A deep learning-based computing resource optimization method and system

Abstract

The present application relates to the technical field of data processing, and particularly relates to a computing power resource optimization method and system based on deep learning, comprising: obtaining running index data of servers and application instances in a data center; abstracting a topological relationship of the servers and the application instances into a time sequence dynamic heterogeneous graph; inputting the time sequence dynamic heterogeneous graph into a space-time graph neural network model, for capturing spatial domain dependence and time sequence evolution characteristics of resources, so as to extract context feature embedding of each node at a future time step; inputting the context feature embedding into a conditional variational autoencoder model, to generate a multi-dimensional probability distribution representing future resource demand; inputting the multi-dimensional probability distribution as a state into a multi-objective deep reinforcement learning model, to output a global optimization scheduling strategy for adjusting computing power allocation of each application instance. The method of the present application realizes comprehensive improvement of data center operation efficiency.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to a method and system for optimizing computing resources based on deep learning. Background Technology

[0002] With the rapid development of cloud computing, big data, and artificial intelligence technologies, the scale and complexity of data centers are increasing daily. The core challenge lies in how to efficiently and intelligently schedule and manage massive computing, storage, and network resources to meet the dynamic and ever-changing performance demands of various upper-layer application instances. Traditional resource scheduling methods mostly rely on preset rules or static thresholds. For example, migration or expansion is triggered when the CPU utilization of a computing node (hereinafter referred to as a server, which includes both physical and virtual servers) exceeds 80%. While such methods are simple and intuitive, they often exhibit significant lag when faced with drastic fluctuations and complex relationships in business loads, making accurate predictive scheduling impossible. Other forecasting methods based on traditional time series analysis models (such as ARIMA) can capture the time-series changes in load to some extent, but they usually treat each server or application instance as an independent analysis object, ignoring the complex topological dependencies between them caused by resource sharing and service calls. This leads to a one-sided understanding of the overall system state, making it difficult to formulate a globally optimal scheduling strategy, and ultimately causing problems such as low resource utilization, excessive energy consumption, and a decline in service quality (Service Level Agreement, SLA).

[0003] To overcome the limitations of traditional methods, academia and industry have begun exploring the application of machine learning, particularly deep learning, to data center resource management. For example, recurrent neural networks (RNNs) or their variants (such as LSTMs) are used to predict future resource demands, or reinforcement learning (RL) is employed to automatically learn scheduling strategies. However, existing solutions still have several shortcomings. First, many predictive models fail to fully utilize the inherent graph structure between data center components and applications, failing to effectively capture the spatial interactions and temporal co-evolution of resources, thus limiting the accuracy of the predictive models. Second, most predictive models only provide deterministic point estimates of future resource demands, failing to quantify the uncertainty of the prediction results themselves. This makes scheduling decisions vulnerable to sudden changes in demand, easily leading to over- or under-allocation of resources due to incorrect estimates. Finally, in the decision optimization phase, how to coordinate and balance multiple conflicting objectives such as service level agreement satisfaction, system energy consumption, resource fragmentation, and migration costs within a unified framework remains a pressing technical challenge. Summary of the Invention

[0004] To address the technical problem of how to improve the operational efficiency of data centers, this invention provides the following solution.

[0005] A deep learning-based method for optimizing computing resources includes the following steps:

[0006] The system acquires operational metrics data for servers and application instances within a data center, including CPU utilization, memory usage, network I / O, and disk I / O. It then abstracts the topological relationships of these servers and application instances into a temporal dynamic heterogeneous graph, where servers and application instances are nodes, the operational metrics data are node attributes, and resource dependencies and communication relationships between nodes are edges. This temporal dynamic heterogeneous graph is input into a spatiotemporal graph neural network model, which integrates graph convolutional networks and recurrent neural networks to capture the spatial dependencies and temporal evolution characteristics of resources, thereby extracting the future performance of each node. The contextual features are embedded at a time step. The embedded contextual features are input into a conditional variational autoencoder model to generate a multidimensional probability distribution representing future resource requirements. The multidimensional probability distribution quantifies the expected value and uncertainty of resource requirements for each node. The multidimensional probability distribution is used as a state and input into a multi-objective deep reinforcement learning model. Its reward function is a weighted function that integrates service level agreement satisfaction, system energy consumption, resource fragmentation degree, and migration cost. By maximizing the cumulative reward, the multi-objective deep reinforcement learning model outputs a global optimization scheduling strategy for adjusting the computing power allocation of each application instance.

[0007] The method of this invention achieves a comprehensive improvement in the operational efficiency of data centers.

[0008] Furthermore, the time-series dynamic heterogeneous graph contains at least two types of nodes, namely server nodes and application instance nodes, and at least two types of edges, representing the resource dependency relationship between the server and the application instance and the communication relationship between the application instances, respectively.

[0009] This invention models the topological relationship between data center servers and application instances as a time-series dynamic heterogeneous graph, which can accurately express the relationship between computing power and applications.

[0010] Furthermore, the spatiotemporal graph neural network model includes: a graph convolutional network for capturing spatial dependencies between nodes by aggregating neighbor node information at each time step; and a recurrent neural network for processing the node feature sequences output by the graph convolutional network at consecutive time steps to capture the temporal evolution characteristics of node states.

[0011] This invention uses a spatiotemporal graph neural network for feature extraction, which can accurately capture the complex spatial dependencies and temporal evolution of resources, thereby significantly improving the prediction accuracy of future resource demand.

[0012] Furthermore, the recurrent neural network is a long short-term memory network or a gated recurrent unit.

[0013] Furthermore, the conditional variational autoencoder model includes: an encoder that embeds and maps the context features into a probability distribution of a latent space; and a decoder that samples from the probability distribution of the latent space and reconstructs a multidimensional probability distribution of multidimensional resources representing future CPU utilization, memory usage, network I / O, and disk I / O.

[0014] This invention introduces a conditional variational autoencoder to probabilistically describe the prediction results, quantifying the uncertainty of resource demand. This enables scheduling decisions to effectively avoid risks caused by point prediction biases and enhances the robustness of the system.

[0015] Furthermore, the service level agreement satisfaction term in the reward function applies a negative reward value when the performance indicators of one or more application instances exceed a preset threshold.

[0016] This invention introduces a negative reward value to make the agent realize that previous scheduling actions have led to adverse consequences.

[0017] Furthermore, the performance metrics are response latency or resource utilization.

[0018] Furthermore, the system energy consumption term in the reward function is inversely proportional to the predicted total energy consumption of all active servers in the data center, wherein the energy consumption of a single server is modeled as an increasing function of server resource utilization.

[0019] Furthermore, the resource fragmentation term in the reward function is inversely proportional to the unevenness of the distribution of unallocated resources on each server, and the unevenness is determined by quantifying the dispersion of available resources among the physical units within the server.

[0020] The present invention also provides a computing resource optimization system based on deep learning, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-mentioned computing resource optimization method based on deep learning.

[0021] In summary, the beneficial effects of this invention are as follows: By modeling the topological relationship between data center servers and application instances as a temporal dynamic heterogeneous graph and employing a spatiotemporal graph neural network for feature extraction, it can accurately capture the complex spatial dependencies and temporal evolution patterns of resources, thereby significantly improving the prediction accuracy of future resource demands. Furthermore, by introducing a conditional variational autoencoder to probabilistically describe the prediction results, the uncertainty of resource demands is quantified, enabling scheduling decisions to effectively avoid risks caused by point prediction biases and enhancing the system's robustness. The multi-objective deep reinforcement learning model used can collaboratively optimize multiple conflicting objectives such as service level agreements, energy consumption, resource fragmentation, and migration costs within a unified framework, finding a globally superior resource allocation strategy. Compared with existing technologies, the method of this invention significantly improves resource utilization and reduces total system energy consumption while ensuring a high level of service quality, achieving a comprehensive improvement in data center operational efficiency. Attached Figure Description

[0022] Figure 1 This is an example of a time-series dynamic heterogeneous graph according to an embodiment of the present invention;

[0023] Figure 2 This is an illustrative example of the temporal evolution of node features according to an embodiment of the present invention;

[0024] Figure 3 This is a schematic diagram illustrating a comparison of resource fragmentation levels according to an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0026] A deep learning-based method for optimizing computing resources includes the following steps:

[0027] S1. Obtain operational metrics data of servers and application instances within the data center, including CPU utilization, memory usage, network I / O, and disk I / O; abstract the topological relationship of servers and application instances within the data center into a time-series dynamic heterogeneous graph, where servers and application instances are nodes, the operational metrics data are node attributes, and the resource dependencies and communication relationships between nodes are edges.

[0028] Specifically, monitoring agents deployed on physical servers and virtual machines, such as Prometheus's Node Exporter, periodically collect runtime metrics data from each server and the containerized application instances running on it, for example, once per minute. At each collection point t, a heterogeneous graph is constructed, where servers (including virtual machines) are one type of node, and application instances are another. The attributes of the nodes are the runtime metrics data collected at that moment (which can be represented as a multi-dimensional vector). If an application instance is deployed on a server, an edge is established from the server to the application instance (representing a resource dependency). If there are API calls or data exchanges between two application instances, an edge is established based on network traffic monitoring data (representing a communication relationship). Continuous collection in this way forms a graph sequence, namely a time-series dynamic heterogeneous graph, denoted by G_t.

[0029] like Figure 1 As shown, in an optional embodiment, the time-series dynamic heterogeneous graph G_t contains at least two types of nodes, namely server nodes and application instance nodes, and at least two types of edges, representing the resource dependency relationship (also known as the bearing relationship) of server nodes to application instances and the communication relationship between application instances, respectively.

[0030] Specifically, the structure of a time-series dynamic heterogeneous graph can accurately depict the complex topology and application dependencies of a data center. For example, a server node can be a physical server, such as server A with 64 CPU cores and 256GB of memory. Meanwhile, a microservice running in the same data center, such as user authentication service instance 1, is represented as an application instance node. If user authentication service instance 1 is deployed on server A, then there will be an edge in the graph pointing from server A node to user authentication service instance 1 node; this edge represents the resource dependency.

[0031] Furthermore, interactions also exist between application instances. Suppose a shopping cart service instance 2 needs to call a user authentication service instance 1 to verify user information. In the time-series dynamic heterogeneous graph, there would be an edge between the shopping cart service instance 2 node and the user authentication service instance 1 node, representing their communication relationship. The weight of this edge can be set according to the actual communication traffic, such as 1000 API calls per second or 50MB of data transfer per minute. By constructing such a time-series dynamic heterogeneous graph containing different types of nodes and edges, a comprehensive understanding of the spatial relationships between computing resources and application instances can be achieved.

[0032] S2, the temporal dynamic heterogeneous graph is input into a spatiotemporal graph neural network model, which integrates graph convolutional networks and recurrent neural networks to capture the spatial dependence and temporal evolution characteristics of resources, thereby extracting the contextual feature embedding of each node in the future time step.

[0033] Specifically, at each time step t, the temporal dynamic heterogeneous graph G_t is processed using a Graph Convolutional Network (GCN) of the spatiotemporal graph neural network model. By aggregating information from neighboring nodes, the feature representation of each node is updated, capturing spatial dependencies such as the influence of a server on all application instances on it, or the associations between communicating application instances. The node feature sequences output by the GCN are fed into the recurrent neural networks (RNNs) of their respective spatiotemporal graph neural network models. The recurrent structure of the RNNs can learn the patterns of CPU utilization changes over time for each node, such as the CPU utilization of a particular application instance, thereby capturing temporal evolution characteristics. The output of the RNNs at the last time step is the context feature embedding containing spatiotemporal information.

[0034] like Figure 2 As shown, the Graph Convolutional Network (GCN) analyzes the graph structure of the entire data center at every point in time, such as 2 PM. For a specific application instance node, such as an online payment service, the GCN aggregates information about its neighboring nodes. These neighboring nodes may include the server node hosting it and the order processing service node communicating with it. If the CPU utilization of the server node is 90%, and the network latency of the order processing service node is 300 milliseconds, the GCN integrates this information to generate a feature vector that reflects the environment and pressure of the payment service at 2 PM.

[0035] The recurrent neural network module, preferably employing a long short-term memory network or a gated recurrent unit, receives feature vectors arranged in chronological order. For example, it processes a sequence of payment service feature vectors generated by a graph convolutional network (GCN) at consecutive time points such as 2:00 PM, 2:05 PM, and 2:10 PM. By analyzing this feature vector sequence, the recurrent neural network module can learn the dynamic change patterns of the system. For instance, it might discover that whenever the server CPU utilization exceeds 85% for ten consecutive minutes, followed by a surge in network traffic, the payment service is highly likely to experience a performance bottleneck in the next five minutes. This combination allows the model to understand both spatial interactions and capture temporal evolution patterns.

[0036] S3, the context features are embedded into a conditional variational autoencoder model to generate a multidimensional probability distribution representing future resource demand. The multidimensional probability distribution quantifies the expected value and uncertainty of resource demand for each node.

[0037] In an optional embodiment, the conditional variational autoencoder model includes: an encoder that embeds and maps the context features into a probability distribution of a latent space; and a decoder that samples from the probability distribution of the latent space and reconstructs a multidimensional probability distribution representing future multidimensional resource demands such as CPU utilization, memory usage, network I / O, and disk I / O.

[0038] The encoder receives contextual feature embeddings from a spatiotemporal graph neural network model, condensing the current and past state information of a specific application instance. Instead of directly outputting a definitive latent representation, the encoder outputs parameters of a probability distribution, typically the mean and variance of a Gaussian distribution. For example, for a database service, its contextual feature embeddings, after being input into the encoder, might output a latent distribution with a mean vector A and a variance vector B. The current state is not entirely determined but fluctuates within a certain range around the mean A, while the variance B quantifies the magnitude of this uncertainty.

[0039] The decoder randomly samples a point from the latent space defined by the mean A and variance B. This sampling process allows the model to generate diverse future predictions. The decoder receives the sampled point as input and outputs a prediction of future resource needs, also in the form of a probability distribution. For example, it might predict that CPU utilization over the next ten minutes follows a normal distribution with a mean of 70% and a standard deviation of 5%, while memory usage follows a normal distribution with a mean of 40GB and a standard deviation of 3GB. This probabilistic prediction is far superior to a single numerical prediction because it not only provides the most likely resource demand value but also the confidence interval of the prediction, offering richer risk assessment information for resource scheduling decisions.

[0040] S4. The multidimensional probability distribution is used as the state and input into a multi-objective deep reinforcement learning model. Its reward function is a weighted function that integrates service level agreement satisfaction, system energy consumption, resource fragmentation degree and migration cost. By maximizing the cumulative reward, the model outputs a global optimization scheduling strategy for adjusting the computing power allocation of each application instance.

[0041] Specifically, a multi-objective deep reinforcement learning model can employ a deep reinforcement learning framework based on the advantage actor / commentator (A2C). This framework can output specific scheduling actions based on the current state, such as reallocating CPU cores and memory size for each application instance, or deciding whether to migrate it to another server. After executing the scheduling action, a scalar reward value is calculated. This reward value is obtained by a weighted sum of multiple sub-items; for example, a positive reward is given if the application response time is below the SLA threshold, if the total system power consumption is reduced, or if server resource allocation is more balanced. Each migration operation is compensated with a fixed negative reward. The multi-objective deep reinforcement learning model tends to output actions that yield higher long-term cumulative rewards, ultimately converging to obtain the optimal global scheduling strategy.

[0042] In an optional embodiment, the service level agreement satisfaction term in the reward function applies a negative reward value when the performance metrics of one or more application instances, such as response latency or resource utilization, exceed a preset threshold.

[0043] Specifically, suppose a critical online transaction application has a Service Level Agreement (SLA) stipulating that 99% of its request response latency must be below 100 milliseconds. This 100 milliseconds is the preset SLA threshold. During reinforcement learning training, after performing a scheduling action, such as migrating an instance of the application to a server with seemingly sufficient resources but a poor network environment, the application's response latency is detected to have risen to 200 milliseconds, violating the SLA. At this point, the reward function calculates a significant negative reward value, such as -50. The magnitude of the negative value can also be correlated with the severity of the violation; for example, the greater the latency exceeds the threshold, the larger the absolute value of the negative reward. If the latency is 250 milliseconds, the negative reward might be -80. This penalty signal is fed back to the reinforcement learning agent, making it aware that the previous scheduling action caused adverse consequences. By continuously receiving such negative feedback through multiple iterations, the agent will learn to avoid scheduling strategies that may lead to service performance degradation and SLA violations, thereby optimizing its decisions to ensure service quality.

[0044] In an optional embodiment, the system energy consumption term in the reward function is inversely proportional to the predicted total energy consumption of all active servers in the data center, wherein the energy consumption of a single server is modeled as an increasing function of server resource utilization.

[0045] Specifically, the energy consumption of a single server is not a simple linear relationship with its resource utilization. An idle server, such as one with zero CPU utilization, still consumes a base power, say 100 watts. As its CPU utilization increases, energy consumption also increases, but the rate of increase may vary. For example, a server's energy consumption model could be a base power of 100 watts plus a term proportional to the square of the CPU utilization. Therefore, when the CPU utilization is 50%, its energy consumption might be 200 watts, while when the CPU utilization reaches 100%, the energy consumption might rise to 350 watts. The system energy consumption term is calculated to predict the energy consumption of every active server in the data center, summing them all up to obtain a total energy consumption prediction. For example, if a scheduling decision results in ten servers being active, each consuming an average of 220 watts, the total energy consumption would be 2200 watts. The system energy consumption term in the reward function is inversely proportional to this total value, for example, it could be set as one divided by 2200. If another scheduling decision, through application consolidation, reduces the total energy consumption to 1,900 watts by requiring only eight servers to be active, then the corresponding reward will increase. This design incentivizes reinforcement learning agents to learn how to concentrate workloads on fewer servers and shut down or put idle servers into low-power modes, thereby achieving energy-efficient operation of the data center.

[0046] In an optional embodiment, the resource fragmentation term in the reward function is inversely proportional to the unevenness of the distribution of unallocated resources on each server, which is determined by quantifying the dispersion of available resources among the physical units within the server.

[0047] Specifically, resource fragmentation is particularly pronounced on a modern server with multiple non-uniform memory access architectures, i.e., NUMA nodes. For example... Figure 3As shown, suppose a server has two NUMA nodes, NUMA node 0 and NUMA node 1, each with 32 CPU cores and 128 GB of memory. If a scheduling decision results in NUMA node 0 having 20 CPU cores and 100 GB of memory remaining, while node 1 has only 2 CPU cores and 10 GB of memory remaining, although the server has a lot of remaining resources, the distribution is extremely uneven. This state is highly fragmented because a large new application requiring 16 cores and 60 GB of memory cannot be fully placed on either NUMA node, and cross-node deployment will result in a significant performance penalty. To quantify this unevenness, the standard deviation or coefficient of variation of the number of available CPU cores and the amount of available memory on each NUMA node can be calculated. In the example above, the distribution of available CPU cores is 20 and 2, with a large standard deviation, indicating high dispersion and severe fragmentation. In contrast, a better schedule might result in both nodes having 11 cores and 55 GB of memory remaining, with a standard deviation of zero, indicating a perfectly balanced resource distribution and low fragmentation. The term in the reward function is inversely proportional to the quantified dispersion. Therefore, a highly fragmented state will receive a very low or even negative reward, while a state with a balanced resource distribution will receive a higher reward. This will guide the agent to consider not only whether the application can be placed at the moment, but also the regularity of the remaining resources after placement, in order to facilitate future resource allocation.

[0048] The present invention also relates to a deep learning-based computing resource optimization system, including a processor and a memory. The memory stores a computer program, and the processor can interact with the memory and call the computer program (e.g., via a bus). The processor then executes the computer program, and when the computer program is executed by the processor, it implements the deep learning-based computing resource optimization method of the above embodiments.

[0049] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium.

[0050] This specification and accompanying drawings are merely illustrative descriptions of the invention as defined in the appended claims. It is obvious that those skilled in the art can make various modifications and variations to this invention without departing from its scope.

Claims

1. A deep learning-based computing resource optimization method, characterized in that, The method comprises the following steps: obtaining running index data of servers and application instances in a data center, including CPU utilization, memory usage, network I / O and disk I / O; abstracting the topological relationship of the servers and the application instances into a time-series dynamic heterogeneous graph, wherein the servers and the application instances are nodes, the running index data are node attributes, and the resource dependency relationship and the communication relationship between the nodes are edges; the time-series dynamic heterogeneous graph comprises at least two types of nodes, namely server nodes and application instance nodes, and at least two types of edges, representing the resource dependency relationship of the servers to the application instances and the communication relationship between the application instances; if an application instance is deployed on a server, an edge is established from the server to the application instance, representing the resource dependency relationship; if there is API calling or data exchange between two application instances, an edge is established according to network flow monitoring data, representing the communication relationship; inputting the time-series dynamic heterogeneous graph into a spatio-temporal graph neural network model, the spatio-temporal graph neural network model fusing a graph convolution network and a recurrent neural network, for capturing the spatial dependency and the time-series evolution characteristics of resources, so as to extract the context feature embedding of each node at a future time step; inputting the context feature embedding into a conditional variational autoencoder model to generate a multi-dimensional probability distribution representing future resource demand, the multi-dimensional probability distribution quantifying the expected value and uncertainty of resource demand of each node; inputting the multi-dimensional probability distribution as a state into a multi-objective deep reinforcement learning model, a reward function of the multi-objective deep reinforcement learning model being a weighted function integrating service level agreement satisfaction, system energy consumption, resource fragmentation level and migration cost, and the multi-objective deep reinforcement learning model outputting a global optimization scheduling strategy for adjusting the computing power allocation of each application instance by maximizing the cumulative reward; the system energy consumption term in the reward function being inversely proportional to the total energy consumption prediction value of all active servers in the data center, wherein the energy consumption of a single server is modeled as an increasing function of the resource utilization of the server; the resource fragmentation level term in the reward function being inversely proportional to the unevenness of the unallocated resource distribution on each server, and the unevenness being determined by quantifying the dispersion degree of the available resources among the physical units inside the server.

2. The method of claim 1, wherein, The spatio-temporal graph neural network model comprises: a graph convolution network for capturing the spatial dependency between nodes by aggregating neighbor node information at each time step; and a recurrent neural network for processing the node feature sequence output by the graph convolution network at consecutive time steps to capture the time-series evolution characteristics of the node state.

3. The method of claim 2, wherein, The recurrent neural network is a long short-term memory network or a gated recurrent unit.

4. The method of claim 1, wherein, The conditional variational autoencoder model comprises: an encoder for mapping the context feature embedding into a probability distribution of a latent space; and a decoder for sampling from the probability distribution of the latent space and reconstructing a multi-dimensional probability distribution representing the multi-dimensional resources of future CPU utilization, memory usage, network I / O and disk I / O.

5. The method of claim 1, wherein, The service level agreement satisfaction term in the reward function is a negative reward value applied when the performance indicator of one or more application instances exceeds a preset threshold.

6. The method of claim 5, wherein, The performance indicator is response delay or resource utilization.

7. A deep learning-based computing resource optimization system, characterized in that, The method comprises a processor and a memory, and the memory stores a computer program. The processor executes the computer program to implement the deep learning-based computing resource optimization method in any one of claims 1 to 6.

Images