Multi-modal network cross-domain scheduling method and system based on storage, conversion and calculation integration

Abstract

The invention belongs to the technical field of distributed computing, and discloses a multi-modal network cross-domain scheduling method and system based on storage, conversion and calculation integration, and the method comprises the steps: mapping a to-be-scheduled task into a specific level in a three-dimensional resource model according to a scheduling task request submitted by a vertical service, and obtaining a three-dimensional resource demand; wherein the three-dimensional resource model comprises a network mode, a storage level and a computing power level required by the task; selecting a member domain meeting a three-dimensional resource requirement from the member domain cluster; taking the three-dimensional resource surplus of each member domain in the candidate domain set as a state variable, selecting one member domain from the candidate domain set to deploy a to-be-scheduled task as an action, and selecting an optimal member domain from the candidate domain set by adopting reinforcement learning; and taking the optimal member domain as a target domain, and allocating a to-be-scheduled task to the target domain for execution, thereby realizing storage-conversion-calculation integrated multi-modal network cross-domain scheduling. The task deployment success rate and the resource use efficiency can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of distributed computing, and more specifically, relates to a multi-modal network cross-domain scheduling method and system based on storage-computation integration. Background Art

[0002] With the continuous growth of the application requirements of intelligent computing, modern application scenarios present significant characteristics such as diversification, storage-computation integration, and resource intensiveness. These trends require the network to no longer be limited to traditional traffic access, distribution, and control functions, but to integrate and schedule computing and storage resources as network elements to provide personalized high-quality services for vertical industries. To meet such multi-modal service requirements, a multi-modal network resource scheduling engine is needed, which can efficiently and real-time allocate heterogeneous resources to achieve intelligent orchestration and scheduling of storage-computation integrated heterogeneous resources across regions, devices, protocols, multi-dimensions, and fine-grainedness.

[0003] The current intelligent computing applications present characteristics of multi-modal, storage-computation integration, and resource intensification. The existing cross-domain scheduling methods face the following challenges:

[0004] (1) Insufficient modal adaptation: The existing cross-domain scheduling methods do not distinguish tasks. When task requirements arrive, the existing methods only allocate resources based on the real-time resource usage of nodes, without fully considering the differential characteristics of task requirements in intelligent computing applications. Specifically, the existing scheduling methods will mix latency-sensitive services and high-throughput services for scheduling, resulting in low success rates of task deployment and resource utilization efficiency.

[0005] (2) Defects in static scheduling: The existing cross-domain scheduling methods rely on static weights or pre-set rules for resource allocation. This method is difficult to adapt to the dynamic changes of resource states between domains, resulting in insufficient resource utilization (on average less than 70%). Summary of the Invention

[0006] Aiming at the above defects or improvement requirements of the existing technology, the present invention provides a multi-modal network cross-domain scheduling method and system based on storage-computation integration, aiming to improve the success rate of task deployment and resource utilization efficiency.

[0007] To achieve the above object, the present invention provides a multi-modal network cross-domain scheduling method based on storage-computation integration, including:

[0008] S1. According to the scheduling task request submitted by the vertical service, map the task to be scheduled to a specific level in the three-dimensional resource model to obtain three-dimensional resource requirements; wherein, the three-dimensional resource model includes the network modality, storage level, and computing power level required by the task;

[0009] S2. Select a member domain that meets the three-dimensional resource requirements from the member domain cluster; among them, the member domains that meet the three-dimensional resource requirements form a candidate domain set; within the same geographical area, a logical domain formed by a resource set with the same network modality, storage level, and computing power level is a member domain;

[0010] S3. Use the remaining three-dimensional resources of each member domain in the candidate domain set as state variables, and use selecting a member domain from the candidate domain set to deploy the task to be scheduled as an action. Use reinforcement learning to select the optimal member domain from the candidate domain set; among them, the reward function of reinforcement learning is to maximize the resource utilization rate of each member domain and reduce load imbalance;

[0011] S4. Use the optimal member domain as the target domain, and allocate the task to be scheduled to the target domain for execution, realizing the cross-domain scheduling of the storage-computation-integration multi-modal network.

[0012] Furthermore, in S3, the selection method of the action is as follows:

[0013] Select the member domain with the highest member domain score from the candidate domain set as the action; among them, the calculation method of the member domain score is as follows:

[0014] Score j =α′×AvgUtilization j -β′×ImBalance j

[0015]

[0016] Among them, Score j is the score of the j-th member domain in the candidate domain set; α′ and β′ are the scoring weights respectively, which are the optimization parameters of reinforcement learning; the set C represents the three-dimensional resources, and |C| is the total number of three-dimensional resources; AvgUtilization j is the average utilization rate of all three-dimensional resources in the j-th member domain; is the utilization rate of the c-th resource in the j-th member domain; ImBalance j is the average imbalance rate of all node three-dimensional resources in the j-th member domain, Node i .Remainning c is the remaining amount of the c-th resource of the i-th node in the j-th member domain, Node i .Tatal c is the total amount of the c-th resource of the i-th node, and std represents the variance calculation operation; is the water level value of the c-th resource in the j-th member domain, and the calculation method is as follows:

[0017]

[0018] Among them, is the total amount of the c-th resource of the i-th node in the j-th member domain, and p is the total number of nodes in the j-th member domain; is the remaining amount of the c-th resource of the i-th node;

[0019] is the maximum remaining amount of the c-th resource in a single node in the j-th member domain; α and β are the total resource weight and the fragmentation penalty term weight respectively.

[0020] Furthermore, the reward function is:

[0021] Reward = γ × (NewUtilization - OldUtilization) - ImBalancePenalty

[0022]

[0023] Among them, γ is the discount factor, taking values between 0 and 1; ImBalancePenalty is the load imbalance value between member domains; NewUtilization is the resource utilization rate after scheduling, and OldUtilization is the resource utilization rate before scheduling.

[0024] Furthermore, in S2, selecting a member domain that meets the three-dimensional resource requirements from the member domain cluster includes:

[0025] Calculating the three-dimensional resource remaining amount of each member domain in the member domain cluster;

[0026] Querying the maximum remaining amount of resources in a single node of each member domain;

[0027] The member domain for which both the three-dimensional resource remaining amount and the maximum remaining amount of resources in a single node meet the three-dimensional resource requirements is the member domain that meets the three-dimensional resource requirements.

[0028] Furthermore, in S4, after allocating the task to be scheduled to the target domain for execution, it further includes:

[0029] Updating the three-dimensional resource remaining amount of the target domain and monitoring the three-dimensional resource remaining amount of each member domain in real time; among them, the updated three-dimensional resource remaining amount of the target domain is equal to the three-dimensional resource remaining amount of the target domain before the task execution minus the three-dimensional resource amount actually used by the task;

[0030] It further includes: synchronizing the three-dimensional resource remaining amount of each member domain, synchronizing the three-dimensional resource requirements of the task, the running state and completion status of the task.

[0031] The present invention also provides a cross-domain scheduling system for a multi-modal network based on integrated memory, transfer, and computing, including:

[0032] A network modality mapping module, configured to map a task to be scheduled to a specific level in a three-dimensional resource model according to a scheduling task request submitted by a vertical service, so as to obtain three-dimensional resource requirements; wherein, the three-dimensional resource model includes the network modality, storage level, and computing power level required by the task.

[0033] A domain scheduler, configured to select a member domain that meets the three-dimensional resource requirements from a member domain cluster; wherein, the member domains that meet the three-dimensional resource requirements form a candidate domain set; within the same geographical area, a logical domain formed by a resource set with the same network modality, storage level, and computing power level is a member domain.

[0034] The domain scheduler is further configured to use the remaining three-dimensional resources of each member domain in the candidate domain set as state variables, and use deploying a task to be scheduled to a member domain selected from the candidate domain set as an action, and select an optimal member domain from the candidate domain set by using reinforcement learning; wherein, the reward function of the reinforcement learning is to maximize the resource utilization rate of each member domain and reduce load imbalance.

[0035] A member domain manager, configured to use the optimal member domain as a target domain, and allocate the task to be scheduled to the target domain for execution, so as to implement cross-domain scheduling of the multi-modal network with integrated memory, transfer, and computing.

[0036] Further, it further includes a domain status monitor, configured to update the remaining three-dimensional resources of the target domain and monitor the remaining three-dimensional resources of each member domain in real time; wherein, the updated remaining three-dimensional resources of the target domain are equal to the remaining three-dimensional resources of the target domain before the task execution minus the three-dimensional resources actually used by the task.

[0037] It further includes a key-value storage module, configured to synchronize the remaining three-dimensional resources of each member domain, synchronize the three-dimensional resource requirements of the task, the running status, and the completion status of the task.

[0038] The present invention also provides a cross-domain scheduling device for a multi-modal network based on integrated memory, transfer, and computing, including a computer-readable storage medium and a processor;

[0039] The computer-readable storage medium is used to store executable instructions;

[0040] The processor is configured to read the executable instructions stored in the computer-readable storage medium and execute the multi-modal network cross-domain scheduling method described in any one of the above.

[0041] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the multimodal network cross-domain scheduling method described in any one of the above is implemented.

[0042] The present invention also provides a computer program product, including a computer program, and when the computer program runs on a computer, the computer is enabled to execute the multimodal network cross-domain scheduling method described in any one of the above.

[0043] Generally speaking, through the above technical solutions conceived by the present invention, the following beneficial effects can be achieved:

[0044] (1) The present invention proposes a joint space model based on three dimensions of network, storage, and computing power. This three-dimensional resource model of transfer, storage, and computing power establishes a mapping relationship between the network mode and resource requirements, and can accurately describe the resource characteristics in a multimodal network environment. By mapping the task requirements into the three-dimensional resource model, the quantification of task resource requirements is realized. The quantified resource requirements enable the differentiation of tasks submitted by vertical services. According to the different requirements of different tasks for the three-dimensional resources, a candidate domain set is screened from the member domain clusters, which improves the success rate of task deployment. Combining reinforcement learning to select the optimal member domain from the candidate domain set, the reward function is designed to improve the overall resource utilization rate and reduce load imbalance as the goal, and the task to be scheduled is deployed in the selected optimal member domain, which can improve the resource utilization rate. The present invention can realize the cross-domain collaborative scheduling of heterogeneous resources (storage / network / computing power), and improve the success rate of task deployment and resource utilization rate.

[0045] (2) Further, a dynamic weight scheduling mechanism based on domain resource water levels is introduced. By accurately calculating the remaining amounts of storage, network, and computing power resources and the maximum single usage amount of each domain, the available resource amount is used as the weight for replica scheduling. This intelligent scheduling method not only realizes the load balancing among member domains, but also ensures that the resource deployment rate of all member domains remains above 95%. Compared with the traditional static weight scheduling, the dynamic strategy of the scheduling engine is more flexible and efficient, and can respond to resource changes in real time, optimizing the system performance and reliability.

[0046] (3) Preferably, in the reinforcement learning process, a reward function is designed based on the resource utilization rate and load imbalance penalty value before and after scheduling, while maximizing the overall resource utilization rate, load balancing is carried out.

[0047] (4) Preferably, the candidate domain set screening method can ensure that the total remaining resources of each member domain in the candidate domain set and the maximum available resources of a single node thereof can meet the three-dimensional resource requirements of the current task to be scheduled, avoiding the problem of scheduling task failure caused by only the total remaining resources of the member domain meeting the resource requirements of the task while the total remaining resources of all single nodes in the member domain do not meet the requirements.

[0048] (5) Further, while realizing elastic resource sharing and load balancing, the present invention also combines a real-time monitoring and automatic adjustment mechanism to ensure the effective utilization of resources and the high availability of the system. Through a comprehensive monitoring function, the domain status is fed back in real time, and the workload distribution is automatically adjusted according to the real-time resource usage of the domain, which can better meet the requirements of the multi-modal network for efficient real-time resource allocation and improve the flexibility and response speed of the overall system.

[0049] Generally speaking, the present invention provides a more intelligent and automated cross-domain resource scheduling technology, which realizes cross-domain collaborative scheduling of heterogeneous resources (storage / network / calculation power), as well as optimization of resource utilization rate and load balancing in a dynamic environment, improves the success rate of task deployment and resource usage efficiency, enhances the flexibility and usability of cross-domain scheduling, and is applicable to cross-domain collaborative scheduling of heterogeneous resources in cloud computing, edge computing, and hybrid cloud environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a cross-domain scheduling method for a multi-modal network based on storage-transfer-computation integration in an embodiment of the present invention;

[0051] Figure 2 is an architecture diagram of a cross-domain scheduling system for a multi-modal network based on storage-transfer-computation integration in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. 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.

[0053] Embodiment 1

[0054] As Figure 1 shown, an embodiment of the present invention provides a cross-domain scheduling method for a multi-modal network based on storage-transfer-computation integration, mainly including:

[0055] S1. Map the task to be scheduled to a specific level in the three-dimensional resource model according to the scheduling task request submitted by the vertical service, and obtain the three-dimensional resource requirements. Among them, the three-dimensional resource model includes the network mode, storage level, and computing power level required by the task.

[0056] S2. Select member domains that meet the three-dimensional resource requirements of the network mode, storage level, and computing power level of the task to be scheduled from n member domains, and form a candidate domain set with the member domains that meet the three-dimensional resource requirements. Among them, a logical domain formed by a resource set with the same network mode identifier (such as IPv6 / latency-sensitive, etc.), storage level (throughput classification), and computing power level (FLOPS classification) within the same geographical area (in the embodiment of the present invention, the error radius ≤ 5 km) is used as a member domain.

[0057] S3. Use the remaining three-dimensional resources of each member domain in the candidate domain set as state variables, and use which member domain to select from the candidate domain set to deploy the task to be scheduled as an action, and use reinforcement learning to select the optimal member domain from the candidate domain set. Among them, the reward function of reinforcement learning is designed to maximize the resource utilization rate of each member domain and reduce load imbalance.

[0058] S4. Use the optimal member domain as the target domain, and allocate the task to be scheduled to the target domain for execution, realizing the cross-domain scheduling of the storage-computation-integration multi-modal network.

[0059] Preferably, before S1, it further includes: initializing the scheduling task request submitted by the vertical service to ensure that the system can correctly receive and parse the task request. Specifically, extract the basic information in the scheduling task request, including the task nature (such as AI inference, data processing, etc.), priority, resource requirements (CPU, Memory, Network, GPU, etc.), and generate a unique identifier (Task ID) for the task to be scheduled for subsequent tracking and management. In the embodiment of the present invention, one implementation method of Task ID is:

[0060] TaskID = hash(Taskname + Priority + ResourceDemand)

[0061] Among them, Taskname is the task name provided by the user, Priority is the priority of the task, ResourceDemand is the task resource requirement; hash represents the hash function.

[0062] Verify the startup parameters of the task to be scheduled (including the port number where the task runs, the startup command) to ensure the integrity and legality of the task parameters. The verification method is that all parameter fields are non-empty and conform to the predefined data type format.

[0063] In S1, the constructed three-dimensional resource model is the network-storage-computing joint space model R:

[0064] R = <N(m), S(l), C(k)>

[0065] Among them, N(m) is the network modality, m ∈ [1, M], where M is the number of network modalities; S(l) is the storage level, l ∈ [1, L], where L is the number of storage levels; C(k) is the computing power level, k ∈ [1, K], where K is the number of computing power levels.

[0066] In the embodiment of the present invention, the network modality is divided into 7 levels, that is, m ∈ [1, 7], and the 7 network types are: Ipv4 network, IPv6 network, identity network, large bandwidth network, earth division network, delay-sensitive network, and high-security network. The storage level is represented by the throughput rate. The throughput rate of 1MB / s to 10TB / s is divided into 7 storage levels, that is, l ∈ [1, 7]. The computing power level is represented by the floating-point operations per second FLOPS. The range of 10^3 to 10^21 FLOPS is divided into 7 computing power levels, that is, k ∈ [1, 7]. The constructed three-dimensional resource model of storage-transfer-computing in the embodiment of the present invention is shown in Table 1.

[0067] Table 1 Three-dimensional resources of storage-transfer-computing

[0068]

[0069] Based on the above constructed three-dimensional resource model of storage-transfer-computing integration, map the task to be scheduled to the resource requirements in three dimensions of storage level, network identifier (M modalities), and computing power level (CPU, GPU performance). In the embodiment of the present invention, for example, for an AI inference task:

[0070] Network requirement: Delay-sensitive type (modality 5)

[0071] Storage requirement: 10GB / s throughput rate (storage level 4)

[0072] Computing power requirement: 1P FLOPS (computing power level 5)

[0073] In S2, select the member domain that meets the three-dimensional resource requirements of network modality, storage level, and computing power level of the task to be scheduled from n member domains, including:

[0074] S21. Calculate the remaining amount of three-dimensional resources in each member domain, including computing resources (CPU, GPU), storage resources (memory Mem, hard disk Storage), network resources, etc.

[0075] Calculate the remaining amount of computing resources:

[0076]

[0077] Among them, RemainningTotalCPU j and RemainningTotalCPU j are respectively the total remaining amounts of CPU and GPU in the j-th member domain Domain j , where j ∈ {1, 2, n}; p represents the total number of nodes Node j in the j-th member domain Domain i ; Node j is the i-th node Node i in the j-th member domain Domain i .RemainningCPU and Node i .RemainningGPU are respectively the total remaining amounts of CPU and GPU on the node Node

[0078] Remaining storage resources:

[0079]

[0080] Among them, RemainningTotalMem j and RemainningTotalStorage j respectively represent the total remaining amounts of memory Mem and hard disk Storage in the j-th member domain Domain j ; Node i .RemainningMem and Node i .RemainningStorage are respectively the total remaining amounts of Mem and Storage on the node Node i .

[0081] Remaining network resources:

[0082]

[0083] Among them, RemainningTotalBandwidth j represents the total remaining amount of bandwidth in the j-th member domain Domain j ; Node i .RemainningBandwidth is the total remaining amount of bandwidth on the node Node i .

[0084] S22. Query the maximum available resource per node MaxAvailableResourcePerNode j in each member domain to prevent scheduling failures caused by resource fragmentation.

[0085]

[0086] Among them, MaxAvailableResourcePerNode j is the maximum remaining resource of a single node in the j-th member domain Domain j .

[0087] S23. Filter out inappropriate domains according to the nature and resource requirements of the task.

[0088] If maxAvailableResourcePerNode j >= Task.ResourceDemand, then Domain j is used as a candidate domain point; otherwise, Domain j is excluded. Among them, Task.ResourceDemand is the three-dimensional resource demand to be scheduled determined in S1, that is, the network mode, storage level, and computing power level of the task to be scheduled.

[0089] If all member domains do not meet the requirements at this time, the currently to-be-scheduled task is added to the waiting queue. After other tasks are completed, when there is an idle member domain that meets the requirements, scheduling is performed.

[0090] The candidate domain set screening method in the embodiments of the present invention can ensure that the total remaining resources of each member domain in the candidate domain set and the maximum available resources of a single node thereof both meet the three-dimensional resource requirements of the currently to-be-scheduled task, avoiding the problem of scheduling task failure caused by only the total remaining resources of the member domain meeting the resource requirements of the task while the total remaining resources of all single nodes in the member domain do not meet the requirements.

[0091] Preferably, in S3, a reinforcement learning model is used to evaluate the state and actions of each domain, and the remaining candidate domains are scored considering the priority of the task, resource constraints, and load balancing between domains.

[0092] In the embodiments of the present invention, the state of the reinforcement learning model is defined as the resource usage conditions (CPU, GPU, Memory, Storage, NetworkBandwidth) of each member domain, that is, the remaining three-dimensional resources during the current scheduling period. The action is defined as which domain to select from the candidate domain set to deploy the task. The reward function is designed with the goal of improving the overall (member domain) resource utilization rate and reducing load imbalance.

[0093] Preferably, the reward function designed in the embodiments of the present invention is as follows:

[0094] Reward=γ×(NewUtilization-OldUtilization)-ImBalancePenalty

[0095]

[0096] Among them, γ is the discount factor, which takes values between 0 and 1; ImBalancePenalty is the load imbalance value between member domains, is the utilization of the cth resource of field j, and std is the variance function. NewUtilization represents the resource utilization after scheduling, and OldUtilization represents the resource utilization before scheduling; where resource utilization is equal to the ratio of the total amount of remaining resources to the total amount of resources.

[0097] In the embodiment of the present invention, in each round of learning, the agent calculates the score of each member domain in the candidate domain set according to the designed scoring formula, and selects the member domain with the largest score as the current action, that is, BestDomain=argmax(Score j ). When designing the scoring formula, a domain resource water level evaluation function was introduced to calculate the resource water level value of each candidate domain and use it as the dynamic weight value of each candidate domain to reduce the impact of resource fragmentation and improve global utilization.

[0098] The scoring formula is:

[0099] Score j =α′×AvgUtilization j -β′×ImBalance j

[0100]

[0101]

[0102] Among them, Score j represents the score of the jth member domain in the candidate domain set; α′ and β′ are score weights, respectively, which are optimization parameters of reinforcement learning, and their initial values are random values; set C represents three-dimensional resources, |C| is the total number of elements in set C, that is, the total number of three-dimensional resources; in the embodiment of the present invention, c∈C={CPU,GPU,Mem,Storae,Bandwidth}; AvgUtilization j Util represents the average utilization of all resources in the jth member domain in the candidate domain set; c Indicates the utilization of the cth resource; ImBalance jrepresents the average imbalance rate of all resources in the j-th member domain of the candidate domain set; std represents the variance operation; Node i .Remainning c represents the remaining amount of the c-th resource in the i-th node (Node i ) in the j-th member domain of the candidate domain set, Node i .Tatal c represents the total amount of the c-th resource in Nodei; is the water level value of the c-th resource in the j-th member domain of the candidate domain set, and the calculation method is:

[0103]

[0104] where, is the total amount of the c-th resource of the i-th node in the j-th member domain of the candidate domain set, p is the total number of nodes in the j-th member domain of the candidate domain set; is the remaining amount of the c-th resource of the i-th node in the j-th member domain of the candidate domain set; is the maximum remaining amount of the c-th resource per node in the j-th member domain of the candidate domain set; α and β are the resource total weight and the fragmentation penalty term weight respectively. In the embodiments of the present invention, α = 0.6 (resource total weight), β = 0.4 (fragmentation penalty term).

[0105] During the reinforcement learning process, the agent selects the member domain with the highest score as the current action according to the scores of each member domain in the current candidate domain set. After the action is executed, the corresponding reward is calculated, and α′ and β′ are adjusted according to the reward value, and the next iteration is carried out. When the preset number of iterations is reached, the agent converges, and the member domain with the highest score at this iteration number is used as the optimal member domain.

[0106] Specifically, according to the startup parameters of the current task, the task to be scheduled is assigned to the target domain for execution.

[0107] As a further design of the present invention, in S4, after the task to be scheduled is assigned to the target domain for execution, it further includes: updating the three-dimensional resource state (three-dimensional resource remaining amount) of the target domain, and recording the allocation situation and resource usage situation of the executed task for subsequent monitoring and adjustment. The update formula is as follows:

[0108] Domain target .RemainingResource

[0109] =Domain target .RemainingResource - AssignedResources

[0110] Among them, Domain target .RemainingResource represents the remaining amount of three-dimensional resources in the target domain, and AssignedResources represents the actual three-dimensional resource requirements of the currently executed task.

[0111] After completing the three-dimensional resource allocation, it also includes synchronizing the resource status of each member domain, synchronizing the basic information, three-dimensional resource requirements, and health status (operation status and completion status of the task) of the task, and maintaining the global view.

[0112] Monitor the remaining amount of three-dimensional resources in each member domain in real time. After the task is completed, update the reward value (Reward) of the reinforcement learning model to further optimize future scheduling decisions.

[0113] After the task is successfully executed, return the result to the vertical service submitted by the user; in case of failure, re-execute the candidate domain set selection and subsequent related steps based on the basic information and three-dimensional resource requirements of the synchronized task.

[0114] Embodiment 2

[0115] As Figure 2 shown, an embodiment of the present invention provides a cross-domain scheduling system for a multi-modal network based on memory-computation-integration, including:

[0116] A network modality mapping module, configured to map the task to be scheduled to a specific level in the three-dimensional resource model according to the scheduling task request submitted by the vertical service, so as to obtain the three-dimensional resource requirements.

[0117] A domain scheduler, configured to select member domains that meet the three-dimensional resource requirements of the network modality, storage level, and computing power level of the task to be scheduled from n member domains, and form a candidate domain set with the member domains that meet the three-dimensional resource requirements; it is also configured to use the remaining amount of three-dimensional resources of each member domain in the candidate domain set as a state variable, and use which member domain to select from the candidate domain set to deploy the task to be scheduled as an action, and use reinforcement learning to select the optimal member domain from the candidate domain set; among them, the reward function of the reinforcement learning is designed to maximize the resource utilization rate of each member domain and reduce the load imbalance.

[0118] A member domain manager, configured to use the optimal member domain as the target domain, and allocate the task to be scheduled to the target domain for execution, so as to implement the cross-domain scheduling of the multi-modal network based on memory-computation-integration.

[0119] The network modality mapping module is also configured to initialize the scheduling task request submitted by the vertical service to ensure that the system can correctly receive and parse the task request.

[0120] The member domain manager is also configured to calculate based on the domain resource water level evaluation function (the resource water level values of each member domain ), and send the resource water level values of each member domain to the domain scheduler, dynamically adjust the weight values of each candidate domain, and then dynamically calculate the scores of each member domain in the candidate domain set to optimize resource allocation.

[0121] Furthermore, it also includes a domain status monitor, which is used to update the three-dimensional resource usage (remaining three-dimensional resources) of the target domain, monitor the three-dimensional resource usage of each domain and the health status of tasks (running status and completion status of tasks) in real time, provide comprehensive monitoring data, and help operation and maintenance personnel quickly conduct fault troubleshooting and performance optimization. This module also supports automated operation and maintenance, automatically adjusts the scheduling strategy according to real-time monitoring data, and improves the high availability and resource utilization efficiency of the system.

[0122] Furthermore, it also includes a key-value storage module. As a key-value storage component, it stores the metadata in the system, ensures the reliability of the system and the consistency of data, and supports the global view and prediction algorithm in a cross-domain environment. Synchronize the three-dimensional resource status of each member domain through the key-value storage module, synchronize the basic information of tasks, three-dimensional resource requirements, and the health status of tasks (running status and completion status of tasks), and maintain the global view.

[0123] Automatically adjust the scheduling strategy according to the real-time monitoring data of the domain status monitor to improve the high availability and resource utilization efficiency of the system. After the task is completed, update the reward value (Reward) of the reinforcement learning model to further optimize future scheduling decisions.

[0124] Among them, the network modality mapping module, domain scheduler, member domain manager, domain status monitor, key-value storage module, etc. are all set in the control domain. The network modality mapping module is set at the access entrance inside the control domain. The access entrance inside the control domain serves as the scheduling request receiving and distribution center of the system, can communicate with other modules, and coordinate and transfer commands. The control domain is also a logical domain.

[0125] For the specific implementation methods of each module, please refer to the description of the corresponding steps in Embodiment 1, which will not be elaborated here.

[0126] The present invention constructs a basic framework for multi-modal network resource scheduling by integrating the resource characteristics of the three dimensions of network, storage, and computing power. This model takes network modality, storage level, and computing power level as key indicators, and defines multiple levels respectively. It can perceive the heterogeneous resource distribution in the multi-modal network and convert it into a quantifiable and comparable parameter system, providing a solid foundation for the subsequent dynamic scheduling mechanism.

[0127] Furthermore, by tightly integrating storage, transmission, and computing, the present invention introduces an integrated storage-transmission-computation dynamic scheduling mechanism, which can flexibly respond to the dynamic changes of domain resources in a cross-domain environment, ensure balanced resource allocation, and improve resource utilization, system performance, and reliability.

[0128] Based on a 5-cluster with a total of 30 nodes (5 member domains in total, and the total number of nodes in the 5 member domains is 30), the method in the embodiment of the present invention was tested. After verification by the test platform, the relevant test results are shown in Table 2:

[0129] Table 2 Test Results

[0130] Index Before improvement (taking kube - scheduler as the benchmark) After improvement Improvement rate Resource utilization rate 78％ 93.2％ 19.5％ Cross - domain latency 120ms 82ms 31.7％ End - to - end latency 218ms 149ms 31.7％

[0131] As can be seen from Table 2, through the integrated storage-transmission-computation multi-modal network cross-domain scheduling method in the embodiment of the present invention, the resource utilization rate, cross-domain delay, and end-to-end delay have all been greatly improved.

[0132] The present invention provides a novel multi-modal network resource scheduling engine, which can efficiently and real-time allocate heterogeneous resources (including storage, network, and computing), and realize the intelligent orchestration and scheduling of integrated storage-transmission-computation heterogeneous resources across regions, devices, protocols, multi-dimensions, and fine-grained. Through this innovative scheduling mechanism, it can provide integrated storage-transmission-computation services for application programs, break the barriers of traditional network operation mechanisms, make full use of the heterogeneous and uneven resources carried by the infrastructure, and provide personalized high-quality services for vertical industries.

[0133] Embodiment 3

[0134] The embodiment of the present invention provides a multi-modal network cross-domain scheduling device based on integrated storage-transmission-computation, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the multi-modal network cross-domain scheduling method based on integrated storage-transmission-computation in Embodiment 1 above.

[0135] The relevant technical solutions are the same as above and will not be elaborated here.

[0136] Embodiment 4

[0137] The embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the multi-modal network cross-domain scheduling method based on integrated storage-transmission-computation in Embodiment 1 above.

[0138] Specifically, the memory may include high-speed random access memory and may also include non-volatile memory, such as a hard disk, internal memory, plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage devices.

[0139] The related technical solutions are the same as above and will not be elaborated here.

[0140] Embodiment 5

[0141] The embodiment of the present application provides a computer program product, including a computer program. When the computer program runs on a computer, it causes the computer to execute the steps of the multi-modal network cross-domain scheduling method based on memory-computation integration in Embodiment 1 above.

[0142] The related technical solutions are the same as above and will not be elaborated here.

[0143] Those skilled in the art can easily understand that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A cross-domain scheduling method for a multi-modal network based on integrated storage, transfer, and computing, characterized in that, Including: S1. Map the task to be scheduled to a specific level in the three-dimensional resource model according to the scheduling task request submitted by the vertical service, so as to obtain the three-dimensional resource requirements. The three-dimensional resource model includes the network mode, storage level, and computing power level required by the task. S2. Select a member domain that meets the three-dimensional resource requirements from the member domain cluster. The member domains that meet the three-dimensional resource requirements form a candidate domain set. In the same geographical area, a logical domain composed of a resource set with the same network mode, storage level, and computing power level is a member domain. S3. Use the remaining three-dimensional resources of each member domain in the candidate domain set as state variables, and use the action of selecting a member domain from the candidate domain set to deploy the task to be scheduled. Reinforcement learning is used to select the optimal member domain from the candidate domain set. The reward function of the reinforcement learning is to maximize the resource utilization rate of each member domain and reduce the load imbalance. S4. Use the optimal member domain as the target domain, and allocate the task to be scheduled to the target domain for execution, so as to achieve cross-domain scheduling of the storage-computation-integration multi-modal network.

2. The multimodal network cross-domain scheduling method according to claim 1, wherein In S3, the selection method of the action is: Select the member domain with the highest member domain score from the candidate domain set as the action. The calculation method of the member domain score is: Score j = α′ × AvgUtilization j - β′ × ImBalance j Among them, Score j is the score of the j-th member domain in the candidate domain set; α′ and β′ are the score weights respectively, which are the optimization parameters of reinforcement learning; the set C represents three-dimensional resources, and |C| is the total number of three-dimensional resources; AvgUtilization j is the average utilization rate of all three-dimensional resources in the j-th member domain; is the utilization rate of the c-th resource in the j-th member domain; ImBalance j is the average imbalance rate of all node three-dimensional resources in the j-th member domain, Node i .Remainning c is the remaining amount of the c-th resource of the i-th node in the j-th member domain, Node i .Tatal c is the total amount of the c-th resource of the i-th node, and std represents the variance calculation operation; is the water level value of the c-th resource in the j-th member domain, and the calculation method is: wherein, is the total amount of the c-th resource of the i-th node in the j-th member domain, and p is the total number of nodes in the j-th member domain; is the remaining amount of the c-th resource of the i-th node; is the maximum remaining amount of the c-th resource in a single node of the j-th member domain; α and β are the total amount of resource weight and the fragmentation penalty term weight respectively.

3. The multimodal network cross-domain scheduling method according to claim 2, wherein The reward function is: Reward = γ × (NewUtilization - OldUtilization) - ImBalancePenalty Where γ is the discount factor, which takes a value between 0 and 1; ImBalancePenalty is the load imbalance value between member domains; NewUtilization is the resource utilization rate after scheduling, and OldUtilization is the resource utilization rate before scheduling.

4. The multimodal network cross-domain scheduling method according to claim 3, wherein, In S2, selecting a member domain that meets the three-dimensional resource requirements from the member domain cluster includes: Calculate the remaining three-dimensional resources of each member domain in the member domain cluster. Query the maximum remaining resources of a single node in each member domain. The member domain whose remaining three-dimensional resources and the maximum remaining resources of a single node both meet the three-dimensional resource requirements is the member domain that meets the three-dimensional resource requirements.

5. The multimodal network cross-domain scheduling method according to claim 4, wherein, In S4, after allocating the task to be scheduled to the target domain for execution, it further includes: Update the remaining three-dimensional resources of the target domain, and monitor the remaining three-dimensional resources of each member domain in real time. The updated remaining three-dimensional resources of the target domain are equal to the remaining three-dimensional resources of the target domain before the task execution minus the three-dimensional resources actually used by the task. It further includes: synchronize the remaining three-dimensional resources of each member domain, synchronize the three-dimensional resource requirements of the task, the running status, and the completion status of the task.

6. A cross-domain scheduling system for a multi-modal network based on integrated memory, transfer, and computing, characterized in that Including: A network mode mapping module, which is used to map the task to be scheduled to a specific level in the three-dimensional resource model according to the scheduling task request submitted by the vertical service, so as to obtain the three-dimensional resource requirements. The three-dimensional resource model includes the network mode, storage level, and computing power level required by the task. A domain scheduler for selecting a member domain that meets the three-dimensional resource requirements from a member domain cluster; wherein, the member domains that meet the three-dimensional resource requirements form a candidate domain set; within the same geographical area, a logical domain formed by a resource set with the same network modality, storage level, and computing power level is a member domain; The domain scheduler is further configured to use the remaining three-dimensional resources of each member domain in the candidate domain set as state variables, use selecting a member domain from the candidate domain set to deploy a task to be scheduled as an action, and use reinforcement learning to select an optimal member domain from the candidate domain set; wherein, the reward function of the reinforcement learning is to maximize the resource utilization rate of each member domain and reduce the load imbalance; A member domain manager for using the optimal member domain as a target domain and allocating the task to be scheduled to the target domain for execution, thereby implementing cross-domain scheduling of a multi-modal network with integrated storage, transfer, and computing.

7. The multi-modal network cross-domain scheduling system according to claim 6, wherein It further includes a domain status monitor for updating the remaining three-dimensional resources of the target domain and real-time monitoring the remaining three-dimensional resources of each member domain; wherein, the updated remaining three-dimensional resources of the target domain are equal to the remaining three-dimensional resources of the target domain before the task execution minus the three-dimensional resources actually used by the task; It further includes a key-value storage module for synchronizing the remaining three-dimensional resources of each member domain, synchronizing the three-dimensional resource requirements of the task, the running status, and the completion status of the task.

8. A cross-domain scheduling device for a multimodal network based on integrated memory, transfer, and computing, characterized in that, It includes a computer-readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the multi-modal network cross-domain scheduling method according to any one of claims 1-5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the multi-modal network cross-domain scheduling method according to any one of claims 1-5.

10. A computer program product, characterized in that, It includes a computer program that, when running on a computer, causes the computer to execute the multi-modal network cross-domain scheduling method according to any one of claims 1-5.

Images