Multi-tenant resource dynamic scheduling system

Abstract

The invention discloses a multi-tenant resource dynamic scheduling system, and relates to the technical field of cloud computing platform multi-tenant resource management, and the system comprises a micro resource slicing engine which is used for abstracting physical resources into physical resource units which can be dynamically combined, and disassembling resource requests on a hardware virtualization layer according to service thread granularity. According to the multi-tenant resource dynamic scheduling system, through refined construction of a physical resource unit topology network and a non-critical path triple judgment mechanism, the risk of damage to the service continuity of low-priority tenants is systematically avoided while a high-priority tenant service level protocol is guaranteed. A demand distribution diagram-driven credit line dynamic calculation and countdown releaser binding strategy is adopted to ensure that a resource scheduling behavior is strictly limited by a future demand peak value of a low-priority tenant, so that deterministic management of a borrowing and returning resource space-time relationship is realized, and the problem of excessive resource invasion caused by extensive preemption in a traditional technology is effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cloud computing platform multi-tenant resource management, in particular to a multi-tenant resource dynamic scheduling system. BACKGROUND

[0002] In a multi-tenant cloud environment, especially in scenarios such as customs declaration enterprises with significant business traffic fluctuations, the dynamic scheduling system needs to allocate computing resources according to the real-time traffic demand of tenants. The existing technology usually triggers resource adjustment based on real-time load or short-term historical data. When a high-priority tenant encounters a sudden traffic, the system tends to quickly seize idle resources or resources allocated to low-priority tenants to meet the key service level agreement. However, this seizure mechanism has serious defects: first, the resource recovery granularity is coarse, often forcing low-priority tenants to lose resources in whole machines or large containers, ignoring their potential business continuity needs; second, the seizure decision only responds to the current pressure of high-priority tenants, without combining traffic pattern prediction to assess the future resource demand peak of low-priority tenants, resulting in excessive occupation of the buffer resources reserved by the latter; third, there is a lack of fine-grained constraints on the degree and duration of resource occupation, and low-priority tenants cannot predict the recovery time after losing resources, nor can they maintain the basic service quality.

[0003] The above problems cause a chain of adverse effects: after the temporary burst demand of high-priority tenants ends, the seized resources often cannot be released or rescheduled in time, causing resource fragmentation and idling; and the service availability of low-priority tenants fluctuates unpredictably due to the disordered deprivation of resources, and even frequently violates the basic service commitment. Ultimately, the system has managed to guarantee the service level agreement of high-priority tenants, but at the expense of overall resource utilization efficiency and low-priority tenant stability, creating an irreconcilable contradiction between multi-tenant fairness and system efficiency. Therefore, there is an urgent need to solve the problem of excessive occupation of low-priority tenant resources and the decline of overall system efficiency caused by extensive resource seizure in the high-priority tenant burst traffic scenario. SUMMARY

[0004] To achieve the above purpose, the present application is implemented by the following technical scheme: a multi-tenant resource dynamic scheduling system, comprising in sequence: a micro-resource slicing engine for abstracting physical resources into dynamically combinable physical resource units, and decomposing resource requests by service thread granularity at the hardware virtualization layer; a double-flow prediction-compensation model connected with the micro-resource slicing engine data, for generating resource seizure credit limits and virtual resource vouchers; a two-level verification arbitrator connected with the double-flow prediction-compensation model control, including a chaos stress test module and a dynamic credit wall module.

[0005] Preferably, the micro-resource slicing engine executes: Identify the criticality of business threads and construct a cross-tenant physical resource unit topology network; When responding to sudden traffic surges from high-priority tenants, physical resource units of non-critical paths are extracted from the topology of low-priority tenants. The determination of a non-critical path includes satisfying the following conditions: a) The service level agreement of the business thread belongs to the preset category; b) The utilization rate of physical resource units is continuously lower than the set threshold; c) There are no strongly dependent related units in the topology.

[0006] Preferably, the dual-stream prediction-compensation model includes: Dynamic Flow Lens Model: Integrating the characteristics of customs declaration cycle, commodity type association, and policy event, it outputs the probability distribution map of resource demand for emergency time windows and buffer time windows through a spatiotemporal convolutional network; Resource lending compensation model: Credit limits are generated based on peak demand within a buffer time window, historical utilization rate of the extracted unit, and business level.

[0007] Preferably, the resource lending compensation model is implemented as follows: Issue virtual resource vouchers that can be exchanged for actual resources to the party whose resources have been seized; Bind a countdown release to the extracted physical resource unit; The release time of the countdown releaser is the peak time of the probability distribution map of resource demand in the buffer time window of low-priority tenants minus a preset safety threshold.

[0008] Preferably, in the two-level verification arbitrator: the chaotic stress test module injects random traffic pulses into the sandbox environment and uses the Monte Carlo algorithm to verify the service level agreement compliance rate; the dynamic credit wall module compares the deviation between the resource occupancy rate and the predicted occupancy rate in real time and triggers the circuit breaker mechanism.

[0009] Preferably, the circuit breaker mechanism includes: freezing new preemption requests from high-priority tenants when the deviation value exceeds a dynamic threshold; initiating a reserve resource pool to take over the preempted unit; the dynamic threshold is calibrated through a reinforcement learning model, and the training data includes historical arbitration records and virtual voucher redemption records.

[0010] Preferably, the construction of the physical resource unit topology network includes: intercepting CPU instruction-level requests through a hardware virtualization layer, marking the data processing path dependencies of business threads, and establishing a topology connection graph with memory pages as the smallest unit.

[0011] Preferably, the virtual resource certificate includes two types of attributes: Resource exchange attribute: Exchange CPU or memory resources according to a set ratio; Time-limited attribute: Valid until the end of the buffer time window.

[0012] Preferably, the micro-resource slicing engine, the dual-stream prediction-compensation model, and the two-level verification arbitrator interact via a bus: the micro-resource slicing engine pushes topology change data to the bus, the dual-stream prediction-compensation model obtains policy execution feedback from the bus, and the two-level verification arbitrator sends a circuit breaker control signal via the bus.

[0013] A resource scheduling method for a multi-tenant dynamic resource scheduling system, executed sequentially: The micro-resource slicing engine constructs a topology network of physical resource units; The dual-flow forecasting-compensation model generates a demand distribution map and credit limit for a buffer time window. Extract non-critical path units within the credit limit and bind them to countdown releasers; The two-level verification arbitrator performs chaos testing and bias circuit breaker verification.

[0014] This invention provides a multi-tenant resource dynamic scheduling system. It has the following beneficial effects: This multi-tenant resource dynamic scheduling system, through the refined construction of the physical resource unit topology network and a triple determination mechanism for non-critical paths, systematically avoids the risk of disrupting the business continuity of low-priority tenants while ensuring the service level agreements of high-priority tenants. It employs a demand distribution map-driven dynamic calculation of credit limits and a countdown release binding strategy to ensure that resource scheduling behavior is strictly limited to the future peak demand of low-priority tenants, achieving deterministic management of the spatiotemporal relationship of resource borrowing and returning, and effectively solving the problem of excessive resource encroachment caused by the preemptive and extensive approach of traditional technologies.

[0015] This multi-tenant resource dynamic scheduling system, based on a dual verification mechanism of chaotic stress testing and dynamic credit wall, combined with seamless takeover technology from the reserve resource pool, constructs a self-correcting closed loop for the entire resource scheduling process. This design not only blocks abnormal preemption behavior when failures are predicted, but also continuously optimizes system decisions through a credit feedback loop, improving overall resource utilization efficiency while maintaining the stability of multi-tenant services, and overcoming the shortcomings of fragmented idle resources and unpredictable service quality of low-priority services. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the module interaction of a multi-tenant resource dynamic scheduling system according to the present invention; Figure 2 This is a timing diagram of the control logic for a multi-tenant resource dynamic scheduling system according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 and Figure 2 This invention provides a technical solution: a multi-tenant resource dynamic scheduling system, comprising: connected in sequence: The micro-resource slicing engine is used to abstract physical resources into dynamically combinable physical resource units and decompose resource requests at the business thread level in the hardware virtualization layer. A dual-stream prediction-compensation model, connected to the micro-resource slicing engine data, is used to generate resource encroachment credit limits and virtual resource certificates; A two-level verification arbitrator, connected to a dual-stream prediction-compensation model control, includes a chaotic stress testing module and a dynamic credit wall module.

[0019] It should be further explained that, in the specific implementation process, the multi-tenant resource dynamic scheduling system operates as follows: The micro-resource slicing engine first intercepts CPU instruction-level resource requests at the hardware virtualization layer, breaks down physical resource units according to the data processing path of the customs declaration business thread, marks the service level protocol type and path dependency relationship of the business thread, and constructs a cross-tenant physical resource unit topology network. When a burst of traffic from a high-priority tenant is detected, the engine scans the low-priority tenant topology network. Only when the target physical resource unit simultaneously meets the following conditions—belonging to a preset low-sensitivity service level protocol category, having a continuous utilization rate lower than a set threshold, and having no strongly dependent related units in the topology network—is it identified as a non-critical path unit and extracted.

[0020] The dual-flow forecasting and compensation model is launched simultaneously: The dynamic flow lens model integrates the periodic peak characteristics in customs declaration history, the flow coupling characteristics of related commodity types, and the demand mutation characteristics triggered by policy events. It generates resource demand probability distribution maps for emergency time windows and buffer time windows through a spatiotemporal convolutional network. The resource lending and compensation model calculates the upper limit of the resource encroachment credit limit based on the peak position of the demand distribution map within the buffer time window, the historical utilization fluctuation range of the extracted unit, and the tenant's business level. After resource extraction within the credit limit, a virtual certificate redeemable for actual CPU / memory resources is immediately issued to the encroached party, and a countdown release device is bound to the extracted unit. The release time is locked at a position offset forward by a preset safety time interval from the peak time point of the demand distribution within the buffer time window.

[0021] The two-level verification arbitrator performs dual interception verification before the scheduling decision takes effect: the chaotic stress test module injects simulated random traffic pulses and resource fragments into the sandbox environment, and repeatedly calculates the service level agreement baseline compliance of low-priority tenants in the scenario of missing topology network units based on the Monte Carlo algorithm. Only when the compliance rate exceeds the set threshold is the request allowed; the dynamic credit wall module compares the deviation between the actual resource occupancy rate and the predicted occupancy rate in real time. When the deviation value exceeds the threshold dynamically adjusted by reinforcement learning, the new preemption request of high-priority tenants is immediately frozen, and the reserve resource pool is called to seamlessly take over the occupied units. The takeover signal is fed back to the dual-stream prediction-compensation model through the bus to update the credit limit.

[0022] Micro-resource slicing engine execution: Identify the criticality of business threads and construct a cross-tenant physical resource unit topology network; When responding to sudden traffic surges from high-priority tenants, physical resource units of non-critical paths are extracted from the topology of low-priority tenants. The determination of a non-critical path includes satisfying the following conditions: a) The service level agreement of the business thread belongs to the preset category; b) The utilization rate of physical resource units is continuously lower than the set threshold; c) There are no strongly dependent related units in the topology.

[0023] It should be further explained that, in the specific implementation process, when the micro-resource slicing engine performs resource scheduling, it first intercepts CPU instruction-level requests through the hardware virtualization layer, parses the data processing path of the customs declaration business thread, and classifies and marks the criticality level according to the preset service level protocol tags associated with the thread. During the construction of the physical resource unit topology network, the utilization rate data of each unit for multiple consecutive collection cycles is recorded, and the instruction jump dependency strength between adjacent units is analyzed. When a high-priority tenant's sudden traffic triggers a resource preemption request, the engine executes the following three judgment logics in sequence, including: First item: Service Level Agreement (SLA) filtering: Select only the unit to which the business thread belongs, marked as "Non-real-time Data Processing" or "Batch Background Job"; The second item: Utilization fluctuation detection: Check if the utilization rate of the target unit is consistently lower than the set threshold in the most recent continuous collection period, and if there are no instantaneous spikes; The third step: Dependency verification: Traverse the adjacent units in the topology that have instruction jump associations with this unit, and confirm that there are no strong dependencies, that is, no data processing chains that must be synchronized and scheduled. A cell is considered a non-critical path extractable cell only if all three conditions are met simultaneously. The extraction operation employs atomic instruction replacement technology, dynamically mapping the cell from its original tenant topology to a high-priority tenant environment after decoupling it, while preserving a snapshot of the original environment's data state. If any of the three conditions are not met, the cell is skipped, and the topology scan continues until a qualified cell is found or the search range is exhausted.

[0024] The dual-stream prediction-compensation model includes: Dynamic Flow Lens Model: Integrating the characteristics of customs declaration cycle, commodity type association, and policy event, it outputs the probability distribution map of resource demand for emergency time windows and buffer time windows through a spatiotemporal convolutional network; Resource lending compensation model: Credit limits are generated based on peak demand within a buffer time window, historical utilization rate of the extracted unit, and business level.

[0025] It should be further explained that, in the specific implementation process, when the dual-flow prediction-compensation model is running, the dynamic flow lens model first accesses the historical declaration database of customs declaration business and extracts three feature vectors: periodic peak feature, which analyzes the quarterly flow pattern based on the enterprise's past declaration records; commodity type association feature, which analyzes the concurrent declaration trend of highly coupled category combinations; and policy event feature, which captures the effective time node of tax rate adjustment or customs clearance rule change announcements in the customs policy database.

[0026] The spatiotemporal convolutional network performs hierarchical fusion processing on three types of features: the first-layer convolutional kernel scans periodic patterns along the time axis, the output layer superimposes commodity association weights, and the final layer injects policy event triggering factors to generate a dual-time period resource demand probability distribution map. The emergency operation time period distribution map is used for real-time response scheduling, while the subsequent business time period distribution maps are input into the resource lending and compensation model.

[0027] When the resource lending compensation model calculates the credit limit, it performs multi-factor coupling judgment, including: when there is a single sharp peak in the distribution map of subsequent business periods, the upper limit of the credit limit = peak demand × decay coefficient; when the distribution map is multi-peaked, the upper limit of the credit limit = the highest peak demand × coupling factor; the decay coefficient is negatively correlated with the historical average utilization rate of the extracted unit, and the coupling factor is positively correlated with the strength of the product type association feature.

[0028] The calculation process incorporates a business level weighting rule: for resource units of tenants with high credit ratings, the credit limit is increased by a fixed percentage; for tenants with service level agreement (SLA) default records, the credit limit is reduced. The final generated credit limit value and the peak location information of the distribution map are simultaneously pushed to the micro-resource slicing engine.

[0029] Resource lending compensation model execution: Issue virtual resource vouchers that can be exchanged for actual resources to the party whose resources have been seized; Bind a countdown release to the extracted physical resource unit; The release time of the countdown releaser is the peak time of the low-priority tenant's buffer time window resource demand probability distribution map minus a preset safety threshold. It should be further explained that, in the specific implementation process, when the resource lending compensation model performs resource scheduling compensation, it first locates the peak time of resource demand based on the buffer time window demand distribution map output by the dual-flow prediction model. The release time of the countdown releaser is set to be a fixed safety interval offset forward from this peak time point. This interval value is determined based on the upper limit of resource preheating time in the historical data of customs declaration business.

[0030] After the micro-resource slicing engine completes the extraction of physical resource units, the model simultaneously performs a dual-track compensation operation: issuing virtual resource certificates to the occupied tenant accounts. These certificates contain resource redemption attributes and time-limited attributes. The resource redemption attributes allow tenants to redeem actual CPU or memory resources according to set rules before the end of the buffer time window, while the time-limited attributes limit the redemption window to the end of the buffer time window. At the same time, a countdown release command is injected into the extracted units. When the system clock reaches the set release time point, the command automatically triggers the units to migrate back to their original topology network positions. The migration process carries atomic snapshot data to ensure the consistency of business status.

[0031] If the demand distribution map of the buffer time window changes shape, such as a shift in the peak position or a sudden change in demand value, the model immediately recalculates the release time point and refreshes the countdown command. The refresh operation must be verified by the chaos test of the two-level verification arbitrator.

[0032] In the two-level verification arbitrator: the chaotic stress testing module injects random traffic pulses into the sandbox environment and uses the Monte Carlo algorithm to verify the service level agreement (SLA) compliance rate; the dynamic credit wall module compares the deviation between the resource occupancy rate and the predicted occupancy rate in real time, triggering the circuit breaker mechanism. It should be further noted that in the specific implementation, the two-level verification arbitrator initiates the interception verification process before the scheduling command takes effect. The chaotic stress testing module first completely clones the current physical resource unit topology in the isolated sandbox environment, and then injects two types of disturbance factors: random traffic pulses simulate the uncertainty fluctuations of customs declaration peaks, and artificially fragmented resource distribution simulates storage fragmentation generated during long-term operation. Multiple rounds of stress calculations are performed based on the Monte Carlo algorithm, with each round randomly combining the intensity and location of the disturbance factors, recording the SLA compliance status of low-priority tenants' critical business threads. Only when the SLA compliance rate exceeds a preset threshold in consecutive calculation rounds is the test considered passed and the scheduling command execution permission unlocked.

[0033] The dynamic credit wall module runs in real time during the scheduling and execution phase, continuously collecting actual resource occupancy rate data streams and comparing them with the predicted occupancy rate benchmark curve provided by the dual-stream prediction model using a sliding window. When the deviation value continuously exceeds the dynamic adjustment threshold within the window period, the circuit breaker mechanism is immediately activated: a high-priority tenant preemption request freeze signal is sent to the micro-resource slicing engine, and at the same time, the reserved equivalent computing units in the reserve resource pool are awakened and take over the load of the occupied physical resource units through hot migration technology.

[0034] During the takeover process, the connectivity of the original topology network data paths is maintained. After the takeover is completed, a credit limit adjustment instruction is sent to the resource lending compensation model. The dynamic adjustment of the threshold update is based on the output of the reinforcement learning model. This model periodically analyzes the correlation between deviation values ​​and service level agreement violation rates in historical arbitration events, and optimizes the threshold generation strategy by combining virtual voucher redemption fulfillment rate data.

[0035] The circuit breaker mechanism includes: freezing new preemption requests from high-priority tenants when the deviation value exceeds a dynamic threshold; initiating a backup resource pool to take over the occupied unit; and adjusting the dynamic threshold using a reinforcement learning model, with training data including historical arbitration records and virtual voucher redemption records. It should be further explained that in the specific implementation process, when the circuit breaker mechanism is activated, the dynamic credit wall module first determines the persistence of the deviation value exceeding the dynamic threshold: if the deviation value is higher than the threshold throughout the continuous monitoring period, a high-priority tenant preemption request freeze command is sent to the micro-resource slicing engine, which immediately interrupts all pending resource extraction request queues. Simultaneously, the backup resource pool takeover procedure is initiated, selecting a backup computing unit that matches the configuration specifications of the occupied physical resource unit. Through real-time memory mirroring synchronization technology, the business load is migrated from the occupied unit to the backup unit, maintaining the input / output connectivity of the original data processing path during the migration process.

[0036] Once the migration completion signal is triggered, the occupied unit is released from its binding relationship with a high-priority tenant and reverts to the idle resource pool awaiting allocation. The takeover event log is pushed to the resource lending compensation model in real time, triggering a credit limit recalibration process: based on the actual takeover duration and the historical utilization data of the occupied unit, the subsequent credit limit cap for the high-priority tenant is adjusted downwards. Dynamic threshold updates are periodically executed by the reinforcement learning model. Training data includes historical deviation records, service level agreement violation timestamps, and virtual resource certificate redemption delay records. The model outputs threshold coefficients, which are written to the dynamic credit wall configuration register.

[0037] The construction of the physical resource unit topology network includes: intercepting CPU instruction-level requests through the hardware virtualization layer, marking the data processing path dependencies of business threads, and establishing a topology connection graph with memory pages as the smallest unit. It should be further explained that, in the specific implementation process, the construction of the physical resource unit topology network begins with the hardware virtualization layer's interception of CPU instruction-level requests. When a customs declaration business thread initiates a resource call, the virtualization layer parses the memory page addressing sequence and cross-core communication path in the instruction stream, and marks the dependency graph of the data processing threads. The specific execution of the three-step mapping logic includes the following: Instruction flow tracing: Track the instruction jump frequency of the business thread, mark frequently accessed memory page clusters as strong dependency units, and mark scattered memory pages associated with isolated computation instructions as weak dependency units; Path dependency modeling: When threads access each other across physical computing nodes, weighted virtual data channels are established in the topology network, and the weight values ​​are positively correlated with the data transmission latency. Minimal unit binding: Using operating system memory pages as the basic unit, memory pages sharing the same L3 cache are aggregated into topological connection subnets, and bidirectional reachable paths are established between units within the subnet.

[0038] The completed topology network presents a multi-layered structure, including: the bottom layer consists of memory page-level physical unit nodes; the middle layer is divided into connecting subnets according to the L3 cache area; and the top layer connects cross-node business threads through weighted virtual channels.

[0039] When the micro-resource slicing engine performs resource extraction, it prioritizes memory page units that exist only in the set of weakly dependent units and have not established cross-node channels, ensuring that topology changes do not cause cascading effects.

[0040] Virtual resource credentials contain two types of attributes: Resource exchange attribute: Exchange CPU or memory resources according to a set ratio; Time-limited attribute: Valid until the end of the buffer time window.

[0041] It should be further explained that, in the specific implementation process, virtual resource vouchers generate two types of binding attributes when they are issued: the resource exchange attribute explicitly limits the voucher holder to exchange actual computing resources at a set ratio through the system resource management interface before the end of the buffer time window. When the exchange operation is triggered, physical resources of the same specifications as the extracted unit in the idle resource pool are allocated first. When there are not enough idle resources, they are added to the waiting queue according to the voucher priority; the timeliness attribute forcibly constrains the exchange window deadline to be strictly aligned with the end of the buffer time window. Vouchers that are not used after the timeout will automatically expire and trigger a recycling event.

[0042] The voucher redemption process employs a dual verification mechanism: first, it verifies whether the validity period is within the valid range; then, it dynamically adjusts the resource injection rate based on the current load of the business threads to avoid secondary resource contention caused by concentrated redemptions. If a redemption request enters the allocation queue due to insufficient resources, the system pushes the estimated waiting time to the applicant and simultaneously reserves the corresponding resource quota. Upon voucher expiration or completion of redemption, the resource lending compensation model updates the credit limit balance and releases the relevant resource locks.

[0043] The micro-resource slicing engine, the dual-stream prediction-compensation model, and the two-level verification arbitrator interact via a bus: the micro-resource slicing engine pushes topology change data to the bus, the dual-stream prediction-compensation model obtains policy execution feedback from the bus, and the two-level verification arbitrator sends circuit breaker control signals via the bus. It should be further explained that, in the specific implementation, the bus acts as a coordination hub between modules, establishing three types of signal transmission channels: After completing the physical resource unit topology change, the micro-resource slicing engine immediately pushes a topology update package containing the unit location mapping table and dependency change flags through the bus data channel; the dual-stream prediction-compensation model continuously monitors the bus policy feedback channel, and when it receives a topology update package, it triggers a differential response based on the change flag type: if it is a dependency change, it recalculates the credit limit of the affected tenants and refreshes the countdown releaser parameters; if it is only a location mapping update, it only adjusts the resource pool pointer of the virtual resource certificate.

[0044] The two-level verification arbitrator sends a circuit breaker instruction set to each module via the bus control channel. This instruction set includes a high-priority tenant request freeze code, a reserve resource pool wake-up flag, and hot migration path parameters. When the arbitrator detects that the circuit breaker condition has been lifted, it sends a unfreeze instruction via the same channel to reset the system state. The bus transmission protocol uses a priority slot mechanism, with the circuit breaker control signal having the highest interrupt privileges, ensuring that the scheduling link is immediately blocked when resource encroachment deviations become uncontrollable.

[0045] A resource scheduling method for a multi-tenant dynamic resource scheduling system, executed sequentially: The micro-resource slicing engine constructs a topology network of physical resource units; The dual-flow forecasting-compensation model generates a demand distribution map and credit limit for a buffer time window. Extract non-critical path units within the credit limit and bind them to countdown releasers; The two-level verification arbitrator performs chaos testing and bias circuit breaker verification. It should be further noted that, in the specific implementation, the resource scheduling method executes the fourth-order core operations sequentially, which include the following: During the physical resource unit topology network construction phase, the micro-resource slicing engine intercepts the CPU instruction stream of the customs declaration business thread through the hardware virtualization layer, parses the memory page access mode and cross-core communication path, marks the sets of strongly dependent units and weakly dependent units, aggregates memory pages based on the L3 cache to form a connection subnet, and finally establishes a multi-layer topology network with virtual data channel weights.

[0046] In the dual-time-window forecasting phase, the dual-flow forecasting-compensation model integrates the characteristics of customs declaration cycle, commodity type association, and policy event characteristics, and generates resource demand distribution maps for emergency operation periods and subsequent business periods through a spatiotemporal convolutional network. The resource lending compensation model generates credit limits based on the peak position of the distribution map for subsequent periods, the historical utilization rate of the extracted units, and the business level, and simultaneously issues virtual resource vouchers with time-sensitive attributes.

[0047] During the resource scheduling execution phase, non-critical path units are extracted by scanning the topology network within the credit limit: atomic decoupling is performed only when the target unit simultaneously meets the three conditions of low sensitivity category of service level agreement, continuous utilization rate below the threshold, and no strong dependency association. After decoupling, a countdown releaser is immediately bound, and the release time is set to the position of the peak point of demand distribution in the subsequent period shifted forward by a fixed safety interval.

[0048] In the dual verification phase, the two-level verification arbitrator initiates a chaotic stress test before scheduling: random traffic pulses and fragmented resources are injected into the topology sandbox replica, and the Monte Carlo algorithm is used to verify the service level agreement baseline compliance rate; during scheduling execution, the dynamic credit wall module compares the resource occupancy rate deviation in real time, and triggers a circuit breaker when the continuous monitoring period exceeds the limit, that is: freezing new requests from high-priority tenants and initiating hot migration takeover of the reserve resource pool, and the takeover completion signal triggers credit limit recalibration.

[0049] It should be further explained that, in the specific implementation process, when the multi-tenant resource dynamic scheduling system starts, the micro-resource slicing engine first captures the CPU instruction stream at the hardware virtualization layer, analyzes the access sequence of memory pages by the customs declaration business threads and the data transmission path across computing cores. Based on the instruction jump frequency, it marks strongly dependent memory page clusters and weakly dependent scattered pages, aggregates memory pages sharing the L3 cache into connection subnets, and associates business threads across physical nodes through weighted virtual channels, forming a multi-layer physical resource unit topology network.

[0050] When a high-priority tenant experiences a sudden surge in traffic triggering resource requests, the engine scans the low-priority tenant topology. Extraction is only performed if the target unit simultaneously meets three conditions: the service level agreement (SLP) of the business thread to which the unit belongs belongs to a preset non-real-time category; its utilization rate is consistently below a set threshold during the continuous collection period; and there are no strongly dependent units in the topology. The extraction process employs atomic instruction replacement technology, preserving a snapshot of the original data state and mapping the unit to a high-priority environment.

[0051] The dual-flow forecasting and compensation model operates synchronously: the dynamic flow lens model accesses the historical customs declaration database to extract three types of features: periodic peak patterns, concurrent trends of related commodity combinations, and policy announcement effective dates. A spatiotemporal convolutional network scans periodic features along the time axis, overlays commodity association weights, and injects policy event factors to generate a probability distribution map of resource demand during emergency operation periods and subsequent business periods.

[0052] The resource lending compensation model generates a credit limit based on the peak position of the subsequent time period distribution map, the historical utilization fluctuation range of the extracted unit, and the tenant's business level, and issues a virtual resource certificate to the party whose resources have been misappropriated. This certificate is limited to redeeming actual computing resources according to set rules before the end of the subsequent business period. At the same time, a countdown release device is bound to the extracted unit, and its release time is strictly set to a position shifted forward by a fixed safety interval from the peak time point of the distribution map. If the distribution map shape changes, the model immediately recalculates the release time and triggers a two-level verification arbitrator for review.

[0053] The two-level verification arbitrator performs a chaotic stress test before the scheduling decision takes effect: The topology is cloned into a sandbox environment, traffic pulses simulating random peak values ​​and artificially created resource fragments are injected, and the Monte Carlo algorithm is used to repeatedly calculate the service level agreement (SLA) baseline compliance of low-priority tenants in resource-scarce scenarios. Scheduling is only authorized when the continuous compliance rate exceeds a preset threshold. During scheduling execution, the dynamic credit wall module collects the actual resource occupancy rate data stream in real time and compares it with the predicted baseline curve using a sliding window.

[0054] When the deviation value continuously exceeds the dynamic threshold within a continuous monitoring period, new preemption requests from high-priority tenants are immediately frozen, and the reserve resource pool is activated. A backup unit in the pool with matching specifications takes over the load of the preempted unit through memory mirroring synchronization technology, maintaining the original data processing path. After takeover, the preempted unit is unbound and released to the idle resource pool, while a credit limit recalibration instruction is sent to the resource lending compensation model. The dynamic threshold is periodically updated by a reinforcement learning model, which generates the threshold coefficient based on historical deviation records, service level agreement violation events, and virtual credential redemption delay data.

[0055] The modules interact via a bus: the micro-resource slicing engine pushes topology change packages to the bus data channel, including cell location mapping tables and dependency change flags. When the dual-stream prediction compensation model detects a dependency change flag, it recalculates the credit limit of the affected tenant and refreshes the releaser parameters; if only the location mapping is updated, it adjusts the credential redemption resource pool pointer. The two-level verification arbitrator sends a circuit breaker instruction set through the control channel, including a request freeze code, a reserve pool wake-up flag, and migration path parameters, with the circuit breaker signal having the highest bus interrupt priority.

[0056] The entire resource scheduling process executes the following four stages in sequence: Topology construction stage: establishing a multi-layer physical resource unit network with weighted channels based on instruction flow analysis; Prediction stage: generating a dual-time period demand distribution map by integrating customs declaration business characteristics, and outputting credit limits and virtual vouchers; Scheduling stage: extracting non-critical path units according to three conditions and binding peak-driven countdown releasers; Verification stage: executing scheduling after verifying the compliance rate through chaos testing, and implementing real-time circuit breaking for deviation exceeding limits using a dynamic credit wall.

[0057] It should be further explained that, in the specific implementation process, the three-stage determination of non-critical paths includes the following: Business nature filtering: Only select thread units that are not real-time data processing to avoid affecting the core business of customs declaration and clearance; Utilization stability verification: The target unit is required to have no instantaneous spikes in low utilization during multiple acquisition cycles to eliminate misjudgments caused by brief idle periods; Dependency chain blocking confirmation: Traverse adjacent cells in the topology to ensure that there are no strongly related data processing chains that must be executed synchronously.

[0058] The countdown release time binding principle is as follows: the release time is determined by shifting the predicted peak demand time forward by a fixed safety interval. This interval is set based on the maximum preheating time of historical resources in customs declaration business. When the predicted peak time shifts, the system automatically recalculates the release time and triggers two levels of verification: chaos testing verifies the service level agreement compliance rate at the new time point, and dynamic credit wall monitors the additional resource overhead generated during the recalculation process.

[0059] The seamless takeover mechanism for the reserve resource pool includes the following: The hot migration process employs real-time memory mirroring synchronization technology: locking the data processing state of the compromised unit; incrementally copying the memory page content to the backup unit; and unbinding the original unit after switching the input / output path to the backup unit. Data path connectivity for business threads is maintained throughout the process, and the takeover duration is included as a factor in the credit limit recalibration calculation.

[0060] Virtual vouchers have dual attribute constraints, including redemption and expiration attributes. The redemption attribute provides resources of the same level as the extracted unit, and when idle resources are insufficient, vouchers are placed in an allocation queue with reserved quotas. The expiration attribute's redemption window deadline is strictly aligned with the end time of subsequent business periods; unredeemed vouchers automatically expire. Redemption requests must pass both expiration validity verification and current load rate adaptation checks.

[0061] The dynamic threshold training data is correlated as follows: The reinforcement learning model establishes a three-layer data mapping: the positive correlation between historical deviation values ​​and service level agreement violation rates; the correlation model between virtual voucher redemption delay and resource fragmentation; and the weighting of the impact of circuit breaker event frequency on overall resource utilization. The output threshold coefficient is written to a hardware register to ensure execution priority.

[0062] By employing a refined construction of the physical resource unit topology network and a triple-determination mechanism for non-critical paths, the system systematically avoids the risk of disrupting the business continuity of low-priority tenants while ensuring the service level agreements of high-priority tenants. A demand distribution map-driven dynamic calculation of credit limits and a countdown release binding strategy ensure that resource scheduling is strictly limited to the future peak demand of low-priority tenants, achieving deterministic management of the spatiotemporal relationship between resource borrowing and returning. This effectively solves the problem of excessive resource encroachment caused by the preemptive and extensive approach of traditional technologies.

[0063] Based on a dual verification mechanism of chaotic stress testing and dynamic credit wall, combined with seamless takeover technology from the reserve resource pool, a self-correcting closed loop is constructed for the entire resource scheduling process. This design not only blocks abnormal preemption behavior when failure is predicted, but also continuously optimizes system decisions through a credit feedback loop, improving overall resource utilization efficiency while maintaining the stability of multi-tenant services, and overcoming the defects of fragmented idle resources and unpredictable service quality of low-priority services.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-tenant resource dynamic scheduling system, characterized in that, Including those connected sequentially: The micro-resource slicing engine is used to abstract physical resources into dynamically combinable physical resource units and decompose resource requests at the business thread level in the hardware virtualization layer. A dual-stream prediction-compensation model, connected to the micro-resource slicing engine data, is used to generate resource encroachment credit limits and virtual resource certificates; A two-level verification arbitrator, connected to a dual-stream prediction-compensation model control, includes a chaotic stress testing module and a dynamic credit wall module.

2. The multi-tenant resource dynamic scheduling system according to claim 1, characterized in that: The micro-resource slicing engine executes: Identify the criticality of business threads and construct a cross-tenant physical resource unit topology network; When responding to sudden traffic surges from high-priority tenants, physical resource units of non-critical paths are extracted from the topology of low-priority tenants. The determination of a non-critical path includes satisfying the following conditions: a) The service level agreement of the business thread belongs to the preset category; b) The utilization rate of physical resource units is continuously lower than the set threshold; c) There are no strongly dependent related units in the topology.

3. The multi-tenant resource dynamic scheduling system according to claim 1, characterized in that: The dual-stream prediction-compensation model includes: Dynamic Flow Lens Model: Integrating the characteristics of customs declaration cycle, commodity type association, and policy event, it outputs the probability distribution map of resource demand for emergency time windows and buffer time windows through a spatiotemporal convolutional network; Resource lending compensation model: Credit limits are generated based on peak demand within a buffer time window, historical utilization rate of the extracted unit, and business level.

4. The multi-tenant resource dynamic scheduling system according to claim 3, characterized in that: The resource lending and compensation model is executed as follows: Issue virtual resource vouchers that can be exchanged for actual resources to the party whose resources have been seized; Bind a countdown release to the extracted physical resource unit; The release time of the countdown releaser is the peak time of the probability distribution map of resource demand in the buffer time window of low-priority tenants minus a preset safety threshold.

5. A multi-tenant resource dynamic scheduling system according to claim 1, characterized in that: In the two-level verification arbitrator: the chaotic stress test module injects random traffic pulses into the sandbox environment and uses the Monte Carlo algorithm to verify the service level agreement compliance rate; the dynamic credit wall module compares the deviation between the resource occupancy rate and the predicted occupancy rate in real time and triggers the circuit breaker mechanism.

6. A multi-tenant resource dynamic scheduling system according to claim 5, characterized in that: The circuit breaker mechanism includes: freezing new preemption requests from high-priority tenants when the deviation value exceeds a dynamic threshold; initiating a reserve resource pool to take over the preempted unit; the dynamic threshold is calibrated through a reinforcement learning model, and the training data includes historical arbitration records and virtual voucher redemption records.

7. A multi-tenant resource dynamic scheduling system according to claim 2, characterized in that: The construction of the physical resource unit topology network includes: intercepting CPU instruction-level requests through the hardware virtualization layer, marking the data processing path dependencies of business threads, and establishing a topology connection graph with memory pages as the smallest unit.

8. A multi-tenant resource dynamic scheduling system according to claim 4, characterized in that: The virtual resource certificate contains two types of attributes: Resource exchange attribute: Exchange CPU or memory resources according to a set ratio; Time-limited attribute: Valid until the end of the buffer time window.

9. A multi-tenant resource dynamic scheduling system according to claim 1, characterized in that: The micro-resource slicing engine, the dual-stream prediction-compensation model, and the two-level verification arbitrator interact via a bus: the micro-resource slicing engine pushes topology change data to the bus, the dual-stream prediction-compensation model obtains policy execution feedback from the bus, and the two-level verification arbitrator sends circuit breaker control signals via the bus.

10. A resource scheduling method based on the system described in any one of claims 1-9, characterized in that, Execute in sequence: The micro-resource slicing engine constructs a topology network of physical resource units; The dual-flow forecasting-compensation model generates a demand distribution map and credit limit for a buffer time window. Extract non-critical path units within the credit limit and bind them to countdown releasers; The two-level verification arbitrator performs chaos testing and bias circuit breaker verification.

Images