Dynamic scheduling method for cloud computing resource pool

Abstract

The invention provides a dynamic scheduling method for a cloud computing resource pool. The dynamic scheduling method comprises the following steps: S1, acquiring a topological structure of the resource pool, multi-dimensional performance index data of each node, running state data of each micro-service instance and dependency relationship data among services in real time; s2, based on the collected data, constructing and dynamically maintaining a lightweight digital twinborn model synchronized with the physical resource pool, the digital twinborn model comprising a node resource model, an application performance model and a dependency graph; according to the method, simulation deduction is carried out in the high-fidelity digital twinborn body, all scheduling strategies are fully verified before deployment, the impact risk of online trial and error on the production environment is completely eradicated, meanwhile, the twinborn body simulates complex service dependence and resource competition, the system can foresee the linkage influence of scheduling actions, and the scheduling efficiency is improved. And spanning from local optimization to global optimization is realized.

Description

Technical Field

[0001] This invention relates to the field of cloud computing technology, and in particular to a dynamic scheduling method for cloud computing resource pools. Background Technology

[0002] With the widespread adoption of cloud-native architectures, large-scale microservice clusters have become the mainstream form of modern application deployment. This architecture brings high flexibility and scalability, but it also presents unprecedented challenges to resource scheduling. The complex call relationships and resource contention among thousands of microservice instances make traditional scheduling methods inadequate.

[0003] Existing resource scheduling technologies mainly suffer from the following bottlenecks: These methods rely on preset thresholds (such as CPU utilization exceeding 80%) for scaling up and down, essentially a reactive approach with delayed responses. They cannot handle sudden surges in traffic and are prone to system instability due to the "herd effect." Rule-making depends on expert experience, making it difficult to adapt to dynamically changing business environments.

[0004] While time series forecasting models such as LSTM can be introduced to predict load trends in advance, these methods have two major problems. First, the predictions themselves contain errors and are difficult to predict "black swan" events. Second, the predictions can only answer "how much the load will become," but cannot answer "what will happen to the system if I take scheduling action A now," lacking the ability to extrapolate the consequences of scheduling actions.

[0005] Reinforcement learning (RL) agents learn optimal policies by interacting with real-world environments. While this approach is theoretically attractive, it faces significant risks in production practice. During the exploration phase, the agent may execute catastrophic scheduling actions (such as incorrectly scaling down core services), directly impacting the stability of online operations. Furthermore, the training process for RL is lengthy and convergence is difficult, and the "black box" nature of its policies makes the results hard to interpret and trust.

[0006] In summary, existing technologies either lack foresight, pose risks of online trial and error, or fail to simulate the interactions between complex systems. Therefore, a dynamic scheduling method for cloud computing resource pools is proposed. Summary of the Invention

[0007] In view of this, embodiments of the present invention provide a dynamic scheduling method for cloud computing resource pools to solve or alleviate the technical problems existing in the prior art, and at least provide a beneficial option.

[0008] The technical solution of this invention is implemented as follows: A dynamic scheduling method for cloud computing resource pools, comprising the following steps: S1. Real-time collection of resource pool topology, multi-dimensional performance metrics of each node, running status data of each microservice instance, and dependency data between services. S2. Based on the collected data, construct and dynamically maintain a lightweight digital twin model that is synchronized with the physical resource pool. The digital twin model includes a node resource model, an application performance model, and a dependency graph. Node resource model: Simulates the capacity and real-time usage of resources such as CPU, memory, and network bandwidth for each physical / virtual node; Application performance model: By fitting historical data through machine learning (such as gradient boosting trees or neural networks), a non-linear relationship is established between microservice instance resource allocation (such as the number of CPU cores) and key performance indicators (such as request latency and throughput). Dependency graph: A directed graph that precisely describes the call relationships, call frequency, and data volume between microservices; S3. When the preset scheduling triggering conditions are met (such as detecting a sudden change in global load), a round of scheduling decision-making process is initiated. S4. On the digital twin model, run an evolutionary computation engine (in this embodiment, a genetic algorithm is used as an example) to find the optimal scheduling strategy: S5. The optimal scheduling strategy is converted into specific execution instructions (such as Kubernetes API calls) and sent to the executors of the physical resource pool for actual deployment; S6. Collect real-world performance data after strategy execution and compare it with the results of the digital twin model. Use the comparison error to calibrate and update the digital twin model (especially the application performance model) online.

[0009] In some embodiments, the application performance model in S2 is trained using historical data. This model describes a non-linear mapping between the resource allocation of a microservice instance and its performance metrics, including CPU and memory allocation, and QPS and latency.

[0010] In some embodiments, an evolutionary computation engine (using a genetic algorithm as an example in this embodiment) is run in S4 to perform the following sub-steps to find the optimal scheduling strategy: S41. Generate an initial population containing multiple candidate scheduling strategies, with each candidate scheduling strategy encoded as a chromosome; S42. Input each candidate scheduling strategy in the population into the digital twin model for simulation. The digital twin model simulates the state evolution of the resource pool in a future time window (such as the next 15 minutes) after the strategy is executed, including resource changes, performance fluctuations and SLA achievement of each service. S43. Based on the simulation results, evaluate each candidate scheduling strategy using a multi-objective fitness function and calculate its fitness value. S44. Based on fitness values, generate a next-generation candidate scheduling strategy population through selection, crossover, and mutation genetic operators; S45. Repeat S42 to S44 until the evolution termination condition is met (such as reaching the maximum number of iterations or the fitness value converges), and output the optimal scheduling policy with the highest fitness value.

[0011] In some embodiments, the chromosome encoding method in S41 is that a chromosome is composed of multiple genes, each gene representing a scheduling operation for a microservice cluster, and its structure is (Service_ID, Action_Type, Target_Value), where Action_Type includes scaling up, scaling down, migration, and resource reconfiguration.

[0012] In some embodiments, the simulation in S42 is based on discrete event simulation, and the digital twin model simulates resource allocation changes, service instance startup and shutdown, network traffic, and the resulting performance chain reactions in a step-by-step manner in the time dimension according to the input candidate scheduling strategy.

[0013] In some embodiments, the multi-objective fitness function in S43 The calculation formula is: in, , , and These are preset weights for SLA satisfaction, cost, energy consumption, and system stability, respectively. The overall SLA score in the simulation results; The total resource cost in the simulation results; The estimated total energy consumption in the simulation results; This is a measure of the number and size of scheduling actions included in a candidate scheduling strategy, used to penalize overly aggressive scheduling.

[0014] In some embodiments, the selection operator in S44 is a tournament selection or a roulette selection; Crossover operators can be single-point or two-point crossovers, used to exchange gene segments from different parent chromosomes; The mutation operator randomly changes the Action_Type or Target_Value of a gene in a chromosome.

[0015] In some embodiments, the scheduling triggering conditions in S3 include the system detecting abnormal fluctuations in macro load, the arrival of a periodic scheduling time point, or the receipt of a high-level business strategy change instruction.

[0016] A dynamic scheduling system for cloud computing resource pools, comprising: The data acquisition module is used to collect real-time data on the topology, performance, status, and dependencies of the resource pool. A digital twin engine is used to build and maintain a lightweight digital twin model that includes a node resource model, an application performance model, and a dependency graph based on collected data. An evolutionary computation engine is used to deduce the optimal scheduling strategy on a digital twin model using evolutionary algorithms when scheduling is triggered. The strategy execution module is used to distribute the optimal scheduling strategy to the physical resource pool for deployment; The feedback calibration module is used to collect real-world performance data after strategy execution and compare it with the simulation results to calibrate and update the digital twin model online.

[0017] In some embodiments: In the above embodiments:

[0018] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions: I. This invention simulates and extrapolates in a high-fidelity digital twin, ensuring that all scheduling strategies are fully validated before deployment, thus completely eliminating the risk of online trial and error impacting the production environment. At the same time, the twin simulates complex service dependencies and resource competition, enabling the system to anticipate the chain effects of scheduling actions and achieve a leap from local optimization to global optimization.

[0019] Second, by utilizing an evolutionary computing engine, this invention can perform an efficient global search among multiple conflicting objectives such as cost, performance, and energy consumption to find an equilibrium solution that is approximately Pareto optimal. Unlike the "black box" model of reinforcement learning, the evolutionary process provides a complete decision-making path, making the final strategy highly interpretable and traceable, and greatly enhancing the credibility and maintainability of the system.

[0020] Third, through the "execution-feedback-calibration" mechanism, the digital twin model can continuously self-correct and optimize based on real-world operational data, ensuring its high consistency with the physical resource pool. This enables the entire scheduling system to adaptively respond to business changes and environmental evolution, achieving long-term and effective intelligent scheduling.

[0021] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0022] 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 only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the present invention; Figure 2 This is a flowchart of the evolutionary computation engine of the present invention; Figure 3 This is a system architecture diagram of the present invention. Detailed Implementation

[0024] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0025] It is important to note that terms such as "first," "second," "symmetric," "array," "set in," and "set with" are used only to distinguish between descriptive and positional descriptions and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features specified with terms such as "first" or "symmetric" may explicitly or implicitly include one or more of that feature; similarly, when the quantity of certain features is not limited by words such as "two" or "three," it should be noted that such features also explicitly or implicitly include one or more features.

[0026] In this invention, unless otherwise explicitly specified and limited, terms such as "installation," "connection," and "fixation" should be interpreted broadly; for example, they can refer to a fixed connection, a detachable connection, or an integral molding; they can refer to a mechanical connection, a direct connection, a welding connection, or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the accompanying drawings and specific circumstances.

[0027] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] like Figures 1-3 As shown, this embodiment of the invention provides a dynamic scheduling method for cloud computing resource pools, including the following steps: S1. The data acquisition module collects the topology of the resource pool, multi-dimensional performance index data of each node, running status data of each microservice instance, and dependency data between services in real time through monitoring probes (such as Prometheus, eBPF) and configuration centers (such as ServiceMesh control plane). S2, the digital twin engine, builds and dynamically maintains a lightweight digital twin model that is synchronized with the physical resource pool based on the collected data. The digital twin model includes a node resource model, an application performance model, and a dependency graph. Node resource model: Simulates the capacity and real-time usage of resources such as CPU, memory, and network bandwidth for each physical / virtual node; Application performance model: The Gradient Boosting Decision Tree (GBDT) model is used. The input features are [CPU_request, Memory_request, QPS], and the output is the predicted P99_Latency. This model is retrained periodically using the latest historical data. Dependency graph: Stored using a graph database, nodes are microservices, edges represent call relationships, and weights can be attached to edges, such as average call latency; S3. When the preset scheduling triggering conditions are met (such as detecting a sudden change in global load), a round of scheduling decision-making process is initiated. Specifically, detecting a global load surge means that when the monitoring module detects, for example, "the total QPS of the entire site increases by more than 50% within 5 minutes," the scheduling process is triggered. S4. On the digital twin model, run an evolutionary computation engine (this example uses a genetic algorithm) to find the optimal scheduling strategy: S5. Transform the optimal scheduling strategy into specific execution instructions (such as Kubernetes API calls) and send them to the executors of the physical resource pool for actual deployment; S6. Collect real-world performance data after strategy execution and compare it with the results of the digital twin model. Use the comparison error to calibrate and update the digital twin model (especially the application performance model) online.

[0029] In this embodiment, specifically, the application performance model in S2 is trained using historical data. This model describes the non-linear mapping relationship between the resource allocation of a microservice instance and its performance metrics. The resource allocation includes CPU and memory, and the performance metrics include QPS and latency.

[0030] In this embodiment, specifically, an evolutionary computation engine (using a genetic algorithm as an example) is run in S4 to perform the following sub-steps to find the optimal scheduling strategy: S41. Generate an initial population containing multiple candidate scheduling strategies, with each candidate scheduling strategy encoded as a chromosome; S42. Input each candidate scheduling strategy in the population into the digital twin model for simulation. The digital twin model simulates the state evolution of the resource pool in a future time window (such as the next 15 minutes) after the strategy is executed, including resource changes, performance fluctuations and SLA achievement of each service. S43. Based on the simulation results, evaluate each candidate scheduling strategy using a multi-objective fitness function and calculate its fitness value. S44. Based on fitness values, generate a next-generation candidate scheduling strategy population through selection, crossover, and mutation genetic operators; S45. Repeat S42 to S44 until the evolution termination condition is met (e.g., 50 generations), and output the optimal scheduling policy with the highest fitness value.

[0031] In this embodiment, specifically, the chromosome encoding method in S41 is that a chromosome is composed of multiple genes, and each gene represents a scheduling operation for a microservice cluster. Its structure is (Service_ID, Action_Type, Target_Value), where Action_Type includes scaling up, scaling down, migration, and resource reconfiguration. The chromosome is defined as a scheduling action sequence through the above settings. For example, a chromosome may be [(svc-A, scale, 5), (svc-B, migrate, node-3), (svc-C, tune_cpu, 2.0)], which represents "scaling service A to 5 instances", "migrating one instance of service B to node 3", and "adjusting the CPU quota of service C to 2.0 cores". N such chromosomes are randomly generated to form the initial population.

[0032] In this embodiment, specifically, the simulation in S42 is based on discrete event simulation. The digital twin model simulates resource allocation changes, service instance start-up and shutdown, network traffic and the resulting performance chain reaction in a step-by-step manner in the time dimension according to the input candidate scheduling strategy. Through the above settings, the simulation (S42) is as follows: for each chromosome (i.e. a candidate strategy) in the population, the digital twin engine performs discrete event simulation. The scheduling action of this strategy is applied at analog clock t=0; Based on the application performance model, calculate the new performance metrics for each service instance after resource changes; Based on the dependency graph, the call traffic and new performance metrics between services are propagated to update the load and latency of downstream services; Advance the analog clock in fixed steps (e.g., 1 second) and repeat the above process until the analog time window ends (e.g., t=900s). Record key metrics throughout the simulation window, such as average SLA satisfaction rate, total resource cost, and scheduling action frequency.

[0033] In this embodiment, specifically, the multi-objective fitness function in S43 The calculation formula is: in, , , and These are preset weights for SLA satisfaction, cost, energy consumption, and system stability, respectively. The overall SLA score in the simulation results; The total resource cost in the simulation results; The estimated total energy consumption in the simulation results; To measure the number and size of scheduling actions included in candidate scheduling policies, and to penalize overly aggressive scheduling, the following multi-objective fitness function is used to calculate the score for each policy based on the above settings: =(1-average SLA violation rate), value range [0, 1]; =(Simulated total resource cost / Baseline cost), the normalized value; =(Simulated total energy consumption / Baseline energy consumption), the normalized value; =(Total number of scheduled actions in the strategy / Maximum number of allowed actions), value range [0, 1]; This function encourages high SLAs and penalizes frequent scheduling that is costly, energy-intensive, and unstable.

[0034] In this embodiment, specifically, the selection operator in S44 is a tournament selection or a roulette selection; Crossover operators can be single-point or two-point crossovers, used to exchange gene segments from different parent chromosomes; The mutation operator randomly changes the Action_Type or Target_Value of a gene in a chromosome.

[0035] In this embodiment, specifically, the selection is done through tournament selection, randomly selecting k individuals, among which the individual with the highest fitness enters the next generation.

[0036] Crossover: For two parent chromosomes, a crossover point is randomly selected, and the gene segments after the crossover point are exchanged to generate two offspring.

[0037] Mutation: With a small probability, randomly select a gene on a chromosome and randomly modify its Action_Type or Target_Value.

[0038] In this embodiment, specifically, the scheduling triggering conditions in S3 include the system detecting abnormal fluctuations in macro load, the arrival of periodic scheduling time points, or the receipt of high-level business strategy change instructions.

[0039] In this embodiment, specifically, the policy execution module translates the optimal policy [(svc-A, scale, 5)] into a call to the Kubernetes API, creates a new Pod, and within 15 minutes after the policy execution, the feedback calibration module collects real SLA data and resource consumption data, compares them with the twin's inference results for the policy, and if it is found that the actual latency is generally higher than the predicted latency, the error data is used to incrementally train the application performance model, thereby correcting the twin and making its next inference more accurate.

[0040] A dynamic scheduling system for cloud computing resource pools, comprising: The data acquisition module is used to collect real-time data on the topology, performance, status, and dependencies of the resource pool; the digital twin engine is used to build and maintain a lightweight digital twin model that includes a node resource model, an application performance model, and a dependency graph based on the collected data. An evolutionary computation engine is used to deduce the optimal scheduling strategy on a digital twin model using evolutionary algorithms when scheduling is triggered. The strategy execution module is used to distribute the optimal scheduling strategy to the physical resource pool for deployment; The feedback calibration module is used to collect real-world performance data after strategy execution and compare it with the simulation results to calibrate and update the digital twin model online.

[0041] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A dynamic scheduling method for cloud computing resource pools, characterized in that, Includes the following steps: S1. Real-time collection of resource pool topology, multi-dimensional performance metrics of each node, running status data of each microservice instance, and dependency data between services. S2. Based on the collected data, construct and dynamically maintain a lightweight digital twin model that is synchronized with the physical resource pool. The digital twin model includes a node resource model, an application performance model, and a dependency graph. S3. When the preset scheduling triggering conditions are met, start a round of scheduling decision-making process; S4. Run an evolutionary computation engine on the digital twin model to find the optimal scheduling strategy: S5. The optimal scheduling strategy is sent to the executor of the physical resource pool for actual deployment; S6. Collect real-world performance data after strategy execution and compare it with the results of the digital twin model. Use the comparison error to calibrate and update the digital twin model online.

2. The dynamic scheduling method for cloud computing resource pools according to claim 1, characterized in that: The application performance model in S2 is trained using historical data. This model describes the non-linear mapping relationship between the resource allocation of microservice instances and their performance metrics. Resource allocation includes CPU and memory, and performance metrics include QPS and latency.

3. The dynamic scheduling method for cloud computing resource pools according to claim 1, characterized in that: An evolutionary computation engine runs in S4, performing the following sub-steps to find the optimal scheduling strategy: S41. Generate an initial population containing multiple candidate scheduling strategies, with each candidate scheduling strategy encoded as a chromosome; S42. Input each candidate scheduling strategy in the population into the digital twin model for simulation and deduction. The digital twin model simulates the state evolution of the resource pool in a future time window after the strategy is executed, including resource changes, performance fluctuations and SLA achievement of each service. S43. Based on the simulation results, evaluate each candidate scheduling strategy using a multi-objective fitness function and calculate its fitness value. S44. Based on fitness values, generate a next-generation candidate scheduling strategy population through selection, crossover, and mutation genetic operators; S45. Repeat S42 to S44 until the evolution termination condition is met, and output the optimal scheduling policy with the highest fitness value.

4. The dynamic scheduling method for cloud computing resource pools according to claim 3, characterized in that: The chromosome encoding method in S41 is that a chromosome is composed of multiple genes, and each gene represents a scheduling operation for a microservice cluster. Its structure is (Service_ID, Action_Type, Target_Value), where Action_Type includes scaling up, scaling down, migration, and resource reconfiguration.

5. The dynamic scheduling method for cloud computing resource pools according to claim 3, characterized in that: The simulation in S42 is based on discrete event simulation. The digital twin model simulates resource allocation changes, service instance start-up and shutdown, network traffic, and the resulting performance chain reactions in a step-by-step manner in the time dimension according to the input candidate scheduling strategy.

6. The dynamic scheduling method for cloud computing resource pools according to claim 3, characterized in that: The multi-objective fitness function in S43 The calculation formula is: in, , , and These are preset weights for SLA satisfaction, cost, energy consumption, and system stability, respectively. The overall SLA score in the simulation results; The total resource cost in the simulation results; The estimated total energy consumption in the simulation results; This is a measure of the number and size of scheduling actions included in a candidate scheduling strategy, used to penalize overly aggressive scheduling.

7. The dynamic scheduling method for cloud computing resource pools according to claim 3, characterized in that: The selection operator in S44 is either a tournament selection or a roulette selection; Crossover operators can be single-point or two-point crossovers, used to exchange gene segments from different parent chromosomes; The mutation operator randomly changes the Action_Type or Target_Value of a gene in a chromosome.

8. The dynamic scheduling method for cloud computing resource pools according to claim 3, characterized in that: The scheduling triggering conditions in S3 include the system detecting abnormal fluctuations in macro load, the arrival of periodic scheduling time points, or the receipt of high-level business strategy change instructions.

9. A dynamic scheduling system for cloud computing resource pools, characterized in that, include: The data acquisition module is used to collect real-time data on the topology, performance, status, and dependencies of the resource pool. A digital twin engine is used to build and maintain a lightweight digital twin model that includes a node resource model, an application performance model, and a dependency graph based on collected data. An evolutionary computation engine is used to deduce the optimal scheduling strategy on a digital twin model using evolutionary algorithms when scheduling is triggered. The strategy execution module is used to distribute the optimal scheduling strategy to the physical resource pool for deployment; The feedback calibration module is used to collect real-world performance data after strategy execution and compare it with the simulation results to calibrate and update the digital twin model online.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images