Search

CN-122001953-A - Modal resource scheduling method and system for multi-modal network environment

CN122001953ACN 122001953 ACN122001953 ACN 122001953ACN-122001953-A

Abstract

The invention relates to a modal resource scheduling method and system for a multi-modal network environment, belonging to the technical field of network architecture and computing network scheduling. The method comprises five steps of mode and service modeling, mode sensing resource abstraction, mode selection and matching, resource scheduling allocation, operation optimization and feedback, wherein the mode and service modeling is used for constructing a mode description model and a service demand model based on mode characteristics and service demands, the mode sensing resource abstraction is used for carrying out unified abstraction on resources such as heterogeneous calculation, storage and network, the mode selection and resource scheduling matching is used for defining a mode adaptation function, a mode candidate set is obtained, then scheduling objective function solution is carried out based on the candidate set, an optimal scheduling scheme is obtained, and the operation optimization and feedback are used for monitoring the model operation state in real time and optimizing the mode scheduling method of the next round.

Inventors

  • GE JING
  • KONG LIFEI
  • CHEN JING
  • WANG DI
  • WANG XINMEI
  • PANG JIARUI

Assignees

  • 山东省计算中心（国家超级计算济南中心）

Dates

Publication Date
20260508
Application Date
20260408

Claims (10)

  1. 1. A modal resource scheduling method facing to a multi-modal network environment is characterized by comprising the following steps: (1) Modeling the modes and the service, and constructing a mode description model and a service demand model based on the characteristics of the modes and the service demand; (2) The method comprises the steps of mode sensing resource abstraction, taking a mode resource unit as an abstract base unit of scheduling resources, fusing bottom heterogeneous software and hardware resources by the mode resource unit, and uniformly abstracting the bottom heterogeneous software and hardware resources into schedulable resource units by combining mode operation capability; (3) The mode selection and matching are carried out, a mode adaptation degree calculation function is defined, the matching degree of the service requirement and the mode capability is represented, if the mode capability is smaller than the service requirement, the mode is judged to be not adapted, and a mode candidate set is obtained; (4) Resource scheduling and distribution, and scheduling objective function solving based on the modal candidate set to obtain an optimal scheduling scheme; (5) And (3) operating optimization and feedback, monitoring the operating state of the model in real time, and optimizing the mode scheduling method of the next round.
  2. 2. The method for scheduling Modal resources in a multi-Modal network environment as claimed in claim 1, wherein in step (1), a Modal description model is used for standardizing the core attribute describing each network mode, including but not limited to protocol type, delay characteristics, network topology, control mode and resource demand characteristics, and the Modal description model is described by using a structured Modal tuple Modal; ; Wherein P represents a protocol type, D represents a delay characteristic, T represents a network topology, W represents a control mode, Representing resource demand characteristics; The Service demand model is used for carrying out standard modeling on performance demands of a Service side, including but not limited to bandwidth demands, maximum tolerance delay, security level, protocol preference and resource demands, and each Service needs to complete Service demand modeling before submitting a scheduling request, and is used as one of inputs of a scheduling algorithm and is described by using a structured modal tuple Service; ; Wherein S w is a bandwidth requirement, which represents a minimum downlink/uplink bandwidth required by a service, S d is a maximum tolerance delay, which represents an allowed time window from scheduling deployment to completion of execution of the service, S s is a security level, which represents a security protection requirement defined based on service importance, S p is a protocol preference, which is used for describing a selection preference of the service to a specific network protocol or mode type, and S r is a resource requirement, which represents a specific requirement of the service to basic resources, including a minimum acceptable CPU core number, memory capacity and required link bandwidth.
  3. 3. The method for scheduling modal resources in a multi-modal network environment as set forth in claim 2, wherein in step (2), the bottom layer of modal resource units aggregates heterogeneous computing resources, storage resources and network resources upward through a mapping mechanism to provide physical support, and the middle layer core encapsulates the resources into structured modal resource unit triples The top layer supports parallel deployment and isolation operation of a multi-mode system on the same platform through capacity binding and instantiation; Each mode resource unit is bound with one or more supported network modes, the type of the mode which can be borne, the resource specification and the real-time running state are recorded, and the scheduling center monitors and updates the mode resource unit resource pool in real time, so that the availability and the state accuracy in the resource allocation process are ensured.
  4. 4. The method for scheduling modal resources in a multi-modal network environment as set forth in claim 3, wherein in step (3), specific: (31) Constructing a modal fitness calculation function; In order to realize accurate matching between the Service and the network modes, a network mode more suitable for Service operation is selected, and a mode adaptation degree calculation function F (Modal, service) is designed for measuring the matching degree between the Service requirement and a certain candidate mode, wherein the function is as follows: ; Wherein P M represents the modal protocol type, S p is protocol preference, D M is modal delay capability, S d is maximum tolerant delay, and weighting factor Dynamically adjusting according to the service priority; Sim (P M ,S P ) represents the matching degree between the modal protocol type and the business preference, and a discrete scoring mechanism is adopted, namely the completed matching is 1.0, no compatibility is 0, and partial compatibility is set to be a numerical value between 0 and 1; Fit (D M ,S d ) measures whether the modal delay capability meets the service requirement, and the calculation mode is as follows: ; The function ensures that the score is higher when the modal delay is lower than the service requirement, and the score is immediately zero once the tolerance value is exceeded, so that the inapplicable mode is eliminated; Res (model, service) measures whether resources (calculation, forwarding, memory, etc.) required by the mode running meet business target resources, and the calculation process is as follows: ; Wherein, res (Modal, service) represents the matching degree score of the mode and the business requirement at the resource level; refers to the amount of available resources to which the modal resource unit is currently bound, Representing the number of CPU cores remained and allocable in the current modal resource unit node; representing the residual allocatable memory capacity of the target modal resource unit at the hardware level currently; Refers to the minimum resource requirements declared by the traffic demand modeling phase to be scheduled, Representing the minimum CPU calculation power required by normal operation of the service; representing the minimum memory guarantee required by normal operation of the business to be scheduled, which is declared in the modeling stage; (32) A resource feasibility determination and weighting factor mechanism; The Res (Service) formula plays a role in the scheduling decision from the underlying resource feasibility to the high-level policy impact, and is specifically expressed in the following three dimensions: If the Res (Modal) result of the mode is smaller than 1.0, the hard constraint judgment means that the physical resource layer has infeasibility, and forced deployment can lead to Service performance reduction or failure; The feasibility guarantee is that the selected mode not only adapts to the service on the characteristics of a logic protocol and a time delay, but also has actual bearing capacity on the physical resource level; The Res (Modal, service) score is used as a core component of the overall function of the fitness, and is formed by a weighting factor Performing quantization control, and directly determining the priority arrangement of a final scheduling scheme; The core regulatory mechanism for the weighting factors is as follows: (311) Definition and function weighting factors , And Weight distribution for controlling three dimensions of protocol matching degree, delay satisfaction degree and resource feasibility, and meeting normalization constraint in multi-target scheduling scene Ensuring that the score has cross-business comparability; (312) Dynamic regulation mechanism, identifying the characteristics of service demand model, regulating, real-time sensing scene and setting Implementing time delay up strategy, reloading calculation scene, and raising height The general service adopts balance configuration weight to search global optimal balance; (313) The structured guiding is that priority guarantees are provided for different-level services, the key tasks are ensured to have certainty in the subsequent Top-k mode candidate set screening, the convergence direction of a scheduling objective function is led, and the optimal mode and resource combination which best accords with expected service experience are screened out; (33) Screening a modal candidate set; Through global evaluation and secondary filtering of dynamic operation indexes, the output Top-k mode candidate set is ensured to have both logic suitability and operation feasibility.
  5. 5. The method for scheduling modal resources in a multi-modal network environment as set forth in claim 4, wherein in step (33), the specific flow is as follows: (331) The global sorting and Top-k mode candidate set selection are carried out, wherein n is the total number of available candidate network modes, descending order is carried out according to F (Modal i , service), i is any numerical value in 1-n, a scale parameter k is preset for reducing the dispatch search space and improving the convergence rate of a subsequent genetic algorithm, and the mode of k bits before sorting is intercepted to form an initial candidate set; (332) And detecting the mode operation state, namely calling real-time state triplet information of the mode resource unit registered in the dispatching center, and checking the current state value of the mode resource unit supported by each mode in the initial candidate set: Health examination, namely eliminating a mode scheme with a fault or unavailable state; The availability judgment is that if the mode is currently in a queuing or distribution state, judging whether the mode exceeds the period of time according to the maximum tolerant delay S d in the service demand model; (333) And filtering the real-time load information, namely combining the resource abstract data perceived by the mode to dynamically calculate the residual capacity of the mode resource unit occupied by the mode, wherein a filtering mechanism comprises the following steps: ① The hard load filtering, namely marking the instantaneous resource load rate of the modal resource unit corresponding to a certain mode as suboptimal or eliminating if the instantaneous resource load rate exceeds a preset safety threshold value so as to prevent service fluctuation caused by resource mismatch; ② The trend prediction filtering refers to historical data in a mode operation scheduling log, and if a mode is predicted to have a burst service flood peak in a future period, the screening weight of the mode operation scheduling log is reduced; (334) And outputting and linking the candidate set, namely, performing adaptation degree calculation and secondary filtering on the state and the load to finally generate a Top-k mode candidate set containing the optimal and suboptimal schemes.
  6. 6. The method for scheduling modal resources in a multi-modal network environment as set forth in claim 5, wherein in step (4), specific: (41) A mode-resource joint scheduling decision algorithm; Comprehensively considering the real-time availability of the mode adaptation degree and the physical resources, executing the deep joint decision, wherein the decision process is not simple resource allocation, but is an optimization process of carrying out strong coupling on the network mode and the bottom layer resources: The multidimensional screening mechanism comprises a scheduling center acquiring a Modal candidate set, firstly performing descending arrangement according to F (Modal i ) values, and simultaneously calling state triplets of Modal resource units in real time Automatically eliminating the option in which the resource allowance is lower than the minimum service requirement or the state is fault/allocation; the construction of the scheduling objective function, namely, the merits of the scheme are measured by constructing a joint scheduling objective function F, and the global optimization of the computing network resource is realized: ; In order to ensure the convergence of the multi-objective optimization problem and the comparability of each dimension index, the variable definition and constraint conditions in the above formula are as follows: in the multi-objective optimized scheduling process, , , The weight coefficients respectively defined as the resource utilization rate, the scheduling delay and the modal adaptation degree are used for reflecting policy bias of the scheduling system under different service scenes, and in the calculation process, the normalization constraint conditions are strictly satisfied Simultaneously, identifying the characteristics of the business demand model and dynamically adjusting the weight; The resource utilization refers to an evaluation value of the balance degree of the overall resources of the physical platform after the implementation of the allocation scheme, and the balance of the allocation scheme is effectively evaluated by maximizing the index, so that single-point overload is avoided, the overall utilization rate of the resources is improved; (42) The method comprises the steps of realizing an algorithm, solving a high-dimensional scheduling problem by adopting an improved multi-objective NSGA-II algorithm, firstly receiving a Top-k mode candidate set and carrying out prefiltering by combining a mode resource unit real-time state, then entering a core solving link of the improved multi-objective NSGA-II algorithm, initializing a population by mixed coding, carrying out fitness evaluation by taking a target function comprising resource utilization rate, delay and joint scheduling, screening a Pareto front solution set by utilizing non-dominant sequencing, and carrying out evolution by a heuristic crossover and mutation operator implementation scheme, and finally outputting an optimal scheduling decision D comprising an optimal mode, a target MRU and configuration parameters when a termination condition is met; (43) The proposal is finally output and recorded; Selecting optimal scheduling decision according to algorithm solving result ; Wherein, the Is the optimal mode selection, namely, after multidimensional adaptation degree evaluation, the selected network mode which is most in line with the preference and time delay requirement of the current service protocol, MRU j is the target resource unit, and the selected specific bottom physical resource unit carrying the service mode, and the bottom physical resource unit not only has sufficient available resources at the current moment, but also is connected with the selected mode on physical topology or logic connection The configuration parameters refer to specific execution parameter sets when a scheduling scheme is issued, including but not limited to CPU logic core numbers and proportions, memory address space size, link bandwidth upper limit and buffer priority allocated for service examples; In the decision execution stage, the configuration issuing is directly carried out on the target physical node according to the content of the optimal scheduling decision, the instantiation loading of the modal container or sandbox environment is completed, and the complete decision chain is synchronously written into a modal operation scheduling log for the subsequent audit and closed loop optimization.
  7. 7. The method for modal resource scheduling in a multi-modal network environment as claimed in claim 6 wherein in step (42), the algorithm specific flow is as follows: (421) The coding mode adopts a mixed coding scheme, namely a mode selection gene, a mode resource unit numbering gene and a resource weight distribution vector; (422) And (3) evaluating the fitness, namely substituting each individual in the population into the joint scheduling objective function to perform score calculation, layering the individuals through a non-dominant ranking algorithm, and reserving an optimal solution set of pareto fronts, wherein the specific ranking and screening logic is as follows: firstly, the algorithm performs layering by comparing the dominance relation of population individuals, wherein if an individual A is not inferior to an individual B on all scheduling targets and at least one index is superior to B, the individual A is judged to be dominated by B; Secondly, the algorithm firstly identifies an individual set which is not dominated by any other individual in the population, marks the individual set as a first level, namely, a pareto front solution set; finally, in the evolutionary selection process, individuals with the front hierarchy are preferentially reserved, and a crowding degree comparison mechanism is introduced into the same hierarchy to ensure the distribution diversity of solution sets in the search space; (423) Heuristic operator design: Exchanging mode resource unit allocation strategies among schemes of the same mode, and searching for a better physical bearing point; And the mutation operation is that mutation operators are introduced to simulate a mode dynamic switching or task cross-node migration strategy, and whether a mode resource unit with higher matching degree exists in a large-scale search space is explored by randomly changing gene positions, so that a local optimal solution is effectively jumped out.
  8. 8. The method for scheduling modal resources facing to multi-modal network environment as set forth in claim 7, wherein in step (5), real-time indexes are collected through a multi-dimensional monitoring layer to provide data support, when QoS violations are detected, an anomaly detection layer is utilized to automatically trigger modal reselection, resource migration or instance copy compensation strategies, meanwhile, a strategy optimization layer feeds scheduling effects back to a decision system to realize model dynamic update and support modal hot plug adaptation, finally, dynamic risk avoidance and system self-evolution are realized, and efficient and stable operation of heterogeneous modes in a changeable environment is ensured, and the method comprises the following specific implementation steps: (51) Providing a modal instantiation deployment and status interface; according to optimal scheduling decisions Specific deployment operations are performed: (511) The environment initialization, namely starting the corresponding modal container, virtual machine or hardware sandbox environment on a target resource unit MRU j to finish loading and initializing a network protocol stack; (512) Dynamic adjustment of resources, namely supporting dynamic adjustment operation on allocated resources in the running process, and performing thermal migration, current limiting or capacity expansion so as to cope with instantaneous flow fluctuation; (513) The operation interface exposure, namely providing a standardized operation state interface, and supporting the operation health degree, the resource occupancy rate and the service processing state of a real-time query mode; (52) Monitoring multi-dimensional mode operation; the monitoring module is used for collecting indexes in a service flow mode at high frequency, providing data support for anomaly detection, and monitoring the indexes comprises the following steps: the service side index is real-time service flow intensity, end-to-end time delay, time delay jitter and packet loss rate; the mode resource side index is the real-time load change, CPU utilization rate, memory occupation and cache hit rate of the mode on the mode resource unit; (53) Anomaly detection and QoS default triggering mechanism; establishing a performance evaluation system based on a service demand model, and if the service performance is monitored to be reduced or QoS violations are generated, automatically triggering the following compensation and rescheduling strategies: (531) The mode reselection, namely triggering a dispatching center to execute the fitness calculation function again and switching to a better mode if the current network mode cannot meet the service preference; (532) Performing migration scheduling of the modal resource units, namely executing dynamic hot migration of the modal instance if the load of the current physical unit is too high or hardware fault early warning occurs; (533) The modal service instance copy, which is to create service copies on a plurality of physical units aiming at high-load service, and to improve the overall processing capacity through transverse expansion; (54) Policy closed loop evolution and system optimization.
  9. 9. The method for modal resource scheduling in a multi-modal network environment as claimed in claim 8, wherein in step (54), to implement self-evolution, a policy closed-loop optimization mechanism is introduced, and the operation actual measurement data is reversely acted on the initial modeling stage: (541) The feedback learning mechanism takes the actual scheduling effect of each round as feedback data and dynamically adjusts the priority weight of the mode candidate set screening in the next round of scheduling; (542) Dynamically updating the evaluation model, namely dynamically correcting the model parameters of the Modal description constructed in the step (1) according to the actually measured operation data periodically or driven by an abnormal trigger mechanism, and specifically, covering or weighting and updating the initial static template parameters in the Modal tuple by utilizing the dynamic indexes acquired by the monitoring module; (543) And supporting the mode hot plug adaptation, namely supporting the dynamic loading or unloading of a novel network mode under the condition of not stopping operation and automatically incorporating the novel network mode into a resource abstraction and scheduling system.
  10. 10. A modal resource scheduling system for a multi-modal network environment, comprising: the construction module is used for constructing a mode description model and a service demand model based on the characteristics of the modes and the service demands; The system comprises a modal sensing resource abstraction module, a scheduling resource abstraction module and a scheduling resource management module, wherein the modal sensing resource abstraction module is used for taking a modal resource unit as an abstraction base unit of the scheduling resource, the modal resource unit fuses bottom heterogeneous software and hardware resources, and the bottom heterogeneous software and hardware resources are unified and abstracted into schedulable resource units by combining modal operation capability; The mode selection and matching module is used for defining a mode adaptation degree calculation function, representing the matching degree of the service requirement and the mode capability, judging that the service requirement is not suitable if the mode capability is smaller than the service requirement, and obtaining a mode candidate set; the resource scheduling distribution module is used for carrying out scheduling objective function solving based on the modal candidate set to obtain an optimal scheduling scheme; And the operation optimization and feedback module is used for monitoring the operation state of the model in real time and optimizing the mode scheduling method of the next round.

Description

Modal resource scheduling method and system for multi-modal network environment Technical Field The invention relates to a modal resource scheduling method and system for a multi-modal network environment, belonging to the technical field of network architecture and computing network scheduling. Background With the continuous development of novel network technologies such as 6G, information Center Network (ICN), network slicing, time Sensitive Network (TSN) and the like, a single network system taking TCP/IP as a core is difficult to meet diversified industry demands, future network systems show a trend of multi-mode coexistence and heterogeneous collaboration, and different network modes have different protocol structures, routing modes, quality of service (QoS) requirements and resource operation characteristics. The existing dispatching system mainly serves a traditional IP network or a single network architecture, focuses more on unified dispatching and management of computing resources, storage resources and network resources, lacks a conceptual recognition and dispatching mechanism of network modes, and is difficult to support the following requirements that the resource characteristics among different modes are large in difference and cannot be uniformly dispatched, the modes and services are not provided with clear adaptation mechanisms, resource mismatching is easy to occur, dynamic perception of mode operation states and resource consumption is lacking, and dispatching efficiency cannot be optimized. In the actual verification and deployment process, multi-mode parallel deployment, hot plug loading, logic isolation and behavior verification cannot be realized, so that the rapid integration and engineering verification capability of a novel network mode are restricted. Therefore, a scheduling method capable of identifying and matching applicable network modes according to service characteristics and reasonably distributing resources such as calculation, forwarding and storage is needed to support efficient operation in a multi-mode network environment. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a modal resource scheduling method and system for a multi-modal network environment, which can perform modal selection according to service characteristics, perform resource matching according to the modal characteristics, realize scheduling deployment and isolation operation of the modalities and improve the resource utilization rate and service suitability. The technical scheme of the invention is as follows: a modal resource scheduling method facing to a multi-modal network environment comprises the following steps: (1) Modeling the modes and the service, and constructing a mode description model and a service demand model based on the characteristics of the modes and the service demand; (2) The method comprises the steps of mode sensing resource abstraction, taking a mode resource unit (Modal Resource Unit, MRU) as an abstract base unit of scheduling resources, fusing underlying heterogeneous software and hardware resources (computing resources, storage resources and forwarding resources) by the mode resource unit, and uniformly abstracting the underlying heterogeneous software and hardware resources into schedulable resource units by combining mode operation capability; (3) The mode selection and matching are carried out, a mode adaptation degree calculation function is defined, the matching degree of the service requirement and the mode capability is represented, if the mode capability is smaller than the service requirement, the mode is judged to be not adapted, and a mode candidate set is obtained; (4) Resource scheduling and distribution, and scheduling objective function solving based on the modal candidate set to obtain an optimal scheduling scheme; (5) And (3) operating optimization and feedback, monitoring the operating state of the model in real time, and optimizing the mode scheduling method of the next round. Preferably, in step (1) according to the present invention, the mode description model is used for standardizing the core attribute describing each network mode, including but not limited to protocol type (such as IPv6, time sensitive network TSN, content centric network ICN, etc.), delay characteristic (real-time or tolerant), network topology (centralized, distributed or slice structure), control mode (centralized control or distributed collaboration) and resource requirement characteristic (required CPU, memory, cache, etc.), and the mode description model is described by using the structured mode tuple modular for calculating the matching degree later; Wherein P represents a protocol type, D represents a delay characteristic (e.g., requirement < = 1ms is real-time requirement, > = 50ms is tolerance), T represents a network topology, W represents a control mode, Representing resource demand characteristics; The Service demand model is used f