Storage and calculation separation method and device based on multi-level cache and intelligent scheduling, and server

By employing a multi-level caching and intelligent scheduling approach to separate storage and computing, resources are dynamically allocated and caching strategies are optimized. This solves the problems of uneven resource utilization and data transmission latency in traditional in-memory computing architectures, thereby improving system performance and resource utilization.

CN121166360APending Publication Date: 2025-12-19SHENZHEN COOCAA NETWORK TECH CO LTD
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202511310681.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-09-15
Publication Date
2025-12-19

AI Technical Summary

Technical Problem

In traditional in-memory computing architectures, storage and computing resources cannot be scaled independently during large-scale data processing, resulting in uneven resource utilization, long data transmission latency, and severely impacting system performance.

Method used

A storage-computing separation approach with multi-level caching and intelligent scheduling is adopted. By collecting user operation behavior data, the cache space of computing nodes is dynamically allocated. The validity of data is judged by combining user historical access, and the caching strategy is optimized. This enables dynamic scheduling and hierarchical storage of data between local cache and remote storage, and real-time monitoring of system indicators to dynamically optimize the scheduling algorithm.

Benefits of technology

It significantly improves system performance, optimizes resource utilization, reduces data transmission latency, and meets the needs of high-performance computing and storage architectures.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121166360A_ABST
    Figure CN121166360A_ABST
Patent Text Reader

Abstract

The invention discloses a storage and calculation separation method and device based on multi-level cache and intelligent scheduling and a server, and belongs to the technical field of data processing, and the method comprises the following steps: collecting operation behavior data of a user, carrying out distributed storage, and dynamically distributing and calculating node cache space; performing classification marking to form marked user data; performing intelligent scheduling layer analysis tasks on the marked user data, and screening and distributing adaptive computing nodes; the computing node receives the analyzed task to check the local cache, obtains corresponding data from the local or request remote storage according to the validity of the local corresponding data, computes the corresponding data, and caches the computed data to the local or remote storage according to the user access probability; and carrying out dynamic scheduling and hierarchical storage on the calculated data between local cache and remote storage according to access frequency and heat factors. According to the invention, the problems of low transmission efficiency and unbalanced resource utilization of the existing storage and calculation separation architecture are solved.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a memory-compute separation method and device based on multi-level cache and intelligent scheduling, a server and a storage medium. BACKGROUND

[0002] In today's era, the amount of data is growing explosively, and application scenarios are increasingly complex and diverse. The traditional memory-compute integrated architecture exposes a series of intractable problems when dealing with large-scale data processing, such as the inability of storage and computing resources to be independently expanded, resulting in uneven resource utilization, and long data transmission delay, which seriously restricts the overall performance of the system.

[0003] Therefore, the prior art still needs to be improved and developed. SUMMARY

[0004] In order to solve the above technical problems, the present application provides a memory-compute separation method and device based on multi-level cache and intelligent scheduling, a server and a storage medium, which aims to solve the problems of low transmission efficiency and uneven resource utilization of the existing memory-compute separation architecture, and proposes a memory-compute separation system and method based on multi-level cache and intelligent scheduling.

[0005] The technical scheme of the present application is as follows: A memory-compute separation method based on multi-level cache and intelligent scheduling, comprising: A user operation behavior data is collected, and the collected operation behavior data is cleaned, integrated, and access frequency and task preference features are extracted; wherein the operation behavior data includes task submission behavior data and data access behavior data; B The operation behavior data is stored in a distributed storage, and the cache space of the computing node is dynamically allocated according to the user operation behavior data; C The operation behavior data stored in the distributed storage is classified and labeled according to the user access mode, and the labeled user data is formed; D The intelligent scheduling layer analyzes the labeled user data, and simultaneously associates the user historical data record and the preference, filters and allocates the adaptive computing node; E The corresponding computing node receives the analyzed task, checks the local cache, judges the data validity in combination with the user historical access, acquires the corresponding data from the local or requests the remote storage according to the validity of the local corresponding data, performs the calculation on the corresponding data, and caches the calculated data in the local or the remote storage according to the user access probability; F The calculated data is dynamically scheduled and stored in layers between the local cache and the remote storage according to the access frequency and the heat factor; The G real-time monitoring system monitors system indicators and user behavior changes, dynamically optimizes the scheduling algorithm, dynamically allocates resources in combination with user task sensitivity and activity, and pre-allocates resources and pre-loads associated data in combination with actual use, and optimizes the cache strategy.

[0006] The method for separating storage and calculation based on multi-level cache and intelligent scheduling, wherein the step of performing intelligent scheduling layer analysis task on the marked user data, simultaneously associating user historical data record and preference, screening and allocating adaptive computing node includes: The marked user data performs intelligent scheduling layer analysis task, obtains computing task, and analyzes the computing task to extract data set and computing resource demand key information required by the computing task; According to the demand key information of the computing task and the current state of the system, the computing node executing the computing task is preliminarily determined, and the task is sent to the corresponding computing node.

[0007] The method for separating storage and calculation based on multi-level cache and intelligent scheduling, wherein the step of the corresponding computing node receiving the parsed task, checking the local cache, judging the data validity in combination with user historical access, obtaining corresponding data for calculation according to the validity of the local corresponding data from the local or remote storage, and caching the calculated data in the local or remote storage according to the user access probability includes: The corresponding computing node receives the parsed computing task, checks whether the required data corresponding to the computing task exists in the local cache; When the local cache exists the required data and the data is valid, the required data is directly read from the local cache for calculation; When the local cache does not exist the required data or the data is invalid, the computing node is controlled to send a data request to the remote storage; the required data is obtained from the remote storage; The required data obtained from the remote is cached to the local cache for calculation, and the cache management strategy is updated according to the characteristics and access mode of the data; And the calculated data is cached in the local or remote storage according to the user access probability.

[0008] The method for separating storage and calculation based on multi-level cache and intelligent scheduling, wherein the step of dynamically scheduling and layering the data between the local cache and the remote storage according to the access frequency and the hotness factor of the calculated data includes: The calculated data is dynamically scheduled and layered between the local cache and the remote storage according to the access frequency and the hotness factor of the data; For frequently accessed hot data, the data is preferentially stored in the local cache, and is replaced and updated according to the cache capacity and the priority of the data; For cold data that is not frequently accessed, it is stored in remote storage.

[0009] The method for storage and calculation separation based on multi-level cache and intelligent scheduling, wherein the step of dynamically optimizing the scheduling algorithm in combination with the user task sensitivity and activity, dynamically allocating resources, and pre-allocating resources and pre-loading associated data in combination with the actual use situation to optimize the cache strategy includes: The control intelligent scheduling layer monitors the load condition, network bandwidth, delay performance index of each computing node, and the read-write performance and data distribution condition of the remote storage in real time. According to the monitoring data, the intelligent scheduling algorithm based on the load balancing scheduling algorithm or the data locality scheduling algorithm is adopted to dynamically adjust the allocation and scheduling strategy of the task, and the task is allocated to the corresponding idle computing node while ensuring the locality access of the data.

[0010] The method for storage and calculation separation based on multi-level cache and intelligent scheduling, wherein the step of the corresponding computing node receiving the parsed task, checking the local cache, judging the data validity in combination with the user historical access, obtaining the corresponding data from the local or requesting the remote storage according to the validity of the local corresponding data, and performing calculation on the corresponding data, and caching the calculated data in the local or remote storage according to the user access probability includes: The corresponding computing node receives the parsed task and obtains the required data according to the scheduling strategy to control the start of the execution of the computing task. During the execution of the task, the computing node dynamically adjusts the cache data and requests new data according to the actual demand. When the computing task is executed, the corresponding computing node returns the calculation result to the user terminal application program, and stores the result data into the remote storage or the local cache according to the need.

[0011] The method for storage and calculation separation based on multi-level cache and intelligent scheduling, wherein the step of dynamically optimizing the scheduling algorithm in combination with the user task sensitivity and activity, dynamically allocating resources, and pre-allocating resources and pre-loading associated data in combination with the actual use situation to optimize the cache strategy includes: Control data analysis and performance evaluation are performed according to the running condition of the system and the execution result of the task to optimize the scheduling algorithm and the cache management strategy. Through the mining and analysis of the historical data, the access mode and resource demand of the future task are predicted, and resource pre-allocation and cache pre-loading are performed in advance.

[0012] A device for storage and calculation separation based on multi-level cache and intelligent scheduling, wherein the device comprises: The data collection module is configured to collect operation behavior data of a user, clean and integrate the collected operation behavior data, and extract access frequency and task preference characteristics. The dynamic allocation module is configured to store the operation behavior data in a distributed manner, and dynamically allocate cache spaces of computing nodes according to the operation behavior data of the user. The classification and labeling module is configured to classify and label the operation behavior data stored in a distributed manner according to characteristics and user access patterns, to form labeled user data. The intelligent scheduling and association module is configured to intelligently schedule the labeled user data to a parsing task, and simultaneously associate user historical data records and preferences, filter and distribute suitable computing nodes. The computing module is configured to receive the parsed task by a corresponding computing node, check a local cache, determine data validity in combination with user historical access, obtain corresponding data from the local or remote storage according to the validity of the local corresponding data, and perform calculation on the corresponding data, and cache the calculated data in the local or remote storage according to a user access probability. The dynamic scheduling module is configured to dynamically schedule and store the calculated data between the local cache and the remote storage according to access frequency and heat factors. The optimization module is configured to monitor system indicators and user behavior changes in real time, dynamically optimize a scheduling algorithm, dynamically allocate resources in combination with user task sensitivity and activity, and pre-allocate resources and pre-load associated data in combination with actual use cases, and optimize a cache strategy.

[0013] A server, comprising a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors, and the one or more programs comprise programs for executing any one of the methods.

[0014] A computer readable storage medium, wherein when instructions in the storage medium are executed by a processor of an electronic device, the electronic device can execute any one of the methods.

[0015] As can be seen from the above, the present application provides a multi-level cache and intelligent scheduling based storage and calculation separation method, device, server and storage medium. The present application focuses on solving the problems of high data transmission delay and insufficient cache management in the existing storage and calculation separation architecture, and proposes a multi-level cache and intelligent scheduling based storage and calculation separation system and method, which aims to significantly improve the performance of the system, optimize resource utilization, provide a more efficient and reliable solution for large-scale data processing, and better meet the urgent needs of high-performance computing and storage architecture in the current data era. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the storage-computation separation method based on multi-level caching and intelligent scheduling according to Embodiment 1 of the present invention.

[0018] Figure 2 The present invention provides a schematic diagram of an embodiment of a storage-compute separation device based on multi-level caching and intelligent scheduling.

[0019] Figure 3 This is a block diagram illustrating the internal structure of the server provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0022] In today's era, data volume is growing explosively, and application scenarios are becoming increasingly complex and diverse. Traditional in-memory computing architectures have exposed a series of thorny problems when dealing with large-scale data processing. For example, storage and computing resources cannot be scaled independently, resulting in uneven resource utilization. Moreover, data transmission latency is relatively long, which seriously restricts the overall performance of the system.

[0023] In view of this, this invention focuses on overcoming the challenges of high data transmission latency and insufficient granular cache management in existing compute-in-memory (CIM) architectures, and innovatively proposes a compute-in-memory (CIM) separation system and method based on multi-level caching and intelligent scheduling. Through this advanced technology and design concept, it aims to significantly improve system performance while optimizing resource utilization, providing a more efficient and reliable solution for large-scale data processing, thereby better meeting the urgent needs of the current data era for high-performance computing and storage architectures.

[0024] To address the aforementioned technical problems, this invention provides a storage-computation separation method based on multi-level caching and intelligent scheduling, as detailed in the following embodiments.

[0025] Example 1 like Figure 1 As shown, an embodiment of the present invention provides a storage-computation separation method based on multi-level caching and intelligent scheduling, comprising the following steps: Step S100: Collect user operation behavior data, clean and integrate the collected operation behavior data, and extract access frequency and task preference features; wherein, the operation behavior data includes task submission behavior data and data access behavior data; In traditional in-memory computing architectures, the lack of precise perception of user behavior often leads to indiscriminate resource allocation, resulting in insufficient resources for high-demand users and idle resources for low-demand users. This step collects and analyzes user behavior data to provide a data basis for subsequent dynamic resource allocation and intelligent task scheduling, fundamentally solving the problem of uneven resource utilization. Specifically, regarding the collection of operational behavior data in this step, the scope of collection includes task submission behavior data and data access behavior data. These two types of data correspond to the user's computing needs and storage needs, respectively. The task submission behavior data includes task submission time, task type (such as offline calculation, real-time analysis), task data volume, task priority setting, task dependency relationship, etc., which can be captured in real time through the system task submission interface. The data access behavior data includes the identifier of the accessed data (such as file ID, database table name), access time, access frequency, access method (such as read operation, write operation), data access duration, etc., which can be collected by storing access logs and data request links. To address potential issues in the collected data, such as noise (e.g., abnormally submitted empty tasks, duplicate access logs), missing values ​​(e.g., some tasks not marked with priority), and inconsistent formats (e.g., different users submitting tasks with different data units), this application embodiment employs data cleaning algorithms (e.g., rule-based outlier removal, interpolation-based missing value filling) to clean the collected operational behavior data. Simultaneously, the two types of data, scattered across the task system and storage system, are integrated into a unified data mart to form a structured operational behavior dataset, avoiding data silos. In the embodiment of the present step, two core features, access frequency and task preference features, are extracted from the integrated data set. The access frequency is the number of accesses to certain data by the user within a specified time window (such as 1 hour, 24 hours), the average access interval, and the quantification of the user's "hotness demand" for data. The task preference feature is analyzed by a classification algorithm (such as decision tree, K-means clustering) to identify the user's task preference (such as a user who submits high-priority real-time analysis tasks for a long time, a user who mainly submits small data volume offline computing tasks), and to provide a basis for matching and adapting the computing node for subsequent matching and adapting.

[0026] Step S200, the operation behavior data is stored in a distributed manner, and the computing node cache space is dynamically allocated according to the operation behavior data of the user. The core pain point of the traditional storage and computing integrated architecture is the binding of storage and computing resources. Storage expansion must be accompanied by computing expansion, and vice versa, resulting in resource waste. The embodiment of the present step realizes independent expansion of storage resources through distributed storage, and realizes on-demand allocation of computing resources through dynamic cache allocation, directly solving the problem of independent expansion of storage and computing resources, and reducing the number of data transmissions between computing nodes and storage nodes. Specifically, in the embodiment of the present step, a distributed file system (such as HDFS) or a distributed database (such as HBase) is used to store the operation behavior data integrated in step S100. The data is sharded according to the user dimension or the data hotness dimension and stored in multiple storage nodes. Each storage node can be independently expanded (such as increasing the hard disk capacity, adding a storage node), without relying on the expansion of the computing node.

[0027] In specific implementation, the present application will establish a distributed storage index to record the storage location and the number of copies of the data shards, so as to ensure that the computing node can quickly locate the data and reduce the data search delay. The dynamic allocation of the computing node cache space of the present application is based on the access frequency feature extracted in step S100. The cache space of each computing node is dynamically adjusted. Specifically, for users (or user groups) with high access frequency, a larger computing node cache space is allocated to temporarily store their high-frequency access data locally, reducing the dependence on remote storage. For users with low access frequency, their cache space is appropriately compressed to release idle resources for high-demand users. The present application uses a cache space dynamic adjustment algorithm (such as a cache eviction policy based on LRU combined with user access frequency weight) to update the cache allocation scheme in real time, ensuring that the cache resources are always tilted towards high-value demand.

[0028] Step S300, classifying and labeling the features and user access patterns of the distributed stored operation behavior data to form labeled user data; The operation behavior data of distributed storage is still in an uncategorized state and cannot be directly used for intelligent scheduling. This step classifies and labels the data to associate it with user access patterns, forms labeled user data, and provides structured decision-making basis for subsequent intelligent scheduling layer analysis tasks, thereby improving scheduling accuracy. Specifically, in this step, based on the access frequency and task preference characteristics extracted in step S100, combined with the data attributes (such as data type and data size) of distributed storage, a machine learning classification model (such as support vector machine and random forest) is used to classify the operation behavior data. In this embodiment, the following classifications can be included. Classification by access pattern, such as high-frequency read-low-frequency write mode corresponding to cold data update and hot data query scenario, high-frequency write-high-frequency read mode corresponding to real-time data processing scenario, and low-frequency access mode corresponding to archived data scenario; classification by user task preference, such as real-time computing preference user, offline analysis preference user, and mixed task preference user. After classification, labeled user data is formed, i.e., access pattern labels and user preference labels are added to each type of operation behavior data, and user unique identification (such as user ID) and data identification are associated to construct four-dimensional labeled user data associated with user-data-access pattern-task preference. For example, high-frequency read data accessed by user A for real-time analysis task is labeled as real-time computing preference + high-frequency read mode label, which facilitates subsequent scheduling layer to quickly identify the matching relationship between data and user demand.

[0029] Step S400, intelligent scheduling layer analysis task is performed on the labeled user data, and user historical data records and preferences are associated to filter and distribute suitable computing nodes; In traditional storage and computing architecture, task allocation often uses polling or random strategy, which leads to mismatch between computing nodes and task demand (such as assigning real-time tasks to nodes with weak performance), and aggravates data transmission delay and resource waste. In this step, by analyzing tasks through intelligent scheduling layer and associating user historical data, accurate matching between tasks and nodes is achieved, and task processing efficiency is improved. In the specific implementation of this step, the intelligent scheduling layer, as the decision-making core of the system, after receiving a new task submitted by a user, first reads the labeled user data corresponding to the task, and analyzes the core requirements of the task, including: Computing demand analysis, such as task type (real-time / offline), required computing resources (CPU core number, memory size), and task deadline; Data demand analysis, such as task-dependent data identification, data access pattern (read / write dominant), and data popularity. Then, the user historical data record is associated with the preference, and specifically, the historical task processing record (such as the execution time length of the historical task, the matching effect of the historical task and the computing node, and the historical cache hit rate) of the user can be called, and the task preference feature extracted in step S100 is combined to further optimize the demand judgment; if the user historical real-time task is executed on a high CPU performance node, the efficiency is improved by 30%, the new real-time task is preferentially matched with the same type of high CPU performance node; if the user historical high-frequency access data has a cache hit rate of 90% on a large cache node, the new task is preferentially matched with the large cache node. Then, the suitable computing node is screened and distributed, and the application screens the suitable node from the computing node resource pool (including nodes with different performance and different cache configuration) based on the task demand analysis result and the user historical preference. Specifically, a node capability evaluation model can be established to quantify the computing capability of each node, such as CPU, memory load rate, and peak computing power; and the cache capability, such as cache capacity and cache hit rate; and the network capability, such as bandwidth and delay to the storage node. Then, the greedy algorithm or the genetic algorithm can be used to select the computing node with the highest matching degree of node capability and task demand, for example, the real-time analysis task is distributed to a node with low load and high network bandwidth, and the offline computing task is distributed to a node with high memory and high load capacity.

[0030] Step S500, the corresponding computing node receives the parsed task and checks the local cache, judges the data validity in combination with the user historical access, obtains the corresponding data from the local or requests the remote storage according to the validity of the local corresponding data, and caches the calculated data according to the user access probability in the local or the remote storage; In the traditional storage-computing separation architecture, when the computing node processes the task, the remote storage data is often requested blindly, resulting in high data transmission delay; at the same time, the cache data is not combined with the user demand, and the cache hit rate is low. In the embodiment of the present application, the data acquisition path is optimized by local cache checking + historical access judgment, and the cache strategy is formulated in combination with the access probability, so as to reduce the delay from the data acquisition and data cache. Specifically, in the embodiment of the present application, after the suitable computing node receives the task distributed by the intelligent scheduling layer, it is first checked whether the data required by the task exists in the local cache, which can be quickly queried through the local cache index (records the identification, storage location and cache time of the cache data), if the data exists, the data validity judgment link is directly entered; if the data does not exist, the remote storage request process is triggered. In the embodiment of the present application, the data validity is determined by combining the user historical access data even if the required data exists in the local cache. Specifically, if the data is static data (such as historical archive data) and there is no data update record in the user historical access, the data is determined to be valid and the local cache data is directly used. If the data is dynamic data such as real-time generated service data, the update time of the local cache data is compared with the update time of the remote storage data. If the local data is not expired (or the user historical access has low time effectiveness requirement for the data), the data is determined to be valid. If the data is expired (or the user has high time effectiveness requirement), the data is determined to be invalid and the latest data needs to be requested from the remote storage. When the data is acquired, the valid data is directly read from the local cache according to the validity determination result, and the invalid data is acquired from the remote storage through the high-speed network. In the embodiment of the present application, the cache is calculated after the calculation, that is, after the calculation, the access probability of the user to the calculation result is predicted based on the access frequency extracted in step S100. The results with high access probability (such as the user historical access to the same calculation result) are cached in the local cache, which is convenient for subsequent rapid calling. The results with low access probability are stored in the remote storage to release the local cache space.

[0031] In step S600, the calculated data is dynamically scheduled and stored in layers between the local cache and the remote storage according to the access frequency and the heat factor. In the traditional storage and calculation separation architecture, the data storage often adopts a single level mode (such as all stored in the remote storage or the local cache, which leads to low storage efficiency of hot data and occupation of valuable cache resources by cold data. In the embodiment of the present application, the intelligent flow of data between the local cache and the remote storage is realized through dynamic scheduling and layered storage, which maximizes the cache utilization rate and the data access efficiency. Specifically, the access of the calculated data (including the local cache data and the remote storage data) is periodically (such as every 30 minutes) counted, and the following indexes are mainly counted: 1) Access frequency, the number of accesses and the number of users accessing the data in the current period are counted, and the higher the frequency, the higher the data heat; 2) Heat factor, that is, the time decay characteristic of the data (such as the recent access frequency is higher than the long-term, and the heat weight is higher), the user priority (such as the data accessed by the high-priority user has higher heat weight), and the comprehensive heat value of the data is calculated through a weighting algorithm. Regarding the execution of dynamic scheduling and layered storage, the data is divided into high-heat data, medium-heat data and low-heat data based on the comprehensive heat value, and the layered storage strategy is executed. Among them, the high-heat data is the data with a comprehensive heat value in the top 20%, which is scheduled to the local cache storage to ensure the shortest access delay. The medium-heat data is the data with a comprehensive heat value between 20% and 80%, which is dynamically allocated between the local cache and the remote storage according to the network status of the computing node and the remote storage (such as temporarily storing in the local cache when the network bandwidth is low, and migrating to the remote storage when the bandwidth is sufficient. The low-heat data is the data with a comprehensive heat value in the last 20%, which is scheduled to the remote storage (such as low-cost object storage) to release the local cache resources. According to the comprehensive heat value, the application dynamically schedules the triggering mechanism, that is, when the data heat value crosses the hierarchical threshold, such as the medium-heat data heat rising to the top 20%, the data migration is automatically triggered from the medium-heat layer to the high-heat layer without manual intervention. Step S700, real-time monitoring of system indicators and user behavior changes, dynamic optimization of scheduling algorithm, dynamic allocation of resources combined with user task sensitivity and activity, and pre-allocation of resources and pre-loading of associated data according to actual use, optimization of cache strategy.

[0032] The dynamic changes of data volume and application scenarios may cause the failure of static scheduling algorithm and resource allocation scheme (such as sudden surge of a certain type of task, which cannot be handled by the original algorithm). The embodiment of the present application ensures that the system always adapts to changing needs through real-time monitoring and dynamic optimization, and further reduces the delay and improves the system response speed through pre-allocation and pre-loading. Specifically, the real-time monitoring of system indicators and user behavior changes includes: System indicator monitoring: real-time collection of indicators such as computing node (CPU load, memory usage, cache hit rate), storage node (storage utilization, IOPS, data transmission delay), network link (bandwidth occupancy, transmission delay) through monitoring tools (such as Prometheus, Grafana); User behavior change monitoring: tracking dynamic changes of user operation behavior data (such as sudden increase in real-time task submission of a user, sudden increase in data access frequency of a certain type of data), identifying behavior mutations through anomaly detection algorithm (such as threshold detection based on standard deviation). Then, based on the monitoring data, the task scheduling algorithm of step S400 is optimized in real time, specifically including: if the monitoring finds that the load of a certain type of computing node is continuously too high, adjusting the node selection weight in the scheduling algorithm to reduce the task allocation proportion of this type of node; if the monitoring finds that the user's demand for real-time task response speed has improved, optimizing the task priority weight in the algorithm to preferentially allocate resources to real-time tasks; using reinforcement learning algorithm to continuously iterate the scheduling strategy with the reward goal of the lowest system overall delay and the highest resource utilization. In this embodiment, for resource pre-allocation, in combination with user task sensitivity (such as high sensitivity of financial transaction task, requiring low delay), activity (such as high task submission frequency of active user), the future demand of the user is predicted based on historical behavior, and for high sensitivity and high activity users, part of CPU and memory resources are pre-allocated in the computing node; for the upcoming business peak period (such as e-commerce promotion), the computing node resource pool is expanded in advance. In the embodiment of the application, for associated data preloading, the dependency relationship of the user task is analyzed (such as the user submitting task A, and task B dependent on the calculation result of A is submitted with high probability), and after the completion of task A, the associated data required by task B (such as the result of task A and the basic data dependent by task B) is preloaded to the local cache of the computing node, so that when the user submits task B, the preloaded data can be directly used, and data transmission delay is avoided. In the embodiment of the application, for cache strategy optimization, the cache strategy of steps S500 and S600 is optimized according to the monitored cache hit rate and data heat change: if the cache hit rate of a certain type of data continuously falls below 50%, the local cache priority of the data is reduced, and the cache occupation is reduced; if the timeliness requirement of a user for a certain type of dynamic data is reduced, the local cache validity period of the data is extended, and the cache hit rate is improved.

[0033] In a further embodiment, the step S400 specifically comprises: S401, the marked user data is analyzed by the intelligent scheduling layer to obtain a computing task, and the computing task is analyzed to extract a data set required by the computing task and key information of computing resource requirement; In the embodiment of this step, the intelligent scheduling layer first analyzes the marked user data into a specific computing task to be performed, and then picks out key information from the task, including a data set required by the task and a computing resource (such as CPU and memory) required to complete the task.

[0034] S402, according to the key information of the computing task requirement and the current state of the system, a computing node for executing the computing task is preliminarily determined, and the task is sent to the corresponding computing node.

[0035] In the embodiment of this step, in combination with the key requirement of the computing task (what data and how many resources are required) and the current situation of the system (such as whether each computing node is busy), the computing nodes that can approximately execute the task are selected, and then the task is sent to these nodes.

[0036] In the embodiment of the application, further, the step S500 specifically comprises: S501, the corresponding computing node receives the analyzed computing task, and checks whether the required data corresponding to the computing task exists in the local cache; In the embodiment of the present step, the computing node responsible for processing the task first receives the parsed computing task, and then checks whether the data required by the task exists in the local cache of the computing node.

[0037] S502, when the local cache exists the required data and the data is valid, the required data is directly read from the local cache for calculation; In the embodiment of the present step, if the data required by the computing task exists in the local cache and the data is valid (for example, not expired and the latest), the data is directly taken from the local cache for calculation.

[0038] S503, when the local cache does not exist the required data or the data is invalid, the computing node sends a data request to the remote storage, and acquires the required data from the remote storage; In the present step, if the local cache does not have the data required by the task or has invalid data, the computing node sends a request to the remote storage to acquire the data required by the task from the remote storage.

[0039] S504, the required data acquired from the remote storage is cached to the local cache for calculation, and the cache management strategy is updated according to the characteristics of the data and the access mode; In the present step, the data acquired from the remote storage is first stored in the local cache of the computing node, and then the data is used for calculation; meanwhile, the rules for managing the cache are updated according to the characteristics of the data (for example, data type and size) and the mode of accessing the data (for example, access frequency and read-write mode).

[0040] S505, and the calculated data is cached in the local or remote storage according to the user access probability.

[0041] In the present step, after the calculation is completed, the possibility of future user access to the calculation result is observed: if the access probability is high, the result is stored in the local cache; if the access probability is low, the result is stored in the remote storage.

[0042] In the further embodiment of the present application, the step S600 specifically comprises: S601, the calculated data is dynamically scheduled and stored in layers between the local cache and the remote storage according to the access frequency and the heat factor of the data; Based on the traditional storage-computing separation architecture, data storage is always one-size-fits-all (either all stored in the local cache or all stored in the remote storage), which leads to high access delay when hot data exists in the remote storage and resource waste when cold data occupies the local cache. In the embodiment of the present step, the intelligent flow and layered storage of data between the local cache and the remote storage are realized by combining the data access frequency and the heat, which not only ensures the fast access of hot data, but also efficiently utilizes the cache resources, solving the efficiency pain point of single-level storage. Specifically, in this embodiment, the evaluation of data access frequency and popularity factors involves first creating a popularity profile of the calculated data, and then performing access frequency statistics. Within a set period (such as 1 hour or 12 hours), the number of times the data is accessed by users or subsequent tasks and the average access interval are statistically analyzed. The more times and the shorter the interval, the higher the access frequency. Then, the popularity factors are added. Specifically, the time attributes of the data (recent access has a higher weight than older access) and the importance of related tasks (data that supports core business tasks has a higher weight) are combined to calculate the comprehensive popularity value of the data and quantify the importance of the data. Then, dynamic scheduling and tiered storage execution are performed. Specifically, data can be allocated to the corresponding storage level and dynamically adjusted based on the overall popularity value. In this embodiment, data with high overall popularity values ​​can be classified as hot data and preferentially allocated to the local cache; data with medium popularity values ​​can be classified as warm data and dynamically allocated according to the system network conditions (e.g., stored remotely when bandwidth is sufficient, and temporarily stored locally when bandwidth is limited); data with low popularity values ​​can be classified as cold data and allocated to remote storage.

[0043] When data popularity changes across a certain level threshold (e.g., warm data popularity rises to hot data range, hot data popularity drops to warm data range), data migration is automatically triggered (e.g., migration from remote storage to local cache, migration from local cache to remote storage), without manual intervention, ensuring that data is always stored in the optimal storage location.

[0044] S602. For frequently accessed hot data, prioritize storing it in the local cache, and replace and update it according to the cache capacity and data priority. Hot data refers to core data that is frequently accessed by users or the system. Storing it in remote storage would frequently cause data transmission delays. At the same time, local cache capacity is limited, and disordered storage would cause important hot data to be squeezed out. This step prioritizes local storage and replaces and updates as needed, which ensures low-latency access to hot data while maximizing the use of limited cache space. In this invention, frequently accessed data is preferentially stored in the local cache. After identifying frequently accessed data (such as real-time order data during a major e-commerce platform promotion or frequently queried user basic information), it is directly stored in the local cache of the computing node. Access frequency tags and priority tags are associated with the stored data (e.g., high priority is set for frequently accessed data supporting real-time transactions, and medium priority is set for ordinary frequently queried data) to facilitate subsequent management. Utilizing the high-speed read / write characteristics of the local cache, subsequent access requests do not need to call remote storage; data can be read directly from the local cache, significantly reducing access latency. In the embodiment of the present step, regarding the replacement and update based on the cache capacity and data priority, when the local cache capacity is about to be full or the hot data is updated, dynamic adjustment is performed according to the rules. Regarding cache replacement, a priority+LRU (Least Recently Used) hybrid strategy is adopted, when the cache space is insufficient, the hot data with low priority and least recently accessed is preferentially eliminated (such as migrating the hot data with medium priority and not accessed within 1 hour to the remote storage), to make room for new high-priority hot data. Regarding data update, if the hot data in the local cache changes (such as user information update, order status change), the local cache data is updated in real time, and the update time is marked; at the same time, according to the data synchronization strategy (such as real-time synchronization of core data, and timing synchronization of ordinary data), the updated data is synchronized to the remote storage, to ensure the consistency of local and remote data.

[0045] S603, for the cold data that is not frequently accessed, is stored in the remote storage.

[0046] In the embodiment of the present application, through data access statistics, cold data that is not frequently accessed is screened out, and the cold data determination standard can be set, such as less than 2 times of access within 72 hours, no access record for 3 consecutive days, or the comprehensive hotness value is lower than the set threshold (such as lower than 20% of all data hotness values). Typical cold data scenarios include: last year's financial report data, user behavior log 3 months ago, historical data of offline business, etc. In the embodiment of the present step, regarding the execution of cold data remote storage, the identified cold data is migrated and stored to the remote storage system (such as object storage, distributed archive storage); before migration, the cold data is compressed and encrypted (such as using GZIP compression to reduce storage occupancy, and using AES encryption to ensure data security), to reduce the space cost and security risk of remote storage; after storage, cold data index is established, to record the storage path, compression format, encryption key, last access time and other information of the cold data, to facilitate subsequent access requirements, to quickly locate and retrieve data; the remote storage system can be independently expanded, without relying on the computing node, to support the storage requirements of a large amount of cold data for a long time, and the storage cost is much lower than the local cache, to realize the low-cost management of cold data.

[0047] In the further embodiment of the present application, the step S700 specifically includes: S701, controlling the intelligent scheduling layer to monitor the load condition, network bandwidth, delay performance index of each computing node, and the read-write performance and data distribution condition of the remote storage in real time; In the embodiment of the present application, the intelligent scheduling layer detects the key situation in the system, on the one hand, detects the state of each computing node, such as whether busy (load situation), network data transmission fast or slow (bandwidth), data transmission delay or not (delay); on the other hand, detects the performance of remote storage, such as how fast the data is read and written, and where the data exists (distribution situation).

[0048] S702, according to the monitoring data, using an intelligent scheduling algorithm based on load balancing or data locality scheduling algorithm, dynamically adjusting the allocation and scheduling strategy of the task, and distributing the task to the corresponding idle computing node, while controlling to ensure the locality access of data.

[0049] In the embodiment of the present application, according to the above monitored situation, the intelligent scheduling algorithm (for example, an algorithm that makes each computing node busy evenly, or an algorithm that makes the task use local data as much as possible) is used to flexibly adjust the rules of task allocation and scheduling. The new task is distributed to the computing node which is relatively idle at present, and the task is allowed to use local data as much as possible without taking too much remote data.

[0050] In the further embodiment of the present application, the step S500 further comprises: S511, the corresponding computing node receives the parsed task, and obtains the required data according to the scheduling strategy, and controls to start executing the computing task; In the embodiment of the present application, the computing node responsible for processing the task first receives the parsed task, and then obtains the data required by the task according to the previously determined scheduling rule, and then starts executing the computing task.

[0051] Specifically, as a starting link of task execution, this step makes the computing node clear what task to do and what data to use, and formally enters the computing process, which is a key bridge connecting task scheduling and actual computing, and ensures smooth transition of the task from the scheduling layer to the execution layer. Specifically, the parsed task is first received, and the intelligent scheduling layer has parsed the user task into a format that can be recognized by the computing node (such as explicitly indicating the task type and computing target), and the corresponding computing node will receive the parsed task to clearly indicate the specific work to be executed.

[0052] Then, the data is obtained according to the scheduling strategy, and the computing node will follow the scheduling strategy formulated by the intelligent scheduling layer (such as preferentially using local cache data, and taking a specified link if remote data is needed), to obtain the necessary data of the task, which can be directly reading the existing effective data in the local cache, or sending a request to the remote storage to obtain the missing data, so as to avoid blind data taking leading to delay. Then, the computing task is started, and when the required data is ready, the computing node calls its own computing resources (CPU, memory, etc.), and starts to perform the computation according to the logic required by the task (such as data operation, model processing, etc.).

[0053] S512, during the task execution, the computing node dynamically adjusts the cached data and requests new data according to actual needs; In the embodiment of the present application, during the task computation, the computing node flexibly adjusts according to actual conditions: for example, if the previously cached data is not enough, the cached content is adjusted, or if the task needs new data, the new data is requested.

[0054] Specifically, regarding the dynamic adjustment of the cached data, the computing node of the present application will check in real time whether the data in the current cache is suitable for the execution progress; for example, if some part of the data in the cache has been used up and is not needed later, this part of the data is cleaned up to release the cache space; if a certain type of data is frequently called, it is preferentially retained in the cache to avoid repeated reading. In the embodiment of the present application, regarding the request for new data, when the computation proceeds to a certain stage, it is found that the existing data is not enough to support the subsequent operation (for example, when processing user order data, associated logistics data needs to be supplemented), the computing node will actively send a new data request to the remote storage or other data node, and after obtaining the required data, the computation is continued to ensure that the task is not interrupted.

[0055] S513, when the computing task is completed, the computing node returns the computation result to the user terminal application program, and stores the result data in the remote storage or the local cache according to needs.

[0056] In the embodiment of the present application, after the computing task is completed, the computing node sends the calculated result to the terminal application (such as a mobile phone APP or computer software) used by the user; at the same time, the storage location of the result data is determined according to the situation, that is, the result data is stored in the remote storage or the local cache.

[0057] Specifically, after the computing task is completed, the computing node sends the final result (such as the order amount calculation result or the data analysis report) to the terminal application program used by the user through the network link, such as the shopping APP on the mobile phone or the office software on the computer, so that the user can intuitively obtain and use the result. The present application also determines the storage location according to the possibility of subsequent access of the result data and the system rules; if it is judged that the user or the subsequent task may frequently access the result (such as the temporary statistical result of real-time business), the result is stored in the local cache for easy and fast retrieval; if the result is probably used only once or needs to be archived for a long time (such as historical bill data), the result is stored in the remote storage, which saves the local cache space and can be safely and permanently saved.

[0058] In further embodiments of the present application, the step S700 is followed by: S701, controlling data analysis and performance evaluation according to the running condition of the system and the execution result of the task, to optimize the scheduling algorithm and the cache management strategy; In embodiments of the present application, on the one hand, real-time data of system running is collected, including CPU load of each computing node, memory usage, network delay, read and write IOPS of remote storage, data transmission bandwidth, and cache hit rate and data replacement frequency; on the other hand, task execution result data is collected, such as task completion time, task failure rate, and data transmission time consumption (local cache reading time consumption vs. remote storage reading time consumption). Then, through data comparison (such as node load difference in different time periods, execution efficiency difference of different task types), correlation analysis (such as the correlation between cache hit rate and data heat, the correlation between task delay and node load), the problem is located. For example, it is found that the execution delay of a certain type of real-time task is relatively high, and further analysis finds that the scheduling algorithm often allocates it to a high-load node; or it is found that the cache hit rate is continuously lower than 80%, which is due to the fact that the cache replacement strategy does not fully consider the long-term access rule of data. In embodiments of the present application, when performance evaluation is performed, evaluation indexes are set, which can be set around the core target of the system (low delay, high resource utilization, high reliability), and quantitative evaluation indexes are set, such as task average execution delay, computing node load balancing degree, cache hit rate, and remote storage access frequency. Then, the actual running data is compared with the preset performance target (such as task average execution delay ≤ 500 ms, cache hit rate ≥ 90%), and the performance of the existing strategy is judged; if the indexes meet the standard, it means that the strategy is suitable for the current system state; if the indexes do not meet the standard, such as task delay exceeding 1000 ms and load balancing degree being lower than 70%, it is marked as an optimization item. In the present application, the strategy optimization can be performed by optimizing the scheduling algorithm, adjusting the algorithm logic according to the evaluated problems, for example, if it is found that the node load is not balanced, the weight distribution of the load balancing algorithm is optimized (such as increasing the weight of the real-time load of the node to avoid allocating new tasks to nodes with load exceeding 80%); if it is found that the task data transmission delay is high, the scheduling logic based on data locality is optimized (such as preferentially allocating tasks to nodes that have cached data required by the task). And the cache management strategy is optimized, specifically, if the cache hit rate is low, the cache replacement rule is adjusted (such as upgrading from pure LRU strategy to LRU+data heat weight strategy to make high-heat data more difficult to be eliminated; if it is found that some cold data occupies the cache for a long time, the cache data screening logic is optimized (such as shortening the cache validity period of cold data, and automatically migrating to remote storage when the validity period expires).

[0059] S702, predict the access mode and resource demand of future tasks through mining and analysis of historical data, and perform resource pre-allocation and cache pre-loading in advance.

[0060] In the embodiment of the application, long-term accumulated historical data is collected, including user task submission time, task type (real-time / off-line), task data volume, task dependent data identifier, and user data access time, access frequency, and access mode (read / write). Then, through time series analysis, it is found that 9:00-11:00 every day is the peak of real-time task submission, and off-line analysis tasks surge after 18:00 every Friday, and the time law of task submission is mastered; through association mining, it is found that 80% of users will submit B type tasks dependent on A type task results within 10 minutes after submitting A type task, and the access frequency of certain data will increase by 5 times at the end of each month, and the task access mode and data heat change law are mastered. Then, future demand prediction is performed, which includes: Access mode prediction, which predicts the task access mode in a future period of time based on historical laws, for example, predicting that there will be a large number of real-time order processing tasks tomorrow from 9:00 to 11:00, and these tasks will frequently access user recent consumption data; or predicting that there will be a large number of monthly financial analysis tasks after 18:00 next Friday, which need to access transaction data in the past 30 days. In the embodiment of the application, the future demand prediction also includes resource demand prediction, which estimates the required resource scale by combining the predicted task volume and task type, for example, predicting that 10 high-CPU-performance computing nodes are needed in addition during the real-time task peak period tomorrow, and the remote storage needs to provide a read IO capability of 2000 times per second for the monthly financial analysis task. In the embodiment of the application, for resource pre-allocation, the system resource configuration is adjusted in advance according to the resource demand prediction; if a task peak is predicted, the idle computing nodes are woken up and added to the resource pool, or the CPU / memory resources of the computing nodes are temporarily expanded; if it is predicted that the read / write pressure of the remote storage is large, the read / write priority of the storage node is adjusted in advance to ensure the data access bandwidth of the core task. In the embodiment of the application, for cache pre-loading, the data frequently accessed in the future is loaded into the local cache of the corresponding computing node in advance according to the access mode prediction; for example, it is predicted that real-time order tasks will frequently access user recent consumption data, so this part of data is migrated from the remote storage to the cache of the computing node responsible for processing order tasks in advance; it is predicted that the financial analysis task needs to access transaction data in the past 30 days, so these data are pre-loaded into the cache of the computing node responsible for the analysis task to avoid reading from the remote storage during task execution, which greatly shortens the data acquisition delay.

[0061] The application will be further described in detail through specific application examples as follows: The application example two specifically provides a memory-computing separation method based on multi-level cache and intelligent scheduling, and comprises the following steps: In step S11, operation behavior data of a user is collected, including operation behavior data such as task submission and data access of the user, and the collected operation behavior data is cleaned, integrated and extracted for features such as access frequency and task preference, and then step S12 is entered.

[0062] In step S12, the collected operation behavior data of the user is stored in a distributed storage, and a cache space of a computing node is dynamically allocated according to the user behavior data, and then step S13 is entered. In step S13, the data stored in step S12 is combined to perform feature and user access mode classification marking data, and marked operation behavior data is formed, and then step S14 is entered. In step S14, the marked data in step S13 is analyzed to perform intelligent scheduling layer analysis tasks, and historical data records and preferences of the user are associated, and adaptive computing nodes are screened and allocated, and then step S15 is entered.

[0063] In step S15, when analyzing the operation behavior data of the user, the computing node checks the local cache, and the data validity is combined with the historical access judgment of the user; if the number is missing or invalid, the remote storage is requested, the data is obtained, and then the data is cached according to the user access probability, and then step S16 is entered.

[0064] In step S16, according to the analysis result of the user data in step S15, the data is stored in layers according to the user access frequency and user behavior characteristics, the associated data is preloaded, and the cache strategy is optimized, and then step S17 is entered.

[0065] In step S17, system indicators and user behavior changes are monitored in real time, the scheduling algorithm is dynamically optimized, resources are dynamically allocated in combination with user task sensitivity and activity, and resources are pre-allocated and preloaded in combination with actual use of the user (for example, a user is particularly active at 7:00-9:00 in the evening, and the system automatically preloads data for the user to facilitate the user behavior).

[0066] Exemplary device As shown in Figure 2 The application example two specifically provides a memory-computing separation method based on multi-level cache and intelligent scheduling, and comprises the following steps: The data collection module 310 is used for collecting operation behavior data of a user, and cleaning, integrating and extracting access frequency and task preference features of the collected operation behavior data; wherein the operation behavior data comprises task submission behavior data and data access behavior data. The dynamic allocation module 320 is configured to store the operation behavior data in a distributed manner and dynamically allocate cache spaces of the computing nodes according to the operation behavior data of the user. The classification and marking module 330 is configured to classify and mark the distributed operation behavior data according to features and user access patterns, to form marked user data. The intelligent scheduling and association module 340 is configured to perform intelligent scheduling layer analysis on the marked user data, associate the user historical data records and preferences, filter and distribute the adaptive computing nodes. The computing module 350 is configured to receive the analyzed tasks of the corresponding computing nodes, check the local cache, combine the user historical access judgment data to determine the validity, obtain the corresponding data from the local or remote storage according to the validity of the local corresponding data, and perform calculation on the data, and cache the calculated data in the local or remote storage according to the user access probability. The dynamic scheduling module 360 is configured to perform dynamic scheduling and hierarchical storage of the data between the local cache and the remote storage according to the access frequency and the heat factor of the calculated data. The optimization module 370 is configured to monitor system indicators and user behavior changes in real time, dynamically optimize the scheduling algorithm, dynamically allocate resources in combination with the user task sensitivity and activity, and in combination with the actual use, pre-allocate resources and pre-load associated data, optimize the cache strategy, and the specific implementation is as described above.

[0067] Based on the above embodiment, the application further provides a server, and a principle block diagram thereof can be shown as Figure 3 The server includes a processor, a memory, a network interface, a display screen and a database connected through a system bus.

[0068] The memory stores one or more programs configured to be executed by the processor to implement the above-mentioned multi-level cache and intelligent scheduling based storage and calculation separation method.

[0069] The server refers to an intelligent computer with data processing capability. The memory can be a flash memory, a hard disk or a cloud storage space, which is used to store program codes and store various data such as user operation behavior data and calculated data. The processor can be a central processing unit, which is used to execute algorithm logic in the program, such as data acquisition, data marking, task analysis and dynamic scheduling steps. The program contains the multi-level cache and intelligent scheduling based storage and calculation separation method.

[0070] In further embodiments, a server of the present embodiments includes a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors, the one or more programs comprising instructions for: A collecting operation behavior data of a user, and cleaning and integrating the collected operation behavior data and extracting access frequency and task preference characteristics; wherein the operation behavior data includes task submission behavior data and data access behavior data; B performing distributed storage on the operation behavior data, and dynamically allocating cache spaces of computing nodes according to the operation behavior data of the user; C performing feature and user access mode classification labeling on the distributed operation behavior data, to form labeled user data; D performing intelligent scheduling layer analysis tasks on the labeled user data, and simultaneously associating user historical data records and preferences, filtering and distributing adaptive computing nodes; E receiving the analyzed tasks by the corresponding computing nodes, checking local caches, combining user historical access judgment data, and according to the effectiveness of the local corresponding data, obtaining corresponding data from the local or requesting remote storage, and performing calculation on the corresponding data, and according to the user access probability, caching the calculated data in the local or remote storage; F performing dynamic scheduling and hierarchical storage of data between the local cache and the remote storage according to the access frequency and the heat factor of the calculated data; G real-time monitoring of system indicators and user behavior changes, dynamic optimization of scheduling algorithms, dynamic allocation of resources combined with user task sensitivity and activity, and combination of actual use cases, pre-allocation of resources and pre-loading of associated data, optimization of cache strategies, as described above.

[0071] The step of performing intelligent scheduling layer analysis tasks on the labeled user data, simultaneously associating user historical data records and preferences, filtering and distributing adaptive computing nodes, includes: Performing intelligent scheduling layer analysis tasks on the labeled user data, obtaining computing tasks, and performing analysis on the computing tasks, extracting data sets and computing resource demand key information required by the computing tasks; According to the demand key information of the computing tasks and the current state of the system, initially determining the computing nodes for executing the computing tasks, and sending the tasks to the corresponding computing nodes.

[0072] The step of receiving the parsed task by the corresponding computing node, checking the local cache, judging the data validity in combination with the user historical access, obtaining the corresponding data from the local or requesting the remote storage according to the validity of the local corresponding data, and caching the calculated data in the local or remote storage according to the user access probability comprises: The corresponding computing node receives the parsed computing task and checks whether the required data corresponding to the computing task exists in the local cache; When the required data exists in the local cache and is valid, the required data is directly read from the local cache for calculation; When the required data does not exist in the local cache or the data is invalid, the computing node is controlled to send a data request to the remote storage, and the required data is obtained from the remote storage; The required data obtained from the remote storage is cached in the local cache for calculation, and the cache management strategy is updated according to the characteristics and access mode of the data; And the calculated data is cached in the local or remote storage according to the user access probability.

[0073] The step of dynamically scheduling and hierarchically storing the calculated data according to the access frequency and heat factors between the local cache and the remote storage comprises: The calculated data is dynamically scheduled and hierarchically stored between the local cache and the remote storage according to the access frequency and heat factors of the data; For frequently accessed hot data, the data is preferentially stored in the local cache, and is replaced and updated according to the cache capacity and the priority of the data; For infrequently accessed cold data, the data is stored in the remote storage.

[0074] The step of monitoring system indicators and user behavior changes in real time, dynamically optimizing the scheduling algorithm, dynamically allocating resources in combination with the user task sensitivity and activity, and optimizing the cache strategy in combination with the actual use and preloading associated data comprises: The intelligent scheduling layer is controlled to monitor the load condition, network bandwidth, delay performance indicators of each computing node, and the read-write performance and data distribution condition of the remote storage in real time; According to the monitoring data, the intelligent scheduling algorithm based on the load balancing scheduling algorithm or the data locality scheduling algorithm is adopted to dynamically adjust the allocation and scheduling strategy of the task, and the task is allocated to the corresponding idle computing node, while the data locality access is ensured.

[0075] The corresponding computing node receives the parsed task, checks the local cache, judges the data validity in combination with the user historical access, obtains the corresponding data from the local or requests the remote storage according to the validity of the local corresponding data, performs the calculation on the corresponding data, and caches the calculated data in the local or the remote storage according to the user access probability. The corresponding computing node receives the parsed task, and obtains the required data according to the scheduling strategy, and controls to start executing the computing task. In the task execution process, the computing node dynamically adjusts the cached data and requests new data according to the actual demand. When the computing task is executed, the corresponding computing node returns the calculation result to the user terminal application program, and stores the result data into the remote storage or the local cache according to the need.

[0076] The step of dynamically optimizing the scheduling algorithm in combination with the user task sensitivity and the activity, dynamically allocating resources, and optimizing the cache strategy in combination with the actual use and the preloaded associated data after the real-time monitoring system index and the user behavior change includes: Controlling to perform data analysis and performance evaluation according to the running situation of the system and the execution result of the task, so as to optimize the scheduling algorithm and the cache management strategy; Through the mining and analysis of the historical data, the access mode and the resource demand of the future task are predicted, and the resource pre-allocation and the cache preloading are performed in advance, which are specifically as described above.

[0077] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0078] As can be seen from the above, the present application provides a memory-compute separation method and device based on multi-level cache and intelligent scheduling, a server and a storage medium. The present application focuses on solving the problems of high data transmission delay and insufficient cache management in the existing memory-compute separation architecture. A memory-compute separation system and method based on multi-level cache and intelligent scheduling are proposed to significantly improve the performance of the system, optimize resource utilization, and provide a more efficient and reliable solution for large-scale data processing, thereby better meeting the urgent needs of high-performance computing and storage architecture in the current data era.

[0079] The above only describes the embodiments of the present application and does not limit the protection scope of the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A memory-compute separation method based on multi-level cache and intelligent scheduling, characterized in that, The application comprises the following steps: Collecting operation behavior data of a user, cleaning and integrating the collected operation behavior data, and extracting access frequency and task preference characteristics; wherein the operation behavior data comprises task submission behavior data and data access behavior data; Storing the operation behavior data in a distributed manner, and dynamically allocating cache spaces of computing nodes according to the operation behavior data of the user; Classifying and marking the operation behavior data stored in a distributed manner according to characteristics and user access patterns, to form marked user data; Intelligently scheduling layer analysis tasks of the marked user data, and associating user historical data records and preferences, to filter and distribute adaptive computing nodes; A corresponding computing node receives the analyzed task, checks a local cache, judges data validity in combination with user historical access, acquires corresponding data from the local cache or remote storage according to the validity of the local corresponding data, performs calculation on the corresponding data, and caches the calculated data in the local cache or remote storage according to user access probability; According to access frequency and heat factors, the calculated data is dynamically scheduled and stored in a hierarchical manner between the local cache and the remote storage; Real-time monitoring of system indicators and user behavior changes, dynamic optimization of scheduling algorithms, dynamic allocation of resources in combination with user task sensitivity and activity, pre-allocation of resources and pre-loading of associated data in combination with actual use, and optimization of cache strategies.

2. The method of claim 1, wherein, The step of intelligently scheduling layer analysis tasks of the marked user data, and associating user historical data records and preferences, to filter and distribute adaptive computing nodes comprises the following steps: Intelligently scheduling layer analysis tasks of the marked user data, obtaining computing tasks, and analyzing the computing tasks to extract data sets and computing resource demand key information required by the computing tasks; According to the demand key information of the computing tasks and the current state of the system, a computing node for executing the computing tasks is preliminarily determined, and the task is sent to the corresponding computing node.

3. The method of claim 1, wherein, The step of the corresponding computing node receiving the analyzed task, checking a local cache, judging data validity in combination with user historical access, acquiring corresponding data from the local cache or remote storage according to the validity of the local corresponding data, performing calculation on the corresponding data, and caching the calculated data in the local cache or remote storage according to user access probability comprises the following steps: The corresponding computing node receives the analyzed computing task, checks whether the required data corresponding to the computing task exists in the local cache; When the local cache contains the required data and the data is valid, the required data is directly read from the local cache for calculation; When the local cache does not contain the required data or the data is invalid, the computing node sends a data request to the remote storage; the required data is acquired from the remote storage; The required data acquired from the remote storage is cached in the local cache for calculation, and the cache management strategy is updated according to the characteristics and access patterns of the data; And the calculated data is cached in the local cache or remote storage according to user access probability.

4. The method of claim 1, wherein, The step of dynamically scheduling and hierarchically storing the calculated data according to access frequency and heat factors between the local cache and the remote storage comprises the following steps: The calculated data is dynamically scheduled and stored in layers between the local cache and the remote storage according to the access frequency and the hotness factor of the data; Hot data frequently accessed is preferentially stored in the local cache, and is replaced and updated according to the cache capacity and the priority of the data; Cold data not frequently accessed is stored in the remote storage.

5. The method of claim 1, wherein, The steps of the real-time monitoring system index and user behavior change, dynamically optimizing the scheduling algorithm, dynamically allocating resources in combination with user task sensitivity and activity, and pre-allocating resources and pre-loading associated data in combination with actual use, and optimizing the cache strategy include: The intelligent scheduling layer controls the load condition, network bandwidth, delay performance index of each computing node, and the read-write performance and data distribution condition of the remote storage in real time; According to the monitoring data, the intelligent scheduling algorithm based on the load balancing scheduling algorithm or the data locality scheduling algorithm is adopted to dynamically adjust the allocation and scheduling strategy of the task, and the task is allocated to the corresponding idle computing node, while the locality access of the data is controlled and ensured.

6. The method of claim 1, wherein, The steps of the corresponding computing node receiving the parsed task, checking the local cache, judging the data validity in combination with the user historical access, obtaining the corresponding data from the local or the remote storage according to the validity of the local corresponding data, and caching the calculated data in the local or the remote storage according to the user access probability include: The corresponding computing node receives the parsed task and obtains the required data according to the scheduling strategy, and controls to start executing the computing task; During the task execution process, the computing node dynamically adjusts the cache data and requests new data according to the actual demand; When the computing task is executed, the corresponding computing node returns the calculation result to the user terminal application program, and stores the result data into the remote storage or the local cache as needed.

7. The method of claim 1, wherein, The steps after the real-time monitoring system index and user behavior change, dynamically optimizing the scheduling algorithm, dynamically allocating resources in combination with user task sensitivity and activity, and pre-allocating resources and pre-loading associated data in combination with actual use, and optimizing the cache strategy include: Data analysis and performance evaluation are performed to optimize the scheduling algorithm and the cache management strategy according to the running condition of the system and the execution result of the task; By mining and analyzing the historical data, the access mode and resource demand of future tasks are predicted, and resource pre-allocation and cache pre-loading are performed in advance.

8. A compute-in-memory device based on multi-level cache and intelligent scheduling, characterized in that, The device includes: A data acquisition module is configured to acquire operation behavior data of a user, clean and integrate the acquired operation behavior data, and extract access frequency and task preference characteristics; wherein the operation behavior data includes task submission behavior data and data access behavior data; A dynamic allocation module is configured to store the operation behavior data in a distributed storage, and dynamically allocate cache spaces of computing nodes according to the operation behavior data of the user; A classification and labeling module is configured to classify and label the operation behavior data stored in a distributed manner, and form labeled user data. The intelligent scheduling and association module is configured to perform intelligent scheduling layer analysis on the marked user data, associate the user historical data record and preference, filter and distribute the adaptive computing nodes; The computing module is configured to receive the analyzed task by the corresponding computing node, check the local cache, judge the data validity in combination with the user historical access, obtain the corresponding data from the local or remote storage according to the data validity of the local corresponding data, perform the calculation on the data, and cache the calculated data in the local or remote storage according to the user access probability; The dynamic scheduling module is configured to perform the dynamic scheduling and hierarchical storage of the data between the local cache and the remote storage according to the access frequency and the heat factor of the calculated data; The optimization module is configured to monitor the system index and the user behavior change in real time, dynamically optimize the scheduling algorithm, dynamically distribute the resources in combination with the user task sensitivity and the activity, pre-distribute the resources and pre-load the associated data in combination with the actual use, and optimize the cache strategy.

9. A server, characterized by The one or more programs stored in the memory and configured to be executed by the one or more processors include programs for executing the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device can execute the method of any one of claims 1-7.

Citation Information

Cited By

  • Data load balancing method and system in data cluster

    CN121387443A

  • A method and system for data load balancing in a data cluster

    CN121387443B

  • Task scheduling system of data stream processor

    CN121807572A

  • A data flow processor task scheduling system

    CN121807572B