Methods, apparatus, storage media and program products for time-sharing computing resources

By using a time-sharing reuse method for computing resources based on multi-dimensional data profiling and activity prediction, the problems of high misjudgment rate and inability of static super-resolution to adapt to dynamic fluctuations in business in existing technologies are solved. This enables refined and efficient utilization of computing resources and low-disturbance scheduling, thereby improving cluster resource utilization and R&D experience.

CN122152512BActive Publication Date: 2026-11-10BEIJING ELECTRONIC DIGITAL INTELLIGENCE TECHNOLOGY CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202610239397.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-02-28
Publication Date
2026-11-10
Estimated Expiration
2046-02-28

AI Technical Summary

Technical Problem

Existing computing resource management solutions suffer from high misjudgment rates and static super-resolution multiples that cannot adapt to dynamic fluctuations in business operations, resulting in low utilization of computing resources and impacting the R&D experience.

Method used

By using multi-dimensional data profiling and activity prediction, the idle status of nodes can be accurately determined. Combined with dynamic scheduling, computing resources can be reused in a time-sharing manner, reducing the misjudgment rate and adapting to business fluctuations.

Benefits of technology

It enables refined and efficient utilization of computing resources, improves the overall utilization rate and scheduling adaptability of cluster resources, reduces the disturbance of resource adjustment, and enhances the R&D experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122152512B_ABST
    Figure CN122152512B_ABST
Patent Text Reader

Abstract

This disclosure presents a method, apparatus, storage medium, and program product for time-sharing reuse of computing resources. The method includes: acquiring multi-dimensional monitoring data of development machines placed on occupied nodes within a target cluster; constructing a resource profile of the placed development machines based on the multi-dimensional monitoring data; predicting the activity probability distribution of the placed development machines based on the resource profile; determining whether the occupied nodes are idle based on the activity probability distribution; when an occupied node is determined to be idle, identifying it as a reclaimable node; when there is a demand for computing resources, selecting a suitable node from the reclaimed nodes and idle nodes; and calling the computing resources of the suitable node. This method can accurately determine node idleness through multi-dimensional data profiling and activity prediction, and, combined with dynamic scheduling, achieve time-sharing reuse of computing resources, effectively reducing the misjudgment rate, adapting to business fluctuations, reducing adjustment disturbances, and achieving refined and efficient resource utilization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure relates to the field of computing power scheduling technology, and in particular to a method, apparatus, storage medium and program product for time-sharing multiplexing of computing resources. Background Technology

[0002] With the rapid development of artificial intelligence technology, the demand for computing resources such as GPUs in AI development platforms continues to rise, and developers often monopolize single or multiple GPUs to carry out complex research and development tasks. However, in actual use, due to users being away from their posts or not shutting down their machines during non-working hours, a large amount of computing resources are left idle, which not only significantly reduces resource utilization but also causes serious waste of computing resources.

[0003] While existing computing resource reuse solutions use NVIDIA-SMI (NVIDIA graphics card management tool) util (utility) / pwr (power consumption) metrics to determine idle status and statically set super-resolution multiples to achieve sharing among multiple lightweight tasks, and employ cold migration or forced shutdown to complete resource adjustments, this approach suffers from several drawbacks. The reliance on a single idle status metric leads to a high false positive rate, the static super-resolution multiple cannot adapt to dynamic business fluctuations, and the recovery time-consuming and experience-impacting effects of cold migration and forced shutdown adjustments are significant. Furthermore, the lack of user behavior prediction and scheduling optimization design makes it difficult to meet the fine-grained requirements for computing resource reuse. Summary of the Invention

[0004] In view of this, the present disclosure provides a method, apparatus, storage medium and program product for time-sharing reuse of computing resources, which can accurately determine node idleness through multi-dimensional data profiling and activity prediction, and complete time-sharing reuse of computing resources by combining dynamic scheduling, effectively reducing the misjudgment rate, adapting to business fluctuations, reducing adjustment disturbances, and achieving refined and efficient utilization of resources.

[0005] In a first aspect, embodiments of this disclosure provide a method for time-sharing reuse of computing resources, employing the following technical solution: Acquire multi-dimensional monitoring data of development machines already placed on occupied nodes within the target cluster; Based on the multi-dimensional monitoring data, a resource profile of the development machine that has been placed is constructed; Based on the resource profile, predict the probability distribution of activity of the deployed development machines; Based on the activity probability distribution, it is determined whether the occupied node is in an idle state; When it is determined that the occupied node is in an idle state, the occupied node is identified as a reclaimed node; When there is a need to call computing resources, a suitable node is selected from the recycled nodes and idle nodes; The computing resources of the adapted node are invoked.

[0006] Optionally, predicting the activity probability distribution of deployed development machines based on the resource profile includes: Extract the profile data of the development machine that has been placed from the resource profile; The profile data is input into the prediction model to obtain the activity probability distribution output by the prediction model. The activity probability distribution includes the activity probability of the placed development machine in at least two look-ahead windows and the corresponding prediction uncertainty index.

[0007] Optionally, the prediction model includes a multi-source heterogeneous input layer, a feature encoding layer, a position encoding layer, a Transformer encoder layer, a shared representation layer, and a multi-task output head; The multi-source heterogeneous input layer is used to align and stitch together the image data of the development machine, generate stitched features, and output them to the feature encoding layer. The feature encoding layer is used to encode the concatenated features, generate a feature token representation sequence, and output it to the position encoding layer; The position encoding layer is used to inject time position encoding into the feature token representation sequence, generate a time-series feature representation sequence, and output it to the Transformer encoder layer; The Transformer encoder layer is used to perform context modeling on the temporal feature representation sequence, generate a context-enhanced vector sequence, and output it to the shared representation layer; The shared representation layer is used to aggregate the context enhancement vector sequence and extract shared features, generate shared latent variables, and output them to the multi-task output head. The multi-task output header is used to generate and output the activity probability of the placed development machine in at least two look-ahead windows and the corresponding prediction uncertainty index based on the shared latent variables.

[0008] Optionally, determining whether the occupied node is in an idle state based on the activity probability distribution includes: Choose between a short lookout window and a long lookout window from at least two lookout windows; When the activity probability of all development machines placed on the occupied node in the short look-ahead window is less than the first activity threshold, and the prediction uncertainty index in the short look-ahead window is less than the uncertainty threshold, the occupied node is determined to be in a short-term idle state. When the probability of activity of all development machines placed on the occupied node in the long look-ahead window is less than the second activity threshold, the occupied node is determined to be in a long-term idle state.

[0009] Optionally, the recycling node includes candidate recycling nodes and recycled nodes; When it is determined that the occupied node is in a short-term idle state, the occupied node is identified as a candidate for recycling. When it is determined that the occupied node is in a long-term idle state, the computing resources of the occupied node are reclaimed; After the computing resources of the occupied node have been reclaimed, the occupied node will be identified as a reclaimed node.

[0010] Optionally, the time-sharing method for computing resources further includes: When any prediction uncertainty index of an occupied node is greater than or equal to the uncertainty threshold, a prediction rollback strategy is triggered for the occupied node. The predictive backoff strategy includes at least one of rule-based idle detection, shortening the look-ahead window, and delaying decision-making.

[0011] Optionally, the time-sharing method for computing resources further includes: Predict the return time distribution of the deployed development machines, and extract the return probability for short time intervals from the return time distribution; When a user operation triggers the power-on startup operation of any of the placed development machines, or when the placed development machine meets the preset parameter conditions, the placed development machine is switched to the development machine to be placed, and it is determined that there is a need to call computing resources. The preset parameter conditions are that the return probability of the placed development machine is greater than a preset generalization threshold, or the activity probability of the placed development machine in the short look-ahead window is greater than a third activity threshold.

[0012] Optionally, when there is a need to call computing resources, selecting suitable nodes from the recycled nodes and idle nodes includes: When there is a need to call computing resources, the original node to which the development machine is to be placed is determined from the reclaimed nodes and the idle nodes; Obtain the affinity, remaining resource score, and interference penalty score of the original node; Determine whether the affinity of the original node is equal to the preset value, whether the remaining resource score is greater than the preset resource threshold, and whether the interference penalty score is less than or equal to the preset score threshold. If so, then the original node will be used as the adapter node; If not, the original node is excluded from the recycled nodes and the idle nodes, and the remaining nodes are candidate nodes; Obtain the comprehensive score of the candidate nodes; From the candidate nodes, the node with the highest comprehensive score is selected as the matching node.

[0013] Optionally, obtaining the comprehensive score of the candidate node includes: Obtain the affinity, timeline complementarity, local hit rate, remaining resource score, and interference penalty score of the candidate nodes; The candidate node's affinity, timeline complementarity, local hit rate, remaining resource score, and interference penalty score are weighted and summed to obtain the candidate node's comprehensive score.

[0014] Optionally, after invoking the computing resources of the adapter node, the method further includes: Schedule the development machine to be placed to the appropriate node.

[0015] Secondly, this disclosure also provides a computing resource time-sharing multiplexing system, which adopts the following technical solution: The data acquisition module is used to acquire multi-dimensional monitoring data of development machines that have been placed on the occupied nodes within the target cluster; The profile building module is used to build a resource profile of the development machine based on the multi-dimensional monitoring data. The probability prediction module is used to predict the activity probability distribution of the placed development machines based on the resource profile. The status judgment module is used to determine whether the occupied node is in an idle state based on the activity probability distribution. The recycling determination module is used to determine the occupied node as a recycling node when it is determined that the occupied node is in an idle state; The node filtering module is used to filter out suitable nodes from the recycled nodes and idle nodes when there is a need to call computing resources; The resource retrieval module is used to invoke the computing resources of the adapted node.

[0016] Thirdly, this disclosure also provides a computer device, which adopts the following technical solution: The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform any of the above-described time-sharing methods for computing resources.

[0017] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing computer instructions for causing a computer to execute any of the above-described time-sharing methods for computing resources.

[0018] Fifthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0019] The computing resource time-sharing reuse method provided in this disclosure constructs a resource profile by collecting multi-dimensional monitoring data from development machines on nodes, and intelligently judges idle status by combining activity probability distribution. Compared with existing technologies that rely solely on NVIDIA-SMI utilization and power consumption indicators, this significantly reduces the false judgment rate of resource idleness. It also incorporates a user behavior prediction mechanism, overcoming the limitations of traditional solutions that lack prediction and optimization design, making resource reuse judgment more accurate and more in line with actual operating patterns. By relying on dynamic prediction results to screen and reclaim nodes and jointly schedule and adapt idle nodes, it replaces the coarse adjustment mode of static over-division multiples, cold migration, and forced shutdown. This dynamically adapts to business load fluctuations, avoiding the problems of time-consuming resource adjustment and recovery that affect the development experience, achieving refined and low-disturbance time-sharing reuse of computing resources, and significantly improving the overall utilization and scheduling adaptability of cluster resources.

[0020] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating the time-sharing multiplexing method for computing resources provided in this embodiment of the disclosure; Figure 2 A flowchart illustrating the method for obtaining the activity probability distribution provided in this embodiment of the disclosure; Figure 3 A flowchart illustrating the idle state determination method provided in this embodiment of the disclosure; Figure 4 A schematic flowchart illustrating a method for determining whether a computing resource call requirement exists, provided in an embodiment of this disclosure. Figure 5 A flowchart illustrating the adaptation node filtering method provided in this embodiment of the disclosure; Figure 6 A flowchart illustrating the comprehensive score acquisition method provided in this embodiment of the disclosure; Figure 7 A schematic diagram of a time-sharing computing resource reuse system provided in this embodiment of the disclosure; Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation

[0023] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0024] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0025] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0026] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0027] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0028] Reference Figure 1This disclosure provides a method for time-sharing reuse of computing resources, including the following steps: S1: Obtain multi-dimensional monitoring data of development machines placed on occupied nodes within the target cluster; S2: Based on multi-dimensional monitoring data, construct a resource profile of the development machines that have been placed; S3: Based on resource profiles, predict the probability distribution of activity of deployed development machines; S4: Based on the activity probability distribution, determine whether the occupied node is in an idle state; S5: When it is determined that an occupied node is in an idle state, the occupied node will be identified as a reclaimable node; S6: When there is a need to call computing resources, select suitable nodes from the recycled nodes and idle nodes; S7: Call upon the computing resources of the adapted node.

[0029] In this context, the target cluster typically refers to a Kubernetes cluster, nodes typically refer to servers, and GPUs exist as hardware resources on the nodes, serving as the core schedulable computing power resource of a Kubernetes node. Based on resource occupancy status and resource reclamation operations, nodes are categorized into occupied nodes, reclaimed nodes, and idle nodes. Occupied nodes are those in the target cluster that have been occupied by Pods / users; reclaimed nodes are those in the target cluster that have been determined to be releaseable or have already been released, and their computing resources can be reassigned by the scheduling system to support new task loads; idle nodes are those in the target Kubernetes cluster that are not occupied by any Pods / users, and both the node and its associated computing resources are in a natively available state; reclaimed nodes automatically become idle nodes after their associated Pods / users release their node occupancy. A DevBox is a cloud computer instance that provides interactive computing services to users, relying on nodes to support resources; a single node typically has 8 / 16 GPU cards, and multiple DevBoxes can be simultaneously supported through resource isolation and time-sharing multiplexing mechanisms to improve the utilization of scarce computing resources. Development machines are categorized into three core states: running, bound, and pending placement. Placed development machines are instances that have completed scheduling decisions and established binding relationships with nodes, encompassing both running and bound states. Running development machines are cloud computer instances that are running normally on nodes and are using computing resources such as GPU / CPU / memory in real time. Bound development machines are cloud computer instances that have been allocated node resource quotas but have not yet started or are in a dormant or suspended state. Pending placement development machines are instances that need to be triggered to start up or request resources and require re-entering the scheduling process to select suitable nodes.

[0030] The time-sharing reuse method for computing resources disclosed herein constructs a resource profile by collecting multi-dimensional monitoring data from development machines on the nodes, and intelligently judges the idle status by combining the activity probability distribution. Compared with the existing technology that only relies on the single utilization rate and power consumption index of NVIDIA-SMI, it significantly reduces the false judgment rate of resource idleness. At the same time, it incorporates a user behavior prediction mechanism, which breaks through the limitations of traditional solutions that lack prediction and optimization design, making the resource reuse judgment more accurate and more in line with the actual operation rules.

[0031] By relying on dynamic prediction results to select and reclaim nodes and coordinate with idle nodes for unified scheduling and adaptation, it replaces the rough adjustment mode of static over-division multiple, cold migration, and forced shutdown. It can dynamically adapt to business load fluctuations, avoid the problems of long resource adjustment and recovery time and impact on R&D experience, realize refined and low-disturbance time-sharing reuse of computing resources, and significantly improve the overall utilization rate and scheduling adaptability of cluster resources.

[0032] In S1, multidimensional monitoring data includes terminal interaction behavior parameters, computing resource usage parameters, user background process parameters, and external activity parameters. Terminal interaction behavior parameters are obtained through a sidecar agent deployed in a container or virtual machine. These parameters are active data related to the teletype machine (TTY) collected by the sidecar agent after monitoring user terminal activity. This includes active event data (timestamps for key presses and command submissions, which are two types of user operation behavior data initiating operation task requests), terminal session duration, and session idle interval. Computational resource usage parameters are collected using Prometheus (a monitoring time-series database) combined with NVIDIA GPU metrics exporter (nvidia / gpuexporter). The statistical granularity is set to a preset time interval (e.g., a coarser one-minute interval suitable for long-term trend analysis). Specifically, this includes GPU utilization (gpu util), streaming multiprocessor utilization (sm util), used video memory (mem used), power consumption (pwr), PCIe transmit traffic (pcie tx), PCIe receive traffic (pcie rx), and also the number of net bytes received by the container network. Network input / output traffic data such as rx and number of bytes sent (net bytestx); user background process parameters are collected by the cloud-native security engine (Falco). These parameters are event data related to the creation and destruction of processes within containers that are subscribed to by the cloud-native security engine (Falco), covering process start and stop events of training scripts and inference services, as well as process operation events related to background running tools / terminal multiplexers (nohup / screen / tmux); external activity parameters are obtained through version control system hooks (VCS Webhook). These parameters are external activity signal data pushed by the code repository, specifically time stream data generated by the distributed version control system (Git) push and commit behavior of the code repository (GitLab / GitHub).

[0033] In S2, a resource profile of the occupied node is constructed by performing time-series and feature extraction and quantitative modeling on the multi-dimensional monitoring data of the development machines placed on the occupied node. The resource profile includes the profile data of the development machines placed on it, including timeline vector, peak and valley summary features, operation mode features and node affinity information.

[0034] The timeline vector refers to a fixed-dimensional time-series feature vector obtained by performing feature fusion and normalization on multidimensional monitoring data of the deployed development machines within a preset observation window, according to equal time intervals. It is used to quantify the continuous activity change pattern of the deployed development machines during the observation period. For example, if the preset observation window is 24 hours and the equal time interval is 15 minutes, a 96-dimensional (because 24 × 60 ÷ 15 = 96) normalized vector is constructed after time-series sampling and feature fusion, denoted as U∈[0,1]. n Where n is the total number of dimensions, i ≤ n, and the i-th timeline vector element U i The normalized value of the comprehensive activity of the multi-dimensional monitoring data of the development machine placed in the i-th equal time period is calculated by feature fusion of multi-dimensional monitoring data such as TTY activity, GPU utilization, network traffic, process creation / destruction events, and external activity signals. Peak-valley summary features are a set of features characterizing the peak patterns and time periodicity of computing resource usage. Specifically, it extracts the top K peak windows with the highest GPU utilization, corresponding to the peak times of user computing resource demand. The average duration of these peak windows, i.e., the average peak width, is calculated to reflect the duration of peak usage. Autocorrelation indicators AC24h and AC7d are calculated using GPU utilization and other data to obtain daily / weekly periodicity indices, measuring the periodicity of computing resource usage within one day and seven days, respectively. Job morphology features refer to the job type proportion characteristics obtained by clustering process and GPU usage patterns. The core is the proportion of training / debugging / inference. By clustering process and GPU usage patterns, the proportion of users performing training, debugging, and inference jobs is determined. Different job types have different usage methods and demands for computing resources, and this feature can be used to understand the user's work patterns and resource requirements. Node affinity information is a set of features used to characterize the historical usage patterns and job startup adaptation efficiency between currently placed development machines and historically placed nodes. Specifically, it includes three core components: the most recent N running nodes, image hit rate, and weight cache hit rate. The most recent N running nodes are historically placed nodes, recording the node identifiers of the most recent N successfully scheduled runs on the currently placed development machine, reflecting the instance's historical node preferences and running distribution patterns. Image hit rate refers to the local hit rate of container images on historically placed nodes, characterizing the local cache hit rate of container images on historically placed nodes. A high image hit rate means reduced time spent pulling images from remote repositories, improving job startup speed. Weight cache hit rate refers to the local hit rate of model cache on historically placed nodes, characterizing the local cache hit rate of model weight cache and hot data on historically placed nodes, used to evaluate the efficiency of loading locally reused model data when starting a computation task.

[0035] Optionally, the profile data also includes short-term regularity features, such as computing resource usage parameters (i.e., GPU / network differential feature sequences) within a preset short time window (e.g., 2 hours), Falco event counts (Falco event counts are the number of various events detected by Falco within a certain time period), external activity signal data, holiday / day / night encoding (encoding time information to distinguish different date types (holidays or weekdays) and time periods (daytime or nighttime), these factors may affect users' activity habits), and contextual features such as session idle time. These short-term regularity features are used to capture recent fine-grained load fluctuations, event mutations, and real-time contextual information of the development machines placed there, providing short-term dynamic basis for activity prediction and user return time estimation.

[0036] After the resource profile is built, the feature store (FeatureStore) is used to store and update the profile features. This module uses the notebook ID as the index key to maintain the time-series features within a sliding 14-day time window and performs incremental updates of the features every 60 seconds. Simultaneously, relying on Kubernetes (K8s) custom resource definitions (CRDs), a custom resource called "Usage Portrait" is created. Key summary information from the resource profile is stored in this custom resource, providing the cluster scheduler with efficient profile information retrieval capabilities and supporting the scheduler in making accurate computing resource scheduling decisions based on the resource profile.

[0037] In S3, refer to Figure 2 The flowchart illustrating the method for obtaining the activity probability distribution, "Predicting the activity probability distribution of deployed development machines based on resource profiling," includes the following steps: S31: Extract profile data of the development machines that have been placed from the resource profile; S32: Input the profile data into the prediction model and obtain the activity probability distribution output by the prediction model. The activity probability distribution includes the activity probability of the development machine placed in at least two look-ahead windows and the corresponding prediction uncertainty index.

[0038] In S31 and S32, in addition to periodically updating the resource profile, it is also necessary to periodically extract the profile data of the deployed development machines from the resource profile. The profile data of a single deployed development machine is then input into the prediction model to complete the prediction of the activity probability distribution. Each deployed development machine is predicted once every minute to enable real-time inference when an event is triggered. The prediction model includes a multi-source heterogeneous input layer, a feature encoding layer, a location encoding layer, a Transformer encoder layer, a shared representation layer, and a multi-task output head.

[0039] The multi-source heterogeneous input layer is used to align and stitch together the profile data of the deployed development machines, generating stitched features and outputting them to the feature encoding layer. This layer receives various multi-source input signals corresponding to the profiles of the deployed development machine resources, performs time stamp alignment processing on these multi-source heterogeneous features, and masks missing values ​​in the feature sequences. Finally, it integrates these features into a unified format temporal feature sequence tensor, along with optional static feature vectors, which serve as the raw input data for the subsequent feature encoding layer. This ensures the integrity, consistency, and temporal validity of the input features, laying the foundation for subsequent feature encoding and model inference.

[0040] The feature encoding layer encodes the concatenated features, generating a sequence of feature token representations which is then output to the positional encoding layer. The core objective of this layer is to transform the original features concatenated from multi-source heterogeneous input layers into a dense representation that the model can learn. First, continuous variables in the concatenated features are normalized to eliminate the influence of dimensions and ensure consistent feature scale. For discrete variables (such as holiday / day / night categories, work type categories, etc.), embedding encoding is used to convert them into low-dimensional dense vectors, achieving numerical representation of discrete information. Subsequently, the normalized continuous numerical features are fused and concatenated with the embedded discrete feature vectors to obtain a comprehensive feature containing multiple types of information. Finally, a fully connected layer linearly projects this comprehensive feature, unifying the heterogeneous features from different sources and dimensions onto a fixed-dimensional high-dimensional feature space, ultimately generating a sequence of feature token representations for each time step, providing adaptive input for subsequent positional encoding layer processing.

[0041] The location encoding layer injects temporal location encoding into the feature token representation sequence, generating a temporal feature representation sequence which is then output to the Transformer encoder layer. This injected temporal location encoding is learnable. Through this injection operation, the relative positional relationships of time steps in the time series data (such as the order and time intervals between different time steps) and periodic patterns such as intraday (e.g., morning, noon, evening) and intraweekly (e.g., weekdays and weekends) are explicitly modeled. This provides the model with a clear temporal semantic reference, enabling it to effectively distinguish the differentiated business meanings corresponding to the same activity level appearing at different time locations (e.g., the same high activity level appearing during weekday morning rush hour and late at night; the former may be normal R&D work, while the latter may be temporary overtime, with significantly different resource scheduling needs). This ensures that the temporal features captured by the model have semantic recognizability that aligns with actual business scenarios. This layer ultimately outputs a temporal feature representation sequence that integrates temporal location information, providing the Transformer encoder layer with input possessing complete temporal semantics.

[0042] The Transformer encoder layer is used to perform contextual modeling on the temporal feature representation sequence, generating a context-enhanced vector sequence and outputting it to the shared representation layer. This layer models long-range dependencies within the complete temporal feature sequence based on a self-attention mechanism, obtaining attention-weighted contextual feature representations. Then, a feedforward network performs nonlinear transformations on the features at each time step to improve feature expressiveness. Simultaneously, residual connections and layer normalization are used to fuse and numerically stabilize the outputs of the attention and feedforward branches. Finally, a context-enhanced vector sequence incorporating global temporal information is output, achieving unified modeling and joint characterization of composite patterns triggered by short-term fluctuations, long-term cycles, and external events.

[0043] The shared representation layer is used to aggregate the context augmentation vector sequence and extract shared features, generating shared latent variables and outputting them to the multi-task output head. This layer aggregates the variable-length context augmentation vector sequence into a fixed-dimensional shared latent vector by selecting features from the last time step, average pooling, attention pooling, or introducing aggregation tokens. The vector is then subjected to feature compression and high-order information extraction by a fully connected layer, ultimately generating a shared latent variable z suitable for multi-task joint prediction.

[0044] The multi-task output head is used to generate and output the activity probability and corresponding prediction uncertainty index of the development machine placed in at least two look-ahead windows based on shared latent variables. This output head includes two core branches: activity probability prediction and uncertainty estimation, which can simultaneously output two types of key prediction results. Specifically, the activity probability prediction head takes the shared latent variable z as input and outputs the activity probability for at least two future look-ahead windows (corely 15 minutes and 30 minutes, expandable to 5 minutes, 20 minutes, etc.) through a lightweight fully connected network, namely Pactive@15, Pactive@30 (and optional Pactive@5, Pactive@20), to depict the probability that the user will remain active within the corresponding time window. To improve the interpretability of the prediction probability and the consistency of threshold decisions, this embodiment uses temperature scaling or ordinal regression to perform posterior calibration on the raw probability output by the model. Simultaneously, the activity probability prediction head outputs the confidence interval [lt, ut] of the corresponding prediction result, intuitively representing the confidence level of the prediction result. The uncertainty estimation head outputs a prediction uncertainty index (entropy), which can be directly calculated from the activity probability of each look-ahead window. This entropy value primarily reflects the ambiguity of the prediction distribution and serves as the confidence gating basis for resource scheduling strategies. A low entropy value indicates high reliability of the prediction results, triggering predictive recovery decisions. Conversely, a high entropy value indicates significant uncertainty in the prediction results, necessitating a rollback mechanism to reduce the risk of misjudgment and ensure the reliability of scheduling decisions.

[0045] In S4, refer to Figure 3The flowchart illustrating the idle state determination method includes the following steps: "Determining whether an occupied node is in an idle state based on the activity probability distribution": S41: Select a short lookout window and a long lookout window from at least two lookout windows; S42: Determine whether the activity probability of all deployed development machines on the occupied node in the short look-ahead window is less than the first activity threshold, and whether the prediction uncertainty index in the short look-ahead window is less than the uncertainty threshold; if the activity probability of all deployed development machines on the occupied node in the short look-ahead window is less than the first activity threshold, and the prediction uncertainty index in the short look-ahead window is less than the uncertainty threshold, then execute step S43; if any deployed development machine on the occupied node does not meet either the condition that the activity probability in the short look-ahead window is less than the first activity threshold or the prediction uncertainty index in the short look-ahead window is less than the uncertainty threshold, then execute step S44. S43: Determine that the occupied node is in a short-term idle state; S44: Determine whether the activity probability of all deployed development machines on the occupied node in the long lookout window is less than the second activity threshold; if the activity probability of all deployed development machines on the occupied node in the long lookout window is less than the second activity threshold, then execute S45; if the activity probability of any deployed development machine on the occupied node in the long lookout window is not less than the second activity threshold, then execute S46. S45: Determines that the occupied node is in a long-term idle state; S46: Determine that the occupied node is in an active state.

[0046] In S41-S46, a short lookahead window refers to a time window that predicts a relatively short future time interval, used to capture the immediate state changes of occupied nodes and support real-time, highly sensitive recycling decisions. A long lookahead window refers to a time window that predicts a relatively long future time interval, used to depict the medium- to long-term state trends of occupied nodes and support predictive, globally optimized resource scheduling decisions. When only two lookahead windows are configured, they are directly divided according to the size of the time span: the window with the smaller value is the short lookahead window, and the window with the larger value is the long lookahead window. For example, if the two lookahead windows are 15 minutes and 30 minutes respectively, the 15-minute window is the short lookahead window, and the 30-minute window is the long lookahead window. When the number of forward windows is greater than 2, short and long forward windows can be selected in the following four ways: First, the extreme value selection method, which selects the window with the smallest time span among all forward windows as the short forward window and the window with the largest time span as the long forward window; second, the uniform interval selection method, which selects the windows in the middle positions of the relatively evenly distributed sequence after sorting the forward windows by time span, respectively, as the short and long forward windows. For example, for the window sequence of 5, 10, 15, 20, 25, 30, 35, 15 and 25 can be selected as the short and long forward windows, respectively; third, the grouping method, which divides all forward windows arranged in ascending order of time span into two groups, with all windows in the first group being short forward windows and all windows in the second group being long forward windows; and fourth, the performance optimization method, which selects the best single or several windows as short and long forward windows based on performance indicators such as the historical prediction success rate of each window.

[0047] The first activity threshold can be set to 0.1, the second activity threshold can be set to 0.2, and the uncertainty threshold can be set to 0.5. These thresholds can be determined using a tiered scheduling strategy based on Service Level Agreements (SLAs). Higher eviction thresholds and longer grace periods are set for development machines used for production-level tasks and long-term training tasks to ensure business stability. For development machines used for general research and development, a more aggressive eviction strategy is adopted to improve the overall utilization of resources.

[0048] In S5, reclaimed nodes include candidate reclaimed nodes and reclaimed nodes. Candidate reclaimed nodes are nodes that meet the criteria for reclaimability but have not yet undergone resource reclamation. These nodes remain in their original occupied state and are not converted to an idle state. They are only included in the schedulable candidate resource pool and are only reclaimed when new resource demands are triggered. Reclaimed nodes are nodes in the target cluster that have been occupied by Pods / users and have completed resource reclamation. These nodes remain in an occupied state and have not been converted to an idle state. Only their idle computing resources are released to a schedulable state.

[0049] When an occupied node is determined to be in a short-term idle state, it is designated as a candidate for reclamation. When an occupied node is determined to be in a long-term idle state, its computing resources are reclaimed. After the computing resources of the occupied node have been reclaimed, it is designated as a reclaimed node. This tiered reclamation strategy, which distinguishes between short-term and long-term idle states, can improve resource scheduling efficiency while ensuring a good user experience. Short-term idle nodes are only marked as candidates for reclamation to address sudden resource demands, while long-term idle nodes have their computing resources reclaimed immediately, achieving a balance between resource utilization and decision reliability.

[0050] Optionally, before performing actual reclamation on the reclaimed node, a snapshot operation needs to be performed on the reclaimed node. In the container layer, the read-only layer reference of OverlayFS (Overlay File System) is combined with the incremental snapshot of the writable layer to persist the incremental snapshot of the writable layer to the distributed shared storage. Then, the computing resource release operation of the reclaimed node is performed.

[0051] Optionally, when any prediction uncertainty indicator of an occupied node is greater than or equal to an uncertainty threshold, a prediction rollback strategy is triggered for the occupied node. The prediction rollback strategy includes at least one of rule-based idle detection, shortening the lookahead window, and delayed decision-making. Rule-based idle detection refers to a deterministic detection method that matches multi-dimensional monitoring data of occupied nodes with preset judgment rules and directly determines whether an occupied node is in a reclaimable idle state based on the matching results. This method uses observable indicators such as GPU utilization, memory usage, and process survival count for joint conditional judgment. When all indicators simultaneously meet the preset threshold, the node is judged as idle; otherwise, it is not idle. This method is often used for safe rollback decisions in scenarios with high prediction uncertainty. Shortening the lookahead window refers to a conservative decision-making approach that, when prediction confidence is insufficient, switches the original lookahead window used for decision-making to a prediction scale with a shorter time span, making the reclamation judgment based only on near-future prediction results. This method can reduce prediction errors and lower the risk of false reclamation and false retention. Delayed decision-making refers to a mechanism that postpones the final recovery decision until the next observation cycle, model inference update, or new monitoring indicator collection is completed. This method offsets the current high uncertainty by introducing more real-time status and the latest prediction information, avoiding irreversible recovery operations when information is insufficient, and improving decision accuracy and business security at the cost of a small delay.

[0052] In S6, refer to Figure 4 The diagram illustrates the process for determining whether a computing resource request exists. The method for determining whether a computing resource request exists includes the following steps: S61: Predict the return time distribution of the deployed development machines and extract the return probability for short time intervals from the return time distribution; S62: When a user operation triggers the power-on startup operation of any placed development machine, or when a placed development machine meets the preset parameter conditions, the placed development machine is switched to the development machine to be placed, and it is determined that there is a need to call computing resources.

[0053] In S61 and S62, regarding the prediction model, the multi-task output head is also used to generate and output the return time distribution of the placed development machine based on shared latent variables. This output head also includes a return time distribution prediction branch, specifically for outputting the return time distribution Treturn∼p(t). This core distribution characterizes the probability distribution of time from the user's current inactive state to their return to the placed development machine and regain activity. This branch discretizes the future time axis according to preset intervals (e.g., dividing it into time intervals of 0–15 minutes, 15–30 minutes, 30–60 minutes, 1–2 hours, etc.), and then outputs the probability of the user returning to the placed development machine and regaining activity within each time interval via the Softmax function, simply referred to as the return probability, thus obtaining the discretized return time probability distribution. This embodiment explicitly uses this return time distribution as one of the core outputs of the model. Its output directly supports the decision-making need of "predicting when the user will return to the placed development machine," providing a crucial basis for subsequent refined computing resource scheduling, task site preservation, and recovery strategy formulation.

[0054] As can be seen from the above, the return time distribution includes the return probability of the development machine placed in at least two time zones. There are two methods for selecting the short time interval: one is to select the one with the smallest start time among all time intervals as the short time interval; the other is to sort all time intervals by their start time from smallest to largest, take the first k time intervals and merge them into a unified short time interval, and then take the weighted average of the return probabilities of the first k time intervals to obtain the return probability of the short time interval.

[0055] If a user action does not trigger the power-on startup of any placed development machine (i.e., the user action triggers the power-on action of a placed development machine), and no placed development machine meets the preset parameter conditions, it is determined that there is no demand for computing resources. The preset parameter conditions are that the return probability of a placed development machine is greater than a preset generalization threshold, or the activity probability of a placed development machine within a short look-ahead window is greater than a third activity threshold.

[0056] The aforementioned determination mechanism can capture in real time the node power-on behavior actively triggered by users and the state changes of occupied nodes, accurately identify computing resource call requirements, and provide timely and reliable triggering basis for subsequent node scheduling and resource allocation.

[0057] Reference Figure 5The flowchart illustrating the method for selecting suitable nodes shows that "when there is a need to call computing resources, suitable nodes are selected from reclaimed nodes and idle nodes," which includes the following steps: S63: When there is a need to call computing resources, determine the original node to place the development machine from the reclaimed nodes and idle nodes; S64: Obtain the affinity, remaining resource score, and interference penalty score of the original node; S65: Determine whether the affinity of the original node is equal to the preset value, whether the remaining resource score is greater than the preset resource threshold, and whether the interference penalty score is less than or equal to the preset score threshold; if the affinity of the original node is equal to the preset value, the remaining resource score is greater than the preset resource threshold, and the interference penalty score is less than or equal to the preset score threshold, then proceed to step S66; if the original node does not meet any of the following conditions, the affinity is equal to the preset value, the remaining resource score is greater than the preset resource threshold, and the interference penalty score is less than or equal to the preset score threshold, then proceed to step S67. S66: Use the original node as the adapter node; S67: Exclude the original nodes from the recycled nodes and idle nodes, and the remaining nodes are candidate nodes; S68: Obtain the comprehensive score of candidate nodes; S69: Select the node with the highest comprehensive score from the candidate nodes as the matching node.

[0058] In S63, the node that was last successfully bound and run can be retrieved from the metadata of the development machine to be placed. This node is the original node.

[0059] In S64-S69, the resource scheduling process adopts the principle of prioritizing source retrieval. It first checks whether the user's original node meets the recovery conditions. Only when the original node is unavailable is a suitable node selected from the candidate nodes. This ensures the continuity and consistency of the user's original environment and allows for efficient use of existing resources when the original node is unavailable, thereby improving the scheduling success rate and resource utilization. The preset value is 1.

[0060] In S68, refer to Figure 6 The flowchart illustrating the comprehensive score acquisition method shows that "acquiring the comprehensive score of candidate nodes" includes the following steps: S681: Obtain the affinity, timeline complementarity, local hit rate, remaining resource score, and interference penalty score of candidate nodes; S682: The candidate node's affinity, timeline complementarity, local hit rate, remaining resource score, and interference penalty score are weighted and summed to obtain the candidate node's comprehensive score.

[0061] Whether it's the original node or the candidate node, the scoring methods for each dimension are based on the same principle. Therefore, the methods for obtaining the scores of these target nodes in each dimension will be described together, as follows: Regarding the method for obtaining affinity, the affinity is 1 for the original node and 10 for the candidate node. Typically, a node usage list is pre-defined, which records the historical invocation status of each node in each user operation task. For example, if node 1 was invoked in the last user operation task, node 1 is considered a hit, and its affinity is 1; if node 1 was not invoked in the last user operation task, node 1 is considered a miss, and its affinity is 0.

[0062] Regarding the method for obtaining timeline complementarity, the first step is to obtain the aggregated timeline baseline vector. The aggregated timeline baseline vector is a fixed-dimensional timeline vector obtained by aggregating the timeline vectors of all deployed development machines for the target node (the original node or candidate node for which scoring is required). It is used to characterize the overall load patterns and resource consumption trends of the target node at different time periods. The aggregated timeline baseline vector is denoted as N∈[0,1]. n , where the i-th timeline vector element N i Let be the overall activity of the target node during the i-th equal time interval. From the vector elements of the timeline vector and the aggregated timeline baseline vector, the minimum value is selected as the overlap rate, and the complement of the overlap rate is taken as the timeline complementarity. This is expressed using a mathematical formula: , ,in, Indicates the overlap rate. Indicates the complementarity of timelines.

[0063] The remaining resource score is calculated by weighted fusion of three core dimensions: GPU availability, memory (MEM) availability, and PCIe bandwidth availability of the target node. PCIe bandwidth availability is defined as the proportion of effective PCIe bandwidth currently available for new task scheduling and invocation on the target node. This dimension is introduced because different user task types (training / debugging / inference) have significantly different requirements for data transmission bandwidth. Model training tasks involve large-scale parameter synchronization and data read / write, requiring strict data throughput capabilities both inside and outside the target node, necessitating high bandwidth availability to avoid communication bottlenecks. Debugging and inference tasks have relatively lower bandwidth dependence. By incorporating PCIe bandwidth availability, the remaining resource score better reflects task characteristics, achieving precise matching of resources and load, and improving scheduling efficiency and task stability.

[0064] The interference penalty score is a penalty factor for the number of overlapping peak periods on the target node. Regarding the method for obtaining the interference penalty score, firstly, the peak window set of the development machine to be placed and the peak window set of the development machines already placed on the target node are obtained. The peak window set refers to the set of time intervals representing the high-load active periods extracted from the timeline profile of the corresponding development machine. It consists of key information such as the Top-K peak window and average peak width of the development machine. It can be maintained and periodically updated by Notebook ID through Feature Store and written to Custom Resource Definition (CRD) for the scheduler to read quickly. The timeline profile is a complete description of the resource usage of the development machine over a continuous period, including resource profiles at different times. Based on the peak window set of the development machine to be placed and the peak window set of the development machines already placed on the target node, the peak window overlap rate between the development machine to be placed and the development machines already placed on the target node is obtained. The peak window overlap rate refers to the percentage of windows that overlap with the peak window set of development machines to be placed on the target node, based on the peak window set of the development machines to be placed. This overlapping window count is then divided by the total number of peak windows of the development machines to be placed. Simply put, it represents the proportion of peak windows of the development machines to be placed that are simultaneously under high load with the peak window of a development machine already placed on the target node. An interference penalty score is obtained based on the peak window overlap rate between the development machines to be placed and all development machines already placed on the target node. The interference penalty score involves iterating through all development machines already placed on the target node. For each development machine, the peak window overlap rate with the development machine to be placed is calculated. If the overlap rate exceeds a preset threshold, it is counted as one peak window conflict, increasing the anti-affinity between the two; otherwise, it is counted as zero. Finally, the total number of conflicts for all development machines is summed, and this total is the interference penalty score. In short, it represents the number of development machines already placed on the target node that have significant peak period overlap with the development machine to be placed.

[0065] The formula for calculating the peak window overlap rate is as follows: In the formula, Indicates the development machine to be placed; This indicates that the development machine has been placed; This indicates the peak window overlap rate between the development machine to be placed and the development machine already placed. This represents the set of peak windows where development machines are to be placed; This represents the set of peak windows where development machines have been placed; This represents the intersection of the peak window set of development machines to be placed and the peak window set of development machines already placed, that is, the set of time windows in which both are simultaneously under high load; The subscript of the summation symbol indicates that each peak window element in the intersection of the peak window sets is traversed. Essentially, it's a time slice; This means counting each peak window element in the intersection of the peak window sets by 1 and summing the results, which equals the number of overlapping peak windows.

[0066] The formula for calculating the interference penalty score is as follows: In the formula, Indicates interference penalty score; This represents the set of all development machines that have been placed on the target node; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This indicates a preset overlap rate threshold, for example, a value of 0.6; This means that if the peak window overlap rate between the development machine to be placed and the development machine already placed exceeds the preset overlap rate threshold, it is counted as 1 conflict; otherwise, it is counted as 0.

[0067] The method for obtaining the local hit rate involves acquiring the local hit rate of the container image on the target node and the local hit rate of the model cache on the target node (i.e., the local hit rate of the model weight cache and hot data on the target node). These two values ​​are then weighted and summed to obtain the local hit rate of the target node. This can be expressed using a mathematical formula: , ,in, Indicates local hit rate. Indicates the weighting coefficient. This indicates the degree to which the container image is locally present on the target node. This indicates the degree of local hit rate of the model cache on the target node. It can be calculated through methods such as the local proportion of the mirror layer and the normalized inversion hit rate of the pull latency; It can be calculated through methods such as weighted file existence, cache hit rate, and load latency normalization. In low-latency, second-level recovery scenarios, it can improve... Weighting is used to enhance the value of deploying resources based on proximity; in scenarios with large image sizes and significant pull overhead, this can improve performance. Weights, making More accurately reflecting the actual startup cost of the target node. The fused [data / materials / etc.] The [0,1] scalar scoring item can be directly used as a scheduling scoring factor in the comprehensive ranking of target nodes.

[0068] The formula for calculating the overall score is as follows: In the formula, express; This represents the weighting coefficient, with a typical weighting coefficient w=[0.35,0.25,0.2,0.15,0.05]. For production and low-latency scenarios, the weighting coefficient can be increased. and ; Indicates affinity; This indicates the score for remaining resources.

[0069] By using a timeline complementary scheduling algorithm, user load is distributed in a staggered manner, and development machines that generate computing tasks at different times are scheduled to the same node, thereby improving the utilization of computing resources and achieving load balancing throughout the day. At the same time, the scheduler prioritizes the original node for instance recovery, and can directly reuse local images and weight caches, skipping the remote fetching and loading process, with instance recovery time of less than 10 seconds.

[0070] In S7, when a candidate node for recycling is selected as an adaptable node, its computing resources are recycled. After the recycling of the adaptable node's computing resources is complete, the recycled computing resources are utilized. When a recycled node is selected as an adaptable node, its recycled and organized computing resources are utilized. When an idle node is selected as an adaptable node, its natively idle computing resources are directly utilized. When the original node is selected as an adaptable node, its computing resources are utilized based on its computing resource status, which reflects whether the original node is a candidate for recycling, a recycled node, or an idle node. After the computing resource utilization is complete, the development machine to be placed is scheduled to the adaptable node.

[0071] In summary, this solution constructs resource profiles by collecting computational resource attribution data from each node in the target cluster. Combined with predictions of future activity probabilities from deployed development machines, it accurately determines the idle status of nodes. By implementing a tiered disposal strategy for idle nodes, it performs real-time snapshots and resource reclamation on nodes predicted to be idle for extended periods, further improving resource utilization. When responding to new resource demands, priority is given to scheduling to the original node to ensure rapid recovery. If the original node becomes unavailable, the optimal replacement node is selected based on node rating, and a timeline complementarity algorithm is used to achieve staggered distribution of user load. The overall solution significantly improves the rationality and efficiency of computational resource allocation, reduces resource waste, provides users with more suitable resource support, and effectively enhances the overall operational performance of the cluster.

[0072] Reference Figure 7 This disclosure provides a time-sharing system for computing resources, including: Data acquisition module 101 is used to acquire multi-dimensional monitoring data of development machines placed on occupied nodes within the target cluster; The profile building module 102 is used to build a profile of the resources that have been placed on the development machine based on multi-dimensional monitoring data; The probability prediction module 103 is used to predict the activity probability distribution of the deployed development machines based on resource profiles. The status judgment module 104 is used to determine whether an occupied node is in an idle state based on the activity probability distribution. The recycling determination module 105 is used to determine the occupied node as a recycling node when it is determined that the occupied node is in an idle state; The node filtering module 106 is used to filter out suitable nodes from recycled nodes and idle nodes when there is a need to call computing resources; Resource Request Module 107 is used to request computing resources from the adapted node.

[0073] The various variations and specific examples of the computing resource time-sharing multiplexing method provided above are also applicable to the computing resource time-sharing multiplexing system provided in this disclosure. Through the foregoing detailed description of the computing resource time-sharing multiplexing method, those skilled in the art can clearly understand the implementation method of the computing resource time-sharing multiplexing system. For the sake of brevity, they will not be described in detail here.

[0074] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0075] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the computing resource time-sharing multiplexing method described in the foregoing embodiments of this disclosure.

[0076] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0077] like Figure 8This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 8 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0078] like Figure 8 As shown, a computer device may include a processor (such as a central processing unit, graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or programs loaded from storage devices into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0079] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although Figure 8 A computer apparatus with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0080] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from ROM. When the computer program is executed by a processor, all or part of the steps of the computing resource time-sharing multiplexing method of embodiments of this disclosure are performed.

[0081] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0082] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When the non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the computing resource time-sharing methods described in the foregoing embodiments of the present disclosure are performed.

[0083] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0084] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0085] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0086] In this disclosure, relational terms such as "first" and "second" are used merely 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. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0087] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0088] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0089] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0090] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0091] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for time-sharing reuse of computing resources, characterized in that, include: Acquire multi-dimensional monitoring data of development machines already placed on occupied nodes within the target cluster; Based on the multi-dimensional monitoring data, a resource profile of the development machine that has been placed is constructed; Based on the resource profile, predict the probability distribution of activity of the deployed development machines; The step of predicting the activity probability distribution of deployed development machines based on the resource profile includes: Extract the profile data of the development machine that has been placed from the resource profile; The profile data is input into the prediction model to obtain the activity probability distribution output by the prediction model, wherein the activity probability distribution includes the activity probability of the placed development machine in at least two forward windows and the corresponding prediction uncertainty index. The prediction model includes a multi-source heterogeneous input layer, a feature encoding layer, a position encoding layer, a Transformer encoder layer, a shared representation layer, and a multi-task output head. The multi-source heterogeneous input layer is used to align and stitch together the image data of the development machine, generate stitched features, and output them to the feature encoding layer. The feature encoding layer is used to encode the concatenated features, generate a feature token representation sequence, and output it to the position encoding layer; The position encoding layer is used to inject time position encoding into the feature token representation sequence, generate a time-series feature representation sequence, and output it to the Transformer encoder layer; The Transformer encoder layer is used to perform context modeling on the temporal feature representation sequence, generate a context-enhanced vector sequence, and output it to the shared representation layer; The shared representation layer is used to aggregate the context enhancement vector sequence and extract shared features, generate shared latent variables, and output them to the multi-task output head. The multi-task output head is used to generate and output the activity probability of the placed development machine in at least two look-ahead windows and the corresponding prediction uncertainty index based on the shared latent variables. Based on the activity probability distribution, it is determined whether the occupied node is in an idle state; When it is determined that the occupied node is in an idle state, the occupied node is identified as a reclaimed node; When there is a need to call computing resources, a suitable node is selected from the recycled nodes and idle nodes; The computing resources of the adapted node are invoked.

2. The time-sharing multiplexing method for computing resources according to claim 1, characterized in that, The step of determining whether the occupied node is in an idle state based on the activity probability distribution includes: Choose between a short lookout window and a long lookout window from at least two lookout windows; When the activity probability of all development machines placed on the occupied node in the short look-ahead window is less than the first activity threshold, and the prediction uncertainty index in the short look-ahead window is less than the uncertainty threshold, the occupied node is determined to be in a short-term idle state. When the probability of activity of all development machines placed on the occupied node in the long look-ahead window is less than the second activity threshold, the occupied node is determined to be in a long-term idle state.

3. The time-sharing multiplexing method for computing resources according to claim 2, characterized in that, The recycling nodes include candidate recycling nodes and already recycled nodes; When it is determined that the occupied node is in a short-term idle state, the occupied node is identified as a candidate for recycling. When it is determined that the occupied node is in a long-term idle state, the computing resources of the occupied node are reclaimed; After the computing resources of the occupied node have been reclaimed, the occupied node will be identified as a reclaimed node.

4. The time-sharing multiplexing method for computing resources according to claim 1, characterized in that, Also includes: When any prediction uncertainty index of an occupied node is greater than or equal to the uncertainty threshold, a prediction rollback strategy is triggered for the occupied node. The predictive backoff strategy includes at least one of rule-based idle detection, shortening the look-ahead window, and delaying decision-making.

5. The time-sharing method for computing resources according to claim 2, characterized in that, Also includes: Predict the return time distribution of the deployed development machines, and extract the return probability for short time intervals from the return time distribution; When a user operation triggers the power-on startup operation of any of the placed development machines, or when the placed development machine meets the preset parameter conditions, the placed development machine is switched to the development machine to be placed, and it is determined that there is a need to call computing resources. The preset parameter conditions are that the return probability of the placed development machine is greater than a preset generalization threshold, or the activity probability of the placed development machine in the short look-ahead window is greater than a third activity threshold.

6. The time-sharing multiplexing method for computing resources according to claim 5, characterized in that, When there is a need to call computing resources, the process of selecting suitable nodes from the recycled nodes and idle nodes includes: When there is a need to call computing resources, the original node to which the development machine is to be placed is determined from the reclaimed nodes and the idle nodes; Obtain the affinity, remaining resource score, and interference penalty score of the original node; Determine whether the affinity of the original node is equal to the preset value, whether the remaining resource score is greater than the preset resource threshold, and whether the interference penalty score is less than or equal to the preset score threshold. If so, then the original node will be used as the adapter node; If not, the original node is excluded from the recycled nodes and the idle nodes, and the remaining nodes are candidate nodes; Obtain the comprehensive score of the candidate nodes; From the candidate nodes, the node with the highest comprehensive score is selected as the matching node.

7. The time-sharing multiplexing method for computing resources according to claim 6, characterized in that, The process of obtaining the comprehensive score of the candidate nodes includes: Obtain the affinity, timeline complementarity, local hit rate, remaining resource score, and interference penalty score of the candidate nodes; The candidate node's affinity, timeline complementarity, local hit rate, remaining resource score, and interference penalty score are weighted and summed to obtain the candidate node's comprehensive score.

8. The time-sharing method for computing resources according to claim 5, characterized in that, After invoking the computing resources of the adapter node, the process further includes: Schedule the development machine to be placed to the appropriate node.

9. A computer device, characterized in that, The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the time-sharing method for computing resources as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the time-sharing method for computing resources as described in any one of claims 1-8.

11. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the steps of the time-sharing method for computing resources as described in any one of claims 1-8.

Citation Information

Patent Citations

  • Cluster processing architecture and method for intelligent analysis of large-scale monitoring nodes

    CN113259451A

  • Method and device for allocating virtual resource, and computer storage medium

    WO2016206441A1

  • Resource scheduling method and apparatus

    WO2025158202A1