A dynamic resource scheduling host hosting system based on an edge computing node

Abstract

The application discloses a kind of dynamic resource scheduling host hosting systems based on edge computing node, through the cross physical machine non-perception computing power slip module, idle server is filtered according to the computing power offset amount of chaos prediction output, builds exclusive memory transmission link and is matched with resource compensation mechanism, realizes non-interruption computing power migration, stable control switching interruption length, energy extinction module real-time back sleep device state, directly filter out sleep server when scheduling, avoid migration hang fault from source, solve the defect that migration interruption length is uncontrollable;Second, through the idle computing power hierarchical energy extinction module adopts two-stage progressive control, short-time idle only stops idle hardware, retains core computing power, long-term idle again whole machine sleep latch, break traditional single power-on / sleep mode, give consideration to energy saving and scheduling response speed, realize load, device state two-way synchronous closed loop, computing power scheduling and hierarchical energy extinction linkage operation, simultaneously solve the short board that existing technology energy saving and business scheduling are difficult to take into account.

Description

Technical Field

[0001] This invention relates to the field of host resource scheduling technology, and in particular to a dynamic resource scheduling host hosting system based on edge computing nodes. Background Technology

[0002] With the large-scale development of cloud hosting and rack hosting services, a large number of heterogeneous physical servers are deployed inside the hosting data center to carry the online computing services of different tenants. Data center computing power migration and energy consumption management face multiple technical challenges.

[0003] Existing data center computing power migration solutions mostly adopt the full memory copy mode, and the duration of business interruption during the migration process is uncontrollable; the transmission phase lacks supporting bandwidth and cache compensation mechanisms, and the synchronization of large amounts of memory data can easily cause network latency and business access lag, making it difficult to achieve seamless switching of tenant services.

[0004] In terms of data center energy consumption management, current energy-saving solutions generally adopt two single management modes: direct hibernation of the entire machine or continuous full power of the entire machine, lacking a tiered and graded power-off strategy. Direct hibernation of the entire machine when the server is under low load and idle will cause excessive wake-up time, making it unable to quickly meet sudden computing power demands. Keeping the entire machine fully powered on for a long time will continuously consume the power of idle hardware modules such as auxiliary expansion cards, redundant network ports, and external storage, resulting in high overall energy consumption in the data center. Summary of the Invention

[0005] In view of this, the present invention proposes a dynamic resource scheduling host hosting system based on edge computing nodes, which can effectively solve the defects of existing technologies, such as uncontrollable service interruption duration during migration and lack of hierarchical power shutdown strategies.

[0006] The technical solution of this invention is implemented as follows:

[0007] A dynamic resource scheduling and hosting system based on edge computing nodes includes:

[0008] Edge computing nodes are distributed and anchored to each rack branch in the hosting data center to build localized computing autonomous domains, isolate heterogeneous physical servers corresponding to different tenant hosting hosts, and continuously collect multi-dimensional business indicators of heterogeneous servers within the domain to obtain the original load data of the entire domain.

[0009] The traffic chaos evolution prediction module is used to deduce the traffic chaos evolution trend based on the original load data of the whole domain, and obtain the offset of the computing power supply and demand of the whole domain.

[0010] The cross-physical machine seamless computing power glide module is used to schedule idle heterogeneous physical servers within the localized computing power autonomous domain based on the global computing power supply and demand offset, and to perform uninterrupted computing power glide operations on the managed computing tasks. At the same time, after the computing power glide operation is completed, the real-time load status of each heterogeneous server is synchronously updated to the idle computing power tiered power shutdown module.

[0011] The idle computing power tiered shutdown module is used to perform two-level hierarchical control of hardware module shutdown and whole machine deep hibernation when the computing power occupancy of heterogeneous physical servers is lower than the preset critical threshold. The status of heterogeneous servers after control is transmitted back to the cross-physical machine seamless computing power sliding module in real time to update the computing power sliding schedulable resource pool.

[0012] As a further optional solution to the aforementioned dynamic resource scheduling and hosting system based on edge computing nodes, the traffic chaos evolution prediction module includes:

[0013] The clutter removal and correction unit is used to filter out false flow clutter in the original load data across the entire domain, based on the phase space neighborhood outlier filtering algorithm and the chaotic trajectory interpolation completion algorithm, to obtain a standardized effective load sequence.

[0014] A chaotic dynamics evolution model is used to capture the nonlinear fluctuation patterns of the standardized payload sequence and iteratively calculates the global computing power supply and demand offset.

[0015] As a further optional solution to the aforementioned dynamic resource scheduling host hosting system based on edge computing nodes, the chaotic dynamic evolution model includes:

[0016] The phase space reconstruction extension sub-unit is used to embed delay parameters and embedding dimensions into the multi-dimensional time series indicators of the standardized payload sequence, reconstruct the phase space of multivariable coupled flow, and generate a continuous time series evolution phase point set;

[0017] The exponent solving subunit is used to extract adjacent trajectory segments of the continuous temporal evolution phase point set frame by frame. The local Lyapunov exponent is solved piecewise by variational method. Based on the positive and negative values ​​of the local Lyapunov exponent, the steady-state interval and chaotic mutation interval of the flow are divided, and nodes with potential sudden increases or decreases in computing power are marked.

[0018] The chaotic attractor contour fitting sub-unit is used to fit the boundary contour of the chaotic attractor in the phase space of multivariable coupled flow using a local weighted fractal interpolation algorithm. Differentiated mapping weights are set in combination with the steady-state interval and the chaotic mutation interval of the flow to establish a nonlinear mapping relationship between load index and computing power requirement.

[0019] The iterative calculation subunit is used to iteratively calculate and output the global computing power supply and demand offset based on the nonlinear mapping relationship between load indicators and computing power requirements.

[0020] As a further optional solution to the aforementioned dynamic resource scheduling host hosting system based on edge computing nodes, the cross-physical machine seamless computing power glide module includes:

[0021] The autonomous region resource addressing unit is used to traverse the real-time status pool of heterogeneous physical servers in each computing power autonomous region based on the global computing power supply and demand offset, filter out dormant devices marked by idle computing power tiered power shutdown modules, match idle target servers that meet the computing power threshold, and output computing power glide addressing path.

[0022] The memory mirroring streaming transfer unit is used to establish a direct memory connection channel between the source server and the target server based on the computing power glide addressing path, and adopts a block incremental memory snapshot transfer mechanism to synchronize the business running memory data segment by segment;

[0023] The slip loss dynamic compensation unit is used to retrieve the reserved bandwidth and distributed cache resources of the autonomous system of edge computing nodes in real time during the process of synchronizing business running memory data, so as to offset the business latency loss caused by memory transmission.

[0024] The load synchronization and distribution unit is used to collect the updated hardware occupancy indicators of the source server and the target server after a single computing power shift operation is completed, and synchronously push the real-time load data of all servers in the data center to the idle computing power tiered shutdown module; at the same time, it receives the device hibernation status returned by the idle computing power tiered shutdown module and periodically refreshes the schedulable server resource pool corresponding to the autonomous system resource addressing unit.

[0025] As a further optional solution to the aforementioned dynamic resource scheduling host hosting system based on edge computing nodes, the memory image streaming transfer unit includes:

[0026] The layered slicing unit is used to divide the source server's runtime memory space into layers according to the business process memory partition, distinguishing between static resident memory segments and dynamic variable memory segments, and generating incremental snapshot slices only for dynamic variable memory segments;

[0027] The time-series streaming push subunit is used to match the source server and the target server according to the computing power sliding addressing path and establish a direct memory connection channel. It also establishes a bidirectional streaming transmission queue based on the direct memory connection channel, sorts the incremental snapshot fragments according to the process call priority, continuously pushes the fragment data serially to the target server, and synchronously records the synchronized memory offset marker.

[0028] The dynamic dirty page replenishment iteration subunit is used to continuously scan newly generated dirty pages in memory on the source server during transmission intervals, and add dirty page data as supplementary fragments to the transmission queue for cyclical pushing until the dirty page generation rate is lower than the transmission threshold.

[0029] The runtime context instantaneous delivery subunit is used to terminate the continuous push of incremental snapshot shards and dirty page supplement shards when the dirty page increment tends to stabilize, and deliver the source server runtime context to the target server in one go.

[0030] As a further optional solution to the aforementioned dynamic resource scheduling host hosting system based on edge computing nodes, the idle computing power tiered power shutdown module includes:

[0031] The load threshold determination unit is used to compare the server computing power occupancy with the preset critical threshold based on the real-time load status data of each heterogeneous server, and to filter out heterogeneous physical servers that meet the no-load power-off condition.

[0032] The hardware partitioning and hierarchical power-off unit is used to perform first-order power-off control on heterogeneous servers that are determined to be idle.

[0033] The whole machine hibernation latch unit is used to perform second-order whole machine deep hibernation control if the heterogeneous server has been idle for a preset resting time after the hardware partition is powered off. It latches the current heterogeneous server hardware configuration identifier and memory snapshot.

[0034] The hibernation state feedback unit is used to collect heterogeneous server marking information that has completed first-order hardware shutdown or second-order whole-machine hibernation in real time, and continuously feed it back to the cross-physical machine seamless computing power sliding module.

[0035] As a further optional solution to the aforementioned dynamic resource scheduling host hosting system based on edge computing nodes, the hardware partitioning and hierarchical power-off unit includes:

[0036] The hardware partition identifier matching subunit is used to retrieve the local hardware partition registry of the heterogeneous physical server and divide the core computing power partition and the auxiliary peripheral partition.

[0037] The partitioned power supply isolation switching subunit is used to maintain continuous power supply to the core computing power partition to preserve the basic operating context of the device, cut off the independent power supply circuit of all auxiliary peripheral partitions, and not interrupt the underlying hardware addressing link of the heterogeneous server.

[0038] The partitioned power supply status archiving subunit is used to generate power-off markers and power cut-off timing records for each auxiliary peripheral partition, which are then synchronously archived to the whole machine hibernation latch unit as a pre-verification credential for second-order whole machine deep hibernation.

[0039] As a further optional solution to the aforementioned dynamic resource scheduling host hosting system based on edge computing nodes, the whole-machine hibernation latch unit includes:

[0040] The pre-state verification subunit is used to read the auxiliary peripheral partition power-off flags archived by the hardware partition hierarchical power-off unit, confirm that the heterogeneous server has completed the first-level hardware partition power-off operation and the continuous idle time has reached the preset static time, which meets the whole machine hibernation trigger conditions.

[0041] The device information snapshot latch subunit is used to collect the hardware configuration identifier, partition power supply record, and simplified memory operation snapshot of the current heterogeneous physical server, and integrate them to generate a hibernation state feature data packet for local persistent storage.

[0042] The whole machine power supply circuit control subunit is used to cut off the power supply to the core computing power partition, and only retain the onboard low power signal detection branch to maintain the wake-up communication link, thus completing the second-order whole machine deep sleep management.

[0043] A dynamic resource scheduling and host hosting method based on edge computing nodes specifically includes:

[0044] Based on the edge computing nodes deployed in each rack branch of the hosting data center in a distributed anchored manner, a local computing power autonomous domain is constructed. The computing power autonomous domain is used to isolate the heterogeneous physical servers corresponding to the hosting hosts of different tenants. Multi-dimensional business indicators of all heterogeneous physical servers in the computing power autonomous domain are continuously collected to generate raw load data of the entire domain.

[0045] The clutter removal and correction process is carried out on the raw load data of the entire domain, a chaotic dynamic evolution model is constructed to capture the nonlinear fluctuation law of the standardized effective load sequence, and the offset of the computing power supply and demand of the entire domain is output through iterative calculation.

[0046] Based on the global computing power supply and demand offset, the idle heterogeneous physical servers within the localized computing power autonomous region are scheduled to perform uninterrupted computing power shifting operations on the managed computing tasks. At the same time, after the computing power shifting operation is completed, the real-time load status of each heterogeneous server is synchronously updated to the idle computing power tiered power shutdown module.

[0047] Based on the real-time load status of each heterogeneous server, the computing power occupancy of the heterogeneous server is compared with the preset critical threshold. Based on the comparison results, two-level hierarchical control is implemented, namely, hardware module power-off and whole machine deep hibernation. The status of the heterogeneous server after control is then transmitted back to the cross-physical machine seamless computing power sliding module in real time to update the computing power sliding schedulable resource pool.

[0048] As a further alternative to the dynamic resource scheduling host hosting method based on edge computing nodes, clutter removal and correction processing is performed on the original load data across the entire domain based on the phase space neighborhood outlier filtering algorithm and the chaotic trajectory interpolation completion algorithm.

[0049] The beneficial effects of this invention are as follows: By setting up a cross-physical machine seamless computing power sliding module, relying on the global computing power supply and demand offset output by the traffic chaos evolution prediction module, idle heterogeneous physical servers in the data center are accurately selected for computing power migration operations. A dedicated transmission link is built throughout the process to complete the synchronous transfer of task memory. At the same time, a resource compensation mechanism is provided to alleviate the business latency caused by data transmission. The computing power sliding operation does not interrupt the tenant's front-end business process, and the interruption time of business switching can be stably controlled. Meanwhile, the idle computing power tiered power-off module will transmit the status of dormant devices back to the cross-physical machine seamless computing power sliding module in real time. During scheduling, dormant devices are directly screened out, avoiding long-term business suspension caused by initiating migration tasks to dormant servers. This avoids the risk of additional business interruption from the source of scheduling and effectively solves the problem of uncontrollable migration interruption time. Secondly, an idle computing power tiered shutdown module is configured. When the server computing power occupancy is lower than a preset critical threshold, a two-level hierarchical management system is implemented, which sequentially performs hardware module shutdown and whole-machine deep hibernation. This differs from the traditional single energy-saving mode that only powers on or puts the whole machine into hibernation. In short-term idle scenarios, only idle hardware partitions are shut down while the core computing power components are powered on. This reduces the energy consumption of the data center while preserving the equipment's ability to quickly take on business. In long-term idle scenarios, whole-machine deep hibernation is then implemented, balancing energy saving and computing power scheduling response speed. In addition, after the computing power is shifted, the server load is updated to the shutdown module. The shutdown module then sends the equipment management status back to the scheduling module to update the schedulable resource pool. This ensures the stable operation of the tiered shutdown strategy and computing power scheduling, making up for the shortcomings of existing technologies that lack a tiered shutdown mechanism and cannot balance energy saving and business scheduling. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of the composition of a dynamic resource scheduling host hosting system based on edge computing nodes according to the present invention.

[0052] Figure 2 This is a flowchart illustrating a dynamic resource scheduling and host hosting method based on edge computing nodes according to the present invention. Detailed Implementation

[0053] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0054] refer to Figures 1 to 2 A dynamic resource scheduling host hosting system based on edge computing nodes includes edge computing nodes, a traffic chaotic evolution prediction module, a cross-physical machine seamless computing power glide module, and an idle computing power tiered power shutdown module, wherein:

[0055] Edge computing nodes are distributed and anchored to each rack branch in the hosting data center to build localized computing autonomous domains, isolate heterogeneous physical servers corresponding to different tenant hosting hosts, and continuously collect multi-dimensional business indicators of heterogeneous servers within the domain to obtain raw load data for the entire domain.

[0056] Specifically, the hosting data center is divided into three rack branches: A, B, and C. Each rack branch has a distributed and anchored edge computing node. Each edge computing node independently builds a localized computing autonomous domain. The heterogeneous physical servers of different tenants are divided into mutually isolated computing autonomous domains. Tenant business data and hardware resources are isolated from each other, and there will be no resource contention or data interference issues.

[0057] Each edge computing node collects three multi-dimensional business indicators in real time: CPU utilization, network concurrent connections, and disk read / write speed of all heterogeneous physical servers within the computing power autonomous domain. The data is continuously aggregated with a collection cycle of 1 second, and integrated to generate raw load data covering the entire rack branch domain. This data is then pushed in real time to the traffic chaos evolution prediction module as the input data source for traffic inference.

[0058] In this way, the distributed anchored deployment of edge computing nodes across racks and branch lines, dividing independent computing autonomous domains by rack, can achieve physical isolation of hardware resources and business operating environments of multi-tenant heterogeneous servers, avoiding mutual interference and resource contention between different tenant businesses, and improving the security and stability of multi-tenant business operations in the hosting data center. Secondly, by collecting multi-dimensional business indicators of servers nearby by rack branch line, the collection link is short and the data transmission latency is low, which can continuously obtain real-time and complete raw load data of the entire domain. This provides accurate and timely underlying data support for subsequent traffic chaos evolution trend inference and computing power supply and demand offset calculation, avoiding the data lag and data loss defects caused by centralized collection.

[0059] A traffic chaos evolution prediction module is used to deduce the traffic chaos evolution trend based on the original load data of the entire domain, and obtain the global computing power supply and demand offset; in some embodiments, the traffic chaos evolution prediction module includes:

[0060] The clutter removal and correction unit is used to filter out false flow clutter in the original load data across the entire domain, based on the phase space neighborhood outlier filtering algorithm and the chaotic trajectory interpolation completion algorithm, to obtain a standardized effective load sequence.

[0061] A chaotic dynamics evolution model is used to capture the nonlinear fluctuation patterns of the standardized payload sequence and iteratively calculates the global computing power supply and demand offset.

[0062] Specifically, the detailed processing methods of the clutter stripping and correction unit are as follows:

[0063] Step 1: Multidimensional indicator phase space mapping: The three-dimensional business indicators of CPU utilization, network concurrent connections, and disk read / write speed collected in each time synchronously in the original load data of the whole domain are used as phase space coordinate axes. The indicator values ​​collected every second are mapped to independent phase points in the phase space of multivariable coupled traffic. All time-series collected data form a continuous evolution phase point trajectory. False traffic generated by interference such as instantaneous pulse read / write of disks in the data center, short-term network jitter of switches, and instantaneous peak values ​​of hardware detection will form isolated outlier phase points that deviate from the main trajectory.

[0064] Step 2: Outlier filtering traversal in phase space neighborhood: Define a neighborhood of fixed size centered on each phase point, and count the number of other normal phase points contained in the neighborhood; if the number of valid phase points in the neighborhood of the current phase point is lower than the preset neighborhood threshold, then the phase point is determined to be a false traffic clutter generated by hardware disturbance, and the outlier phase point is directly removed from the time series dataset, filtering out all pulse-type and transient interference clutter in one go.

[0065] Step 3: Chaotic Trajectory Interpolation Timing Completion: After clutter removal, the original timing sequence will have data gaps. Extract the slope and amplitude of the evolution trajectory of multiple consecutive normal phase points before and after the gaps. Use the chaotic trajectory interpolation completion algorithm to fit interpolation phase points that conform to the evolution law of business traffic. Use the interpolation phase points to fill the timing gaps corresponding to the filtered clutter to ensure that the entire load timing sequence is continuous without breaks.

[0066] Step 4: Sequence Standardization Output: Perform uniform dimension normalization on the completed time-series load data to eliminate the differences in the numerical magnitudes of the three types of indicators: CPU, network, and disk. Generate a standardized effective load sequence and transmit it to the chaotic dynamics evolution model.

[0067] The chaotic dynamics evolution model receives the standardized payload sequence output by the clutter stripping and correction unit, captures the nonlinear fluctuation law of the standardized payload sequence, iteratively calculates the global computing power supply and demand offset, and sends the offset to the cross-physical machine seamless computing power sliding module as the scheduling basis.

[0068] Thus, by employing a phase space neighborhood outlier identification mechanism, the true fluctuation phases of the service can be accurately distinguished from spurious clutter caused by instantaneous hardware interference. Compared to traditional one-dimensional time-series mean filtering and threshold truncation noise reduction methods, it avoids misjudging normal sudden changes in service traffic as clutter and removes them, thus fully preserving the nonlinear fluctuation characteristics of the service. Combined with a chaotic trajectory interpolation completion mechanism, it automatically repairs time-series gaps after removing interference clutter, preventing service time-series breaks and data loss due to filtering operations. Simultaneously, it standardizes and unifies the dimensions of metrics, eliminating differences in the magnitude of multi-dimensional indicators and providing a pure data source for the backend chaotic dynamics evolution model. Clean, continuous, and standardized input data significantly reduces the error in computing power prediction at the data source level, avoiding clutter-induced misoperation of computing power scheduling. Secondly, the raw load data across the entire domain is purified by a clutter stripping and correction unit, and the traffic evolution trend is deduced through a chaotic dynamics evolution model. The two-level process linkage ensures the accuracy of computing power supply and demand offset calculation. Based on the accurate prediction results, computing power glide scheduling can be carried out, which can reduce unnecessary cross-server computing power migration and reduce the frequency of idle computing power staged power shutdown modules accidentally triggering whole machine hibernation, thereby simultaneously improving the computing power utilization rate and energy-saving management effect of the data center.

[0069] In some embodiments, the chaotic dynamic evolution model includes:

[0070] The phase space reconstruction extension sub-unit is used to embed delay parameters and embedding dimensions into the multi-dimensional time series indicators of the standardized payload sequence, reconstruct the phase space of multivariable coupled flow, and generate a continuous time series evolution phase point set;

[0071] The exponent solving subunit is used to extract adjacent trajectory segments of the continuous temporal evolution phase point set frame by frame. The local Lyapunov exponent is solved piecewise by variational method. Based on the positive and negative values ​​of the local Lyapunov exponent, the steady-state interval and chaotic mutation interval of the flow are divided, and nodes with potential sudden increases or decreases in computing power are marked.

[0072] The chaotic attractor contour fitting sub-unit is used to fit the boundary contour of the chaotic attractor in the phase space of multivariable coupled flow using a local weighted fractal interpolation algorithm. Differentiated mapping weights are set in combination with the steady-state interval and the chaotic mutation interval of the flow to establish a nonlinear mapping relationship between load index and computing power requirement.

[0073] The iterative calculation subunit is used to iteratively calculate and output the global computing power supply and demand offset based on the nonlinear mapping relationship between load indicators and computing power requirements.

[0074] Specifically, the phase space reconstruction extension sub-unit: For the three multi-dimensional time-series indicators of CPU usage, network concurrency, and disk read / write in the standardized effective load sequence, it embeds latency parameters and embedding dimensions adapted to the hosting data center business scenario, maps one-dimensional time-series data to a high-dimensional space, and builds a multi-variable coupled traffic phase space; it reads standardized load data second by second to generate one-to-one corresponding phase points, and connects all time-series phase points in sequence to form a complete and continuous time-series evolution phase point set, fully reproducing the running trajectory of data center business traffic changing over time.

[0075] The exponent calculation sub-unit traverses the continuous temporal evolution phase point set frame by frame, extracting adjacent trajectory segments that connect each other; it uses the variational method to calculate the local Lyapunov exponent corresponding to each trajectory segment, and divides the phase space region according to the positive or negative exponent: the region with a negative exponent is defined as the steady-state interval of the flow, and the region with a positive exponent is defined as the chaotic mutation interval; it simultaneously locates the boundary inflection point of the two types of intervals, and marks the inflection point as a potential node of sudden increase or decrease in computing power, providing the interval division basis for subsequent differential weight calculation.

[0076] Chaotic attractor contour fitting sub-unit: A local weighted fractal interpolation algorithm is used to fit the boundary contour of the chaotic attractor that runs through the steady-state and abrupt change intervals in the phase space of multivariate coupled flow. Differentiated mapping weights are configured for the steady-state interval and chaotic abrupt change interval of the flow divided by the exponential solution sub-unit. The abrupt change interval is assigned a higher weight to amplify the impact of sudden flow computing power demand, while the steady-state interval is configured with a gentle weight to match the stable business load. Based on the differentiated weights, a nonlinear mapping relationship between the load index value and the server computing power demand is constructed.

[0077] Iterative Calculation Subunit: Calls the load-computing power nonlinear mapping relationship output by the chaotic attractor contour fitting subunit, and uses a fixed short time window to iteratively deduce the trend of the change of the business load of the entire data center in subsequent periods. It compares the total amount of available computing power resources in the current data center, iteratively calculates and outputs the supply and demand offset of computing power in the entire region, and sends the offset to the cross-physical machine seamless computing power sliding module as the benchmark parameter for computing power scheduling.

[0078] Thus, by extending the one-dimensional load time series into a high-dimensional phase space trajectory, the coupling relationship between multi-dimensional business indicators is fully preserved, breaking through the limitation of traditional single linear time series analysis that cannot capture the linkage changes of multiple indicators, and fully restoring the true evolution characteristics of business traffic. Secondly, by dividing the traffic into steady-state and sudden-change intervals through local Lyapunov exponents and marking the computing power change nodes, it is possible to accurately distinguish between stable business conditions and sudden fluctuation conditions, locate the time nodes of traffic surges and drops in advance, avoid the defects of traditional prediction models that ignore traffic changes and computing power scheduling lags, and reserve computing power scheduling operation windows in advance. In addition, by using local weighted fractal interpolation to accurately fit the contour of chaotic attractors, and setting differentiated mapping weights for the two types of traffic intervals, it can adaptively amplify the impact of sudden traffic on computing power demand, accurately establish the nonlinear load-computing power correspondence, and avoid the problems of underestimating computing power demand for sudden traffic and overestimating computing power demand for stable load, thus significantly improving the accuracy of computing power demand matching.

[0079] A cross-physical machine seamless computing power glide module is used to schedule idle heterogeneous physical servers within a localized computing power autonomous region based on the global computing power supply and demand offset, and to perform uninterrupted computing power glide operations on managed computing tasks. Simultaneously, after the computing power glide operation is completed, the real-time load status of each heterogeneous server is synchronously updated to the idle computing power tiered shutdown module. In some embodiments, the cross-physical machine seamless computing power glide module includes:

[0080] The autonomous region resource addressing unit is used to traverse the real-time status pool of heterogeneous physical servers in each computing power autonomous region based on the global computing power supply and demand offset, filter out dormant devices marked by idle computing power tiered power shutdown modules, match idle target servers that meet the computing power threshold, and output computing power glide addressing path.

[0081] The memory mirroring streaming transfer unit is used to establish a direct memory connection channel between the source server and the target server based on the computing power glide addressing path, and adopts a block incremental memory snapshot transfer mechanism to synchronize the business running memory data segment by segment;

[0082] The slip loss dynamic compensation unit is used to retrieve the reserved bandwidth and distributed cache resources of the autonomous system of edge computing nodes in real time during the process of synchronizing business running memory data, so as to offset the business latency loss caused by memory transmission.

[0083] The load synchronization and distribution unit is used to collect the updated hardware occupancy indicators of the source server and the target server after a single computing power shift operation is completed, and synchronously push the real-time load data of all servers in the data center to the idle computing power tiered shutdown module; at the same time, it receives the device hibernation status returned by the idle computing power tiered shutdown module and periodically refreshes the schedulable server resource pool corresponding to the autonomous system resource addressing unit.

[0084] In some embodiments, the memory image streaming unit includes:

[0085] The layered slicing unit is used to divide the source server's runtime memory space into layers according to the business process memory partition, distinguishing between static resident memory segments and dynamic variable memory segments, and generating incremental snapshot slices only for dynamic variable memory segments;

[0086] The time-series streaming push subunit is used to match the source server and the target server according to the computing power sliding addressing path and establish a direct memory connection channel. It also establishes a bidirectional streaming transmission queue based on the direct memory connection channel, sorts the incremental snapshot fragments according to the process call priority, continuously pushes the fragment data serially to the target server, and synchronously records the synchronized memory offset marker.

[0087] The dynamic dirty page replenishment iteration subunit is used to continuously scan newly generated dirty pages in memory on the source server during transmission intervals, and add dirty page data as supplementary fragments to the transmission queue for cyclical pushing until the dirty page generation rate is lower than the transmission threshold.

[0088] The runtime context instantaneous delivery subunit is used to terminate the continuous push of incremental snapshot shards and dirty page supplement shards when the dirty page increment tends to stabilize, and deliver the source server runtime context to the target server in one go.

[0089] Specifically, the autonomous system resource addressing unit reads the supply and demand offset of computing power across the entire system in real time, traverses the real-time status pool of heterogeneous physical servers maintained by each computing power autonomous system in the data center, synchronously records the hardware partition power-off and whole-machine hibernation flags pushed by the idle computing power tiered power-off module in the status pool, and directly filters out unschedulable devices with hibernation flags during the addressing process; compares the remaining computing power resources of the remaining online servers one by one, selects idle target servers whose computing power reserves meet the business computing power threshold, generates a computing power sliding addressing path containing the source server, target server, and rack communication link information, and sends it to the memory mirror streaming transfer unit.

[0090] Memory Mirroring Streaming Transfer Unit: Receives the computing power glide addressing path output by the Autonomous System Resource Addressing Unit, and establishes a point-to-point dedicated memory direct connection channel between the source server and the target server based on the server address and rack communication link parameters within the path. It adopts a block-based incremental memory snapshot transfer mechanism, generating incremental fragments only for dynamically variable memory segments of the server, serially synchronizing them segment by segment. During the synchronization interval, it continuously scans for newly added dirty memory pages and adds them to the transfer queue until the dirty page generation rate stabilizes, at which point it hands over the process running context, without terminating the tenant's business process throughout the entire process.

[0091] The dynamic compensation unit for slippage loss runs continuously throughout the entire process of memory sharding synchronization by the memory image streaming transfer unit. It retrieves the pre-allocated reserved bandwidth pool and edge distributed cache resources of the corresponding computing power autonomous region in real time. It dynamically allocates reserved bandwidth according to the backlog of the sharding transmission queue and preloads the basic memory image to be synchronized to the nearest edge cache node. This offsets the network congestion and service access latency loss caused by large-scale memory sharding transmission and maintains stable front-end service access.

[0092] Load Synchronization and Distribution Unit: After the completion of a single computing power shift operation, the unit automatically collects the hardware occupancy indicators of the source server after releasing computing power and the hardware occupancy indicators of the target server after taking on business, integrates them to generate a real-time load dataset of all servers in the data center and pushes it to the idle computing power tiered shutdown module; at the same time, it periodically receives the hardware shutdown and whole-machine hibernation device markers returned by the idle computing power tiered shutdown module, refreshes the schedulable server resource pool stored in the autonomous region resource addressing unit, and ensures that the hibernation devices are automatically shielded in the next round of addressing operations.

[0093] Thus, by pre-screening dormant and deactivated servers during the addressing phase, there is no need to establish transmission links for unavailable devices, avoiding the problems of prolonged service suspension and migration failure caused by initiating migration tasks to dormant servers. Based on the global computing power supply and demand offset, idle servers with sufficient computing power are accurately matched, reducing invalid migration operations from the scheduling source and improving the efficiency of computing power sliding scheduling. Secondly, block-based incremental memory snapshot synchronization replaces full-machine memory copying, transmitting only dynamic incremental memory data, significantly reducing the amount of data transmitted in memory. Service switching only has a very short switching window during the final context handover, which can stably control the duration of service interruption and achieve seamless computing power sliding for tenant services. A dirty page recycling mechanism is also included to ensure service operation. The entire process ensures complete synchronization of in-memory data, preventing any loss of task data. Furthermore, the transmission process leverages reserved bandwidth within the autonomous computing domain and distributed cache resources to offset transmission latency in real time, avoiding network congestion and service disruptions caused by large-scale concurrent memory transmissions. This guarantees smooth access for tenant services during the computing power migration process and further mitigates the impact of migration operations on front-end services. Additionally, after computing power migration is complete, the server load supply to the idle power-off module is updated synchronously for energy-saving determination. The power-off module sends back the hibernation status to refresh the scheduling resource pool, ensuring that computing power scheduling and tiered power-off management are linked and matched. Through the coordinated action of these modules, low-interruption, low-lag cross-machine computing power migration is achieved, while the idle power-off module precisely manages the energy consumption of idle servers in the data center.

[0094] An idle computing power tiered shutdown module is used to perform two-level hierarchical management of hardware module shutdown and whole-machine deep hibernation when the computing power occupancy of heterogeneous physical servers is lower than a preset critical threshold. The module also transmits the managed heterogeneous server status back to the cross-physical-machine seamless computing power sliding module in real time to update the schedulable computing power sliding resource pool. In some embodiments, the idle computing power tiered shutdown module includes:

[0095] The load threshold determination unit is used to compare the server computing power occupancy with the preset critical threshold based on the real-time load status data of each heterogeneous server, and to filter out heterogeneous physical servers that meet the no-load power-off condition.

[0096] The hardware partitioning and hierarchical power-off unit is used to perform first-order power-off control on heterogeneous servers that are determined to be idle.

[0097] The whole machine hibernation latch unit is used to perform second-order whole machine deep hibernation control if the heterogeneous server has been idle for a preset resting time after the hardware partition is powered off. It latches the current heterogeneous server hardware configuration identifier and memory snapshot.

[0098] The hibernation state feedback unit is used to collect heterogeneous server marking information that has completed first-order hardware shutdown or second-order whole-machine hibernation in real time, and continuously feed it back to the cross-physical machine seamless computing power sliding module.

[0099] Specifically, the load threshold determination unit: The unit is pre-configured locally with a unified no-load critical threshold adapted to the hosting data center business scenario. It continuously receives the server load messages of the entire data center pushed by the load synchronization distribution unit at a period of 2 seconds, parses the messages to extract the real-time computing power utilization index of a single heterogeneous physical server; sets up multi-period verification logic, when the server computing power utilization is lower than the preset critical threshold for 3 consecutive collection periods, it determines that the server is in a stable no-load state and meets the power-off trigger condition; generates a no-load trigger command carrying the server's unique device ID and sends it to the hardware partition and hierarchical power-off unit.

[0100] Hardware partitioned power-off unit: Receives the idle server device ID and trigger command issued by the load threshold determination unit, performs first-order mild power-off control on the target idle server, cuts off the power supply to the idle auxiliary hardware of the server, and keeps the core computing components powered on to preserve the basic operating context; after the hardware partitioned power-off is completed, it generates a receipt message and pushes it to the whole machine hibernation latch unit.

[0101] The whole machine hibernation latch unit: After receiving the first-level power-off receipt, it starts the idle time timer and continuously monitors the server's idle status; if the cumulative idle time after the hardware partition is powered off reaches the preset idle time threshold, it executes the second-level whole machine deep hibernation control, synchronously latches the server hardware configuration identifier and memory simplified snapshot, and retains the basic data required for the device to be quickly woken up; after the whole machine hibernation is completed, it pushes the receipt message to the hibernation status feedback unit.

[0102] Hibernation Status Feedback Unit: Receives two types of feedback messages from the hardware partitioning and hierarchical power-off unit and the whole-machine hibernation latch unit in real time. Generates differentiated status markers for servers that have only completed the first-level hardware power-off or completed the second-level whole-machine hibernation. At a period of 1 second, continuously feeds back the marker information carrying the device ID and power-off level to the cross-physical machine seamless computing power sliding module. This allows the autonomous region resource addressing unit within the module to periodically refresh the schedulable server resource pool. During computing power sliding addressing matching, servers with power-off markers are directly filtered out.

[0103] Thus, by employing a multi-cycle continuous sampling and verification mechanism to determine server idleness, the problem of false power-off triggering caused by instantaneous load fluctuations can be avoided. Load data is synchronously pushed by the computing power sliding module, and the scheduling process and energy-saving process share the same real-time load data source. The idleness judgment benchmark is unified and stable, which greatly reduces the probability of false judgment and missed judgment of idle status, and makes the power-off control actions conform to the actual business load change pattern of the data center. Secondly, the progressive layered control logic of first-order hardware partition power-off and second-order whole-machine deep hibernation is adopted, breaking the single energy-saving mode of existing technologies that can only power on the whole machine or put the whole machine into hibernation. In short-term idle scenarios, only idle hardware is shut down and the core computing power operation capability is retained. While reducing the power consumption of ineffective hardware in the data center, it can quickly take over sudden computing power demand without the need for a whole-machine wake-up process. Only long-term continuously idle devices will perform whole-machine hibernation and lock snapshots, balancing the energy-saving benefits of the data center and the computing power scheduling response speed.

[0104] In some embodiments, the hardware partitioning and hierarchical power-off unit includes:

[0105] The hardware partition identifier matching subunit is used to retrieve the local hardware partition registry of the heterogeneous physical server and divide the core computing power partition and the auxiliary peripheral partition.

[0106] The partitioned power supply isolation switching subunit is used to maintain continuous power supply to the core computing power partition to preserve the basic operating context of the device, cut off the independent power supply circuit of all auxiliary peripheral partitions, and not interrupt the underlying hardware addressing link of the heterogeneous server.

[0107] The partitioned power supply status archiving subunit is used to generate power-off markers and power cut-off timing records for each auxiliary peripheral partition, which are then synchronously archived to the whole machine hibernation latch unit as a pre-verification credential for second-order whole machine deep hibernation.

[0108] Specifically, the hardware partition identification matching subunit receives the unique device ID of the idle server and retrieves the hardware partition registry stored in the server's local flash memory. The registry pre-enters the power supply branch information of all hardware components of the server. The subunit completes the hardware classification based on the registry, assigning the CPU and main memory to the core computing power partition, and uniformly assigning the extended computing power board, redundant network port, external storage, and auxiliary heat dissipation module to the auxiliary peripheral partition. The partition division result is then output to the partition power supply isolation switching subunit.

[0109] Partition power supply isolation switching subunit: Reads the partition division result output by the hardware partition identifier matching subunit, controls the independent power supply circuit switch inside the server to perform isolation switching operation; continuously conducts power supply to the power supply branch of the core computing power partition, retains the server's current memory running data, underlying hardware addressing links and other basic operating context; simultaneously cuts off the independent power supply circuits corresponding to all auxiliary peripheral partitions, only shutting down idle peripheral hardware, without interrupting the server's underlying hardware communication and addressing links, ensuring that the equipment still has the basic operating conditions to quickly undertake business computing power.

[0110] Partition power supply status archiving subunit: After the partition power supply isolation switch is completed, the power failure mark and power cut-off operation timestamp sequence record of each auxiliary peripheral partition are collected and integrated to generate partition power failure status data packets; the data packets are archived and stored locally, and simultaneously pushed to the whole machine hibernation latch unit as a pre-verification credential for subsequent judgment on whether the second-order whole machine deep hibernation control is triggered.

[0111] Thus, by relying on the server's local hardware partition registry to achieve standardized hardware classification, the system accurately distinguishes between components carrying core computing power and idle peripherals. The partitioning logic is unified and adaptable to various heterogeneous server hardware configurations, avoiding accidental disconnection of core computing power hardware during power outages. This provides a clear and reliable hardware operation boundary for tiered power shutdown. Secondly, a partitioned independent power supply isolation switching mechanism is adopted, cutting off power to only idle peripherals while maintaining continuous power to the core computing power partition. Compared to a direct power-off solution for the entire machine, this significantly reduces the ineffective energy consumption of idle servers. At the same time, it preserves the device's basic operating context and underlying hardware addressing links, enabling rapid response to sudden increases in computing power scheduling demands without requiring a complete machine restart or long-duration wake-up, balancing energy saving and computing power scheduling response speed. In addition, the system archives and stores peripheral power-off flags and timing records, and simultaneously pushes them to the machine's hibernation latch unit as a verification basis. This establishes a linkage verification logic between first-level partition power-off and second-level machine hibernation, ensuring that machine hibernation is only triggered after hardware partition power shutdown is completed, strictly following the progressive order of two-level hibernation control.

[0112] In some embodiments, the whole-machine hibernation latch unit includes:

[0113] The pre-state verification subunit is used to read the auxiliary peripheral partition power-off flags archived by the hardware partition hierarchical power-off unit, confirm that the heterogeneous server has completed the first-level hardware partition power-off operation and the continuous idle time has reached the preset static time, which meets the whole machine hibernation trigger conditions.

[0114] The device information snapshot latch subunit is used to collect the hardware configuration identifier, partition power supply record, and simplified memory operation snapshot of the current heterogeneous physical server, and integrate them to generate a hibernation state feature data packet for local persistent storage.

[0115] The whole machine power supply circuit control subunit is used to cut off the power supply to the core computing power partition, and only retain the onboard low power signal detection branch to maintain the wake-up communication link, thus completing the second-order whole machine deep sleep management.

[0116] Specifically, the pre-state verification subunit continuously monitors the auxiliary peripheral partition power-off flags and power supply timing archive records pushed by the hardware partition power-off unit. First, it verifies the power-off flags in the data packet to confirm that the target heterogeneous physical server has fully executed the first-order hardware partition power-off operation. The unit's built-in timer continuously counts the idle time after the hardware partition power-off and compares the cumulative time with the preset idle time threshold. Only when both conditions of "first-order partition power-off completed" and "idle time reaches the target" are met is the server determined to meet the whole-machine hibernation trigger condition, and a hibernation start command is issued to the downstream subunit.

[0117] Device Information Snapshot Latch Subunit: Receives the hibernation start command issued by the pre-state verification subunit, and synchronously collects three types of device operation data: server hardware configuration identifier, partition power supply record archived by the hardware partition and hierarchical power-off unit, and current server simplified memory operation snapshot; integrates and encapsulates the three types of data into an independent hibernation state feature data packet, and writes it to the server's local non-volatile storage medium to complete persistent storage.

[0118] The whole machine power supply circuit control subunit reads the snapshot latching subunit's receipt signal for snapshot storage and performs hierarchical control operations on the server power supply circuit: it cuts off all power supply branches of the core computing power partition composed of CPU and main memory, and keeps only the onboard low-power signal detection branch continuously powered to maintain the communication link for the wake-up command issued by the data center scheduling side without interruption; after completing the power supply circuit switch, it generates a whole machine hibernation completion receipt message and pushes it to the hibernation state feedback unit to mark the server as entering the second-level whole machine deep hibernation state.

[0119] Thus, by setting up dual pre-verification logic, the entire machine can only be allowed to hibernate after a first-level hardware partition power outage and a specified idle time, strictly following the progressive control sequence of light partition power outage and deep entire machine hibernation. This prevents the disorderly operation of servers directly hibernating without performing partition power outages, ensuring the orderly implementation of the two-level energy-saving strategy and differentiating between short-term idle and long-term idle conditions. Secondly, hardware configuration, partition power supply records, and simplified memory snapshots are pre-locked and persistently stored locally. When the server receives a wake-up command later, it can directly read the snapshot data packet to quickly restore the hardware power supply configuration and business operation status, significantly shortening the wake-up time after the entire machine hibernates. At the same time, the complete partition power outage history is retained, facilitating the data center operation and maintenance to trace the entire process of equipment energy consumption control.

[0120] A dynamic resource scheduling and host hosting method based on edge computing nodes specifically includes:

[0121] Based on the edge computing nodes deployed in each rack branch of the hosting data center in a distributed anchored manner, a local computing power autonomous domain is constructed. The computing power autonomous domain is used to isolate the heterogeneous physical servers corresponding to the hosting hosts of different tenants. Multi-dimensional business indicators of all heterogeneous physical servers in the computing power autonomous domain are continuously collected to generate raw load data of the entire domain.

[0122] The clutter removal and correction process is carried out on the raw load data of the entire domain, a chaotic dynamic evolution model is constructed to capture the nonlinear fluctuation law of the standardized effective load sequence, and the offset of the computing power supply and demand of the entire domain is output through iterative calculation.

[0123] Based on the global computing power supply and demand offset, the idle heterogeneous physical servers within the localized computing power autonomous region are scheduled to perform uninterrupted computing power shifting operations on the managed computing tasks. At the same time, after the computing power shifting operation is completed, the real-time load status of each heterogeneous server is synchronously updated to the idle computing power tiered power shutdown module.

[0124] Based on the real-time load status of each heterogeneous server, the computing power occupancy of the heterogeneous server is compared with the preset critical threshold. Based on the comparison results, two-level hierarchical control is implemented, namely, hardware module power-off and whole machine deep hibernation. The status of the heterogeneous server after control is then transmitted back to the cross-physical machine seamless computing power sliding module in real time to update the computing power sliding schedulable resource pool.

[0125] Preferably, clutter stripping and correction processing is performed on the original load data across the entire domain based on the phase space neighborhood outlier filtering algorithm and the chaotic trajectory interpolation completion algorithm.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic resource scheduling and hosting system based on edge computing nodes, characterized in that, include: Edge computing nodes are distributed and anchored to each rack branch in the hosting data center to build localized computing autonomous domains, isolate heterogeneous physical servers corresponding to different tenant hosting hosts, and continuously collect multi-dimensional business indicators of heterogeneous servers within the domain to obtain the original load data of the entire domain. The traffic chaos evolution prediction module is used to deduce the traffic chaos evolution trend based on the original load data of the whole domain, and obtain the offset of the computing power supply and demand of the whole domain. The cross-physical machine seamless computing power glide module is used to schedule idle heterogeneous physical servers within the localized computing power autonomous domain based on the global computing power supply and demand offset, and to perform uninterrupted computing power glide operations on the managed computing tasks. At the same time, after the computing power glide operation is completed, the real-time load status of each heterogeneous server is synchronously updated to the idle computing power tiered power shutdown module. The idle computing power tiered shutdown module is used to perform two-level hierarchical control of hardware module shutdown and whole machine deep hibernation when the computing power occupancy of heterogeneous physical servers is lower than the preset critical threshold. The status of heterogeneous servers after control is transmitted back to the cross-physical machine seamless computing power sliding module in real time to update the computing power sliding schedulable resource pool.

2. The dynamic resource scheduling and host hosting system based on edge computing nodes according to claim 1, characterized in that, The flow chaotic evolution prediction module includes: The clutter removal and correction unit is used to filter out false flow clutter in the original load data across the entire domain, based on the phase space neighborhood outlier filtering algorithm and the chaotic trajectory interpolation completion algorithm, to obtain a standardized effective load sequence. A chaotic dynamics evolution model is used to capture the nonlinear fluctuation patterns of the standardized payload sequence and iteratively calculates the global computing power supply and demand offset.

3. The dynamic resource scheduling and host hosting system based on edge computing nodes according to claim 2, characterized in that, The chaotic dynamic evolution model includes: The phase space reconstruction extension sub-unit is used to embed delay parameters and embedding dimensions into the multi-dimensional time series indicators of the standardized payload sequence, reconstruct the phase space of multivariable coupled flow, and generate a continuous time series evolution phase point set; The exponent solving subunit is used to extract adjacent trajectory segments of the continuous temporal evolution phase point set frame by frame. The local Lyapunov exponent is solved piecewise by variational method. Based on the positive and negative values ​​of the local Lyapunov exponent, the steady-state interval and chaotic mutation interval of the flow are divided, and nodes with potential sudden increases or decreases in computing power are marked. The chaotic attractor contour fitting sub-unit is used to fit the boundary contour of the chaotic attractor in the phase space of multivariable coupled flow using a local weighted fractal interpolation algorithm. Differentiated mapping weights are set in combination with the steady-state interval and the chaotic mutation interval of the flow to establish a nonlinear mapping relationship between load index and computing power requirement. The iterative calculation subunit is used to iteratively calculate and output the global computing power supply and demand offset based on the nonlinear mapping relationship between load indicators and computing power requirements.

4. The dynamic resource scheduling and host hosting system based on edge computing nodes according to claim 1, characterized in that, The cross-physics machine seamless computing power glide module includes: The autonomous region resource addressing unit is used to traverse the real-time status pool of heterogeneous physical servers in each computing power autonomous region based on the global computing power supply and demand offset, filter out dormant devices marked by idle computing power tiered power shutdown modules, match idle target servers that meet the computing power threshold, and output computing power glide addressing path. The memory mirroring streaming transfer unit is used to establish a direct memory connection channel between the source server and the target server based on the computing power glide addressing path, and adopts a block incremental memory snapshot transfer mechanism to synchronize the business running memory data segment by segment; The slip loss dynamic compensation unit is used to retrieve the reserved bandwidth and distributed cache resources of the autonomous system of edge computing nodes in real time during the process of synchronizing business running memory data, so as to offset the business latency loss caused by memory transmission. The load synchronization and distribution unit is used to collect the updated hardware occupancy indicators of the source server and the target server after a single computing power shift operation is completed, and synchronously push the real-time load data of all servers in the data center to the idle computing power tiered shutdown module; at the same time, it receives the device hibernation status returned by the idle computing power tiered shutdown module and periodically refreshes the schedulable server resource pool corresponding to the autonomous system resource addressing unit.

5. The dynamic resource scheduling and host hosting system based on edge computing nodes according to claim 4, characterized in that, The memory image streaming transfer unit includes: The layered slicing unit is used to divide the source server's runtime memory space into layers according to the business process memory partition, distinguishing between static resident memory segments and dynamic variable memory segments, and generating incremental snapshot slices only for dynamic variable memory segments; The time-series streaming push subunit is used to match the source server and the target server according to the computing power sliding addressing path and establish a direct memory connection channel. It also establishes a bidirectional streaming transmission queue based on the direct memory connection channel, sorts the incremental snapshot fragments according to the process call priority, continuously pushes the fragment data serially to the target server, and synchronously records the synchronized memory offset marker. The dynamic dirty page replenishment iteration subunit is used to continuously scan newly generated dirty pages in memory on the source server during transmission intervals, and add dirty page data as supplementary fragments to the transmission queue for cyclical pushing until the dirty page generation rate is lower than the transmission threshold. The runtime context instantaneous delivery subunit is used to terminate the continuous push of incremental snapshot shards and dirty page supplement shards when the dirty page increment tends to stabilize, and deliver the source server runtime context to the target server in one go.

6. The dynamic resource scheduling and host hosting system based on edge computing nodes according to claim 1, characterized in that, The idle computing power staged power extinction module includes: The load threshold determination unit is used to compare the server computing power occupancy with the preset critical threshold based on the real-time load status data of each heterogeneous server, and to filter out heterogeneous physical servers that meet the no-load power-off condition. The hardware partitioning and hierarchical power-off unit is used to perform first-order power-off control on heterogeneous servers that are determined to be idle. The whole machine hibernation latch unit is used to perform second-order whole machine deep hibernation control if the heterogeneous server has been idle for a preset resting time after the hardware partition is powered off. It latches the current heterogeneous server hardware configuration identifier and memory snapshot. The hibernation state feedback unit is used to collect heterogeneous server marking information that has completed first-order hardware shutdown or second-order whole-machine hibernation in real time, and continuously feed it back to the cross-physical machine seamless computing power sliding module.

7. The dynamic resource scheduling and host hosting system based on edge computing nodes according to claim 6, characterized in that, The hardware partitioning and hierarchical power-off unit includes: The hardware partition identifier matching subunit is used to retrieve the local hardware partition registry of the heterogeneous physical server and divide the core computing power partition and the auxiliary peripheral partition. The partitioned power supply isolation switching subunit is used to maintain continuous power supply to the core computing power partition to preserve the basic operating context of the device, cut off the independent power supply circuit of all auxiliary peripheral partitions, and not interrupt the underlying hardware addressing link of the heterogeneous server. The partitioned power supply status archiving subunit is used to generate power-off markers and power cut-off timing records for each auxiliary peripheral partition, which are then synchronously archived to the whole machine hibernation latch unit as a pre-verification credential for second-order whole machine deep hibernation.

8. The dynamic resource scheduling and hosting system based on edge computing nodes according to claim 7, characterized in that, The overall machine hibernation latch unit includes: The pre-state verification subunit is used to read the auxiliary peripheral partition power-off flags archived by the hardware partition hierarchical power-off unit, confirm that the heterogeneous server has completed the first-level hardware partition power-off operation and the continuous idle time has reached the preset static time, which meets the whole machine hibernation trigger conditions. The device information snapshot latch subunit is used to collect the hardware configuration identifier, partition power supply record, and simplified memory operation snapshot of the current heterogeneous physical server, and integrate them to generate a hibernation state feature data packet for local persistent storage. The whole machine power supply circuit control subunit is used to cut off the power supply to the core computing power partition, and only retain the onboard low power signal detection branch to maintain the wake-up communication link, thus completing the second-order whole machine deep sleep management.

9. A dynamic resource scheduling and host hosting method based on edge computing nodes, characterized in that, Specifically, it includes: Based on the edge computing nodes deployed in each rack branch of the hosting data center in a distributed anchored manner, a local computing power autonomous domain is constructed. The computing power autonomous domain is used to isolate the heterogeneous physical servers corresponding to the hosting hosts of different tenants. Multi-dimensional business indicators of all heterogeneous physical servers in the computing power autonomous domain are continuously collected to generate raw load data of the entire domain. The clutter removal and correction process is carried out on the raw load data of the entire domain, a chaotic dynamic evolution model is constructed to capture the nonlinear fluctuation law of the standardized effective load sequence, and the offset of the computing power supply and demand of the entire domain is output through iterative calculation. Based on the global computing power supply and demand offset, the idle heterogeneous physical servers within the localized computing power autonomous region are scheduled to perform uninterrupted computing power shifting operations on the managed computing tasks. At the same time, after the computing power shifting operation is completed, the real-time load status of each heterogeneous server is synchronously updated to the idle computing power tiered power shutdown module. Based on the real-time load status of each heterogeneous server, the computing power occupancy of the heterogeneous server is compared with the preset critical threshold. Based on the comparison results, two-level hierarchical control is implemented, namely, hardware module power-off and whole machine deep hibernation. The status of the heterogeneous server after control is then transmitted back to the cross-physical machine seamless computing power sliding module in real time to update the computing power sliding schedulable resource pool.

10. The dynamic resource scheduling and host hosting method based on edge computing nodes according to claim 9, characterized in that, The clutter removal and correction process is performed on the original load data of the entire domain based on the phase space neighborhood outlier filtering algorithm and the chaotic trajectory interpolation completion algorithm.

Images