Calculation power network workload resource mode prediction method and device, equipment and medium

Abstract

The invention belongs to the field of cloud computing, and provides a computing power network work load resource mode prediction method, device, equipment and medium, and the method comprises the steps: determining a shape set according to a work load set, carrying out the processing through a membership function, obtaining the membership degree of the work load to a mode fragment with the combined multi-resource utilization rate, and obtaining the mode fragment with the combined multi-resource utilization rate; further determining time and space coupled workload distribution of the computing power network; determining a first-order dynamic neighbor load set according to the working load, and determining a non-stationary fuzzy set according to the first-order dynamic neighbor load set and the fuzzy set; determining fuzzy representation of competition of a working load resource mode in the target computing power network by adopting an r-hop space tuple and a link loss factor; and predicting the mode of the workload resource of the target computing power network in the next time slot by adopting a multi-order adaptive cross entropy loss function. The method has the beneficial effect that the prediction accuracy of the working load resource mode of the computing power network is improved.

Description

Technical Field

[0001] This invention relates to the field of cloud computing technology, and in particular to a method, apparatus, device, and medium for predicting the workload resource patterns of a computing power network. Background Technology

[0002] The fragmented utilization and uneven distribution of computing resources in computing networks can easily lead to a dual contradiction: local server overload and global computing resource idleness. Accurately predicting the resource occupancy patterns of workloads helps computing networks proactively allocate computing resources, avoiding resource delays caused by server rotation time and ensuring user service quality constraints. The utilization rate of computing resources during workload operation exhibits multi-dimensional spatiotemporal coupling characteristics. Spatially, distributed workloads are subject to competition and sharing interference for physical resources from other workloads on neighboring computing nodes, resulting in topological changes in resource utilization. Temporally, factors such as intrinsic computing logic and exogenous data access patterns jointly affect the resource utilization rate of workloads, exhibiting uncertain abrupt changes and nonlinear time-varying behavior.

[0003] The statistical distribution of resource occupancy trajectories at different time periods exhibits significant heterogeneity, and the non-stationary nature of workload resource patterns poses a severe challenge to traditional time series forecasting methods.

[0004] 1. The collaborative prediction dimension of multi-dimensional resource types is missing. Existing studies mostly focus on the single dimension of CPU, ignoring the dynamic coupling effect of heterogeneous resources such as GPU and memory. Its cross-modal correlation modeling faces the disaster of feature dimension and the problem of quantifying competitive relationships.

[0005] 2. Insufficient ability to represent nonstationarity. For example, traditional methods of autoregressive moving average models are limited by the assumption of linear dynamics and cannot capture endogenous nonstationary mutations. While deep models such as Long Short-Term Memory Networks (LSTM) alleviate long-term dependence through implicit encoding, they still suffer from mutation response lag and variance estimation bias.

[0006] 3. Lack of spatial topological correlation modeling. For example, the mainstream time series model of LSTM ignores the spatial topological dependence formed between nodes due to the sharing of physical resources by workload scheduling strategies, and the prediction results deviate from the actual dynamic evolution of the topology.

[0007] While graph learning has demonstrated powerful capabilities in jointly representing spatiotemporal multivariate variables in other fields, its low-order modeling paradigm based on paired edges struggles to effectively represent the complex relationships between resource utilization trajectories of different workloads in computing networks. The multimodal dynamic evolution of workload resource utilization is driven by both internal subtask scheduling uncertainties and state switching disturbances caused by the load of topologically connected nodes. Graph learning, due to its inherent black-box nature, cannot explicitly quantify the propagation path of uncertainty in non-stationary processes and lacks interpretability for multimodal causal evolutionary reasoning. Summary of the Invention

[0008] Aimed at at least solving one of the technical problems existing in the prior art, the present invention provides a method, apparatus, device and medium for predicting the workload resource patterns of computing networks, thereby improving the accuracy of predicting the workload resource patterns of computing networks.

[0009] One aspect of the present invention provides a method for predicting the workload resource patterns of a computing network, characterized in that it includes:

[0010] Obtain the workload set of the target computing power network, and determine the shape set based on the workload set. The workload set is used to characterize the resource utilization rate index of the workload during its life cycle, and the shape set is used to characterize the pattern fragment of the joint utilization of multiple resources. Each shape set has a corresponding fuzzy set.

[0011] The shape set is processed by a membership function to obtain the membership degree of the workload to the pattern fragment of the multi-resource utilization joint, and then the temporal and spatial coupled workload distribution of the target computing power network is determined by the hybrid hypergraph.

[0012] A first-order dynamic neighbor load set is determined based on the workload. A non-stationary fuzzy set is determined based on the first-order dynamic neighbor load set and the fuzzy set, wherein the non-stationary fuzzy set is used to capture fluctuations caused by changes in the first-order dynamic neighbor load set and time variations.

[0013] The nonstationary fuzzy Markov model is determined based on the nonstationary fuzzy set, and the dynamic transition probability is determined by softmax through the nonstationary fuzzy Markov model.

[0014] The variable-order Markov model is determined by dynamic transition probabilities, and based on the hybrid hypergraph, the following is adopted: The jump space tuple and link loss factor determine the fuzzy representation of the competition for workload resource patterns in the target computing power network.

[0015] Based on the fuzzy representation of the competition between the variable-order Markov model and the workload resource pattern, a multi-order adaptive cross-entropy loss function is used to predict the workload resource pattern of the target computing network in the next time slot, and the prediction results are obtained.

[0016] According to the aforementioned method for predicting the workload resource patterns of a computing power network, a workload set of the target computing power network is obtained, and a shape set is determined based on the workload set, including:

[0017] The workload set is represented as Each workload Within their respective life cycles Resource utilization metrics include CPU utilization, GPU utilization, memory utilization, disk read throughput, and disk write throughput. These metrics are expressed as... , Total workload, These are integers starting from 1, used to uniquely identify different stable operating mode segments. Fragment of a multidimensional resource utilization joint model Time span, shape set It allows the extraction of any number of representative stable patterns with any position and any statistical characteristics from time series.

[0018] According to the aforementioned computing power network workload resource pattern prediction method, the shape set is processed using a membership function to obtain the membership degree of the workload to a pattern fragment jointly defined by multiple resource utilization rates. Then, the temporal and spatially coupled workload distribution of the target computing power network is determined through a hybrid hypergraph, including:

[0019] The fuzzy set of the shape set Quantization is performed using a membership function, where the membership function... Indicates workload For pattern fragments degree of membership, domain Represents workload set All workloads in, i.e. ;

[0020] Each time slot workload Mapped to the universe of discourse All fuzzy sets Unified fuzzy set The membership degrees of the fuzzy sets representing the shape set are compared dynamically on a time-slot basis, and the maximum value is used to filter out each time slot. The most prominent pattern feature, among which the unified fuzzy set for:

[0021]

[0022] in, Indicates time slot workload Workload The entire load flow;

[0023] Using Hybrid Hypergraph Construct a temporally and spatially coupled workload distribution in the target computing power network, where Hyperedges represent geographically ubiquitous computing power nodes. Connect to the scheduling center, which is used for workloads. Allocate computing power nodes;

[0024] Directed pairs of edges are used to determine the received physical constraints on the data flow path. parameter vector Represents computing power nodes To computing nodes The normalized uplink and downlink bandwidth asymmetric weight parameter vector, ,in, This represents the Z-score normalization function, which is independently normalized for each link parameter. , , and These represent bandwidth, latency, packet loss rate, and link availability probability, respectively. It is a vertex-hyperedge incidence matrix used to show the node coverage relationship of each hyperedge. It is a vertex-to-vertex weighted adjacency matrix representing the link quality between computing nodes in the underlying fiber optic link of a computing power network. ,in

[0025] .

[0026] According to the aforementioned computing network workload resource pattern prediction method, a first-order dynamic neighbor load set is obtained based on the workload. Based on the first-order dynamic neighbor load set and the fuzzy set, a non-stationary fuzzy set is determined. The non-stationary fuzzy set is used to capture fluctuations caused by changes in the first-order dynamic neighbor load set and time variations, including:

[0027] The first-order dynamic neighbor load set determined by the workload is:

[0028]

[0029] in, Indicates neighbor load At any time slot within its own life cycle This indicates that the time windows of the two overlap, that is... exist Active within time slots;

[0030] Non-stationary fuzzy sets are:

[0031]

[0032] in, Indicates the superedge and super edge The size of the normalized intersection; This represents the bipolar Sigmoid function. Represents the fuzzy modulation operator. This represents vector concatenation. Indicates neighbor load Membership vector on all fuzzy sets It is a learnable fuzzy relation matrix used to promote or inhibit the interaction between coding pattern loads.

[0033] According to the aforementioned method for predicting the workload resource pattern of computing power networks, a non-stationary fuzzy Markov model is obtained based on non-stationary fuzzy sets. The dynamic transition probability is then determined using softmax through the non-stationary fuzzy Markov model, including:

[0034] Based on the state space of the workload pattern fragment Each state Corresponding to a non-stationary fuzzy set Non-stationary fuzzy sets Through membership function Dynamically adjust the load. In the time slot State The degree of membership;

[0035] The state transition path of the workload over time is determined by the transition fuzzy probability matrix, which is:

[0036]

[0037] in, and They represent the states respectively. arrive Number of transitions and states within one step Number of observations The parameter is neural networks, The total number of states. Indicates the current state For the membership vector of all fuzzy sets, Represents the exponential transformation function. Used to capture time correlation A smoothing statistic is used to calculate the frequency of transfers in order to capture historical transfer trends.

[0038] Based on the aforementioned method for predicting the workload resource patterns of computing networks, a variable-order Markov model is obtained by determining the dynamic transition probability, and based on the hybrid hypergraph, a... The hop space tuple and link loss factor determine a fuzzy representation of the contention for workload resource patterns in the target computing power network, including:

[0039] The variable-order Markov model is:

[0040]

[0041] In the variable-order Markov model, the co-occurrence frequency is calculated in units of state tuples. It is a state tuple, and , For state space of Cartesian product of order 1, representing all possible Cartesian products. Order state combination, From the continuous states in the workload sequence dynamic order composition; For indicator functions, Let be the decay rate parameter, and , Indicates the current time slot. For historical time slot index;

[0042] The Jump space tuple representation for each hyperedge ,That Jumping inner neighbor super-edge set through computing power nodes Skip all superedges of the inner join:

[0043]

[0044]

[0045] in, Setting a non-zero value to 1 indicates the existence of a connection; shield The neighbors outside the jump; This represents the Hadema product, which is element-wise multiplication; for Each neighbor superedge in the jump space tuple Calculate from arrive path Link loss factor :

[0046]

[0047] in, Adjacency matrix Middle node arrive The weighted sum of the link quality vectors;

[0048] Through non-stationary fuzzy sets Add to the membership function The fuzzy representation of the competition for workload resource patterns in the target computing power network is determined by the jump space tuple.

[0049] .

[0050] According to the aforementioned method for predicting the workload resource patterns of a computing power network, a fuzzy representation of the competition between a variable-order Markov model and workload resource patterns is obtained. A multi-order adaptive cross-entropy loss function is then used to predict the workload resource patterns of the target computing power network in the next time slot, yielding prediction results, including:

[0051] Based on a nonstationary fuzzy Markov model, the dynamic transition probability matrix parameters are optimized using gradient descent, and the workload is... Next time slot Fuzzy patterns of workloads The multi-order adaptive cross-entropy loss function is used. Make a prediction:

[0052]

[0053] in, It is the length of the workload's lifecycle. Is it a non-stationary fuzzy set? The probabilistic simplex and the multi-order adaptive cross-entropy loss function for:

[0054]

[0055] in, Historical time state tuples in historical workload trajectories The prior probability, This represents the predicted conditional probability of the next state. This indicates that the work is in the state space. Distribution on, The conditional probability of the distribution. and It is the regularization sparse coefficient. This indicates the calculation of the 2nd norm of a matrix.

[0056] Another aspect of the present invention provides a computing network workload resource pattern prediction device, comprising:

[0057] The first module is used to obtain the workload set of the target computing power network and determine the shape set based on the workload set. The workload set is used to characterize the resource utilization rate index of the workload during its life cycle, and the shape set is used to characterize the pattern fragment of the joint utilization of multiple resources. Each shape set has a corresponding fuzzy set.

[0058] The second module is used to process the shape set through a membership function to obtain the membership degree of the workload to the pattern fragment of the multi-resource utilization joint, and then determine the temporal and spatial coupled workload distribution of the target computing power network through the hybrid hypergraph.

[0059] The third module is used to determine the first-order dynamic neighbor load set based on the workload, and to determine the non-stationary fuzzy set based on the first-order dynamic neighbor load set and the fuzzy set, wherein the non-stationary fuzzy set is used to capture the fluctuations caused by changes in the first-order dynamic neighbor load set and time changes.

[0060] The fourth module is used to determine the non-stationary fuzzy Markov model based on the non-stationary fuzzy set, and to determine the dynamic transition probability using softmax based on the non-stationary fuzzy Markov model.

[0061] The fifth module is used to determine the variable-order Markov model through dynamic transition probabilities, and, based on the hybrid hypergraph, employ... The jump space tuple and link loss factor determine the fuzzy representation of the competition for workload resource patterns in the target computing power network.

[0062] The sixth module is used to predict the workload resource pattern of the target computing network in the next time slot based on the fuzzy representation of the competition between the variable-order Markov model and the workload resource pattern, and to obtain the prediction result.

[0063] Another aspect of the present invention provides an electronic device, including a processor and a memory;

[0064] The memory is used to store programs;

[0065] The processor executes the program to implement the method as described above.

[0066] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the methods described above.

[0067] The beneficial effects of this invention are as follows: Based on resource pattern fragments of historical loads, a shape-based fuzzy set is designed to map the multidimensional resource utilization index of the workload to the universe of discourse through a non-stationary membership function. Based on the topology of the workload deployed in the computing network and the quality relationship of the fiber optic links on the all-optical base of the computing network, a hybrid fuzzy hypergraph including hyperedges and paired edges is designed. Further considering the interference of workloads residing in empty common locations, a neighborhood-aware non-stationary membership function and a non-stationary fuzzy Markov model are used. The dynamic transition probability matrix of the softmax model is used to integrate non-stationary membership, historical state contribution probability, and time factors. The dynamic transition probability matrix accurately predicts the uncertainty distribution of the state of the workload in the next time slot. By capturing the spatial neighbor workloads within multiple hops on the hybrid hypergraph network, the first-order Markov model is extended to a variable order to more completely observe the long-range association of workloads within a longer historical window and the complex state transition dependencies caused by dynamic neighbor interference. Finally, by using a multi-order adaptive cross-entropy loss function to update the membership function parameters, fuzzy relation matrix, and probability transition parameter matrix, the accuracy of computing network workload resource pattern prediction is improved. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the computing power network workload resource pattern prediction process according to an embodiment of the present invention.

[0069] Figure 2 This is a schematic diagram of a computing power network workload resource pattern prediction device according to an embodiment of the present invention. Detailed Implementation

[0070] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings. Throughout the description, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" can be used interchangeably. Terms such as "first," "second," etc., are used only to distinguish technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the sequential relationship of the indicated technical features. In the following description, the consecutive reference numerals for method steps are for ease of review and understanding. Adjusting the implementation order of steps, in conjunction with the overall technical solution of the present invention and the logical relationship between the various steps, will not affect the technical effect achieved by the technical solution of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0071] refer to Figure 1This includes, but is not limited to, steps S100~S600:

[0072] S100: Obtain the workload set of the target computing power network. Based on the workload set, determine the shape set. The workload set is used to characterize the resource utilization rate index of the workload during its life cycle, and the shape set is used to characterize the pattern fragment of the joint utilization of multiple resources. Each shape set has a corresponding fuzzy set.

[0073] In some embodiments, the workload set in the computing network is represented as Each workload Within their respective life cycles The main monitoring metrics are five resource utilization indicators: CPU utilization, GPU utilization, memory utilization, disk read throughput, and disk write throughput. Define shape set It is a multidimensional resource utilization joint pattern segment with a stable statistical distribution in the workload time series. ,in, It is an integer starting from 1, used to uniquely identify different stable mode fragments, each time span Shape sets are variable and can be expressed independently. It allows the extraction of any number, location, and statistical characteristics of representative stable patterns from time series without pre-constraining the length or number of segments.

[0074] For each shape set Define its fuzzy set And it is quantified by the membership function. Indicates workload For pattern fragments degree of membership, domain Indicates all workloads ,Right now .

[0075] S200 processes the shape set using a membership function to obtain the membership degree of the workload to the pattern fragments of the multi-resource utilization joint, and then determines the temporal and spatial coupled workload distribution of the target computing power network through a hybrid hypergraph.

[0076] In some embodiments, workload Throughout the life cycle The resource utilization mode is not static; it switches between different modes, affecting each workload. life cycle Divided into A fixed-length unit time slot The time slot length is Workload In each time slot The 5-dimensional resource utilization rate index is expressed as follows: Each time slot workload Mapped to the universe of discourse All fuzzy sets Unified fuzzy set The membership degree of the fuzzy set representing the shape set is compared dynamically in each time slot. The most significant pattern feature of the time slot is selected by taking the maximum value, that is, matching its dominant resource usage pattern.

[0077]

[0078] in, Indicates time slot workload Workload The entire load flow.

[0079] Using Hybrid Hypergraph Constructing a temporally and spatially coupled workload distribution in a computing power network. Hyperedges represent geographically ubiquitous computing power nodes. Connecting to the scheduling center is the workload. The allocated computing power nodes. Because the workload requires data transfer across nodes, directed pairs of edges are used to explicitly define the physical constraints received along the data flow path. parameter vector Represents computing power nodes To computing nodes The normalized uplink and downlink bandwidth asymmetric weight parameter vector, ,in, This represents the Z-score normalization function, which is independently normalized for each link parameter. , , and These represent bandwidth, latency, packet loss rate, and link availability probability, respectively. It is a vertex-hyperedge incidence matrix used to indicate the coverage relationship between each hyperedge and a node. It is a vertex-to-vertex weighted adjacency matrix representing the link quality between computing nodes in the underlying fiber optic link of a computing power network. .

[0080]

[0081] S300 determines a first-order dynamic neighbor load set based on the workload, and determines a non-stationary fuzzy set based on the first-order dynamic neighbor load set and the fuzzy set. The non-stationary fuzzy set is used to capture fluctuations caused by changes in the first-order dynamic neighbor load set and time variations.

[0082] In some embodiments, workloads residing in a shared spatiotemporal location within a computing network influence each other. The neighbor load is formally defined through the vertex cover relationship between hyperedges as follows: the first-order dynamic neighbor load set is defined as the load... life cycle Any time slot within All of the following Other loads share at least one computing node.

[0083]

[0084] in, Indicates neighbor load Any time slot within its own life cycle. This indicates that the time windows of the two overlap, that is... exist Active within the time slot.

[0085] Considering the non-stationary nature of workload resource pattern trajectories, based on fuzzy sets Design nonstationary fuzzy sets It aims to capture small fluctuations in non-stationary fuzzy sets that vary with dynamic neighbor load sets and time. Non-stationary membership function of time-varying neighbor perception The definition is as follows.

[0086]

[0087] in, Indicates the superedge and super edge The normalized intersection size. This represents the bipolar Sigmoid function. This represents the fuzzy modulation operator (Einstein product). This represents vector concatenation. Indicates neighbor load Membership vector on all fuzzy sets It is a learnable fuzzy relation matrix that encodes the promoting or inhibiting effects between loads.

[0088] S400: Determine the non-stationary fuzzy Markov model based on the non-stationary fuzzy set, and use softmax to determine the dynamic transition probability through the non-stationary fuzzy Markov model.

[0089] In some embodiments, for predicting workload In the next time slot Pattern fragments based on non-stationary fuzzy sets Design a nonstationary fuzzy Markov model The transition fuzzy transition matrix of the Markov model tracks the changing trend of workload patterns, and the probability estimation of the membership function helps to handle the random uncertainty in the workload trajectory. The stationarity assumption (fixed transition probabilities) of the traditional Markov model is difficult to adapt to dynamic load environments. The non-stationary fuzzy Markov model captures the evolutionary characteristics of patterns over time and with the influence of neighbors through non-stationary fuzzy sets. Let the state space of the workload pattern segment be... Each state Corresponding to a non-stationary fuzzy set Its membership function Dynamic adjustment to reflect load In the time slot State The degree of membership. Transition fuzzy probability matrix. It specifies the state transition path of the workload over time. Indicates from Going to status The probability, and They represent the states respectively. arrive Number of transitions and states within one step The number of observations, since the state transition probability depends not only on the current state, but also on factors such as time slots. Neighbor load Related, based solely on historical co-occurrence frequency calculate Insufficient to capture the dynamic evolution of workloads, it suffers from pattern drift: static transition probabilities cannot adapt to long-term changes in load patterns, and spatial effects are ignored: the competition or synergistic effects of neighboring loads are not explicitly modeled. The dynamic transition probability calculation formula based on softmax is shown below.

[0090]

[0091] in, The parameter is The neural network takes dynamic features as input and outputs unnormalized scores of state transitions. (Denominator) Ensure probability normalization. This represents the total number of states. Indicates the current state Membership vector for all fuzzy sets. This represents the exponential transformation function. Its purpose is to capture time correlations, and certain patterns are more likely to shift during specific time periods (such as peak periods). A smoothing statistic used to calculate the frequency of transfers, capturing historical transfer trends.

[0092] S500, using dynamic transition probabilities to determine the variable-order Markov model, and based on the hybrid hypergraph, employing... The jump space tuple and link loss factor determine the fuzzy representation of the competition for workload resource patterns in the target computing power network.

[0093] It should be noted that the fixed-order first-order Markov model assumes that the next state depends entirely on the current state and is independent of earlier states. This leads to a history truncation bias when predicting periodic or bursty points in workload patterns. The variable-order Markov model, by observing a longer historical window of states, captures the complex state transition dependencies in workload patterns caused by long-range correlations and dynamic neighbor interference. The original workload sequence contains continuous states... Composition of dynamic order The time state tuple below , For state space of Cartesian product of order 1, representing all possible Cartesian products. Order-state combination. Corresponding transition frequency statistics. and They represent from tuples Transition to state The frequency of occurrence and its total frequency. Unlike the first-order Markov model, the variable-order transition probability matrix... Co-occurrence frequency is calculated in units of state tuples.

[0094]

[0095] in, This is an indicator function that counts the frequency of conditions being met. (State tuple) It appeared frequently in the early stages, but gradually disappeared due to the evolution of load patterns. Traditional frequency statistics This will consistently overestimate the transition probability, leading to prediction bias. By introducing an exponentially decaying weight, the frequency of historical transitions decays over time. The decay rate parameter controls how quickly the weights of old data decrease. For the current time slot, This is a historical time slot index.

[0096] Because the competition for computing resources has a multi-level transmission characteristic, a resource pressure transmission chain is formed sequentially between computing nodes. A decrease in disk IOPS of a downstream computing node will, in turn, affect the cache hit rate of the upstream node. This can be mitigated by establishing... Jump space tuples improve the information completeness of workload pattern prediction for each hyperedge. ,That Jump inner neighbor superedge set is through computing power nodes Skip all superedges of the inner join. The definition is as follows.

[0097]

[0098] in, Setting a non-zero value to 1 indicates the existence of a connection. shield The neighbors outside the jump. This represents the Hadema product, which is element-wise multiplication. Each neighbor superedge in the jump space tuple Calculate from arrive path Link loss factor .

[0099]

[0100] in, Adjacency matrix Middle node arrive The weighted sum of the link quality vectors. Along paths with multiple paths All adjacent node pairs Link quality Perform a series of products and take all the results. The highest quality among the skip paths (maximum) The path (value) represents the impact of the link.

[0101] In nonstationary fuzzy sets In the membership function, introduce The influence of hopping neighbors allows for a fuzzy representation of dynamic resource competition as follows:

[0102] .

[0103] S600 uses a multi-order adaptive cross-entropy loss function to predict the workload resource pattern of the target computing network in the next time slot based on the fuzzy representation of the competition between the variable-order Markov model and the workload resource pattern.

[0104] The nonstationary fuzzy Markov model constructed based on the above steps By optimizing the dynamic transition probability matrix parameters through gradient descent, workload can be accurately predicted. Next time slot Fuzzy patterns of workloads .

[0105]

[0106] in, It is the length of the workload's lifecycle. Is it a non-stationary fuzzy set? The probability simplex is used. Combining the multi-order properties of the dynamic transition probability matrix and the optimization of fuzzy membership parameters, a multi-order adaptive cross-entropy loss function is designed as shown below.

[0107]

[0108] in, Historical time state tuples in historical workload trajectories The prior probability, This represents the predicted conditional probability of the next state. It does not refer to a specific pattern, but rather to the state space. Distribution on, That is the corresponding actual conditional probability. and It is the regularization sparse coefficient. This represents the calculation of the 2nd norm of a matrix. Membership function. The design employs Gaussian membership functions, whose mean, variance, and... and Updates are obtained during the backpropagation process. This indicates that Markov models allow the use of Time state tuple of order Predict the state of the next time slot.

[0109] Figure 2 This is a schematic diagram of a computing network workload resource pattern prediction device according to an embodiment of the present invention. The device includes a first module 210, a second module 220, a third module 230, a fourth module 240, a fifth module 250, and a sixth module 260.

[0110] The system comprises the following modules: First, it acquires the workload set of the target computing network and determines shape sets based on the workload set. The workload set represents the resource utilization rate of the workload over its lifecycle, and the shape sets represent pattern segments of multiple resource utilization. Each shape set has a corresponding fuzzy set. Second, it processes the shape sets using a membership function to obtain the membership degree of the workload to the pattern segments of multiple resource utilization, and then determines the temporally and spatially coupled workload distribution of the target computing network using a hybrid hypergraph. Third, it determines a first-order dynamic neighbor load set based on the workload, and determines a non-stationary fuzzy set based on the first-order dynamic neighbor load set and the fuzzy sets. The non-stationary fuzzy set is used to capture fluctuations caused by changes in the first-order dynamic neighbor load set and time variations. Fourth, it determines a non-stationary fuzzy Markov model based on the non-stationary fuzzy set, and uses softmax to determine the dynamic transition probability using the non-stationary fuzzy Markov model. Fifth, it determines a variable-order Markov model based on the dynamic transition probability, and uses softmax based on the hybrid hypergraph. The jump space tuple and link loss factor determine the fuzzy representation of the competition for workload resource patterns in the target computing power network; the sixth module is used to predict the workload resource patterns of the target computing power network in the next time slot based on the variable-order Markov model and the fuzzy representation of the competition for workload resource patterns, and obtain the prediction results.

[0111] For example, with the cooperation of the first, second, third, fourth, fifth, and sixth modules in the device, the embodiment device can implement any of the aforementioned methods for predicting the workload resource patterns of a computing power network. Specifically, it obtains the workload set of the target computing power network, determines a shape set based on the workload set, where the workload set characterizes the resource utilization rate index of the workload over its lifecycle, and the shape set characterizes a pattern fragment of joint multi-resource utilization. Each shape set has a corresponding fuzzy set. The shape set is then processed using a membership function to obtain the workload's relationship to the joint multi-resource utilization. The membership degree of the pattern fragments is determined, and then the temporal and spatial coupling workload distribution of the target computing power network is determined through the hybrid hypergraph. A first-order dynamic neighbor load set is determined based on the workload. A non-stationary fuzzy set is determined based on the first-order dynamic neighbor load set and the fuzzy set, wherein the non-stationary fuzzy set is used to capture fluctuations caused by changes in the first-order dynamic neighbor load set and time variations. A non-stationary fuzzy Markov model is determined based on the non-stationary fuzzy set, and the dynamic transition probability is determined using softmax based on the non-stationary fuzzy Markov model. A variable-order Markov model is determined based on the dynamic transition probability, and, based on the hybrid hypergraph, a variable-order Markov model is determined using softmax. The jump space tuple and link loss factor determine the fuzzy representation of the competition for workload resource patterns in the target computing power network. Based on the variable-order Markov model and the fuzzy representation of the competition for workload resource patterns, the multi-order adaptive cross-entropy loss function is used to predict the workload resource patterns of the target computing power network in the next time slot, and the prediction results are obtained. The beneficial effects of this invention are as follows: Based on resource pattern fragments of historical loads, a shape-based fuzzy set is designed to map the multidimensional resource utilization index of the workload to the universe of discourse through a non-stationary membership function. Based on the topology of the workload deployed in the computing network and the quality relationship of the fiber optic links on the all-optical base of the computing network, a hybrid fuzzy hypergraph including hyperedges and paired edges is designed. Further considering the interference of workloads residing in empty common locations, a neighborhood-aware non-stationary membership function and a non-stationary fuzzy Markov model are used. The dynamic transition probability matrix of the softmax model is used to integrate non-stationary membership, historical state contribution probability, and time factors. The dynamic transition probability matrix accurately predicts the uncertainty distribution of the state of the workload in the next time slot. By capturing the spatial neighbor workloads within multiple hops on the hybrid hypergraph network, the first-order Markov model is extended to a variable order to more completely observe the long-range association of workloads within a longer historical window and the complex state transition dependencies caused by dynamic neighbor interference. Finally, by using a multi-order adaptive cross-entropy loss function to update the membership function parameters, fuzzy relation matrix, and probability transition parameter matrix, the accuracy of computing network workload resource pattern prediction is improved.

[0112] This invention also provides an electronic device, which includes a processor and a memory;

[0113] The memory stores the program;

[0114] The processor executes a program to perform the aforementioned computing network workload resource pattern prediction method; the electronic device has the function of carrying and running the software system for predicting computing network workload resource patterns provided in the embodiments of the present invention, such as a personal computer, minicomputer, mainframe, workstation, network or distributed computing environment, standalone or integrated computer platform, or communicating with charged particle tools or other imaging devices, etc.

[0115] This invention also provides a computer-readable storage medium storing a program that is executed by a processor to implement the computing power network workload resource pattern prediction method described above.

[0116] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented in the embodiments of this invention. Alternative embodiments are contemplated, in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0117] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned computing network workload resource pattern prediction method.

[0118] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, considering the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed in the embodiments of the invention, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0119] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0120] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can include, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0121] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0122] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0123] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0124] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0125] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for predicting workload resource patterns in a computing network, characterized in that, include: Obtain the workload set of the target computing power network, and determine the shape set based on the workload set. The workload set is used to characterize the resource utilization rate index of the workload during its life cycle, and the shape set is used to characterize the pattern fragment of the joint utilization of multiple resources. Each shape set has a corresponding fuzzy set. The shape set is processed by a membership function to obtain the membership degree of the workload to the pattern fragment of the multi-resource utilization joint, and then the temporal and spatial coupled workload distribution of the target computing power network is determined by the hybrid hypergraph. A first-order dynamic neighbor load set is determined based on the workload. A non-stationary fuzzy set is determined based on the first-order dynamic neighbor load set and the fuzzy set, wherein the non-stationary fuzzy set is used to capture fluctuations caused by changes in the first-order dynamic neighbor load set and time variations. The nonstationary fuzzy Markov model is determined based on the nonstationary fuzzy set, and the dynamic transition probability is determined by softmax through the nonstationary fuzzy Markov model. The variable-order Markov model is determined by dynamic transition probabilities, and based on the hybrid hypergraph, the following is adopted: The jump space tuple and link loss factor determine the fuzzy representation of the competition for workload resource patterns in the target computing power network. Based on the fuzzy representation of the competition between the variable-order Markov model and the workload resource pattern, a multi-order adaptive cross-entropy loss function is used to predict the workload resource pattern of the target computing network in the next time slot, and the prediction results are obtained.

2. The method for predicting the workload resource pattern of a computing network according to claim 1, characterized in that, The process of obtaining the workload set of the target computing power network and determining the shape set based on the workload set includes: The workload set is represented as Each workload Within their respective life cycles Resource utilization metrics include CPU utilization, GPU utilization, memory utilization, disk read throughput, and disk write throughput. These metrics are expressed as... , Total workload, These are integers starting from 1, used to uniquely identify different stable operating mode segments. Fragment of a multidimensional resource utilization joint model Time span, shape set It allows the extraction of any number of representative stable patterns with any position and any statistical characteristics from time series.

3. The method for predicting computing network workload resource patterns according to claim 2, characterized in that, The process of processing the shape set using a membership function to obtain the membership degree of the workload to the pattern fragments of the multi-resource utilization joint, and then determining the temporally and spatially coupled workload distribution of the target computing power network through a hybrid hypergraph, includes: The fuzzy set of the shape set Quantization is performed using a membership function, where the membership function... Indicates workload For pattern fragments degree of membership, domain Represents workload set All workloads in, i.e. ; Each time slot workload Mapped to the universe of discourse All fuzzy sets Unified fuzzy set The membership degrees of the fuzzy sets representing the shape set are compared dynamically on a time-slot basis, and the maximum value is used to filter out each time slot. The most prominent pattern feature, among which the unified fuzzy set for: ； in, Indicates time slot workload Workload The entire load flow; Using Hybrid Hypergraph Construct a temporally and spatially coupled workload distribution in the target computing power network, where Hyperedges represent geographically ubiquitous computing power nodes. Connect to the scheduling center, which is used for workloads. Allocate computing power nodes; Directed pairs of edges are used to determine the received physical constraints on the data flow path. parameter vector Represents computing power nodes To computing nodes The normalized uplink and downlink bandwidth asymmetric weight parameter vector, ,in, This represents the Z-score normalization function, which is independently normalized for each link parameter. , , and These represent bandwidth, latency, packet loss rate, and link availability probability, respectively. It is a vertex-hyperedge incidence matrix used to show the node coverage relationship of each hyperedge. It is a vertex-to-vertex weighted adjacency matrix representing the link quality between computing nodes in the underlying fiber optic link of a computing power network. ,in 。 4. The method for predicting the workload resource pattern of a computing network according to claim 3, characterized in that, The first-order dynamic neighbor load set is determined based on the workload. A non-stationary fuzzy set is determined based on the first-order dynamic neighbor load set and the fuzzy set. The non-stationary fuzzy set is used to capture fluctuations caused by changes in the first-order dynamic neighbor load set and time variations, including: The first-order dynamic neighbor load set determined by the workload is: ； in, Indicates neighbor load At any time slot within its own life cycle This indicates that the time windows of the two overlap, that is... exist Active within time slots; Non-stationary fuzzy sets are: ； in, Indicates the superedge and super edge The size of the normalized intersection; This represents the bipolar Sigmoid function. Represents the fuzzy modulation operator. This represents vector concatenation. Indicates neighbor load Membership vector on all fuzzy sets It is a learnable fuzzy relation matrix used to promote or inhibit the interaction between coding pattern loads.

5. The method for predicting the workload resource pattern of a computing network according to claim 4, characterized in that, The step of determining a non-stationary fuzzy Markov model based on a non-stationary fuzzy set, and then using softmax to determine the dynamic transition probability through the non-stationary fuzzy Markov model, includes: Based on the state space of the workload pattern fragment Each state Corresponding to a non-stationary fuzzy set Non-stationary fuzzy sets Through membership function Dynamically adjust the load. In the time slot State The degree of membership; The state transition path of the workload over time is determined by the transition fuzzy probability matrix, which is: ； in, and They represent the states respectively. arrive Number of transitions and states within one step Number of observations The parameter is neural networks, The total number of states. Indicates the current state For the membership vector of all fuzzy sets, Represents the exponential transformation function. Used to capture time correlation A smoothing statistic is used to calculate the frequency of transfers in order to capture historical transfer trends.

6. The method for predicting computing network workload resource patterns according to claim 5, characterized in that, The variable-order Markov model is determined by dynamic transition probabilities, and based on the hybrid hypergraph, a hybrid hypergraph is used. The hop space tuple and link loss factor determine a fuzzy representation of the contention for workload resource patterns in the target computing power network, including: The variable-order Markov model is: ； In the variable-order Markov model, the co-occurrence frequency is calculated in units of state tuples. It is a state tuple, and , For state space of Cartesian product of order 1, representing all possible Cartesian products. Order state combination, From the continuous states in the workload sequence dynamic order composition; For indicator functions, Let be the decay rate parameter, and , Indicates the current time slot. For historical time slot index; The Jump space tuple representation for each hyperedge ,That Jumping inner neighbor super-edge set through computing power nodes Skip all superedges of the inner join: ； ； in, Setting a non-zero value to 1 indicates the existence of a connection; shield The neighbors outside the jump; This represents the Hadema product, which is element-wise multiplication; for Each neighbor superedge in the jump space tuple Calculate from arrive path Link loss factor : ； in, Adjacency matrix Middle node arrive The weighted sum of the link quality vectors; Through non-stationary fuzzy sets Add to the membership function The fuzzy representation of the competition for workload resource patterns in the target computing power network is determined by the jump space tuple. 。 7. The method for predicting the workload resource pattern of a computing network according to claim 6, characterized in that, The method, based on the fuzzy representation of the competition between variable-order Markov models and workload resource patterns, uses a multi-order adaptive cross-entropy loss function to predict the workload resource pattern of the target computing network in the next time slot, obtaining prediction results, including: Based on a nonstationary fuzzy Markov model, the dynamic transition probability matrix parameters are optimized using gradient descent, and the workload is... Next time slot Fuzzy patterns of workloads The multi-order adaptive cross-entropy loss function is used. Make a prediction: ； in, It is the length of the workload's lifecycle. Is it a non-stationary fuzzy set? The probabilistic simplex and the multi-order adaptive cross-entropy loss function for: ； in, Historical time state tuples in historical workload trajectories The prior probability, This represents the predicted conditional probability of the next state. This indicates that the work is in the state space. Distribution on, The conditional probability of the distribution. and It is the regularization sparse coefficient. This indicates the calculation of the 2nd norm of a matrix.

8. A computing power network workload resource pattern prediction device, characterized in that, include: The first module is used to obtain the workload set of the target computing power network and determine the shape set based on the workload set. The workload set is used to characterize the resource utilization rate index of the workload during its life cycle, and the shape set is used to characterize the pattern fragment of the joint utilization of multiple resources. Each shape set has a corresponding fuzzy set. The second module is used to process the shape set through a membership function to obtain the membership degree of the workload to the pattern fragment of the multi-resource utilization joint, and then determine the temporal and spatial coupled workload distribution of the target computing power network through the hybrid hypergraph. The third module is used to determine the first-order dynamic neighbor load set based on the workload, and to determine the non-stationary fuzzy set based on the first-order dynamic neighbor load set and the fuzzy set, wherein the non-stationary fuzzy set is used to capture the fluctuations caused by changes in the first-order dynamic neighbor load set and time changes. The fourth module is used to determine the non-stationary fuzzy Markov model based on the non-stationary fuzzy set, and to determine the dynamic transition probability using softmax based on the non-stationary fuzzy Markov model. The fifth module is used to determine the variable-order Markov model through dynamic transition probabilities, and, based on the hybrid hypergraph, employ... The jump space tuple and link loss factor determine the fuzzy representation of the competition for workload resource patterns in the target computing power network. The sixth module is used to predict the workload resource pattern of the target computing network in the next time slot based on the fuzzy representation of the competition between the variable-order Markov model and the workload resource pattern, and to obtain the prediction result.

9. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the computing power network workload resource pattern prediction method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program, which is executed by a processor to implement the computing power network workload resource pattern prediction method as described in any one of claims 1-7.

Images