A function filling model based on resource fragment space-time feature perception and a method thereof

Abstract

The application discloses a function filling model based on source fragment space-time feature perception and a method thereof, wherein an unallocated resource predictor periodically sends the unallocated resource amount of a node in a future period of time to a phase space according to a discrete stable period and a local dense period; an unused resource predictor collects the size of non-periodic unused resources available at the current time after predicting the periodic unused resources by a deep neural learning model, and uniformly sends the size to the phase space; and a scheduling unit sends an optimal scheduling scheme to the phase space according to the "unallocated" resource state, the "partially used" resource state and the function request resource amount. The application can effectively perceive the change of the resource fragments generated by the virtual machine service in a future period of time, and schedule the function to the fragments that can meet the required resources in the execution time, thereby reducing resource competition.

Description

Technical fields:

[0001] This invention belongs to the field of hybrid deployment technology for different applications in the public cloud environment of cloud computing to improve resource utilization, and particularly relates to a function filling model and method based on resource fragment spatiotemporal feature perception. Background technology:

[0002] Cloud data centers have become a critical infrastructure for the modern digital economy. However, with the high cost of building large-scale data centers, the utilization rate of data center resources is generally low, resulting in significant resource waste. For example, data released by Google Cloud, Microsoft Cloud, and Alibaba Cloud shows that their average CPU utilization is between 25% and 60%. To "reduce costs and increase efficiency," cloud service providers often adopt hybrid deployment technologies to improve data center resource utilization. However, resource competition and interference between hybrid applications cause performance degradation in cloud services, especially a significant increase in tail latency for latency-sensitive interactive services.

[0003] To address the performance degradation of cloud services during application hybrid deployment, existing work has proposed two hybrid scheduling and deployment technologies: "active control" and "feedback regulation." The "active control" hybrid deployment technology pre-characterizes the application's execution characteristics and establishes a predictive model to accurately predict disruptions. The "feedback regulation" hybrid scheduling and deployment technology monitors the runtime performance of cloud services in real time, and when application performance exceeds critical limits, the management controller intervenes in the resource allocation process to achieve rapid recovery of cloud service performance. However, these solutions heavily rely on offline analysis or real-time monitoring of applications. In a public cloud environment, tenant applications exist as "black boxes" within virtual machines, making it difficult for cloud service providers to accurately characterize their execution characteristics.

[0004] Within public cloud data centers, on the one hand, the fragmented "unallocated resources" of servers after virtual machine scheduling exhibit relatively stable and locally concentrated characteristics, while resources allocated to virtual machines but not yet fully utilized—"allocated but unused resources"—show partial periodicity. For example, Microsoft's average lifespan for "unallocated resources" is 61.5 days, and the average time interval for CPU increases and decreases is 17.8 hours, reflecting the relative stability of unallocated resource fluctuations. Furthermore, 75% of the services running on virtual machines within the data center exhibit both periodicity and constancy, meaning that resources allocated to virtual machines but not used in a timely manner show partial periodicity. On the other hand, the emerging serverless computing model has developed rapidly in recent years. The basic computational tasks in serverless computing—functions—exhibit significant small size and short cycle characteristics. 90% of Azure function requests require less than 400MB of memory, 50% of applications allocate a maximum of 170MB of memory at runtime, 50% of calls have an average execution time of less than 1 second, and 96% of functions have an average execution time of less than 60 seconds. Therefore, utilizing fragmented server resources for serverless function computing can significantly improve the resource utilization of data center resources and increase system throughput. Summary of the Invention:

[0005] To address the problems of existing technologies, this invention proposes a function filling model based on the spatiotemporal characteristics of resource fragments. Applied to runtime systems in public cloud environments, this invention can effectively perceive the size changes of resource fragments generated by virtual machine services over a future period and schedule functions to fragments that can meet their resource requirements during execution, thereby reducing resource contention. Furthermore, as an independent runtime system, this invention is not strongly coupled with the platform, thus it can run on most serverless computing platforms. In hybrid deployments, it ensures that the resource requirements of various applications are met during their execution time, thereby reducing resource contention after hybrid deployment and significantly improving resource utilization and effective throughput while maintaining service quality.

[0006] The present invention solves its practical problem by adopting the following technical solution:

[0007] A function filling model based on resource fragmentation spatiotemporal feature awareness is disclosed. This model is applied to a public cloud system containing virtual machines. The model includes a spatiotemporal predictor of unallocated resources, a spatiotemporal predictor of allocated but unused resources, a periodic classifier for VM services, a resource phase space, and a scheduling unit. Wherein:

[0008] The unallocated resource spatiotemporal predictor is used to analyze the virtual machine creation time interval of a node and predict the fluctuation time point of the next unallocated resource to obtain the amount of unallocated resources in the future.

[0009] The unused resource spatiotemporal predictor is used to predict the amount of unused resources generated by periodic services in the future and to collect the amount of non-periodic unused resources generated by non-periodic services every second.

[0010] The periodic classifier is used to divide the virtual machine resources on the node into periodic categories, and the division results affect the prediction range of the spatiotemporal predictor of the unused resources.

[0011] The phase space is used to record the amount of unallocated resources, periodically allocated but unused resources in the future, and non-periodic allocated but unused resources in the current time.

[0012] The scheduling unit outputs the optimal matching resource scheduling scheme based on the amount of unallocated resources, allocated but unused resources, and function-requested resources.

[0013] Furthermore, the scheduling unit includes a function profiler and a function scheduler; wherein:

[0014] The function profiler obtains the required resource amount at any time during the execution time by profileing the function, and sends the profiled function to the function scheduler for scheduling.

[0015] The function scheduler analyzes the phase space resource stability of each global node and selects the node with the best stability and sufficient resource supply during the function execution time to deploy the function service.

[0016] This invention can also be implemented using the following technical solutions:

[0017] A function filling model based on resource fragment spatiotemporal feature awareness includes the following steps:

[0018] The unallocated resource predictor periodically sends the amount of unallocated resources for a future period to the phase space based on the virtual machine time interval during discrete stable periods and local dense periods;

[0019] The unused resource predictor, after predicting periodically unused resources using a deep neural learning model, collects the size of available non-periodic resources at the current moment and sends them uniformly to the phase space.

[0020] The scheduling unit sends the optimal scheduling scheme to the phase space according to the function scheduling request, based on the resource status of "unallocated", "allocated but unused" and the amount of resource requested by the function.

[0021] Furthermore, the unallocated resource predictor will periodically send the amount of unallocated resources over a future period to the phase space process:

[0022] The unallocated resource spatiotemporal predictor initializes the local dense creation trajectory input of all virtual machines on the node;

[0023] The unallocated resource spatiotemporal predictor analyzes the virtual machine creation time interval on the current node to determine whether the current time is in a discrete stable period or a local dense period. When the unallocated resource spatiotemporal predictor is in a local dense period, it extracts the dense creation trajectory with the closest average interval time, calculates the shortest creation time, the longest creation time, and the average resource quantity of the trajectory, sets the time range from the shortest creation time to the longest creation time as the time range for the next unallocated resource fluctuation, and sets the average resource quantity as the amount of reduction in the next unallocated resource.

[0024] The unallocated resource spatiotemporal predictor converts the corresponding amount of unallocated resources into non-periodic allocated but unused resources near the predicted virtual machine creation time. After the predicted time has passed and no new real virtual machines are created, the corresponding periodically allocated but unused resources are converted back into unallocated resources.

[0025] Furthermore, the unused resource predictor predicts the phase space process of unused resource transmission:

[0026] The unused resource spatiotemporal predictor only predicts the periodic resources that have been divided by the periodic classifier. It predicts the size of the periodic resources at each future time scale through a deep neural learning model and calculates the variance between the predicted value and the value consumed by the virtual machine during execution. When the variance is greater than a threshold, the corresponding periodic resource will be converted into an aperiodic resource.

[0027] Furthermore, the scheduling unit sends the optimal scheduling scheme to the phase space according to the function scheduling request, based on the "unallocated" resource status, the "allocated but unused" resource status, and the amount of resource requested by the function:

[0028] The function profiler executes each newly arrived function to obtain the amount of resources used and the execution time of the function, and schedules function requests with two-dimensional required resource tags.

[0029] The function scheduler schedules function requests with required resource tags in the node phase space where the stability is optimal and the unallocated resources plus periodically allocated unused resources are greater than the resources required during the function execution time. Functions with execution times less than 1 second are scheduled in the non-periodic allocation of unused resources where the resource quantity is sufficient.

[0030] Beneficial effects:

[0031] To mitigate the service quality degradation caused by the co-location of virtual machine applications and function applications, previous research proposed monitoring virtual machine IPC fluctuations to adjust resource allocation to function services, and then allocating the current available resources to function services accordingly. However, this approach only considers the amount of resources available at the current moment during scheduling, ignoring the potential impact on virtual machine service quality from resource contention throughout the entire execution time. To overcome these shortcomings, this invention provides a function filling technique based on resource spatiotemporal feature awareness, which significantly improves resource utilization and effective throughput while ensuring that the original service tail latency meets the target.

[0032] Compared to existing work, this invention considers meeting the resource requirements of function applications during execution time during scheduling. By classifying and predicting allocable fragmented resources, the size of fluctuating fragmented resources over a future period is obtained. Functions are then scheduled to resource fragments where their resource requirements are met at any execution time, thereby reducing the possibility of future resource contention. Evaluation results show that, compared to state-of-the-art technologies, this invention can improve CPU utilization by 19% and effective throughput by 47% while ensuring LC service 99th percentile latency compliance. Attached image description:

[0033] Figure 1 This is a schematic diagram of a function filling model structure based on resource fragment spatiotemporal feature perception according to the present invention. Detailed Implementation

[0034] like Figure 1As shown, this invention provides a function filling model based on resource fragment spatiotemporal feature perception, which consists of core concepts such as resource fragment feature classification and spatiotemporal prediction, and function spatiotemporal filling. The system consists of a node-level unallocated resource spatiotemporal predictor, an allocated but unused resource spatiotemporal predictor, a periodic classifier, and a resource phase space, as well as a cluster-level function characterizer and function scheduler. The unallocated resource spatiotemporal predictor analyzes the fluctuation intervals of unallocated resources on nodes and uses a statistical model to predict the next fluctuation time point of unallocated resources, thereby calculating the amount of unallocated resources in the future. The allocated but unused resource spatiotemporal predictor uses a deep learning model to predict the amount of periodic resources in the future and collects and reports the amount of non-periodic resources every second. The periodic classifier uses the Fast Fourier Transform method to divide the virtual machine services on the nodes into periodic categories, and the division results affect the prediction range of the spatiotemporal predictor. The phase space records the amount of available resources at each moment from the current time point to a future period, which is used by the scheduler to schedule function services. The function profiler obtains the required amount of resources at any moment during the execution time of the function by profile the function, and sends the profiled function to the function scheduler for scheduling. The function scheduler analyzes the phase space resource stability of each node globally, selects the node with the best stability and sufficient resource supply during the function execution time to deploy function services.

[0035] 1. Periodic Classifier

[0036] The periodic classifier collects a segment of virtual machine resource utilization data using virt-top and performs a Fast Fourier Transform (FFT) on this data to obtain its spectrogram. The spectrogram is analyzed to determine if it simultaneously possesses the following characteristics: 1. The spectrogram contains significant outliers, with outliers differing in magnitude from the mode group; 2. The spectrogram signal is discrete; 3. The range of points near the domain boundary in the spectrogram has a set of numbers of equal magnitude and close to zero. If all these characteristics are met, the original service is considered periodic and predictable; otherwise, it is considered a non-periodic service.

[0037] 2. Unallocated Resource Spatiotemporal Predictor

[0038] The unallocated resource spatiotemporal predictor analyzes the virtual machine creation time interval graph to determine whether the current interval is approximately equal to the previous interval and less than a threshold time. If so, the average interval is set as the next virtual machine creation time, and the average virtual machine resource size in the two intervals is set as the amount of unallocated resources to be reduced next time. Near the predicted virtual machine creation time, the corresponding amount of unallocated resources is converted into non-periodic allocated unused resources. If no new real virtual machines are created after the predicted time, the corresponding periodically allocated unused resources are converted back into unallocated resources.

[0039] Spatiotemporal distribution prediction of "unallocated" fragment resources:

[0040] "Unallocated" fragmented resources refer to the resources remaining on the server after virtual machine scheduling is completed but not allocated to any virtual machine. Their fluctuation time series exhibits typical characteristics of "relative stability and localized density." Based on these characteristics, statistical analysis and modeling can be performed separately for the discrete stable period and the localized dense period. During the discrete stable period, virtual machine scheduling requests are relatively sparse, making prediction difficult; therefore, no prediction method is designed. During the localized dense period, statistical methods are used to predict the request arrival time in dense regions, and the size of "unallocated" fragmented resources in the next future phase space window is calculated as the prediction result and reported to the scheduler.

[0041] 3. Spatiotemporal predictor of allocated but unused resources

[0042] The unused resource spatiotemporal predictor only predicts periodic resources identified by the periodic classifier. It uses an LSTM model to predict the size of periodic resources at each future time scale, while simultaneously collecting the actual resource fluctuation trajectory and comparing it with the predicted value. If the variance exceeds a threshold, the corresponding periodic resource is converted to an aperiodic resource, and the periodic classifier is requested to re-divide the service periodicity. After predicting the periodic resources, the unused resource spatiotemporal predictor collects the size of the available aperiodic resources at the current moment and sends it uniformly to the phase space.

[0043] "Allocated but unused" resources refer to resources that have been allocated to virtual machines but have not yet been used by applications within the virtual machines. Different applications have different resource usage patterns. To improve the accuracy of prediction, a Fast Fourier Transform is first used to distinguish whether resources are periodic (i.e., predictable). At regular intervals, predictions are made for all predictable services for a future period, and the future windows of periodic "allocated but unused" fragmented resources are reported to the scheduler. At the same time, non-periodic "allocated but unused" fragmented resources are collected every second and reported to the scheduler.

[0044] 4. Phase space

[0045] Phase space records the amount of unallocated resources in the future, the amount of periodically allocated but unused resources, and the amount of non-periodic allocated but unused resources at the current time.

[0046] 6. A function-filling hybrid scheduling method based on fragmented resource profiling is proposed. Whenever the scheduler receives a function scheduling request, it determines the optimal scheduling scheme based on the "unallocated" resource status, the "allocated but unused" resource status, and the amount of resources requested by the function. A reinforcement learning method is proposed to establish a resource status-scheduling decision learning model, aiming to maximize long-term effective throughput for resource scheduling.

[0047] 5.1 Function Characterizer

[0048] The function profiler monitors the resource usage of each newly arrived function during execution and records the execution time after completion. It then assigns a two-dimensional resource tag to the function, which includes both resource usage and execution time. For subsequent identical functions, the corresponding tag is directly assigned and the function is sent to the function scheduler for scheduling.

[0049] 6.2 Function Scheduler

[0050] The function scheduler calculates the variance of the phase space view of each server globally. The variance represents the resource stability, and the nodes are sorted in descending order of stability. After receiving a well-defined function request, the function scheduler schedules functions with an execution time greater than 1 second into the phase space of nodes with optimal stability and unallocated resources plus periodically allocated unused resources that exceed the resources required for the function's execution time. Functions with an execution time less than 1 second are scheduled into non-periodic allocated unused resources that meet the resource requirements.

[0051] This invention is not limited to the embodiments described above. The above description of specific embodiments is intended to illustrate and explain the technical solutions of this invention. The specific embodiments described above are merely illustrative and not restrictive. Without departing from the spirit and scope of the claims, those skilled in the art can make many specific modifications based on the teachings of this invention, and these modifications all fall within the scope of protection of this invention.

Claims

1. A function filling system based on resource fragment spatio-temporal feature perception, the filling system is applied to a public cloud containing a virtual machine, characterized in that: The filling system comprises an unallocated resource time-space predictor, a partially unused resource time-space predictor, a periodic classifier, a resource phase space, and a scheduling unit; wherein: The unallocated resource time-space predictor is used to analyze the virtual machine creation time interval prediction of a node to obtain the unallocated resource amount in the future time at the fluctuation time point of the next unallocated resource; The partially unused resource time-space predictor is used to predict the partially unused resource amount of the periodic virtual machine service in the future time and collect the non-periodic partially unused resource amount of the non-periodic virtual machine every second; The periodic classifier is used to divide the virtual machine resources on the node into periods, and affect the prediction range of the partially unused resource time-space predictor according to the division result; The phase space is used to record the unallocated resource amount, the periodic partially unused resource amount, and the non-periodic partially unused resource amount at the current time in the future period of time; The scheduling unit outputs the optimal matching resource scheduling scheme according to the unallocated resource amount, the partially unused resource amount, and the function request resource amount.

2. The function filling system based on resource fragment space-time feature perception according to claim 1, characterized in that: The scheduling unit comprises a function describer and a function scheduler; wherein: The function describer obtains the required resource amount at any time in the execution time of the function by describing the function, and sends the described function to the function scheduler for scheduling; The function scheduler selects the node with the optimal stability and sufficient resource supply in the function execution time by analyzing the resource stability of the global nodes in the phase space.

3. A function filling method based on resource fragment space-time feature perception, characterized in that, The filling method is realized based on the system of claims 1-2, comprising the following steps: The unallocated resource predictor predicts the unallocated resource amount of the node in the future period of time in the discrete stable period and the local dense period according to the virtual machine time interval, and periodically sends it to the phase space; The partially unused resource predictor collects the available non-periodic resource size at the current time after predicting the periodic partially unused resource by a deep neural learning system, and uniformly sends it to the phase space; The scheduling unit sends the optimal scheduling scheme to the phase space according to the "unallocated" resource state, "partially unused" resource state, and function request resource amount of the function scheduling request.

4. The function filling method based on resource fragment space-time feature perception according to claim 3, characterized in that, The unallocated resource predictor periodically sends the unallocated resource amount in the future period of time to the phase space process: The unallocated resource time-space predictor inputs all the local dense creation trajectories of the virtual machines on the node for initialization; The unallocated resource time-space predictor analyzes the virtual machine creation time interval on the current node to determine whether the current time is in the discrete stable period or the local dense period, extracts the dense creation trajectory closest to the average interval time when the unallocated resource time-space predictor is in the local dense period, calculates the shortest creation time, the longest creation time, and the average resource amount of the trajectory, sets the shortest creation time to the longest creation time as the time range of the next unallocated resource fluctuation, and sets the average resource amount as the next unallocated resource reduction amount; The unallocated resource spatiotemporal predictor converts the corresponding amount of unallocated resources into non-periodic allocated but unused resources near the predicted virtual machine creation time. After the predicted time has passed and no new real virtual machines are created, the corresponding periodically allocated but unused resources are converted back into unallocated resources.

5. The method of claim 3, wherein, The unused resource predictor predicts the phase space process of unused resource transmission: The unused resource spatiotemporal predictor only predicts periodic resources that have been divided by the periodic classifier. It uses a deep neural learning system to predict the size of periodic resources at each future time scale and calculates the variance between the predicted value and the value consumed by the virtual machine during execution. When the variance is greater than a threshold, the corresponding periodic resource will be converted into an aperiodic resource.

6. A function filling method based on resource fragment spatiotemporal feature perception as described in claim 3, characterized in that, The process by which the scheduling unit sends the optimal scheduling scheme to the phase space according to the function scheduling request, based on the "unallocated" resource state, the "allocated but unused" resource state, and the amount of resource requested by the function: The function profiler executes each newly arrived function to obtain the amount of resources used and the execution time of the function, and schedules function requests with two-dimensional required resource tags. The function scheduler schedules function requests with required resource tags in the node phase space where the stability is optimal and the unallocated resources plus periodically allocated unused resources are greater than the resources required during the function execution time. Functions with execution times less than 1 second are scheduled in the non-periodic allocation of unused resources where the resource quantity is sufficient.

Images