Resource scheduling method and apparatus based on computing cluster twin modeling

Abstract

The present invention relates to the field of resource scheduling, and in particular to a resource scheduling method and apparatus based on computing cluster twin modeling. The method mainly comprises: constructing a time sequence matrix on the basis of available resource information of a computing cluster, and constructing a digital twin model for resource state on the basis of the time sequence matrix; extracting a time sequence from the digital twin model for resource state to construct a state space, solving a probability transfer matrix of the state space, and constructing a digital twin model for resource probability transfer on the basis of the probability transfer matrix; for the resource requirements, constructing a feasible solution set for resource allocation on the basis of the digital twin model for resource probability transfer, obtaining the feasible solution set for resource allocation on the basis of the feasible solution set, and forming a scheduling policy library on the basis of the feasible solution set; and receiving resource requests in parallel, extracting a corresponding feasible solution from the scheduling policy library, and allocating resources to each resource request by taking the feasible solution as an allocation policy. The present invention can realize unified quantitative description of real-time states of various resources, and can also realize dynamic intelligent optimal scheduling of the resources.

Description

A resource scheduling method and device based on computing cluster twin modeling

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority from the following six patent applications:

[0003] (1) A Chinese patent application entitled “A resource scheduling method and device based on computing cluster twin modeling”, filed with the Chinese Patent Office on September 1, 2023, with application number 2023111324997; Technical Field

[0004] The present invention relates to the field of resource scheduling, and in particular to a resource scheduling method and device based on computing cluster twin modeling. Background Art

[0005] Digital transformation is accelerating across all industries. High-performance, highly reliable communication networks and large-scale computing clusters are crucial infrastructure supporting this transformation. Large-scale computing clusters, including servers, switches, routers, and other devices, provide computing, storage, traffic, latency, throughput, and other information processing and data communication capabilities for digital applications across various industries. Different business applications have varying resource requirements, necessitating efficient resource scheduling methods and systems to achieve optimal or near-optimal resource allocation. Traditional data center resource allocation methods include round-robin scheduling, maximum carrier-to-interference ratio (CCIR), and proportional fairness, each with its own advantages and disadvantages.

[0006] Round-robin scheduling gives all resource requesters equal access to the resources they need. This scheduling algorithm is highly fair, but it doesn't comprehensively consider target factors such as latency, priority, and efficiency, nor does it consider the weighting of these factors. The maximum carrier-to-interference ratio scheduling algorithm prioritizes resource requesters with the highest carrier-to-interference ratio when allocating resources. This allows for high throughput, but if a resource requester has a low carrier-to-interference ratio, it will struggle to gain an advantage in resource allocation, resulting in lower fairness. The proportional fairness algorithm combines the advantages of both round-robin and maximum carrier-to-interference ratio algorithms, achieving fair resource allocation by dynamically adjusting priorities. However, if a resource requester currently has a high carrier-to-interference ratio and obtains the relevant resources, its priority is lowered at the next moment, giving other requesters a greater chance of being allocated resources. The proportional fairness algorithm balances efficiency and fairness, but also has a relatively high algorithmic complexity. Each of the three algorithms mentioned above has its own advantages and disadvantages in terms of efficiency, fairness, and computational complexity. When resource resources are limited, the appropriate algorithm can be selected based on the specific business scenario.

[0007] With the development and deployment of various new business applications, resource scheduling algorithms for large-scale computing clusters are facing increasing pressure. Traditional round-robin scheduling algorithms, maximum load-to-interference ratio algorithms, and proportional fairness algorithms struggle to fully cover the heterogeneous resources of large-scale computing clusters and cannot achieve optimal resource allocation across multiple resource dimensions and evaluation objectives. Furthermore, many large-scale business applications currently require resources for a long time, potentially requiring a week, two weeks, or even months to complete statistical analysis and training. This requires resource scheduling algorithms to not only consider current resource status but also be able to predict future resource utilization. Digital twin technology is currently being researched and applied in military, aerospace, marine, and automotive fields. This technology builds simulation models based on physical operational mechanisms, reflecting the operational laws of physical entities. It can replay operational commands in a virtual space, present the entity's status in real time, and predict changing trends. This digital twin virtual model can better support physical operational scheduling, health assessment, and fault location, and support the optimization and improvement of various physical processes.

[0008] In summary, how to implement computing cluster resource simulation modeling based on digital twin technology, accurately reflect the real-time status of resources such as computing, storage, traffic, latency, and throughput, and model and predict the resource change mechanism to achieve optimal or near-optimal scheduling of computing cluster resources is a key technical issue that urgently needs to be studied and solved.

[0009] Application Contents

[0010] In response to the above defects or improvement needs of the prior art, the present invention solves the problem of efficient and optimized resource allocation in large-scale cluster resource scheduling.

[0011] The embodiment of the present invention adopts the following technical solutions:

[0012] In the first aspect, the present invention provides a resource scheduling method based on computing cluster twin modeling, specifically: constructing a timing matrix according to the available resource information of the computing cluster, and constructing a real-time state twin model of the available resources of the computing cluster based on the timing matrix; extracting time series from the real-time state twin model of the available resources of the computing cluster to construct a state space, obtaining the probability transfer matrix of the state space, and constructing a probability transfer twin model of the available resources of the computing cluster based on the probability transfer matrix; constructing a feasible solution set for resource allocation based on the probability transfer twin model according to resource requirements, and obtaining a feasible solution set for resource allocation based on the feasible solution set, and forming a scheduling policy library for the available resources of the computing cluster based on the feasible solution set; receiving resource requests in parallel, extracting corresponding feasible solutions from the scheduling policy library, and allocating resources to each resource request using the feasible solution as the allocation strategy.

[0013] Preferably, the method of constructing a timing matrix based on the available resource information of the computing cluster specifically includes: collecting the available status information of the requestable resources in the computing cluster, and for each type of requestable resource, constructing the current available status information obtained at each collection moment into a timing matrix in chronological order.

[0014] Preferably, the real-time status twin model of the available resources of the computing cluster is constructed based on the timing matrix, specifically including: uniformly modeling the timing matrix into a three-dimensional space based on the time, location and type dimensions, and embedding each timing matrix into the three-dimensional space of the real-time status twin model along the time dimension to construct a real-time status twin model of the available resources.

[0015] Preferably, the embedding of each timing matrix into the three-dimensional space of the real-time state twin model along the time dimension specifically includes: mapping the timing matrix to a two-dimensional area according to the type and position dimensions, embedding the mapped two-dimensional area into the three-dimensional space according to the type dimension, and setting the specific element values ​​in the three-dimensional space according to the resource type measurement unit corresponding to the timing matrix.

[0016] Preferably, the method of extracting time series from the real-time state twin model of the available resources of the computing cluster to construct a state space and obtaining the probability transfer matrix of the state space specifically includes: the time series statistical laws represented by the time series matrix of each type of resource in the real-time state twin model, extracting the time series vector of each type of resource, constructing a state space set according to the characteristics of the requested resources, and statistically analyzing the state space set to obtain the transfer matrix model of each type of resource.

[0017] Preferably, the extraction of the time series vector of each type of resource specifically includes: extracting all position and type dimension values ​​when the resource probability is transferred, searching for the time series vector corresponding to the position and type dimension values ​​in the real-time state twin model, and intercepting the time series vector fragments of each type of available resources according to the needs of the upper-level business.

[0018] Preferably, the probability transfer twin model of the available resources of the computing cluster is constructed based on the probability transfer matrix, specifically including: traversing and obtaining the time series vector fragments of the available resources corresponding to all positions and types in the probability transfer matrix, calculating the probability transfer values ​​based on the time series vector fragments, and assigning the calculated probability transfer values ​​to the corresponding elements in the probability transfer twin model.

[0019] Preferably, the construction of a set of feasible solutions for resource allocation based on the probabilistic transfer twin model specifically includes: constructing a multidimensional space based on the user who makes the resource request, the time when the resource request is made, and at least one objective function; obtaining a feasible solution corresponding to the user and time in the dimension corresponding to each objective function, and taking the feasible solutions in all objective function dimensions as a set of feasible solutions for resource allocation.

[0020] Preferably, the method of obtaining a set of feasible solutions for resource allocation based on a set of feasible solutions specifically includes: judging the dominance relationship of each feasible solution in different objective function dimensions based on a probabilistic transition twin model, obtaining a feasible solution that is optimal in any dimension and no worse than the remaining changing states in other dimensions, and using the obtained feasible solutions as a set of feasible solutions.

[0021] On the other hand, the present invention provides a resource scheduling device based on computing cluster twin modeling, specifically: including at least one processor and a memory, at least one processor and the memory are connected through a data bus, the memory stores instructions that can be executed by at least one processor, and after the instructions are executed by the processor, they are used to complete the resource scheduling method based on computing cluster twin modeling in the first aspect.

[0022] Compared with existing technologies, the advantages of this embodiment are: It uses the available resource information of a computing cluster to construct a real-time state twin model and a probabilistic transition twin model of available resources. Based on these models, it constructs a scheduling policy library for available resources and allocates resources to each resource request based on feasible solutions in the scheduling policy library. This enables a unified quantitative description of the real-time state of various resources in the computing cluster, simulates and models the probabilistic transition mechanism of remaining available resources, and realizes dynamic and intelligent resource optimization scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0024] FIG1 is a flow chart of a resource scheduling method based on computing cluster twin modeling provided by an embodiment of the present invention;

[0025] FIG2 is a schematic diagram of twin modeling of the real-time status of available resources in a computing cluster according to an embodiment of the present invention;

[0026] FIG3 is a schematic diagram of a twin modeling of probabilistic transfer of available resources in a computing cluster according to an embodiment of the present invention;

[0027] FIG4 is a schematic diagram of multi-objective intelligent optimization scheduling of computing cluster resources according to an embodiment of the present invention;

[0028] FIG5 is a schematic diagram of a process of obtaining a suitable solution from an allocation solution when computing cluster resource scheduling according to an embodiment of the present invention;

[0029] FIG6 is a schematic diagram of key functional modules of a computing cluster twin modeling and resource scheduling system according to an embodiment of the present invention;

[0030] FIG7 is a schematic diagram of the structure of a resource scheduling device based on computing cluster twin modeling provided by an embodiment of the present invention;

[0031] The reference numerals are as follows: 11: processor; 12: memory. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0033] The present invention is an architecture of a specific functional system. Therefore, the specific embodiments mainly illustrate the functional logical relationship between the various structural modules, and do not limit the specific software and hardware implementation methods.

[0034] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0035] Example 1:

[0036] In the process of allocating heterogeneous resources in large-scale clusters, how to achieve optimal or near-optimal resource allocation, thereby most efficiently allocating heterogeneous resources to the user groups requesting them, is a technical problem that currently needs to be solved. To address this technical problem, this embodiment proposes a computing cluster twin modeling and resource scheduling method. This method constructs a real-time digital twin model of the available resources' status and a probabilistic digital twin model of the available resources' transitions in a virtual space. It then constructs a set of feasible resource allocation solutions based on Pareto improvement and other methods, forming a resource allocation strategy library. This method enables intelligent and efficient scheduling of large-scale computing cluster resources.

[0037] As shown in Figure 1, the resource scheduling method based on computing cluster twin modeling provided by an embodiment of the present invention has the following specific steps:

[0038] Step 101: Build a timing matrix based on the available resource information of the computing cluster, and build a real-time status twin model of the available resources of the computing cluster based on the timing matrix.

[0039] Collect information on the various types of available resources in the computing cluster, construct a time series matrix of available resource information based on the time of collection, and build a real-time state twin model (Digital Twin Model for Resource State, abbreviated as DTMRS) of the computing cluster's available resources based on the three dimensions of time, location, and type. In actual implementation, available resources include computing resources, memory, external storage, data transmission latency, throughput, etc. In actual resource allocation, the real-time state twin model can be regarded as a quantitative description of the state of available physical resources at each moment in virtual space, and available resources can be allocated based on this model.

[0040] Step 102: Extract the time series from the real-time state twin model of the available resources of the computing cluster to construct the state space, obtain the probability transfer matrix of the state space, and construct the probability transfer twin model of the available resources of the computing cluster based on the probability transfer matrix.

[0041] Because the real-time state twin model of a computing cluster's available resources is constructed from a time-series matrix of available resource information, the real-time states at different time points can be extracted from the real-time state twin model to form a time series. This constructs a probabilistic transition state space and statistically calculates the transition matrix model for each resource type. Furthermore, a probabilistic transition twin model of the computing cluster's available resources is constructed based on location, type, current state, and next state. In actual resource allocation, the probabilistic transition twin model can be viewed as the subsequent state after allocating available resources using different allocation strategies. Based on these different subsequent states, it is possible to determine which allocation strategy produces the best subsequent state.

[0042] Step 103: Based on the resource demand, a feasible solution set for resource allocation is constructed based on the probabilistic transfer twin model, and a feasible solution set for resource allocation is obtained based on the feasible solution set. Based on the feasible solution set, a scheduling strategy library for the available resources of the computing cluster is formed.

[0043] The probabilistic transition twin model includes subsequent states generated by multiple possible allocation strategies, and each subsequent state is a feasible solution for resource allocation. However, in the actual allocation process, some allocation strategies may lead to unreasonable resource allocation in subsequent states or fail to meet business needs, so not every allocation strategy can be accepted and used. Therefore, it is also necessary to construct a set of feasible solutions for resource allocation based on the probabilistic transition digital twin model based on the computing cluster resource requirements proposed by users to form an intelligent scheduling strategy library for heterogeneous resources. Each feasible solution in the scheduling strategy library is the best allocation strategy for a certain resource request. In actual implementation, the feasible solution set can use an allocation algorithm that can achieve a global optimal solution under multiple constraints, such as a greedy algorithm, dynamic programming, etc. In the preferred solution, a Pareto improvement can be used to obtain a stable solution set, and the stable solution set is used as the feasible solution set for resource allocation.

[0044] Step 104: Receive resource requests in parallel, extract corresponding feasible solutions from the scheduling policy library, and allocate resources to each resource request using the feasible solutions as allocation strategies.

[0045] After obtaining the scheduling policy library, the corresponding feasible solution in the scheduling policy library can be selected as the allocation strategy for resource allocation based on the content of the resource request. Because both the real-time state twin model and the probabilistic transition twin model contain resource states of multiple types in multiple locations, they can respond to multiple resource requests simultaneously. Therefore, in actual implementation, user resource requests can be received in parallel and the requested resources can be allocated to multiple user groups. When allocating resources, different allocation objectives can also be selected based on different business needs, such as minimizing cost, maximizing processing volume, minimizing computing time, and maximizing remaining available resources.

[0046] After steps 101 to 104 provided in this embodiment, the modeling of the computing cluster twin model can be completed, and the optimal allocation of available resources can be achieved based on the real-time state twin model and the probabilistic transfer twin model.

[0047] To construct a time series matrix based on the available resource information of the computing cluster, it is necessary to collect the availability status information of the requestable resources in the computing cluster. For each type of requestable resource, the current availability status information obtained at each collection moment is chronologically constructed into a time series matrix. In practice, it is necessary to collect information on all types of available resources in the computing cluster, including computing resources, storage resources, and network resources. This information can be collected at a set interval or based on the status reported by the device when a resource status change occurs.

[0048] After constructing the timing matrix, a real-time status twin model of the available resources in the computing cluster can be constructed based on the timing matrix. Specifically, the timing matrix is ​​uniformly modeled as a three-dimensional space based on the time, location, and type dimensions. Each timing matrix is ​​embedded into the three-dimensional space of the real-time status twin model along the time dimension to construct the real-time status twin model of the available resources. Since the timing matrix is ​​a two-dimensional structure, when constructing the model, each timing matrix needs to be embedded into the model's three-dimensional space along the time dimension. Specifically, the timing matrix is ​​mapped to a two-dimensional region based on the type and location dimensions. The mapped two-dimensional region is embedded into the three-dimensional space based on the type dimension. The specific element values ​​in the three-dimensional space are set according to the resource type measurement unit corresponding to the timing matrix. In specific implementation, the uniformly modeled real-time status twin model includes the three dimensions of time, location, and type. For each type of resource in the computing cluster, the current available resource status obtained at each acquisition moment is formed into a timing matrix and embedded into the model along the time dimension, thus constructing a digital twin model of the real-time status of the available resources in the computing cluster. The digital twin model realizes the quantitative description of the status of various available physical resources in the virtual space, where each type of resource can be measured in appropriate units.

[0049] To obtain the status after different allocation strategies are used, it is necessary to extract time series from the real-time state twin model of the available resources in the computing cluster to construct a state space and calculate the probabilistic transition matrix of the state space. Specifically, the time series statistical patterns represented by the time series matrix of each resource type in the real-time state twin model are used to extract the time series vectors of each resource type. A state space set is constructed based on the characteristics of the requested resource. The state space set is statistically analyzed to obtain the transition matrix model for each resource type. The specific method for extracting the time series vectors for each resource type is to extract all the location and type dimension values ​​during the resource probabilistic transition, search the real-time state twin model for the time series vectors corresponding to these values, and then extract a segment of the time series vectors for each available resource type based on the needs of the upper-layer business. In actual implementation, the time series vectors of each resource type can be extracted based on the time series statistical patterns. A state space set is constructed based on the characteristics of the user-requested resource, and the transition matrix model for each resource type is obtained through statistical analysis. A digital twin model of the probabilistic transition of the available resources in the computing cluster is then constructed based on the location, type, current state, and next state. The current state is the current real-time state, and the next state is the state at the next moment after the state transition. Both states are in the real number domain. The transition matrix for each resource type is a square matrix, where the matrix elements represent the probability of transition due to changes in the availability of that type of resource.

[0050] In this embodiment, the probability transfer twin model of the available resources of the computing cluster is used to describe the changing mechanism of the available resource situation in the virtual space. After obtaining the probability transfer matrix, the probability transfer twin model of the available resources of the computing cluster can be constructed according to the probability transfer matrix. Specifically: traverse and obtain the time series vector fragments of the available resources corresponding to all positions and types in the probability transfer matrix, calculate the probability transfer value according to the time series vector fragments, and assign the calculated probability transfer value to the corresponding element in the probability transfer twin model. When allocating resources according to the resource requirements proposed by the user, the feasible probability of the user's heterogeneous resource request can be calculated based on the probability transfer digital twin model of the available resources of the computing cluster according to the principle of time priority, select the user request with a larger probability and allocate resources first.

[0051] Furthermore, in actual implementation, different allocation strategies can be used to meet different business needs. Therefore, different objective functions can be defined based on different user needs, and different allocation requirements can be matched using different objective functions. Specifically, a multidimensional space is constructed based on the user making the resource request, the time of the resource request, and at least one objective function. A feasible solution corresponding to the user and time is obtained in each dimension corresponding to the objective function, and the feasible solutions in all objective function dimensions are used as the set of feasible solutions for resource allocation. Specifically, feasible solutions can be generated using the following method: Based on a probabilistic transition twin model, the dominance relationship between each feasible solution in different objective function dimensions is determined. A feasible solution that is optimal in any dimension and not inferior to the remaining variation states in other dimensions is obtained, and the obtained feasible solutions are used as the set of feasible solutions. Using different objective functions, different feasible solutions can be obtained, thereby constructing a set of feasible resource allocation solutions, forming a library of intelligent scheduling strategies for heterogeneous resources in computing clusters. The scheduling strategy library includes a set of allocation solutions that can be provided for each type of resource in the computing cluster in response to user requests. In the preferred solution, Pareto improvement can be used to obtain a set of feasible solutions, and the stable solution of the Pareto set can be used as the feasible solution for resource allocation. The stable solution of the Pareto set means that in the change from one allocation state to another, without making any state worse, at least one item condition becomes better. Therefore, the overall optimal allocation strategy can be obtained.

[0052] After receiving user resource requests in parallel, the user's resource requests can be formed into a resource request queue, and the request queue can be matched with the resource scheduling policy library to obtain the corresponding feasible solution, thereby obtaining an available resource allocation strategy that matches the resource request and business needs.

[0053] Furthermore, to avoid resource allocation failures or resource conflicts caused by unmet resource demands, when a resource request cannot be met, it is moved to the end of the resource request queue or to a separate waiting queue, where it will be allocated again when conditions are met. In a specific scenario, users are allocated their requested resources based on the maximum remaining available resources. If the cumulative number of a particular resource type requested by a group of users exceeds the available resources for that type, the user's request is moved to the end of the resource request queue and re-queued to await resource allocation.

[0054] Based on the above-mentioned resource scheduling method based on computing cluster twin modeling, in actual implementation, the specific modeling and allocation process can be completed according to actual needs. The following provides some modeling methods or allocation methods in specific scenarios. It is understandable that the following methods do not limit the scope of protection of this solution. In actual scenarios, you can refer to the following solutions for implementation, or adjust or expand the following methods according to actual needs.

[0055] In step 101, the available computing cluster resources collected include computing, memory, external storage, latency, and throughput. In a specific scenario, the computing cluster can collect resource availability information for resources that are open to users and available for user requests, and build a real-time digital twin model. As shown in Figure 2, the top row is a collection of computing cluster resources. During implementation, positive natural numbers can be used to identify various types of resources. For example, Table 1 below provides a quantitative representation method for the entire resource set. The computing resources that the central processing unit can provide are quantified as 1, the memory is quantified as 2, and so on.

[0056] Table 1 Quantitative description methods for various types of resources in computing clusters

[0057] In Figure 2, the location dimension refers to the specific location of available resources within a computing cluster. A computing cluster includes various physical devices, such as servers, switches, and routers, each of which provides different available resources. This dimension allows the available resources of a computing cluster to be fully embedded in a four-dimensional space. For example, if physical resources are numbered by layer, starting from left to right and from top to bottom, and the first cabinet has 6 devices, the second cabinet has 8 devices, and the third cabinet has 7 devices, then the value of the position dimension I2 corresponding to the second device in the third cabinet is 6 + 8 + 2 = 16. The value of the time dimension I1 corresponds to the available resource collection period. For example, if the collection period is 30 seconds, then a value of 1 in the time dimension I1 represents the first collection moment, and a value of 2 represents the second collection moment. The time interval between the two time dimension I1 values ​​is 30 seconds. During implementation, the computing cluster twin modeling and resource scheduling system is responsible for matching the specific values ​​of each dimension with the actual content they represent. In Figure 2, each type of resource first constructs its own time series matrix, which is then embedded into the bottom three-dimensional space based on the type dimension I3. The timing matrix corresponding to the computing resources is A c , the timing matrix corresponding to the memory resource is A s , the timing matrix corresponding to the external memory is A h , the timing matrix corresponding to the delay resource is A t , the timing matrix corresponding to the throughput resource is A u . Take the computing resource timing matrix as A c For example, this is an MxN matrix, where the number of rows, M, corresponds to the location dimension, I2, of the real-time digital twin model, and the number of columns, N, corresponds to the time dimension, I1. According to Table 1, the value of the type dimension, I3, for computing resources is 1. Following this approach, the timing matrix of all resources can be embedded into the three-dimensional space of the real-time digital twin model.

[0058] The specific element values ​​in the three-dimensional space of the real-time state digital twin model of the computing cluster are set according to the resource type measurement unit. For example, computing resources can be measured using the peak floating point operations per second (TFLOPS). For example, the available resource value of the computing resource at position 16 of the computing cluster at the 8th moment is 20TFlops, and the specific element value of the real-time state digital twin model is expressed as t dtmrs , then the value of the model element at position 16, at the 8th moment, and of type computing resource is expressed as t dtmrs (8,16,1)=20.

[0059] Figure 3 shows the probabilistic transfer model of available computing cluster resources used in this embodiment. This model accurately maps the evolution of heterogeneous computing cluster resources in virtual space. The probabilistic transfer model for available resources includes four dimensions: location (I1), type (I2), current state (I3), and next state (I4).

[0060] Based on the aforementioned twin model of the real-time status of available resources in the computing cluster constructed using the three dimensions of time, location, and type, we can extract the time series corresponding to specific locations and types of resources and calculate the transition probability of each next-step state corresponding to a certain current-step state. For example, for computing resources at location 16 and type 1, with a utilization rate of 0% to 100%, the state space set is defined as S = {0, 1, 2, ..., 100}. For example, the following Table 2 gives the time series from time point t1 to t 11 The corresponding specific location and type resource status, at time point t n , time t in Table 2 n The corresponding state is the state of this step, time t n+1 The corresponding state is the next step state.

[0061] Table 2. Time series of specific locations and types of resources in the computing cluster.

[0062] The time series in Table 2 contains 11 states corresponding to 11 time points, with a total of 10 state transitions. Among them, there are 2 transitions from state 80 to state 60, accounting for 20%. That is, the four-dimensional space element value of position 16, type 1, current state 86, and next state 60 is 0.2. The digital twin model for the probability transfer of available resources in the computing cluster is represented by DTMRP (Digital Twin Model for Resource Probability Transfer), and the specific element value of the model is represented by t dtmrp , then the value of the model element at position 16, at the 8th moment, and of type computing resource is expressed as t dtmrs (16,1,86,1)=20.

[0063] The digital twin model of the real-time status of available resources in a computing cluster includes three dimensions: time, location, and type. The digital twin model of the probabilistic transition of available resources includes four dimensions: location, type, current state, and next state. A step-by-step calculation method can be used to calculate the probabilistic transition values ​​in four dimensions. In the implementation process, all values ​​for the location and type dimensions of the resource probabilistic transition digital twin model are extracted. The time series vectors corresponding to these two dimensions are found in the digital twin model of the real-time status of available resources. Time series vector segments are then intercepted based on upper-layer business needs.

[0064] Furthermore, to more accurately calculate the transition probability values ​​of four-dimensional space elements, a larger time series can be selected, containing more state transitions and more accurately reflecting the statistical patterns of state transitions in the computing cluster's resources. For example, a time series vector containing 10,000 state elements can be intercepted for calculation. The calculated transition probability values ​​are assigned to the corresponding elements in the probabilistic transition digital twin model. By traversing and calculating the time series of available resources at all locations and types, a probabilistic transition digital twin model of the computing cluster's available resources can be constructed.

[0065] In order to meet different resource requests and business needs, various types of resources can be efficiently allocated to users. In the specific implementation process, the objective function can be defined according to user needs, such as minimizing the cost of various resources and minimizing the user business processing time. Figure 4 is a multidimensional space, where dimensions I1, I2, ..., I N Represents N kinds of objective functions. Dimension I N+1 Represents the user, dimension I N+2 Indicates time. These two dimensions indicate that a user makes a resource request at a certain moment. According to user needs, dimensions I1, I2, ..., I N Various feasible solutions for resource allocation are given. When performing resource allocation, a suitable solution can be selected from these feasible solutions. To clearly illustrate the process of selecting a suitable solution from the allocation solutions, the three objective function dimensions in Figure 5 are used as an example. For example, the three objective function dimensions I1, I2, and I3 represent the minimum cost, maximum processing capacity, and minimum computation time of the computing cluster resource allocation solution, respectively. Therefore, in this example, the three objective function evaluation methods are abstracted as min, max, and min, that is, seeking the minimum, maximum, and minimum. In the three-dimensional space of Figure 5, solution S1 Pareto dominates solution S2. This is because solution S1 is smaller than solution S2 in dimension I1, larger in dimension I2, and smaller in dimension I3. However, solutions S3 and S4 do not form a Pareto dominance relationship because solution S3 is larger than solution S4 in dimension I2, while solution S4 is smaller than solution S4 in dimensions I1 and I3. During implementation, this example can obtain all stable solution sets with dominant capabilities based on Figure 5, and use the stable solution sets as feasible solution sets for the intelligent scheduling strategy library of heterogeneous resources in the computing cluster to guide the allocation of heterogeneous resources.

[0066] Based on the method proposed in this embodiment, a computing cluster twin modeling and resource scheduling system can also be established to receive user resource requests in parallel, extract resource allocation solutions from the scheduling policy library, and allocate resources to user groups with the goal of maximizing the remaining available resources.

[0067] In order to realize the corresponding functions of different steps in the method, the computing cluster twin modeling and resource scheduling system can include different functional modules, as shown in Figure 6, including: user request management module, scheduling strategy construction module, and computing cluster resource scheduling module.

[0068] The user request management module receives resource requests submitted by users and builds a user resource request queue. In a large-scale computing cluster environment, multiple users may receive resource requests at the same time, because the user request management module has the ability to process multiple user resource requests in parallel.

[0069] The scheduling strategy construction module is used to collect the current available status of various resources in the computing cluster, including computing, memory, external storage, latency, throughput, and other resources. It is also used to generate a timing matrix for each resource and arrange the timing matrices of all resource types in an orderly manner to build a real-time state digital twin model. It is also used to create a transition probability matrix for each resource type and construct a probabilistic transition digital twin model based on four dimensions: location, type, current state, and next state. At the same time, based on the real-time state twin model of the computing cluster's available resources, the probabilistic transition twin model, and the multi-objective resource optimization model, a stable solution set with dominant capabilities is constructed to form an intelligent scheduling strategy library.

[0070] The computing cluster resource scheduling module is used to determine the set of feasible allocation solutions for each resource type, using Pareto improvement to obtain a stable solution set, and to create an intelligent resource scheduling policy library. During the operation of the scheduling system, the scheduling policy library is maintained in both memory and external storage, and remains consistent. The memory policy library is used for scheduling matching, while the external storage policy library is used for backup and recovery. It is also used to receive user resource requests in parallel and create request queues. It is also used to match the scheduling policy library with the user request queues, allocating the requested resources to each user group based on the goal of maximizing the remaining available resources. In actual implementation, the computing cluster resource scheduling module can traverse the user resource request queue and allocate resources to each user based on the intelligent scheduling policy library. If the scheduling policy library has multiple feasible solutions for a particular user resource request, computing cluster resources are allocated to that user based on the goal of maximizing the remaining available resources. The goal of this allocation is to preserve as many resources as possible to meet subsequent user resource requests.

[0071] The resource scheduling method based on computing cluster twin modeling provided in this embodiment can solve the problem that the existing technical solutions cannot uniformly quantify the real-time available status of complex types of resources in large-scale computing clusters. It can comprehensively and accurately measure the current available status of various types of resources from both time and space levels, and can construct a digital twin model of available status in virtual space to reflect the available status of various types of physical entity resources in real time, and support quantitative modeling in the process of heterogeneous resource allocation. It can solve the problem that the existing technical solutions cannot accurately describe the transfer mechanism of complex types of resources in computing clusters. It can obtain the transfer law of available resources based on time series data statistics, construct a digital twin model of the probability transfer of available resources in virtual space, and construct the optimal solution set or approximate optimal solution for resource allocation based on user resource requests. It can realize intelligent and efficient scheduling of heterogeneous resources in large-scale computing clusters, accurately map the change law of entity available resources in virtual space based on the digital twin model, guide the creation of resource optimization allocation plan, and improve the efficiency of computing cluster resource allocation.

[0072] Example 2:

[0073] Based on the resource scheduling method based on computing cluster twin modeling provided in the above embodiment 1, the present invention also provides a resource scheduling device based on computing cluster twin modeling that can be used to implement the above method. As shown in Figure 7, it is a schematic diagram of the device architecture of an embodiment of the present invention. The resource scheduling device based on computing cluster twin modeling in this embodiment includes one or more processors 11 and a memory 12. Among them, Figure 7 takes one processor 11 as an example.

[0074] The processor 11 and the memory 12 may be connected via a bus or other means. FIG7 takes the bus connection as an example.

[0075] The memory 12 is a non-volatile computer-readable storage medium for a resource scheduling method based on computing cluster twin modeling, and can be used to store non-volatile software programs, non-volatile computer executable programs, and modules, such as the resource scheduling method based on computing cluster twin modeling in Example 1. The processor 11 executes the various functional applications and data processing of the resource scheduling device based on computing cluster twin modeling by running the non-volatile software programs, instructions, and modules stored in the memory 12, thereby implementing the resource scheduling method based on computing cluster twin modeling in Example 1.

[0076] The memory 12 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state memory device. In some embodiments, the memory 12 may optionally include a memory remotely located relative to the processor 11, and such remote memory may be connected to the processor 11 via a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0077] The program instructions / modules are stored in the memory 12 and, when executed by one or more processors 11, execute the resource scheduling method based on computing cluster twin modeling in the above-mentioned embodiment 1, for example, execute the various steps shown in Figure 1 described above.

[0078] Those skilled in the art will understand that all or part of the steps in the various methods of the embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a disk or an optical disk, etc.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A resource scheduling method based on computing cluster twin modeling, characterized in that: Specifically include: A timing matrix is ​​constructed based on the available resource information of the computing cluster, and a real-time status twin model of the available resources of the computing cluster is constructed based on the timing matrix; Extract the time series from the real-time state twin model of the available resources of the computing cluster to construct the state space, obtain the probability transfer matrix of the state space, and construct the probability transfer twin model of the available resources of the computing cluster based on the probability transfer matrix; According to resource requirements, a feasible solution set for resource allocation is constructed based on the probabilistic transfer twin model, and a feasible solution set for resource allocation is obtained based on the feasible solution set. Based on the feasible solution set, a scheduling strategy library for available resources of the computing cluster is formed; Receive resource requests in parallel, extract corresponding feasible solutions from the scheduling strategy library, and use the feasible solutions as allocation strategies to allocate resources to each resource request.

2. The resource scheduling method based on computing cluster twin modeling according to claim 1 is characterized in that: The step of constructing a timing matrix according to the available resource information of the computing cluster specifically includes: The available status information of the requestable resources in the computing cluster is collected. For each type of requestable resource, the current available status information obtained at each collection moment is constructed into a time series matrix in chronological order.

3. The resource scheduling method based on computing cluster twin modeling according to claim 1 is characterized in that: The real-time status twin model of available resources of the computing cluster is constructed based on the time series matrix, specifically including: Based on the dimensions of time, location and type, the timing matrix is ​​uniformly modeled as a three-dimensional space, and each timing matrix is ​​embedded into the three-dimensional space of the real-time status twin model along the time dimension to construct a real-time status twin model of available resources.

4. The resource scheduling method based on computing cluster twin modeling according to claim 3 is characterized in that: The embedding of each time series matrix into the three-dimensional space of the real-time state twin model along the time dimension specifically includes: The time series matrix is ​​mapped to a two-dimensional area according to the type and position dimensions, and the mapped two-dimensional area is embedded in the three-dimensional space according to the type dimension. The specific element values ​​in the three-dimensional space are calculated according to the corresponding values ​​of the time series matrix. Resource type measurement unit settings.

5. The resource scheduling method based on computing cluster twin modeling according to claim 1 is characterized in that: The step of extracting the time series from the real-time state twin model of the available resources of the computing cluster to construct the state space and obtaining the probability transfer matrix of the state space specifically includes: The time series statistical laws represented by the time series matrix of each type of resource in the real-time state twin model are used to extract the time series vector of each type of resource. A state space set is constructed according to the characteristics of the requested resource. The transfer matrix model of each type of resource is obtained by statistically analyzing the state space set.

6. The resource scheduling method based on computing cluster twin modeling according to claim 5 is characterized in that: The extracting of the time series vector of each type of resource specifically includes: Extract all the location and type dimension values ​​when resources are transferred probabilistically, find the time series vectors corresponding to the location and type dimension values ​​in the real-time state twin model, and intercept the time series vector fragments of each type of available resources according to the needs of the upper-level business.

7. The resource scheduling method based on computing cluster twin modeling according to claim 1 is characterized in that: The probabilistic transfer twin model of available resources of the computing cluster is constructed according to the probabilistic transfer matrix, specifically including: Traverse and obtain the time series vector fragments of available resources of all positions and types in the probability transfer matrix, calculate the probability transfer value according to the time series vector fragments, and assign the calculated probability transfer value to the corresponding element in the probability transfer twin model.

8. The resource scheduling method based on computing cluster twin modeling according to claim 1 is characterized in that: The feasible solution set for resource allocation based on the probability transfer twin model specifically includes: constructing a multidimensional space according to a user who makes a resource request, a time when the resource request is made, and at least one objective function; Obtain feasible solutions corresponding to users and time in the dimension corresponding to each objective function, and use the feasible solutions in all objective function dimensions as a set of feasible solutions for resource allocation.

9. The resource scheduling method based on computing cluster twin modeling according to claim 8 is characterized in that: The step of obtaining a feasible solution set for resource allocation according to the feasible solution set specifically includes: Based on the probability transition twin model, the dominant relationship of each feasible solution in different objective function dimensions is judged to obtain the feasible solution that is optimal in any dimension and no worse than the remaining change states in other dimensions, and the obtained feasible solution is taken as the feasible solution set.

10. The resource scheduling method based on computing cluster twin modeling according to claim 1, characterized in that: The step of obtaining a feasible solution set for resource allocation according to the feasible solution set specifically includes: The Pareto improvement is used to obtain the stable solution set, which is used as the feasible solution set for resource allocation.

11. The resource scheduling method based on computing cluster twin modeling according to claim 4, characterized in that: The real-time state twin model of unified modeling includes three dimensions: time, location, and type. The mapped two-dimensional area is embedded in the three-dimensional space according to the type dimension. The specific element values ​​in the three-dimensional space are set according to the resource type measurement unit corresponding to the timing matrix, including: For various types of resources in the computing cluster, the current available resource status obtained at each acquisition moment forms a time series matrix and is embedded into the model along the time dimension to construct a digital twin model of the real-time status of available resources in the computing cluster; among them, each type of resource uses corresponding measurement units.

12. The resource scheduling method based on computing cluster twin modeling according to claim 6, characterized in that: The step of intercepting the time series vector segments of each type of available resources according to the upper layer service needs specifically includes: Based on the statistical laws of time series of various resources, extract their respective time series vectors, build a state space set according to the characteristics of user requested resources, and obtain the transfer matrix model of various resources through statistical analysis; A digital twin model of the probabilistic transfer of available resources in a computing cluster is constructed based on location, type, current state, and next state. The current state is the real-time state at the current moment, and the next state is the state at the next moment after the state transfer. Both states are in the real number domain. The transfer matrix of each type of resource is constructed as a square matrix, and each matrix element in the square matrix represents the probability of transfer due to changes in the availability of this type of resource.

13. The resource scheduling method based on computing cluster twin modeling according to claim 7 is characterized in that assigning the calculated probability transfer value to the corresponding element in the probability transfer twin model specifically includes: When allocating resources based on the resource demands proposed by users, the digital twin model of the probability transfer of available resources in the computing cluster is used to calculate the feasibility probability of user heterogeneous resource requests according to the principle of time priority, and user requests with higher probabilities are selected and resources are allocated preferentially.

14. The resource scheduling method based on computing cluster twin modeling according to claim 9, characterized in that: The step of obtaining a feasible solution that is optimal in any dimension and not inferior to other change states in other dimensions, and taking the obtained feasible solution as a feasible solution set, specifically includes: Pareto improvement is used to obtain a set of feasible solutions, and the stable solution of the Pareto set is used as the feasible solution for resource allocation; wherein, the stable solution of the Pareto set means that in the change from one allocation state to another, at least one item becomes better without making any state worse.

15. The resource scheduling method based on computing cluster twin modeling according to claim 1, characterized in that: When a resource request cannot be satisfied, the method further includes: The request is moved to the end of the resource request queue, or to a separate waiting queue, and allocated after the conditions are met.

16. The resource scheduling method based on computing cluster twin modeling according to claim 1, characterized in that: The method further comprises: A larger time series is selected so that the time series contains more state transitions, so as to more accurately reflect the statistical laws of the state transitions of computing cluster resources.

17. A resource scheduling device based on computing cluster twin modeling, characterized in that: It includes at least one processor and a memory, and the at least one processor and the memory are connected via a data bus. The memory stores instructions that can be executed by the at least one processor, and after being executed by the processor, the instructions are used to complete the resource scheduling method based on computing cluster twin modeling as described in any one of claims 1-16.

Images