Energy efficiency sensing and load distribution method and system for distributed network acceleration nodes

Abstract

This application provides a method and system for energy efficiency perception and load allocation of distributed network acceleration nodes, relating to the field of communication technology. The system inputs the server performance profile characteristics of each service server into an energy efficiency prediction model to determine the Nth estimated energy efficiency of each service server under different loads. The Nth estimated energy efficiency, Nth total service access volume, and Nth existing load of each service server under different loads are then input into a service access volume allocation model to allocate service sub-access volumes to each service server. Based on the Nth actual energy efficiency of multiple service servers, the Nth average actual energy efficiency of the multiple service servers is determined. If the Nth average energy efficiency is less than the (N-1)th average energy efficiency, and the difference between the Nth average energy efficiency and the (N-1)th average energy efficiency is greater than a set difference threshold, a reinforcement learning model is used to update the network parameters of the service access volume allocation model.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a method and system for energy efficiency sensing and load distribution of distributed network acceleration nodes. Background Technology

[0002] A distributed network is a network architecture consisting of interconnected computer systems located in different places. There is no central node in the network; the nodes cooperate to share resources and process tasks. Acceleration nodes are key components in distributed networks, used to optimize network performance and accelerate data transmission.

[0003] Among them, the energy efficiency of each server in the distributed network (i.e., the ratio between energy consumption and actual workload) is related to the operating cost of the distributed network and the user experience, and is an indicator that technical personnel are particularly concerned about.

[0004] Currently, in some scenarios, it's possible to keep the total number of accesses to a distributed network constant per unit time (such as batch processing scenarios where scheduled tasks execute at a fixed frequency, and registration systems controlling the number of registrations at a fixed rate). In these scenarios, DNS round-robin can be used to balance the load across servers in the distributed network, thereby improving resource utilization, reducing latency and energy consumption, and ultimately reducing the energy efficiency of the distributed network. However, even with these methods, the energy efficiency of the distributed network remains relatively high. Summary of the Invention

[0005] This application provides a method and system for energy efficiency sensing and load distribution of distributed network acceleration nodes, which is used to solve the problem of high energy efficiency of each server in a distributed network under the scenario where the total number of accesses per unit time remains unchanged in the prior art.

[0006] Firstly, this application provides an energy efficiency awareness and load distribution method for distributed network acceleration nodes, applied to distributed network acceleration nodes that communicate with multiple subordinate service servers. The method provided in this application includes:

[0007] Step 1: In the Nth time slice, obtain the server type of each business server, the environmental data of each business server in the Nth time slice, and the hardware health status of each business server, where N is an integer greater than or equal to 2.

[0008] Step 2: For each business server, construct server performance profile features based on the server type, the environmental data of the business server in the Nth time slice, and the hardware health status of the business server.

[0009] Step 3: Input the server performance profile features of each business server into the pre-trained energy efficiency prediction model to determine the Nth estimated energy efficiency of each business server under different loads. The energy efficiency prediction model is trained by inputting multiple first training samples into the first neural network. Each first training sample includes historical server performance profile features and the corresponding historical Nth estimated energy efficiency.

[0010] Step 4: Obtain the Nth total service access volume of the terminal device set and the Nth existing load of each service server;

[0011] Step 5: Input the Nth estimated energy efficiency, Nth total service access volume, and Nth existing load of each service server under different loads into the pre-trained service access volume allocation model, so that the service access volume allocation model allocates service sub-access volume to each service server according to the pre-configured network parameters. The service access volume allocation model is trained by inputting multiple second training samples into the second neural network. Each second training sample includes the historical Nth estimated energy efficiency of each historical service server under different loads, the historical Nth total service access volume of the historical terminal device set, and the historical Nth existing load of each historical service server.

[0012] Step 6: Detect the Nth actual energy efficiency of each service server, and determine the Nth average actual energy efficiency of multiple service servers based on the Nth actual energy efficiency of each service server.

[0013] Step 7: If the Nth average energy efficiency of multiple service servers is less than the N-1th average energy efficiency of multiple service servers, and the difference between the Nth average energy efficiency and the N-1th average energy efficiency is greater than the set difference threshold, then the network parameters of the service access volume allocation model are updated using a reinforcement learning model based on the Nth average energy efficiency.

[0014] Step 8: Increment the value of N by 1, return to step 1, until the difference between the average energy efficiency of the Nth and the average energy efficiency of the (N-1)th is less than or equal to the set difference threshold.

[0015] Step 9: Repeat steps 1-5 until the distributed network acceleration node is powered down.

[0016] In some implementations, before obtaining the Nth total service access volume of the terminal device set and the Nth existing load of each service server, the method provided in this application further includes:

[0017] For each load on each business server, obtain the historical energy efficiency of the business server under load for multiple consecutive time slices within a previously preset duration.

[0018] Based on the Deep Beliefs Network (DBN), extract the time-series dependency of the historical energy efficiency of the business server under load for multiple consecutive time slices within a previously preset duration.

[0019] Based on the historical energy efficiency of the business server under load for multiple consecutive time slices within a previously preset duration, and the temporal dependency of the historical energy efficiency of the business server under load for multiple consecutive time slices within a previously preset duration, the energy efficiency variation characteristics of the business server under load are obtained.

[0020] For each business server, the energy efficiency change features under each load are input into a pre-trained energy efficiency prediction model to predict the Nth calibrated estimated energy efficiency of the business server under each load in the Nth time slice. The energy efficiency prediction model is trained by inputting multiple third training samples into a long short-term memory network. Each third training sample includes the historical energy efficiency change features of the historical business server under each load and the corresponding historical Nth calibrated estimated energy efficiency of the historical business server under each load in the historical Nth time slice.

[0021] For each business server, calibrate the Nth estimated energy efficiency under the corresponding load based on the Nth calibration estimated energy efficiency under each load.

[0022] In some implementations, for each service server, the Nth estimated energy efficiency under the corresponding load is calibrated based on the Nth calibration estimated energy efficiency under each load, including:

[0023] According to the formula calibrate the Nth estimated energy efficiency under the corresponding load, where, For the j-th load of the i-th service server, estimate the N-th energy efficiency before calibration in the N-th time slice. For the j-th load of the i-th service server, estimate the energy efficiency for calibration in time slice N. The calibrated energy efficiency of the j-th load of the i-th service server in the N-th time slice is estimated. The first weighting coefficient, It is the second weighting coefficient, and .

[0024] In some implementations, obtaining the hardware health status of each service server includes:

[0025] Obtain the wear level of the solid-state drives, the accuracy of the operation logs, and the stability of the power supply voltage for each business server;

[0026] The wear and tear of the solid-state drives (SSDs) of each business server, the accuracy of the operation logs, and the stability of the power supply voltage are input into a pre-trained hardware health determination model to determine the hardware health of each business server. The hardware health determination model is trained by inputting multiple fourth training samples into a third neural network. Each fourth training sample includes the historical wear and tear of the SSDs of the historical business servers, the accuracy of the historical operation logs, the stability of the historical power supply voltage, and the corresponding historical actual hardware health.

[0027] In some implementations, the network parameters of the traffic allocation model are the state of the reinforcement learning model, updating the network parameters of the traffic allocation model is the action of the reinforcement learning model, and the Nth average energy efficiency is the reward of the reinforcement learning model, with a higher reward for a lower Nth average energy efficiency.

[0028] Secondly, this application also provides an energy efficiency sensing and load distribution system for distributed network acceleration nodes, applied to distributed network acceleration nodes, which are respectively connected to multiple subordinate business servers. The system provided by this application includes:

[0029] The data acquisition unit is used to acquire, in the Nth time slice, the server type of each of the multiple business servers, the environmental data of each business server in the Nth time slice, and the hardware health status of each business server, where N is an integer greater than or equal to 2.

[0030] The feature construction unit is used to construct server performance profile features for each business server based on the server type, the environmental data of the business server in the Nth time slice, and the hardware health status of the business server.

[0031] The energy efficiency determination unit is used to input the server performance profile features of each business server into a pre-trained energy efficiency prediction model to determine the Nth estimated energy efficiency of each business server under different loads. The energy efficiency prediction model is trained by inputting multiple first training samples into a first neural network. Each first training sample includes historical server performance profile features and the corresponding historical Nth estimated energy efficiency.

[0032] The data acquisition unit is also used to acquire the Nth total service access volume of the terminal device set and the Nth existing load of each service server;

[0033] The access volume allocation unit is used to input the Nth estimated energy efficiency, Nth total service access volume, and Nth existing load of each service server under different loads into the pre-trained service access volume allocation model, so that the service access volume allocation model allocates service sub-access volume to each service server according to the pre-configured network parameters. The service access volume allocation model is trained by inputting multiple second training samples into the second neural network. Each second training sample includes the historical Nth estimated energy efficiency of each historical service server under different loads, the historical Nth total service access volume of the historical terminal device set, and the historical Nth existing load of each historical service server.

[0034] The energy efficiency detection unit is used to detect the Nth actual energy efficiency of each business server, and determine the Nth average actual energy efficiency of multiple business servers based on the Nth actual energy efficiency of each business server.

[0035] The parameter update unit is used to update the network parameters of the service access volume allocation model based on the Nth average energy efficiency of multiple service servers if the Nth average energy efficiency of multiple service servers is less than the N-1th average energy efficiency of multiple service servers, and the difference between the Nth average energy efficiency and the N-1th average energy efficiency is greater than a set difference threshold. The unit increments the value of N by 1 and returns to step 1 until the difference between the Nth average energy efficiency and the N-1th average energy efficiency is less than or equal to the set difference threshold.

[0036] The business execution unit is used to execute steps 1-5 in a loop until the distributed network acceleration node is powered down.

[0037] In some implementations, the data acquisition unit is also used to acquire, for each load of each service server, the historical energy efficiency of the service server under load for multiple consecutive time slices within a previously preset duration.

[0038] The apparatus provided in this application further includes: a relation extraction unit, used to extract the time-series dependency of the historical energy efficiency of the business server under load for multiple consecutive time slices within a previously preset duration, based on the deep belief DBN network;

[0039] The feature acquisition unit is used to obtain the energy efficiency change features of the business server under load based on the historical energy efficiency of the business server under load for multiple consecutive time slices within a previous preset duration, and the temporal dependency of the historical energy efficiency of the business server under load for multiple consecutive time slices within a previous preset duration.

[0040] The energy efficiency prediction unit is used to input the energy efficiency change characteristics under each load into a pre-trained energy efficiency prediction model for each service server, so as to predict the Nth calibrated estimated energy efficiency of the service server under each load in the Nth time slice. The energy efficiency prediction model is trained by inputting multiple third training samples into a long short-term memory network. Each third training sample includes the historical energy efficiency change characteristics of the historical service server under each load and the corresponding historical Nth calibrated estimated energy efficiency of the historical service server under each load in the historical Nth time slice.

[0041] The energy efficiency calibration unit is used to calibrate the Nth estimated energy efficiency under each load for each business server, based on the Nth calibration estimated energy efficiency under each load.

[0042] In some implementations, the energy efficiency calibration unit is specifically used to perform a calculation based on a formula. calibrate the Nth estimated energy efficiency under the corresponding load, where, For the j-th load of the i-th service server, estimate the N-th energy efficiency before calibration in the N-th time slice. For the j-th load of the i-th service server, estimate the energy efficiency for calibration in time slice N. The calibrated energy efficiency of the j-th load of the i-th service server in the N-th time slice is estimated. The first weighting coefficient, It is the second weighting coefficient, and .

[0043] In some implementations, the data acquisition unit is specifically used to acquire the wear level of the solid-state drive (SSD) of each business server, the accuracy of the operation logs, and the stability of the power supply voltage; input the wear level of the SSD of each business server, the accuracy of the operation logs, and the stability of the power supply voltage into a pre-trained hardware health determination model to determine the hardware health of each business server. The hardware health determination model is trained by inputting multiple fourth training samples into a third neural network. Each fourth training sample includes the historical wear level of the SSD of a historical business server, the accuracy of historical operation logs, the stability of historical power supply voltage, and the corresponding historical actual hardware health.

[0044] In some implementations, the network parameters of the traffic allocation model are the state of the reinforcement learning model, updating the network parameters of the traffic allocation model is the action of the reinforcement learning model, and the Nth average energy efficiency is the reward of the reinforcement learning model, with a higher reward for a lower Nth average energy efficiency.

[0045] This application provides a method and system for energy efficiency sensing and load allocation of distributed network acceleration nodes. For each business server, a server performance profile is constructed based on the server type, environmental data of the business server in the Nth time slice, and the hardware health status of the business server. Understandably, different server types have different processing capabilities and energy efficiencies; the processing capability of the same business server varies under different environmental conditions (such as temperature and / or humidity). For example, excessively high temperatures can cause the business server to reduce its frequency, and different temperatures can cause different degrees of frequency reduction, thus reducing the computing power of the business server. Excessively high humidity can corrode the hardware of the business server, and different humidity levels can cause different degrees of corrosion, leading to a decrease in the processing capability of the business server. Furthermore, different hardware health statuses of business servers correspond to different processing capabilities. Therefore, the server performance profile constructed in the above manner has high reliability.

[0046] The server performance profile features of each business server are input into a pre-trained energy efficiency prediction model to determine the Nth estimated energy efficiency of each business server under different loads. Because the server performance profile features are highly reliable, and because the energy efficiency prediction model is trained by inputting multiple first training samples into a first neural network, with each first training sample including historical server performance profile features and the corresponding historical Nth estimated energy efficiency, the reliability of determining the Nth estimated energy efficiency of each business server under different loads is high.

[0047] Furthermore, the Nth total service access volume of the terminal device set and the Nth existing load of each service server are obtained. Moreover, server energy efficiency varies under different loads (e.g., at 0%-60% load, each additional 10% load only results in a 6% increase in power consumption; at 60%-80% load, each additional 10% load results in a 12%-15% increase in power consumption; and at loads greater than 80%, each additional 5% load results in a 15% increase in power consumption). Therefore, rationally allocating the Nth total service access volume to various service servers (i.e., assigning different loads to different service servers) can result in lower energy efficiency for each service server.

[0048] Furthermore, the estimated Nth energy efficiency, total Nth service access volume, and existing Nth load of each service server under different loads are input into a pre-trained service access volume allocation model. This model allocates service sub-access volumes to each service server based on pre-configured network parameters. Since the service access volume allocation model is trained by inputting multiple second training samples into a second neural network, each second training sample includes the historical estimated Nth energy efficiency of each historical service server under different loads, the historical total Nth service access volume of the historical terminal device set, and the historical existing Nth load of each historical service server. In this way, the total Nth service access volume can be rationally allocated to each service server.

[0049] Step 6: Detect the Nth actual energy efficiency of each service server, and determine the Nth average actual energy efficiency of multiple service servers based on the Nth actual energy efficiency of each service server.

[0050] Step 7: If the Nth average energy efficiency of multiple service servers is less than the (N-1)th average energy efficiency of multiple service servers, and the difference between the Nth and (N-1)th average energy efficiency is greater than a set difference threshold, it indicates that the Nth average energy efficiency of the multiple service servers still has room for reduction. Therefore, based on the Nth average energy efficiency, the network parameters of the service access volume allocation model are updated using a reinforcement learning model. The value of N is incremented by 1, and the process returns to Step 1 until the difference between the Nth and (N-1)th average energy efficiency is less than or equal to the set difference threshold. This ensures that the average energy efficiency of the multiple service servers is minimized.

[0051] Finally, steps 1-5 are executed repeatedly until the distributed network acceleration node is powered down. By operating each business server at its lowest average energy efficiency, the operating costs of the distributed network are saved while ensuring its efficiency and performance. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a schematic diagram illustrating the interaction between the distributed network acceleration node provided in this application embodiment and a set of terminal devices and multiple service servers;

[0054] Figure 2 A flowchart illustrating the energy efficiency sensing and load allocation method for distributed network acceleration nodes provided in this application embodiment;

[0055] Figure 3 This is a functional block diagram of the energy efficiency sensing and load distribution system for distributed network acceleration nodes provided in an embodiment of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments made by those skilled in the art under the guidance of these embodiments are within the scope of protection of this application.

[0057] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0058] Explanation of technical terms used in this patent application:

[0059] Business server energy efficiency refers to the ratio between the energy consumed by a business server when providing computing services and the actual effective work completed. It reflects the efficiency with which the server converts electrical energy into useful computing power.

[0060] The technical solutions of this application and how they solve the aforementioned technical problems will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0061] This application provides an energy efficiency awareness and load distribution method for distributed network acceleration nodes, applicable to distributed network acceleration nodes. For example... Figure 1As shown, the distributed network acceleration nodes communicate with multiple subordinate business servers and a collection of terminal devices. In scenarios such as batch processing jobs where scheduled tasks execute at a fixed frequency and registration systems controlling registration volume at a fixed rate, the total number of accesses from the terminal device collection to the multiple business servers remains constant per unit time. Figure 2 As shown, the method provided in this application embodiment includes:

[0062] Step 1: In the Nth time slice, obtain the server type of each business server, the environmental data of each business server in the Nth time slice, and the hardware health status of each business server, where N is an integer greater than or equal to 2.

[0063] It should be noted that the duration of the Nth time slice mentioned above can be, but is not limited to, 1 minute, 2 minutes, 5 minutes, 10 minutes, etc., and is not limited here. The business server is used to process tasks corresponding to batch processing jobs, or to process registration business, etc., and is not limited here.

[0064] Furthermore, obtaining the hardware health status of each business server can be specifically implemented as follows: Obtain the wear level of the solid-state drive (SSD) of each business server, the accuracy of the operation logs, and the stability of the power supply voltage; input these data into a pre-trained hardware health status determination model to determine the hardware health status of each business server. Specifically, the hardware health status determination model is trained by inputting multiple fourth training samples into a third neural network. Each fourth training sample includes the historical wear level of the SSD of a historical business server, the accuracy of historical operation logs, the stability of historical power supply voltage, and the corresponding historical actual hardware health status. This approach ensures high reliability of the obtained hardware health status.

[0065] Step 2: For each business server, construct a server performance profile based on the server type, the environmental data of the business server in the Nth time slice, and the hardware health status of the business server.

[0066] Understandably, different types of server business servers have different processing capabilities and energy efficiency. The processing capability of the same business server varies under different environmental conditions (such as temperature and / or humidity). For example, excessively high temperatures can cause the business server to reduce its frequency, and the degree of frequency reduction varies with different temperatures, thus reducing the server's processing capability. Excessive humidity can corrode the hardware of the business server (e.g., high humidity can oxidize copper foil, solder joints, and connectors on circuit boards, leading to increased contact resistance; high humidity can also increase dust accumulation on circuit boards; and high humidity can cause the circuit board surface to absorb trace amounts of moisture, reducing insulation resistance and causing a slight increase in leakage current, thus increasing static power consumption). Different humidity levels cause varying degrees of corrosion to the business server hardware, leading to a decrease in processing capability. Furthermore, different levels of hardware health in a business server result in different processing capabilities. Therefore, the server performance profile characteristics constructed using the above methods are highly reliable.

[0067] Step 3: Input the server performance profile features of each business server into the pre-trained energy efficiency prediction model to determine the Nth estimated energy efficiency of each business server under different loads. The energy efficiency prediction model is trained by inputting multiple first training samples into the first neural network. Each first training sample includes historical server performance profile features and the corresponding historical Nth estimated energy efficiency.

[0068] Because the server performance profile features are highly reliable, and because the energy efficiency prediction model is trained by inputting multiple first training samples into a first neural network, with each first training sample including historical server performance profile features and the corresponding historical Nth estimated energy efficiency, the reliability of determining the Nth estimated energy efficiency of each business server under different loads can be highly accurate.

[0069] It should be noted that the method provided in the embodiments of this application may further include:

[0070] Step A: For each load of each business server, obtain the historical energy efficiency of the business server under load for multiple consecutive time slices within a previously preset duration.

[0071] For example, if the Nth time slice is 9:00-9:05, then the historical energy efficiency of load j1 under business server i1 during the previous day (i.e., the previous preset duration) during multiple consecutive time slices such as 8:55-9:00, 8:50-8:55, and 8:45-8:50; and the historical energy efficiency of load j2 under business server i1 during the previous day (i.e., the previous preset duration) during multiple consecutive time slices such as 8:55-9:00, 8:50-8:55, and 8:45-8:50, etc.

[0072] And the historical energy efficiency of load j1 under business server i2 during the previous day (i.e., the previous preset duration) for multiple consecutive time slices such as 8:55-9:00, 8:50-8:55, 8:45-8:50, etc.; the historical energy efficiency of load j2 under business server i2 during the previous day (i.e., the previous preset duration) for multiple consecutive time slices such as 8:55-9:00, 8:50-8:55, 8:45-8:50, etc., etc.

[0073] Step B: Based on the Deep Belief Network (DBN), extract the time-series dependency of the historical energy efficiency of the business server under load for multiple consecutive time slices within a previously preset time period.

[0074] For example, extract the time-series dependency of historical energy efficiency of service server i1 under load j1 for a previous preset duration and multiple consecutive time slices; extract the time-series dependency of historical energy efficiency of service server i1 under load j2 for a previous preset duration and multiple consecutive time slices; extract the time-series dependency of historical energy efficiency of service server i2 under load j1 for a previous preset duration and multiple consecutive time slices; extract the time-series dependency of historical energy efficiency of service server i2 under load j2 for a previous preset duration and multiple consecutive time slices, and so on.

[0075] Step C: Based on the historical energy efficiency of the business server under load for multiple consecutive time slices within the previous preset duration, and the temporal dependency of the historical energy efficiency of the business server under load for multiple consecutive time slices within the previous preset duration, obtain the energy efficiency change characteristics of the business server under load.

[0076] Step C is based on the same principle as steps A and B above, and will not be repeated here.

[0077] Step D: For each service server, input the energy efficiency variation characteristics under each load into the pre-trained energy efficiency prediction model to predict the Nth calibrated estimated energy efficiency of the service server under each load at the Nth time slice. Understandably, the Nth calibrated estimated energy efficiency obtained based on steps A-D above has high reliability.

[0078] The energy efficiency prediction model is trained by inputting multiple third training samples into a long short-term memory network. Each third training sample includes the historical energy efficiency change characteristics of the historical service server under each load and the corresponding historical Nth time slice, and the historical Nth calibration predicted energy efficiency of the historical service server under each load.

[0079] Step E: For each service server, calibrate the Nth estimated energy efficiency under each load, based on the Nth calibration estimated energy efficiency. This further improves the reliability of the Nth estimated energy efficiency under the corresponding load.

[0080] For example, it can be based on the formula calibrate the Nth estimated energy efficiency under the corresponding load, where, For the j-th load of the i-th service server, estimate the N-th energy efficiency before calibration in the N-th time slice. For the j-th load of the i-th service server, estimate the energy efficiency for calibration in time slice N. The calibrated energy efficiency of the j-th load of the i-th service server in the N-th time slice is estimated. The first weighting coefficient, It is the second weighting coefficient, and .For example, ; .

[0081] Step 4: Obtain the Nth total service access volume of the terminal device set and the Nth existing load of each service server.

[0082] It's important to note that server energy efficiency varies under different loads. For example, with a load of 0%-60%, every additional 10% of load only results in a 6% increase in power consumption; with a load of 60%-80%, every additional 10% of load results in a 12%-15% increase in power consumption; and with a load exceeding 80%, every additional 5% of load results in a 15% increase in power consumption. Therefore, rationally distributing the Nth total service traffic across various service servers (i.e., assigning different loads to different service servers) can result in lower energy efficiency for each service server.

[0083] Step 5: Input the estimated energy efficiency of each service server under different loads, the total service access volume of the Nth load, and the existing load of each service server into the pre-trained service access volume allocation model, so that the service access volume allocation model allocates service sub-access volume to each service server according to the pre-configured network parameters.

[0084] The business access volume allocation model is obtained by inputting multiple second training samples into the second neural network for training. Each second training sample includes the historical Nth estimated energy efficiency of each historical business server under different loads, the historical Nth total business access volume of the historical terminal device set, and the historical Nth existing load of each historical business server.

[0085] Since the service access volume allocation model is trained by inputting multiple second training samples into a second neural network, each second training sample includes the historical Nth estimated energy efficiency of each historical service server under different loads, the historical Nth total service access volume of the historical terminal device set, and the historical Nth existing load of each historical service server. In this way, the Nth total service access volume can be reasonably allocated to each service server.

[0086] Step 6: Detect the Nth actual energy efficiency of each service server, and determine the Nth average actual energy efficiency of multiple service servers based on the Nth actual energy efficiency of each service server.

[0087] Step 7: If the Nth average energy efficiency of multiple service servers is less than the N-1th average energy efficiency of multiple service servers, and the difference between the Nth average energy efficiency and the N-1th average energy efficiency is greater than the set difference threshold, then the network parameters of the service access volume allocation model are updated using a reinforcement learning model based on the Nth average energy efficiency.

[0088] This allows for the lowest possible average energy efficiency across multiple business servers.

[0089] Step 8: Increment the value of N by 1, return to step 1, until the difference between the average energy efficiency of the Nth and the average energy efficiency of the (N-1)th is less than or equal to the set difference threshold.

[0090] Step 9: Repeat steps 1-5 until the distributed network acceleration node is powered down.

[0091] For example, the network parameters of the service access volume allocation model are the state of the reinforcement learning model, updating the network parameters of the service access volume allocation model is the action of the reinforcement learning model, and the Nth average energy efficiency is the reward of the reinforcement learning model, and the lower the Nth average energy efficiency, the higher the reward.

[0092] Understandably, when each business server operates at its lowest average energy efficiency, the operating costs of the distributed network are saved while ensuring the efficiency and performance of the distributed network.

[0093] In summary, the energy efficiency perception and load allocation method for distributed network acceleration nodes provided in this application constructs a server performance profile for each business server based on its server type, environmental data at time slice N, and hardware health. Understandably, different server types have different processing capabilities and energy efficiencies; the processing capability of the same business server varies under different environmental conditions (such as temperature and / or humidity). For example, excessively high temperatures can cause the business server to reduce its frequency, and different temperatures can cause different degrees of frequency reduction, thus reducing the server's computing power. Excessively high humidity can corrode the hardware of the business server, and different humidity levels can cause different degrees of corrosion, leading to a decrease in processing power. Furthermore, different hardware health levels of business servers correspond to different processing capabilities. Therefore, the server performance profile constructed in the above manner has high reliability.

[0094] The server performance profile features of each business server are input into a pre-trained energy efficiency prediction model to determine the Nth estimated energy efficiency of each business server under different loads. Because the server performance profile features are highly reliable, and because the energy efficiency prediction model is trained by inputting multiple first training samples into a first neural network, with each first training sample including historical server performance profile features and the corresponding historical Nth estimated energy efficiency, the reliability of determining the Nth estimated energy efficiency of each business server under different loads is high.

[0095] Furthermore, the Nth total service access volume of the terminal device set and the Nth existing load of each service server are obtained. Moreover, server energy efficiency varies under different loads (e.g., at 0%-60% load, each additional 10% load only results in a 6% increase in power consumption; at 60%-80% load, each additional 10% load results in a 12%-15% increase in power consumption; and at loads greater than 80%, each additional 5% load results in a 15% increase in power consumption). Therefore, rationally allocating the Nth total service access volume to various service servers (i.e., assigning different loads to different service servers) can result in lower energy efficiency for each service server.

[0096] Furthermore, the estimated Nth energy efficiency, total Nth service access volume, and existing Nth load of each service server under different loads are input into a pre-trained service access volume allocation model. This model allocates service sub-access volumes to each service server based on pre-configured network parameters. Since the service access volume allocation model is trained by inputting multiple second training samples into a second neural network, each second training sample includes the historical estimated Nth energy efficiency of each historical service server under different loads, the historical total Nth service access volume of the historical terminal device set, and the historical existing Nth load of each historical service server. In this way, the total Nth service access volume can be rationally allocated to each service server.

[0097] Step 6: Detect the Nth actual energy efficiency of each service server, and determine the Nth average actual energy efficiency of multiple service servers based on the Nth actual energy efficiency of each service server.

[0098] Step 7: If the Nth average energy efficiency of multiple service servers is less than the (N-1)th average energy efficiency of multiple service servers, and the difference between the Nth and (N-1)th average energy efficiency is greater than a set difference threshold, it indicates that the Nth average energy efficiency of the multiple service servers still has room for reduction. Therefore, based on the Nth average energy efficiency, the network parameters of the service access volume allocation model are updated using a reinforcement learning model. The value of N is incremented by 1, and the process returns to Step 1 until the difference between the Nth and (N-1)th average energy efficiency is less than or equal to the set difference threshold. This ensures that the average energy efficiency of the multiple service servers is minimized.

[0099] Finally, steps 1-5 are executed repeatedly until the distributed network acceleration node is powered down. By operating each business server at its lowest average energy efficiency, the operating costs of the distributed network are saved while ensuring its efficiency and performance.

[0100] Please see Figure 3 This application also provides an energy efficiency sensing and load distribution system for distributed network acceleration nodes, applied to distributed network acceleration nodes, which communicate with multiple subordinate service servers. It should be noted that the basic principle and technical effects of the energy efficiency sensing and load distribution system for distributed network acceleration nodes provided in this application are the same as those in the above embodiments. For the sake of brevity, any parts not mentioned in this application can be referred to the corresponding content in the above embodiments. Figure 3 As shown, the system provided in this application embodiment includes a data acquisition unit, a feature construction unit, an energy efficiency determination unit, an access volume allocation unit, an energy efficiency detection unit, a parameter update unit, and a business execution unit, wherein...

[0101] The data acquisition unit is used to acquire, in the Nth time slice, the server type of each of the multiple business servers, the environmental data of each business server in the Nth time slice, and the hardware health status of each business server, where N is an integer greater than or equal to 2.

[0102] The feature construction unit is used to construct server performance profile features for each business server based on the server type, the environmental data of the business server in the Nth time slice, and the hardware health status of the business server.

[0103] The energy efficiency determination unit is used to input the server performance profile features of each business server into a pre-trained energy efficiency prediction model to determine the Nth estimated energy efficiency of each business server under different loads. The energy efficiency prediction model is trained by inputting multiple first training samples into a first neural network. Each first training sample includes historical server performance profile features and the corresponding historical Nth estimated energy efficiency.

[0104] The data acquisition unit is also used to acquire the Nth total service access volume of the terminal device set and the Nth existing load of each service server;

[0105] The access volume allocation unit is used to input the Nth estimated energy efficiency, Nth total service access volume, and Nth existing load of each service server under different loads into the pre-trained service access volume allocation model, so that the service access volume allocation model allocates service sub-access volume to each service server according to the pre-configured network parameters. The service access volume allocation model is trained by inputting multiple second training samples into the second neural network. Each second training sample includes the historical Nth estimated energy efficiency of each historical service server under different loads, the historical Nth total service access volume of the historical terminal device set, and the historical Nth existing load of each historical service server.

[0106] The energy efficiency detection unit is used to detect the Nth actual energy efficiency of each business server, and determine the Nth average actual energy efficiency of multiple business servers based on the Nth actual energy efficiency of each business server.

[0107] The parameter update unit is used to update the network parameters of the service access volume allocation model based on the Nth average energy efficiency of multiple service servers if the Nth average energy efficiency of multiple service servers is less than the N-1th average energy efficiency of multiple service servers, and the difference between the Nth average energy efficiency and the N-1th average energy efficiency is greater than a set difference threshold. The unit increments the value of N by 1 and returns to step 1 until the difference between the Nth average energy efficiency and the N-1th average energy efficiency is less than or equal to the set difference threshold.

[0108] The business execution unit is used to execute steps 1-5 in a loop until the distributed network acceleration node is powered down.

[0109] In some implementations, the data acquisition unit is also used to acquire, for each load of each service server, the historical energy efficiency of the service server under load for multiple consecutive time slices within a previously preset duration.

[0110] The apparatus provided in this application further includes: a relation extraction unit, used to extract the time-series dependency of the historical energy efficiency of the business server under load for multiple consecutive time slices within a previously preset duration, based on the deep belief DBN network;

[0111] The feature acquisition unit is used to obtain the energy efficiency change features of the business server under load based on the historical energy efficiency of the business server under load for multiple consecutive time slices within a previous preset duration, and the temporal dependency of the historical energy efficiency of the business server under load for multiple consecutive time slices within a previous preset duration.

[0112] The energy efficiency prediction unit is used to input the energy efficiency change characteristics under each load into a pre-trained energy efficiency prediction model for each service server, so as to predict the Nth calibrated estimated energy efficiency of the service server under each load in the Nth time slice. The energy efficiency prediction model is trained by inputting multiple third training samples into a long short-term memory network. Each third training sample includes the historical energy efficiency change characteristics of the historical service server under each load and the corresponding historical Nth calibrated estimated energy efficiency of the historical service server under each load in the historical Nth time slice.

[0113] The energy efficiency calibration unit is used to calibrate the Nth estimated energy efficiency under each load for each business server, based on the Nth calibration estimated energy efficiency under each load.

[0114] In some implementations, the energy efficiency calibration unit is specifically used to perform a calculation based on a formula. calibrate the Nth estimated energy efficiency under the corresponding load, where, For the j-th load of the i-th service server, estimate the N-th energy efficiency before calibration in the N-th time slice. For the j-th load of the i-th service server, estimate the energy efficiency for calibration in time slice N. The calibrated energy efficiency of the j-th load of the i-th service server in the N-th time slice is estimated. The first weighting coefficient, It is the second weighting coefficient, and .

[0115] In some implementations, the data acquisition unit is specifically used to acquire the wear level of the solid-state drive (SSD) of each business server, the accuracy of the operation logs, and the stability of the power supply voltage; input the wear level of the SSD of each business server, the accuracy of the operation logs, and the stability of the power supply voltage into a pre-trained hardware health determination model to determine the hardware health of each business server. The hardware health determination model is trained by inputting multiple fourth training samples into a third neural network. Each fourth training sample includes the historical wear level of the SSD of a historical business server, the accuracy of historical operation logs, the stability of historical power supply voltage, and the corresponding historical actual hardware health.

[0116] In some implementations, the network parameters of the traffic allocation model are the state of the reinforcement learning model, updating the network parameters of the traffic allocation model is the action of the reinforcement learning model, and the Nth average energy efficiency is the reward of the reinforcement learning model, with a higher reward for a lower Nth average energy efficiency.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for energy efficiency sensing and load allocation of distributed network acceleration nodes, characterized in that, The method, applied to distributed network acceleration nodes, which are communicatively connected to multiple subordinate service servers, includes: Step 1: In the Nth time slice, obtain the server type of each of the plurality of service servers, the environmental data of each service server in the Nth time slice, and the hardware health status of each service server, where N is an integer greater than or equal to 2; Step 2: For each of the business servers, construct server performance profile features based on the server type of the business server, the environmental data of the business server in the Nth time slice, and the hardware health status of the business server; Step 3: Input the server performance profile features of each of the business servers into the pre-trained energy efficiency prediction model to determine the Nth estimated energy efficiency of each of the business servers under different loads. The energy efficiency prediction model is obtained by inputting multiple first training samples into the first neural network for training. Each first training sample includes historical server performance profile features and the corresponding historical Nth estimated energy efficiency. Step 4: Obtain the Nth total service access volume of the terminal device set and the Nth existing load of each of the service servers; Step 5: Input the Nth estimated energy efficiency of each service server under different loads, the Nth total service access volume, and the Nth existing load of each service server into the pre-trained service access volume allocation model, so that the service access volume allocation model allocates service sub-access volume to each service server according to the pre-configured network parameters. The service access volume allocation model is trained by inputting multiple second training samples into a second neural network. Each second training sample includes the historical Nth estimated energy efficiency of each historical service server under different loads, the historical Nth total service access volume of the historical terminal device set, and the historical Nth existing load of each historical service server. Step 6: Detect the Nth actual energy efficiency of each of the service servers, and determine the Nth average actual energy efficiency of the multiple service servers based on their respective Nth actual energy efficiency; Step 7: If the Nth average energy efficiency of the multiple service servers is less than the (N-1)th average energy efficiency of the multiple service servers, and the difference between the Nth average energy efficiency and the (N-1)th average energy efficiency is greater than a set difference threshold, then the network parameters of the service access volume allocation model are updated using a reinforcement learning model based on the Nth average energy efficiency. Step 8: Increment the value of N by 1, return to step 1, until the difference between the Nth average energy efficiency and the (N-1)th average energy efficiency is less than or equal to the set difference threshold. Step 9: Repeat steps 1-5 until the distributed network acceleration node is powered down.

2. The method according to claim 1, characterized in that, Before obtaining the Nth total service access volume of the terminal device set and the Nth existing load of each of the service servers, the method further includes: For each load of each of the aforementioned service servers, obtain the historical energy efficiency of the service server under the load for multiple consecutive time slices within a previously preset duration; Based on the Deep Belief Network (DBN), the historical energy efficiency of the service server under the load for multiple consecutive time slices within a previously preset time period is extracted, showing the time-series dependency relationship. Based on the historical energy efficiency of the service server under the load for multiple consecutive time slices within a previously preset time period, and the temporal dependence of the historical energy efficiency of the service server under the load for multiple consecutive time slices within a previously preset time period, the energy efficiency variation characteristics of the service server under the load are obtained. For each of the aforementioned service servers, the energy efficiency variation characteristics under each load are input into a pre-trained energy efficiency prediction model to predict the Nth calibrated estimated energy efficiency of the service server under each load in the Nth time slice. The energy efficiency prediction model is trained by inputting multiple third training samples into a long short-term memory network. Each third training sample includes the historical energy efficiency variation characteristics of the historical service server under each load and the corresponding historical Nth calibrated estimated energy efficiency of the historical service server under each load in the historical Nth time slice. For each of the aforementioned service servers, the Nth estimated energy efficiency under the corresponding load is calibrated based on the Nth calibration estimated energy efficiency under each load.

3. The method according to claim 2, characterized in that, For each of the aforementioned service servers, based on the Nth estimated energy efficiency under each load, the Nth estimated energy efficiency is calibrated, including: According to the formula calibrate the Nth estimated energy efficiency under the corresponding load, where, For the j-th load of the i-th service server, calculate the N-th estimated energy efficiency before calibration in the N-th time slice. For the j-th load of the i-th service server, estimate the energy efficiency for the N-th calibration in the N-th time slice. The calibrated estimated energy efficiency for the j-th load of the i-th service server in the N-th time slice. The first weighting coefficient, It is the second weighting coefficient, and .

4. The method according to claim 1, characterized in that, Obtaining the hardware health status of each business server includes: The wear level of the solid-state drive, the accuracy of the operation logs, and the stability of the power supply voltage of each of the aforementioned business servers are obtained. The wear level of the solid-state drive (SSD) of each business server, the accuracy of the operation logs, and the stability of the power supply voltage are input into a pre-trained hardware health determination model to determine the hardware health of each business server. The hardware health determination model is trained by inputting multiple fourth training samples into a third neural network. Each fourth training sample includes the historical wear level of the SSD of the historical business server, the accuracy of the historical operation logs, the stability of the historical power supply voltage, and the corresponding historical actual hardware health.

5. The method according to any one of claims 1-4, characterized in that, The network parameters of the service access volume allocation model are the state of the reinforcement learning model, updating the network parameters of the service access volume allocation model is the action of the reinforcement learning model, the Nth average energy efficiency is the reward of the reinforcement learning model, and the lower the Nth average energy efficiency, the higher the reward.

6. An energy efficiency sensing and load distribution system for distributed network acceleration nodes, characterized in that, The system is applied to distributed network acceleration nodes, which are communicatively connected to multiple subordinate service servers. The system includes: The data acquisition unit is used to acquire, in the Nth time slice, the server type of each of the plurality of service servers, the environmental data of each of the service servers in the Nth time slice, and the hardware health status of each service server, where N is an integer greater than or equal to 2. The feature construction unit is used to construct server performance profile features for each of the business servers based on the server type of the business server, the environmental data of the business server in the Nth time slice, and the hardware health status of the business server. An energy efficiency determination unit is used to input the server performance profile features of each of the business servers into a pre-trained energy efficiency prediction model to determine the Nth estimated energy efficiency of each of the business servers under different loads. The energy efficiency prediction model is obtained by inputting multiple first training samples into a first neural network for training. Each first training sample includes historical server performance profile features and the corresponding historical Nth estimated energy efficiency. The data acquisition unit is also used to acquire the Nth total service access volume of the terminal device set and the Nth existing load of each of the service servers; The access volume allocation unit is used to input the Nth estimated energy efficiency of each of the service servers under different loads, the Nth total service access volume, and the Nth existing load of each of the service servers into a pre-trained service access volume allocation model, so that the service access volume allocation model allocates service sub-access volume to each of the service servers according to pre-configured network parameters. The service access volume allocation model is obtained by training multiple second training samples into a second neural network. Each second training sample includes the historical Nth estimated energy efficiency of each historical service server under different loads, the historical Nth total service access volume of the historical terminal device set, and the historical Nth existing load of each of the historical service servers. An energy efficiency detection unit is used to detect the Nth actual energy efficiency of each of the service servers, and to determine the Nth average actual energy efficiency of the multiple service servers based on the Nth actual energy efficiency of each of the multiple service servers. The parameter update unit is configured to update the network parameters of the service access volume allocation model based on the Nth average energy efficiency of the multiple service servers if the Nth average energy efficiency is less than the N-1th average energy efficiency of the multiple service servers, and the difference between the Nth average energy efficiency and the N-1th average energy efficiency is greater than a set difference threshold; increment the value of N by 1, and return to step 1 until the difference between the Nth average energy efficiency and the N-1th average energy efficiency is less than or equal to the set difference threshold. The business execution unit is used to execute steps 1-5 in a loop until the distributed network acceleration node is powered off.

7. The system according to claim 6, characterized in that, The data acquisition unit is also used to acquire, for each load of each of the business servers, the historical energy efficiency of the business server under the load for multiple consecutive time slices within a previously preset duration. The system further includes: a relation extraction unit, used to extract the time-series dependency of the historical energy efficiency of the business server under the load for multiple consecutive time slices within a previous preset period, based on the deep belief DBN network; The feature obtaining unit is used to obtain the energy efficiency change characteristics of the service server under the load based on the historical energy efficiency of the service server under the load for multiple consecutive time slices within a previous preset duration, and the temporal dependency of the historical energy efficiency of the service server under the load for multiple consecutive time slices within a previous preset duration. An energy efficiency prediction unit is used to input the energy efficiency change characteristics under each load into a pre-trained energy efficiency prediction model for each of the service servers, so as to predict the Nth calibrated estimated energy efficiency of the service server under each load in the Nth time slice. The energy efficiency prediction model is trained by inputting multiple third training samples into a long short-term memory network. Each third training sample includes the historical energy efficiency change characteristics of the historical service server under each load and the corresponding historical Nth calibrated estimated energy efficiency of the historical service server under each load in the historical Nth time slice. An energy efficiency calibration unit is used to calibrate the Nth estimated energy efficiency under each load for each of the said service servers, based on the Nth estimated energy efficiency under each load.

8. The system according to claim 7, characterized in that, The energy efficiency calibration unit is specifically used to perform the calibration according to the formula. calibrate the Nth estimated energy efficiency under the corresponding load, where, For the j-th load of the i-th service server, calculate the N-th estimated energy efficiency before calibration in the N-th time slice. For the j-th load of the i-th service server, estimate the energy efficiency for the N-th calibration in the N-th time slice. The calibrated estimated energy efficiency for the j-th load of the i-th service server in the N-th time slice. The first weighting coefficient, It is the second weighting coefficient, and .

9. The system according to claim 6, characterized in that, The data acquisition unit is specifically used to acquire the wear level of the solid-state drive (SSD) of each business server, the accuracy of the operation logs, and the stability of the power supply voltage; input the wear level of the SSD of each business server, the accuracy of the operation logs, and the stability of the power supply voltage into a pre-trained hardware health determination model to determine the hardware health of each business server. The hardware health determination model is trained by inputting multiple fourth training samples into a third neural network. Each fourth training sample includes the historical wear level of the SSD of a historical business server, the accuracy of historical operation logs, the stability of historical power supply voltage, and the corresponding historical actual hardware health.

10. The system according to any one of claims 6-9, characterized in that, The network parameters of the service access volume allocation model are the state of the reinforcement learning model, updating the network parameters of the service access volume allocation model is the action of the reinforcement learning model, the Nth average energy efficiency is the reward of the reinforcement learning model, and the lower the Nth average energy efficiency, the higher the reward.

Images