not mark the order of priority; when the computing power demander marks the order of priority, the weights are set in combination with the number of keywords; the node recommendation index calculation unit adjusts the calculation formula of the node remaining computing power service quality index based on the set keyword weight to calculate the node recommendation index; the node recommendation scheme determination unit arranges the calculated remaining computing power service quality index of each node from high to low according to the numerical value and numbers them in the order of arrangement when there is no computing power service requirement, and retains a preset number of computing power nodes as recommended schemes according to the number; when it is determined that there is a computing power service requirement, the calculated node recommendation index is arranged from high to low according to the numerical value and numbered according to the order of arrangement, and retains a preset number of computing power nodes as recommended schemes according to the number, and the number i becomes the i-th recommended scheme; the node recommendation scheme output unit is used to output the node recommendation scheme to the client.

[0029] The technical effects and advantages of the present invention are as follows:

[0030] The present invention sets up an intelligent classification module to classify computing power chips and computing power service scenarios based on the existing computing power background, sets up a service scenario computing power evaluation module to evaluate the computing power of the classified computing power service scenarios, automatically backs up the specific evaluation algorithm, and provides an evaluation method for computing power computing power of different computing power service scenarios. At the same time, it expands the scope of computing power service scenarios, which is conducive to improving the applicability of the computing power service industry. A computing power node data collection module is set to collect static configuration data, dynamic resource data and historical service data of different computing power nodes, which improves the collection range of original parameters and is conducive to improving the accuracy of computing power measurement. A computing power node recommendation module is set to screen computing power nodes based on the identified computing power demand scenarios, and the screened computing power nodes are sorted based on the computing power service requirements and the node remaining computing power service quality index and a recommendation plan is generated, providing a new computing power node recommendation method, which is conducive to achieving accurate demand matching for computing power demanders. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a system structure diagram of the present invention.

[0032] Figure 2 This is a flow chart of the system operation of the present invention. DETAILED DESCRIPTION

[0033] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0034] like Figure 1 The embodiment shown provides a hardware chip computing power measurement system based on computing power nodes, including an intelligent classification module, a service scenario computing power evaluation module, a computing power node data collection module, a computing power node data processing module, a computing power node comprehensive evaluation module, a computing power demand information identification module, a computing power node recommendation module, and a database. The intelligent classification module, the service scenario computing power evaluation module, and the computing power node data processing module are connected in sequence, the computing power node data collection module, the computing power node data processing module, and the computing power node comprehensive evaluation module are connected in sequence, the computing power node comprehensive evaluation module and the computing power demand information identification module are both connected to the computing power node recommendation module, and all modules in the system are connected to the database.

[0035] The intelligent classification module classifies computing power chips and computing power service scenarios based on the existing computing power background.

[0036] Furthermore, the intelligent classification module includes a computing power chip classification unit and a computing power service scenario classification unit. The computing power chip classification unit divides the types of computing power chips into four categories: central processing unit (CPU), graphics processing unit (GPU), embedded neural network processor (NPU), and tensor processing unit (TPU); the computing power service scenario classification unit divides computing power service scenarios into three types: general computing power, artificial intelligence computing power, and supercomputing computing power. The general computing power is mainly carried by the central processing unit (CPU), and is mainly aimed at general software applications and executing logical operations. The artificial intelligence computing power is mainly carried by the embedded neural network processor (NPU), tensor processing unit (TPU), and graphics processing unit (GPU), and is mainly aimed at AI applications, with the characteristics of simple logic, intensive calculations, and high concurrency. The supercomputing computing power is mainly carried by the central processing unit (CPU) and graphics processing unit (GPU), and is mainly aimed at scientific computing and industrial simulation scenarios, with complex calculations and high requirements for calculation accuracy.

[0037] The service scenario computing capacity evaluation module evaluates the computing capacity of the classified computing service scenarios and automatically backs up the specific evaluation algorithm.

[0038] Furthermore, the service scenario computing capacity evaluation module includes a general computing capacity evaluation unit, an artificial intelligence computing capacity evaluation unit, a super computing capacity evaluation unit, and an automatic backup unit. The general computing capacity evaluation unit is based on the number n of CPUs of the i-th model. cai 、CPU core number m cai , single core main frequency h cai , and the number of floating-point operations per CPU cycle f cai Calculate the general computing power coefficient X TS The specific formula is: n xca The number of CPU models used in general computing power evaluation; the number of NPUs of the i-th model n based on the artificial intelligence computing power evaluation unit ai , the precision d of the jth data type processed sij , the working frequency h when processing the jth data type precision gij , the number of GPUs of the kth model n gak , the number of instructions executed in the clock cycle f gak 、GPU core number m gak , GPU core operating frequency h gak , the number of TPUs of the first model n tl , and the amount of computation f processed by the TPU in a single clock cycle tl Calculate the AI ​​computing power coefficient X RZ The specific formula is:

[0039] Among them, a1, a2, and a3 are weight coefficients, a1>0, a2>0, a3>0, n xa 、n xdi 、n xga 、n xt They are the number of NPU models, the number of data type precision types of the i-th NPU model, the number of GPU models when evaluating artificial intelligence computing power, and the number of TPU models; the supercomputing computing power evaluation unit is based on the number n of i-th CPU models cbi 、CPU core number m cbi , single core main frequency h cbi 、Number of floating-point operations per CPU cycle f cbi , the number of GPUs of type j n gbj , the number of instructions executed in the clock cycle f gbj 、GPU core number m gbj , and the GPU core operating frequency h gbj Calculate the supercomputing power coefficient X CS The specific formula is: where n xcb 、n xgb The number of CPU models and the number of GPU models used in supercomputing computing power evaluation are b1 and b2, respectively, and b1>0 and b2>0. The automatic backup unit is used to automatically back up the calculation formulas for computing power in different service scenarios.

[0040] The computing power node data collection module collects static configuration data, dynamic resource data and historical service data of different computing power nodes.

[0041] Furthermore, the computing power node data collection module includes a node static configuration data collection unit, a node dynamic resource data collection unit, a node historical service data collection unit, and a data output unit. The node static data collection unit is used to collect the model, quantity, performance parameters, node storage capacity, storage bandwidth, node external memory read and write operations per second, node memory capacity, node memory bandwidth, and node network connection bandwidth of the processor contained in the node; the node dynamic data collection unit is used to collect the node's CPU idle rate, GPU idle rate, NPU idle rate, TPU idle rate, node remaining storage capacity, node remaining memory capacity, data transmission delay, number of node service connections, and throughput rate; the node historical service data collection unit is used to collect the total computing power service time, fault handling time, and abnormality duration of each historical computing power service; the data output unit is used to send the collected node computing power data to the computing power node data processing module.

[0042] The computing power node data processing module processes the collected computing power node data and calculates the node static configuration performance excellence coefficient, the node remaining service capacity coefficient and the node computing power service stability control coefficient respectively.

[0043] Furthermore, the computing power node data processing module includes a data receiving unit, a static configuration performance evaluation unit, a dynamic remaining service support capability evaluation unit, a node computing power service stability evaluation unit, and a data output unit. The data receiving unit is used to receive the collected node static configuration data, dynamic resource data, and historical service data; the static configuration performance evaluation unit calculates the node static configuration performance excellence coefficient X based on the collected static configuration data of the i-th computing power node. PZi The dynamic remaining service support capability evaluation unit calculates the node remaining service capability coefficient Y based on the collected dynamic data of the i-th computing power node αi The node computing power service stability evaluation unit calculates the node computing power service stability control coefficient X based on the collected historical computing power service data of the i-th computing power node wdi ; The data output unit is used to send the calculated node static configuration performance excellence coefficient, node remaining service capacity coefficient and node computing power service stability control coefficient to the computing power node comprehensive evaluation module.

[0044] Furthermore, the specific data processing process in the computing power node data processing module is as follows:

[0045] A1. Based on the model, quantity, and performance parameters of the processors contained in the i-th computing power node, calculate the general computing power coefficient X that the node can provide. TSi , artificial intelligence computing power coefficient X RZi And the supercomputing power coefficient X CSi ;

[0046] A2, based on the node storage capacity r of the collected i-th computing power node ci , storage bandwidth v ci And the number of read and write operations per second of the off-node memory m cdi Compute node storage capacity coefficient X CCi , the specific formula is: Based on the node memory capacity r of the collected i-th computing power node hi and node memory bandwidth v hi Compute node cache capacity coefficient X HCi , the specific formula is: Based on the node network connection bandwidth v of the collected i-th computing power node wi Calculate the node communication capability coefficient X Ti , the specific formula is: X Ti =vwi 2 +v wi +1;

[0047] A3. Compute node static configuration performance excellence coefficient X PZi , the specific formula is:

[0048] A4. Based on the collected CPU idle rate θ of the i-th computing power node ci , GPU idle rate θ gi , NPU idle rate θ ai , TPU idle rate θ ti , node remaining storage capacity r yci , node remaining memory capacity r yhi Calculate the computing power resource surplus rate α syi , the specific formula is: e is a natural constant, based on the data transmission delay T of the collected i-th computing power node ai Number of connections to node services m li Calculate the communication quality control capability coefficient α txi , the specific formula is: Based on the collected throughput θ of the i-th computing power node ei Calculate the business processing capacity coefficient α ywi , the specific formula is: ywi =∫(θ ei ) 2 d(θ ei );

[0049] A5. Calculate the node's remaining service capacity coefficient Y αi , the specific formula is: αi =ln(α syi +e)+ln(α txi +e)+ln(α ywi +e), e is a natural constant;

[0050] A6. The total computing service time T based on the collected historical computing service of the i-th computing node zij , Troubleshooting time gij And the abnormal duration T yij Computing node computing power service stability control coefficient X wdi , the specific formula is: n fwi is the number of historical computing power services collected for the i-th computing power node, and e is a natural constant.

[0051] The computing power node comprehensive evaluation module calculates the node remaining computing power service quality index based on the node static configuration performance excellence coefficient, the node remaining service capacity coefficient and the node computing power service stability control coefficient, and then automatically backs up the calculation results.

[0052] Furthermore, the computing power node comprehensive evaluation module includes a data receiving unit, a node remaining computing power service quality index calculation unit and an automatic backup unit. The data receiving unit is used to receive the calculated node static configuration performance excellence coefficient, node remaining service capacity coefficient and node computing power service stability control coefficient of different computing power nodes; the node remaining computing power service quality index calculation unit is based on the received node static configuration performance excellence coefficient X of the i-th computing power node. PZi , node remaining service capacity coefficient Y αi And the node computing power service stability control coefficient X wdi Compute node remaining computing power service quality index Y sei The specific formula is: e is a natural constant; the automatic backup unit is used to automatically back up the calculated node residual computing power service quality index of each computing power node.

[0053] The computing power demand information identification module identifies the computing power demand information of the computing power demander in the order of application of the computing power demand order of the computing power demander, and confirms the computing power demand scenario and the computing power demander's requirements for computing power services.

[0054] The computing power node recommendation module screens the computing power nodes based on the identified computing power demand scenarios, sorts the screened computing power nodes based on the computing power service requirements and the node remaining computing power service quality index, and generates a recommendation plan.

[0055] Furthermore, the computing power node recommendation module includes a computing power node condition screening unit, a computing power service requirement judgment unit, a computing power service requirement keyword extraction unit, a keyword weight setting unit, a node recommendation index calculation unit, a node recommendation scheme determination unit, and a node recommendation scheme output unit. The computing power node condition screening unit screens the existing computing power nodes based on the identified computing power demand scenario of the computing power demander, retains the computing power nodes containing the same computing power service scenario, and screens out the computing power nodes that do not contain the same computing power service scenario; the computing power service requirement judgment unit is used to judge whether the computing power demander has computing power service requirements, and if so, sends a computing power service requirement prompt to the computing power service requirement keyword extraction unit; if not, sends a no computing power service requirement prompt to the node recommendation scheme determination unit; the computing power service requirement keyword extraction unit is used to extract the computing power service requirement keywords of the computing power demander; the keyword weight setting unit is used to set The extracted computing power service requirement keywords are weighted. When the computing power demander does not mark the order of precedence, the keywords are weighted equally. When the computing power demander marks the order of precedence, the weights are set in combination with the number of keywords. The node recommendation index calculation unit adjusts the calculation formula of the node remaining computing power service quality index based on the set keyword weight to calculate the node recommendation index. When there is no computing power service requirement, the node recommendation scheme determination unit arranges the calculated remaining computing power service quality index of each node from high to low according to the numerical value and numbers them in the order of arrangement, and retains a preset number of computing power nodes as recommended schemes according to the number. When it is determined that there is a computing power service requirement, the calculated node recommendation index is arranged from high to low according to the numerical value and numbered in the order of arrangement, and retains a preset number of computing power nodes as recommended schemes according to the number, and the node numbered i becomes the i-th recommended scheme. The node recommendation scheme output unit is used to output the node recommendation scheme to the client.

[0056] What needs to be specifically explained in this embodiment is that the computing power service requirement keywords used are in the form of collected computing power node data or calculated computing power node indicators. The computing power service requirement keywords are divided into three types: node static configuration requirements, dynamic resource requirements, and service stability requirements, and the extracted computing power service requirement keywords are matched with the above-mentioned divided computing power service requirements.

[0057] In this embodiment, it is specifically necessary to explain that, in order to facilitate understanding of the principle of setting keyword weights, a specific solution is provided for illustration, including the following steps:

[0058] B1. Assume that six keywords are extracted, namely general computing power, storage capacity, storage bandwidth, memory capacity, memory bandwidth, and throughput;

[0059] B2. The computing power demander did not mark the order of priority, so the weights of the six keywords are

[0060] B3. Assuming that the order of keywords extracted after the computing power demander marks the order is: general computing power, throughput, memory capacity, memory bandwidth, storage capacity, storage bandwidth, then the weight of general computing power is The weight of throughput is The weight of memory capacity is The weight of memory bandwidth is The weight of storage capacity is The weight of storage bandwidth is

[0061] Specifically, in this embodiment, a method for calculating a node recommendation index is provided, which includes the following steps:

[0062] C1. Count the number of keywords n belonging to node static configuration data, dynamic resource data, and service stability data respectively. v 、n y 、n z And the corresponding keyword weight ε i , ε j , ε k ;

[0063] C2. Calculate the keyword weights ε of node static configuration data, dynamic resource data, and service stability data respectively. v , ε y , ε z , the specific formula is as follows:

[0064] C3, computing node recommendation index Y Ti , the specific formula is:

[0065] It should be specifically noted in this embodiment that the preset values ​​and weight coefficients used are selected based on actual needs and are not limited to specific values ​​here.

[0066] The database is used to store data of all modules in the system.

[0067] like Figure 2 This embodiment provides an operation process of a hardware chip computing power measurement system based on computing power nodes, including the following steps:

[0068] S1: Classify computing chips and computing service scenarios based on the existing computing power background;

[0069] S2: Evaluate the computing power of the classified computing service scenarios and automatically back up the specific evaluation algorithm;

[0070] S3: Collects static configuration data, dynamic resource data, and historical service data of different computing power nodes;

[0071] S4: Process the collected computing power node data and calculate the node static configuration performance excellence coefficient, node remaining service capacity coefficient, and node computing power service stability control coefficient respectively;

[0072] S5: Calculates the node's remaining computing power service quality index based on the node's static configuration performance excellence coefficient, the node's remaining service capacity coefficient, and the node's computing power service stability control coefficient, and automatically backs up the calculation results.

[0073] S6: Identify the computing power demand information of the computing power demanders in the order in which they applied for the computing power demand orders, and confirm the computing power demand scenarios and the computing power demanders' requirements for computing power services;

[0074] S7: Filter computing power nodes based on the identified computing power demand scenarios, sort the filtered computing power nodes based on the computing power service requirements and the node remaining computing power service quality index, and generate recommended solutions.

[0075] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A hardware chip computing power measurement system based on computing power nodes, characterized by: include: Intelligent classification module: classifies computing chips and computing service scenarios based on the existing computing power background; Service scenario computing capacity evaluation module: evaluates the computing capacity of classified computing service scenarios and automatically backs up specific evaluation algorithms; Computing power node data collection module: collects static configuration data, dynamic resource data, and historical service data of different computing power nodes; Computing power node data processing module: processes the collected computing power node data and calculates the node static configuration performance excellence coefficient, node remaining service capacity coefficient, and node computing power service stability control coefficient; Computing power node comprehensive evaluation module: Calculates the node's remaining computing power service quality index based on the node's static configuration performance excellence coefficient, the node's remaining service capacity coefficient, and the node's computing power service stability control coefficient, and automatically backs up the calculation results; Computing power demand information identification module: This module identifies the computing power demand information of computing power demanders in the order in which they apply for computing power demand orders, confirms the computing power demand scenarios and the computing power demanders' requirements for computing power services; Computing power node recommendation module: This module selects computing power nodes based on the identified computing power demand scenarios, sorts the selected computing power nodes based on the computing power service requirements and the node remaining computing power service quality index, and generates a recommendation plan.

2. The hardware chip computing power measurement system based on computing power nodes according to claim 1, characterized in that: The intelligent classification module includes a computing power chip classification unit and a computing power service scenario classification unit. The computing power chip classification unit divides the types of computing power chips into four categories: central processing unit (CPU), graphics processing unit (GPU), embedded neural network processor (NPU), and tensor processing unit (TPU); the computing power service scenario classification unit divides computing power service scenarios into three categories: general computing power, artificial intelligence computing power, and supercomputing computing power. The general computing power is mainly carried by the central processing unit (CPU), and is mainly aimed at general software applications and executing logical operations. The artificial intelligence computing power is mainly carried by the embedded neural network processor (NPU), tensor processing unit (TPU), and graphics processing unit (GPU). It is mainly aimed at AI applications and is characterized by simple logic, intensive calculations, and high concurrency. The supercomputing computing power is mainly carried by the central processing unit (CPU) and graphics processing unit (GPU). It is mainly aimed at scientific computing and industrial simulation scenarios, with complex calculations and high requirements for calculation accuracy.

3. The hardware chip computing power measurement system based on computing power nodes according to claim 1, characterized in that: The service scenario computing capacity evaluation module includes a general computing capacity evaluation unit, an artificial intelligence computing capacity evaluation unit, a super computing capacity evaluation unit, and an automatic backup unit. The general computing capacity evaluation unit is based on the number n of CPUs of the i-th model. cai 、CPU core number m cai , single core main frequency h cai , and the number of floating-point operations per CPU cycle f cai Calculate the general computing power coefficient X TS The specific formula is: n xca The number of CPU models used in general computing power evaluation; the number of NPUs of the i-th model n based on the artificial intelligence computing power evaluation unit ai , the precision d of the jth data type processed sij , the working frequency h when processing the jth data type precision gij , the number of GPUs of the kth model n gak , the number of instructions executed in the clock cycle f gak 、GPU core number m gak , GPU core operating frequency h gak , the number of TPUs of the first model n tl , and the amount of computation f processed by the TPU in a single clock cycle tl Calculate the AI ​​computing power coefficient X RZ The specific formula is: Among them, a1, a2, and a3 are weight coefficients, a1>0, a2>0, a3>0, n xa 、n xdi 、n xga 、n xt They are the number of NPU models, the number of data type precision types of the i-th NPU model, the number of GPU models when evaluating artificial intelligence computing power, and the number of TPU models; the supercomputing computing power evaluation unit is based on the number n of i-th CPU models cbi 、CPU core number m cbi , single core main frequency h cbi 、Number of floating-point operations per CPU cycle f cbi , the number of GPUs of type j n gbj , the number of instructions executed in the clock cycle f gbj 、GPU core number m gbj , and the GPU core operating frequency h gbj Calculate the supercomputing power coefficient X CS The specific formula is: where n xcb 、n xgb The number of CPU models and the number of GPU models used in supercomputing computing power evaluation are b1 and b2, respectively, and b1>0 and b2>0. The automatic backup unit is used to automatically back up the calculation formulas for computing power in different service scenarios.

4. The hardware chip computing power measurement system based on computing power nodes according to claim 1, characterized in that: The computing power node data collection module includes a node static configuration data collection unit, a node dynamic resource data collection unit, a node historical service data collection unit, and a data output unit. The node static data collection unit is used to collect the model, quantity, performance parameters, node storage capacity, storage bandwidth, number of read and write operations per second of the node's processor, node memory capacity, node memory bandwidth, and node network connection bandwidth of the node; the node dynamic data collection unit is used to collect the node's CPU idle rate, GPU idle rate, NPU idle rate, TPU idle rate, node remaining storage capacity, node remaining memory capacity, data transmission delay, number of node service connections, and throughput rate; the node historical service data collection unit is used to collect the total computing power service time, fault handling time, and abnormality duration of each historical computing power service; The data output unit is used to send the collected node computing power data to the computing power node data processing module.

5. The hardware chip computing power measurement system based on computing power nodes according to claim 1, characterized in that: The computing power node data processing module includes a data receiving unit, a static configuration performance evaluation unit, a dynamic remaining service support capability evaluation unit, a node computing power service stability evaluation unit, and a data output unit. The data receiving unit is used to receive the collected node static configuration data, dynamic resource data, and historical service data; the static configuration performance evaluation unit calculates the node static configuration performance excellence coefficient X based on the collected static configuration data of the i-th computing power node. PZi The dynamic remaining service support capability evaluation unit calculates the node remaining service capability coefficient Y based on the collected dynamic data of the i-th computing power node αi The node computing power service stability evaluation unit calculates the node computing power service stability control coefficient X based on the collected historical computing power service data of the i-th computing power node wdi ; The data output unit is used to send the calculated node static configuration performance excellence coefficient, node remaining service capacity coefficient and node computing power service stability control coefficient to the computing power node comprehensive evaluation module.

6. The hardware chip computing power measurement system based on computing power nodes according to claim 5, characterized in that: The specific data processing process in the computing power node data processing module is as follows: A1. Based on the model, quantity, and performance parameters of the processors contained in the i-th computing power node, calculate the general computing power coefficient X that the node can provide. TSi , artificial intelligence computing power coefficient X RZi And the supercomputing power coefficient X CSi ; A2, based on the node storage capacity r of the collected i-th computing power node ci , storage bandwidth v ci And the number of read and write operations per second of the off-node memory m cdi Compute node storage capacity coefficient X CCi , the specific formula is: Based on the node memory capacity r of the collected i-th computing power node hi and node memory bandwidth v hi Compute node cache capacity coefficient X HCi , the specific formula is: Based on the node network connection bandwidth v of the collected i-th computing power node wi Calculate the node communication capability coefficient X Ti , the specific formula is: X Ti =v wi 2 +v wi +1; A3. Compute node static configuration performance excellence coefficient X PZi , the specific formula is: A4. Based on the collected CPU idle rate θ of the i-th computing power node ci , GPU idle rate θ gi , NPU idle rate θ ai , TPU idle rate θ ti , node remaining storage capacity r yci , node remaining memory capacity r yhi Calculate the computing power resource surplus rate α syi , the specific formula is: e is a natural constant, based on the data transmission delay T of the collected i-th computing power node ai Number of connections to node services m li Calculate the communication quality control capability coefficient α txi , the specific formula is: Based on the collected throughput θ of the i-th computing power node ei Calculate the business processing capacity coefficient α ywi , the specific formula is: ywi =∫(θ ei ) 2 d(θ ei ); A5. Calculate the node's remaining service capacity coefficient Y αi , the specific formula is: αi =ln(α syi +e)+ln(α txi +e)+ln(α ywi +e), e is a natural constant; A6. The total computing service time T based on the collected historical computing service of the i-th computing node zij , Troubleshooting time gij And the abnormal duration T yij Computing node computing power service stability control coefficient X wdi , the specific formula is: n fwi is the number of historical computing power services collected for the i-th computing power node, and e is a natural constant.

7. The hardware chip computing power measurement system based on computing power nodes according to claim 1, characterized in that: The computing power node comprehensive evaluation module includes a data receiving unit, a node remaining computing power service quality index calculation unit and an automatic backup unit. The data receiving unit is used to receive the calculated node static configuration performance excellence coefficient, node remaining service capacity coefficient and node computing power service stability control coefficient of different computing power nodes; the node remaining computing power service quality index calculation unit is based on the received node static configuration performance excellence coefficient X of the i-th computing power node. PZi , node remaining service capacity coefficient Y αi And the node computing power service stability control coefficient X wdi Compute node remaining computing power service quality index Y sei The specific formula is: e is a natural constant; the automatic backup unit is used to automatically back up the calculated node residual computing power service quality index of each computing power node.

8. The hardware chip computing power measurement system based on computing power nodes according to claim 1, characterized in that: The computing power node recommendation module includes a computing power node condition screening unit, a computing power service requirement judgment unit, a computing power service requirement keyword extraction unit, a keyword weight setting unit, a node recommendation index calculation unit, a node recommendation scheme determination unit, and a node recommendation scheme output unit. The computing power node condition screening unit screens the existing computing power nodes based on the identified computing power demand scenario of the computing power demander, retains the computing power nodes containing the same computing power service scenario, and screens out the computing power nodes that do not contain the same computing power service scenario; the computing power service requirement judgment unit is used to judge whether the computing power demander has computing power service requirements, and if so, sends a notification with computing power service requirements. A prompt is sent to the computing power service requirement keyword extraction unit, and if no computing power service requirement exists, a prompt of no computing power service requirement is sent to the node recommendation scheme determination unit; the computing power service requirement keyword extraction unit is used to extract the computing power service requirement keywords of the computing power demander; the keyword weight setting unit is used to set the weight of the extracted computing power service requirement keywords, and when the computing power demander does not mark the order, the keywords are set with the same weight, and when the computing power demander marks the order, the weight is set in combination with the number of keywords; the node recommendation index calculation unit adjusts the calculation formula of the node remaining computing power service quality index based on the set keyword weight to calculate the node recommendation index; When there is no computing power service requirement, the node recommendation scheme determination unit will arrange the calculated remaining computing power service quality index of each node from high to low according to the numerical value and number them in the order of arrangement, and retain a preset number of computing power nodes as recommended schemes according to the numbering. When it is determined that there is a computing power service requirement, the calculated node recommendation index will be arranged from high to low according to the numerical value and numbered in the order of arrangement, and retain a preset number of computing power nodes as recommended schemes according to the numbering, and the one numbered i becomes the i-th recommended scheme; the node recommendation scheme output unit is used to output the node recommendation scheme to the client.

Images