Computing storage split server control method for low latency interaction

Abstract

The application belongs to the technical field of server control, and particularly relates to a computing and storage split type server control method for low-delay interaction, which constructs a split type architecture with separated computing and storage, monitors interaction requests in real time, extracts type features, data volume space-time distribution features and priority decay coefficients to construct a request feature vector, dynamically allocates communication bandwidth based on the features, corrects the allocation result in combination with the priority decay coefficients, predicts data access probability through an LSTM model fused with an attention mechanism, preloads target data shards to a multi-level cache according to access heat, collects computing node data in real time to perform dynamic health score, and selects an optimal target node to migrate interaction requests and associated data based on a real-time communication delay matrix when the score is lower than a threshold. The application significantly improves the low-delay interaction performance of the server, improves bandwidth utilization and data access efficiency, and guarantees stable operation of the system through dynamic bandwidth allocation, intelligent preloading and node health management.

Description

Technical Field

[0001] This invention belongs to the field of server control technology, and particularly relates to a control method for a computing and storage-separated server for low-latency interaction. Background Technology

[0002] With the rapid development of technologies such as 5G communication, artificial intelligence, and the Internet of Things, various real-time interactive scenarios place extremely high demands on the response speed and data processing efficiency of servers. Traditional servers mostly adopt an integrated computing and storage architecture, with data processing and storage relying on local hardware resources. In high-concurrency, high-volume data interaction scenarios, resource contention can easily lead to increased latency, making it difficult to meet the low latency requirements of milliseconds or even microseconds.

[0003] To address this issue, a separate computing and storage architecture has gradually become a research hotspot. This architecture achieves independent resource expansion and flexible scheduling by physically separating the computing and storage modules. However, existing control methods for separate servers still have the following shortcomings: In terms of communication bandwidth allocation, traditional methods mostly adopt static allocation strategies, allocating bandwidth resources based on fixed request types or priorities, which cannot dynamically adapt to the spatiotemporal fluctuations of request data volume and the characteristics of priority decay. For example, using the same bandwidth allocation logic for real-time interactive requests and batch processing requests can easily lead to delays in critical interactions due to insufficient bandwidth; at the same time, it does not consider the pulse-like growth characteristics of data volume, making it difficult to cope with transmission congestion caused by sudden traffic surges.

[0004] Regarding data preloading mechanisms, existing technologies mostly rely on simple access frequency statistics to determine whether to preload, lacking comprehensive analysis of multi-dimensional characteristics such as the popularity of related requests and time decay patterns, resulting in insufficient preloading accuracy. Some methods employ fixed preloading strategies without dynamically adjusting the loading rate based on real-time bandwidth resources and computing node load, which may consume excessive system resources and negatively impact the processing efficiency of current requests.

[0005] When selecting migration targets, existing methods often only consider topological distance or communication latency as a single factor, without taking into account the real-time health status and load redundancy of candidate nodes. This may lead to new nodes quickly becoming overloaded after migration, failing to achieve true load balancing. Furthermore, existing methods handle the type characteristics of interaction requests and data volume distribution characteristics in a relatively simple manner during feature extraction and model prediction, lacking deep semantic parsing and spatiotemporal correlation modeling, resulting in insufficient accuracy in subsequent resource scheduling decisions.

[0006] Therefore, how to construct a control method that can dynamically perceive request characteristics, accurately allocate communication resources, intelligently predict data access needs, and realize health management and efficient migration of computing nodes is a key technical issue for improving the low-latency interactive performance of computing and storage split servers. Summary of the Invention

[0007] The purpose of this invention is to provide a control method for a separate computing and storage server for low-latency interaction, which can be used to build a system that can dynamically sense request characteristics, accurately allocate communication resources, and intelligently predict data access needs, thereby achieving health management and efficient migration of computing nodes.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for controlling a compute-storage split-type server for low-latency interaction includes the following steps: S1: Construct a split architecture that includes computing modules and storage modules. The computing modules and storage modules establish a communication connection through a high-speed interconnect network. This architecture pre-sets an initial communication protocol and resource scheduling interface. S2: Real-time monitoring of interaction requests between the computing module and the storage module; dynamically extracting the type characteristics, spatiotemporal distribution characteristics of data volume, and priority attenuation coefficient of interaction requests through the resource scheduling interface; constructing a request feature vector. S3: Determine the basic bandwidth allocation weight based on the type characteristics. The basic weight of real-time interactive requests is higher than that of batch processing requests. Combine the spatiotemporal distribution characteristics of data volume to calculate the bandwidth demand fluctuation coefficient. Allocate elastic bandwidth pool resources to requests whose data volume increases in a pulse manner over time. S4: Based on the priority decay coefficient, the allocation result is corrected. For requests whose priority decays over time, the bandwidth ratio is dynamically reduced according to the decay rate. The allocation result is synchronized to the interaction interface between the computing module and the storage module through the initial communication protocol. S5: Obtain historical access data from the server, extract multi-dimensional features including access frequency, related request popularity, and time decay factor, input them into an LSTM model with an attention mechanism to predict the access probability within a future preset time window, and execute step S6 when the access frequency is greater than the preset probability threshold. S6: Based on the allocation results corrected by S4, the target data is preloaded into the multi-level cache of the computing module by being sharded according to access popularity; S7: Real-time acquisition of specified data from each computing node in the computing module during the data processing process, and dynamic health score of each computing node based on the specified data. When the health score is lower than the preset threshold, the optimal target computing node is selected based on the real-time communication delay matrix to perform interactive requests and migration of related data.

[0009] Preferably, the specific process of dynamically extracting the type characteristics, spatiotemporal distribution characteristics of data volume, and priority attenuation coefficient of the interaction request through the resource scheduling interface in step S2 to construct the request feature vector is as follows: S21: Automatically identify the business scenario tags corresponding to interactive requests through deep learning models and map them into high-dimensional semantic vectors; S22: Combining sliding time windows and spatial clustering algorithms, the discrete request data volume is transformed into a distribution matrix with spatiotemporal correlation. The real-time request data volume is sliced ​​and sampled in units of windows, recording the amount of interaction data between each computing node and storage node within each window. The DBSCAN spatial clustering algorithm is used to cluster the sampled data, dividing nodes with similar geographical or topological distances into the same spatial cluster. The average data volume, data volume variance, and interaction frequency of node pairs within the cluster are calculated. Using the time window as the row index and the spatial cluster as the column index, the statistical data within the cluster are filled in chronological order to form a two-dimensional distribution matrix. The matrix element value is the standardized data volume of the corresponding spatial cluster within the time window. S23: Combine the SLA level of the request initiator, historical response time deviation, and current system load pressure to construct a dynamic attenuation function to calculate the priority attenuation coefficient; S24: The three types of features are adaptively weighted into a unified request feature vector through a feature fusion network, and the vector dimension is dynamically adjusted according to the request complexity.

[0010] Preferably, the specific process of step S3 is as follows: S31: Based on the high-dimensional semantic vector of the type features in step S2, the pre-set business scenario template vector is matched with the cosine similarity. The scenario template with the highest matching degree corresponds to the pre-set basic weight range. The specific weight value is dynamically determined within the range in combination with the real-time parameters of the request. S32: Calculate the bandwidth demand fluctuation coefficient by combining the spatiotemporal distribution characteristics of data volume: Extract the time series data of the spatial cluster to which the target request belongs from the spatiotemporal distribution matrix of data volume, calculate the data volume growth rate of three consecutive sliding windows, and determine the pulse growth when the standard deviation of the growth rate exceeds the preset fluctuation threshold and there are two adjacent windows where the data volume increase exceeds 50%. Calculate the bandwidth demand fluctuation coefficient K = (current window data volume − historical average data volume) / historical average data volume, and allocate temporary resources of K × base bandwidth from the elastic bandwidth pool according to the K value. The total capacity of the elastic bandwidth pool is 20%-30% of the total bandwidth, and a token bucket mechanism is used for dynamic scheduling.

[0011] Preferably, the specific process of correcting the allocation result based on the priority attenuation coefficient in step S4, and dynamically reducing the bandwidth proportion of requests whose priority decays over time according to the attenuation rate, is as follows: S41: Calculate the difference in priority decay coefficients for a request over two consecutive sliding time windows: Δλ = λ(t1) − λ(t2); Where λ(t1) is the priority decay coefficient of the sliding time window t1, λ(t2) is the priority decay coefficient of the sliding time window t2, t2>t1, and when Δλ>0, it is determined that the priority decays with time, and the decay rate is... v =Δλ·(t2−t1); S42: Determine the bandwidth proportion adjustment coefficient based on the attenuation rate. η =1− k · v , k This is a proportionality coefficient, ranging from 0.1 to 0.5, which is dynamically adjusted based on the total bandwidth utilization. S43: Multiply the initial bandwidth value after basic bandwidth allocation and elastic bandwidth pool resource allocation by the adjustment coefficient η to obtain the corrected bandwidth allocation result.

[0012] Preferably, the specific process for extracting multidimensional features including access frequency, associated request popularity, and time decay factor in step S5 is as follows: S51: Count the total number of times the target data has been accessed in the past N sliding time windows. N is dynamically set according to the data type and the value range is 10-30. S52: By constructing a request association graph, identify other requests that have data dependencies or temporal relationships with the current storage access request, count the number of accesses and data volume of associated requests in the most recent sliding window, and calculate the heat value. Popularity score = Number of related requests / Total number of requests + Data volume of related requests / Total data volume; S53: Based on the difference ΔT between the last access time and the current time of the target data, the following calculation is performed using an exponential decay function: f (ΔT)=e −μ·ΔT , μ This is the attenuation coefficient, which is dynamically adjusted according to the data lifecycle, and its value ranges from 0.01 to 0.1.

[0013] Preferably, the specific process of the LSTM model using the multi-dimensional feature input fusion attention mechanism in step S5 to predict the access probability within the future preset time window is as follows: S54: The extracted access frequency, associated request popularity, and time decay factor are standardized. The feature values ​​are transformed into a distribution with a mean of 0 and a standard deviation of 1 through z-score normalization, forming a feature sequence with uniform dimensions. S55: The feature sequence is divided into fixed-length input windows in chronological order, with the window size consistent with the sliding time window, which serve as the input to the LSTM model; the LSTM model processes the input sequence, captures the long-term temporal dependencies of the features through a gating mechanism, and outputs the hidden state sequence; S56: Introduce a self-attention mechanism to weight the hidden state sequence, calculate the attention weight of the hidden state at each time step, the weight value is positively correlated with the contribution of that time step to the prediction result, after normalizing the weights through the softmax function, the hidden state and the corresponding weight are weighted and summed to obtain the context vector. S57: Input the context vector into the fully connected layer, and output the access probability within the future preset time window through the sigmoid activation function, with a value range of [0, 1].

[0014] Preferably, the specific process of preloading the target data into the multi-level cache of the computing module based on the access popularity in step S6 is as follows: S61: Calculate the access popularity value based on the historical access frequency, recent access time and related request popularity of each data segment of the target data, and divide the target data into multiple data fragments according to the popularity value from high to low. S62: Based on the communication bandwidth determined in step S4, allocate a corresponding transmission bandwidth quota to each segment. The higher the popularity value of the segment, the higher the bandwidth quota allocated. S63: During preloading, the most popular fragments are transferred to the highest-level cache of the computing module first, the next most popular fragments are transferred to the next highest-level cache, and so on. During the preloading process, an intelligent gap-aware asynchronous loading strategy is adopted. By monitoring the CPU idle time slots and memory bandwidth utilization of the current request in real time, the preloading rate is dynamically adjusted to maximize the use of idle resources while avoiding blocking the current request.

[0015] Preferably, the specific process of step S7 is as follows: S71: Collects specified data including load rate, memory fragmentation level, temperature data and hardware aging factor in real time at sampling intervals of 50-100ms; S72: Standardize the collected data to the range [0, 1]. The higher the value, the better the status. Perform dynamic health scoring for each computing node based on the specified data. The initial weight allocation is: load rate 0.3, memory fragmentation degree 0.2, temperature data 0.2, hardware aging coefficient 0.3. S73: When the health score is lower than the preset threshold, the migration process is triggered. A node topology graph is constructed through a graph neural network, and the comprehensive cost of each candidate node is calculated. S74: Call the bandwidth allocation resources in step S4 to reserve a dedicated bandwidth channel for the migration process, accounting for 10%-20% of the total bandwidth, and migrate interaction requests and associated data to the optimal target node.

[0016] The beneficial effects of this invention include: 1. Dynamic allocation is achieved based on the type characteristics of interactive requests, the spatiotemporal distribution characteristics of data volume, and the priority decay coefficient, breaking through the limitations of traditional static allocation strategies. By assigning higher basic weights to real-time interactive requests, bandwidth resources for critical scenarios can be prioritized, reducing real-time interaction latency; by allocating elastic bandwidth pool resources in conjunction with the pulse-like growth characteristics of data volume, burst traffic demands can be quickly responded to, avoiding transmission congestion; the introduction of a priority decay coefficient to dynamically adjust the bandwidth ratio can prevent expired high-priority requests from occupying redundant resources, significantly improving the utilization efficiency of bandwidth resources and enhancing the overall response speed of the communication link.

[0017] 2. By extracting multi-dimensional features such as access frequency, associated request popularity, and time decay factor, and utilizing an LSTM model with an attention mechanism to predict access probability, the accuracy of preloading judgment is significantly improved. The strategy of preloading data into multi-level caches based on access popularity, combined with intelligent gap-aware asynchronous loading technology, ensures rapid response to frequently accessed data while avoiding interference from the preloading process on current request processing. This reduces the average data access latency by 40%-50%, making it particularly suitable for high-concurrency real-time interactive scenarios.

[0018] 3. Dynamic health scoring comprehensively considers load rate, memory fragmentation, temperature, and hardware aging coefficient, more accurately reflecting the true state of nodes and reducing unnecessary migrations caused by misjudgments based on a single indicator. The method of selecting the optimal target node based on a graph neural network combined with a real-time communication latency matrix balances communication latency, node health, and topology distance, improving the load balancing of the system after migration by more than 25%. Meanwhile, the reservation of dedicated bandwidth channels ensures the high efficiency and stability of the migration process, avoiding the impact of data migration on normal interactions.

[0019] 4. By employing a deep learning model to perform high-dimensional semantic parsing of request type features, and combining a sliding time window with spatial clustering algorithms to process the spatiotemporal distribution characteristics of data volume, the accuracy and relevance of feature representation are significantly improved, providing a reliable basis for subsequent resource scheduling decisions. Overall, this method can comprehensively optimize the low-latency interaction performance of compute-storage split servers, meeting the high real-time and high reliability requirements of emerging technology scenarios such as 5G and artificial intelligence. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the separate computing and storage server control method for low-latency interaction according to the present invention.

[0021] Figure 2 This is a schematic diagram of the separate computing and storage architecture for low-latency interaction according to the present invention. Detailed Implementation

[0022] The following is in conjunction with the appendix Figures 1-2The present invention will be further described in detail below: Example 1 See appendix Figure 1 As shown, a control method for a compute-storage split server for low-latency interaction includes the following steps: S1: Construct a split architecture that includes a computing module and a storage module. The computing module and the storage module establish a communication connection through a high-speed interconnect network. The computing module includes multiple computing nodes, and the storage module includes multiple storage nodes. This architecture pre-sets an initial communication protocol and resource scheduling interface to provide a basic execution environment for subsequent interactive control.

[0023] S2: Real-time monitoring of interaction requests between the computing module and the storage module. Dynamically extracts the type characteristics, spatiotemporal distribution characteristics of data volume, and priority attenuation coefficient of the interaction requests through the resource scheduling interface, and constructs a request feature vector as the basis for bandwidth allocation and data processing decisions.

[0024] S3: Determine the basic bandwidth allocation weight based on the type characteristics. The basic weight of real-time interactive requests is higher than that of batch processing requests. Combine the spatiotemporal distribution characteristics of data volume to calculate the bandwidth demand fluctuation coefficient. Allocate elastic bandwidth pool resources to requests whose data volume increases in a pulse manner over time.

[0025] S4: Based on the priority decay coefficient, the allocation result is corrected. For requests whose priority decays over time, the bandwidth ratio is dynamically reduced according to the decay rate. The allocation result is synchronized to the interaction interface between the computing module and the storage module through the initial communication protocol.

[0026] S5: Obtain historical access data from the server, extract multi-dimensional features including access frequency, related request popularity, and time decay factor, input them into an LSTM model with an attention mechanism to predict the access probability within a preset time window in the future, and execute step S6 when the access frequency is greater than the preset probability threshold.

[0027] S6: Based on the allocation results corrected by S4, the target data is preloaded into the multi-level cache of the computing module by being sharded according to access popularity.

[0028] S7: Real-time acquisition of specified data from each computing node in the computing module during the data processing process, and dynamic health score of each computing node based on the specified data. When the health score is lower than the preset threshold, the optimal target computing node is selected based on the real-time communication delay matrix to perform interactive requests and migration of related data.

[0029] In this embodiment, the specific process of dynamically extracting the type characteristics, spatiotemporal distribution characteristics of data volume, and priority attenuation coefficient of the interaction request through the resource scheduling interface in step S2 to construct the request feature vector is as follows: S21: Automatically identify the business scenario tags corresponding to the interactive request through a deep learning model and map them into high-dimensional semantic vectors. This deep learning model can be a pre-trained lightweight BERT model.

[0030] S22: Combining sliding time windows and spatial clustering algorithms, the discrete request data volume is transformed into a distribution matrix with spatiotemporal correlation. A fixed-duration sliding time window is set, and the window length is adaptively adjusted according to the historical request response time distribution. The default value is 50-500ms. The real-time request data volume is sliced ​​and sampled in units of windows, and the amount of interactive data between each computing node and storage node in each window is recorded.

[0031] The DBSCAN spatial clustering algorithm is used to cluster the sampled data, dividing nodes that are geographically or topologically close into the same spatial cluster, and calculating the average data volume, data volume variance and interaction frequency of node pairs within the cluster.

[0032] Using time windows as row indices and spatial clusters as column indices, statistical data within clusters are filled in chronological order to form a two-dimensional distribution matrix. The matrix element values ​​are the standardized data amounts of the corresponding spatial clusters within the time window, which are then mapped to the [0,1] interval through min-max normalization.

[0033] S23: By integrating the SLA level of the request initiator, historical response time deviations, and current system load, a dynamic attenuation function is constructed to calculate the priority attenuation coefficient. The specific formula for the dynamic attenuation function is as follows: λ(t)=α· SLA w +β·(1−Δ t / Δ t max )+γ·(1− Load / Load max ); Where λ(t) is the priority decay coefficient at time t, and its value ranges from [0, 1]. SLA w The SLA level is assigned a weight, with weight ranges of 0.8-1.0, 0.5-0.7, and 0.2-0.4 corresponding to SLA levels from high to low; Δ t Δ represents the deviation between the current request and the historical average response time. t max The preset maximum allowable deviation threshold; Load This represents the current system's average load rate. Load max The maximum load threshold of the system is defined as follows: α, β, and γ are weighting coefficients, and α+β+γ=1. Their values ​​are dynamically optimized based on the real-time system state using the gradient descent algorithm, so that the priority decay coefficient evolves in real time with the system state.

[0034] S24: The feature fusion network adaptively weights the three types of features into a unified request feature vector. The vector dimension is dynamically adjusted according to the complexity of the request, thereby achieving both accuracy and flexibility in feature representation.

[0035] Example 2 Based on Example 1, the specific process of step S3 is as follows: S31: Based on the high-dimensional semantic vector of the type features in step S2, a preset business scenario template vector is matched using cosine similarity. Real-time interaction templates include tags for millisecond-level response requirements and bidirectional high-frequency interactions, while batch processing templates include tags for hourly task cycles and one-way large-volume data transmission. The scenario template with the highest matching degree corresponds to a preset basic weight range: 0.6-0.8 for real-time interaction and 0.2-0.4 for batch processing. The specific weight value is then dynamically determined within this range, taking into account the real-time parameters of the request (including timeout thresholds).

[0036] S32: Calculate the bandwidth demand fluctuation coefficient by combining the spatiotemporal distribution characteristics of data volume: Extract the time series data of the spatial cluster to which the target request belongs from the spatiotemporal distribution matrix of data volume, calculate the data volume growth rate of three consecutive sliding windows, and determine the pulse growth when the standard deviation of the growth rate exceeds the preset fluctuation threshold and there are two adjacent windows where the data volume increase exceeds 50%.

[0037] The volatility coefficient K is calculated as (current window data volume − historical average data volume) / historical average data volume. Based on the K value, temporary resources of K × base bandwidth are allocated from the elastic bandwidth pool. The total capacity of the elastic bandwidth pool is 20%-30% of the total bandwidth, and a token bucket mechanism is used for dynamic scheduling.

[0038] In step S4, the allocation result is corrected based on the priority attenuation coefficient. The specific process of dynamically reducing the bandwidth ratio for requests whose priority decays over time according to the attenuation rate is as follows: S41: Calculate the difference in priority decay coefficients for a request over two consecutive sliding time windows: Δλ = λ(t1) − λ(t2); Where t2>t1, when Δλ>0, it is determined that the priority decays over time, and the decay rate is... v =Δλ(t2−t1).

[0039] S42: Determine the bandwidth proportion adjustment coefficient based on the attenuation rate. η =1− k · v k is a proportionality coefficient, with a value ranging from 0.1 to 0.5, which is dynamically adjusted according to the total bandwidth utilization.

[0040] S43: Multiply the initial bandwidth value after basic bandwidth allocation and elastic bandwidth pool resource allocation by the adjustment coefficient η to obtain the corrected bandwidth allocation result. The corrected bandwidth ratio shall not be lower than the preset minimum threshold to avoid excessive restriction of requests due to attenuation.

[0041] The specific process of extracting multidimensional features, including access frequency, popularity of associated requests, and time decay factor, in step S5 is as follows: S51: Count the total number of times the target data has been accessed in the past N sliding time windows. N is dynamically set according to the data type, with a value range of 10-30. The weighted access frequency is calculated in combination with the window weight, with the weight of recent windows being higher than that of distant windows.

[0042] S52: By constructing a request correlation graph, identify other requests that have data dependencies or temporal correlations with the current storage access request, count the number of accesses and data volume of these related requests in the most recent sliding window, and calculate the heat value. Popularity value = Number of accesses for related requests / Total number of requests + Data volume of related requests / Total data volume.

[0043] S53: Based on the difference ΔT between the last access time and the current time of the target data, the following calculation is performed using an exponential decay function: f (ΔT)=e −μ·ΔT , μ This is the attenuation coefficient, which is dynamically adjusted according to the data lifecycle, and its value ranges from 0.01 to 0.1.

[0044] The specific process of the LSTM model using the multi-dimensional feature input fusion attention mechanism in step S5 to predict the access probability within the future preset time window is as follows: S54: The extracted access frequency, associated request popularity, and time decay factor are standardized. The feature values ​​are transformed into a distribution with a mean of 0 and a standard deviation of 1 through z-score normalization, forming a feature sequence with uniform dimensions. S55: The feature sequence is divided into fixed-length input windows in chronological order, with the window size consistent with the sliding time window, which serve as the input to the LSTM model; the LSTM model processes the input sequence, captures the long-term temporal dependencies of the features through a gating mechanism, and outputs the hidden state sequence; S56: Introduce a self-attention mechanism to weight the hidden state sequence, calculate the attention weight of the hidden state at each time step, the weight value is positively correlated with the contribution of that time step to the prediction result, after normalizing the weights through the softmax function, the hidden state and the corresponding weight are weighted and summed to obtain the context vector. S57: Input the context vector into the fully connected layer, and output the access probability within the future preset time window through the sigmoid activation function, with a value range of [0, 1].

[0045] Example 3 Based on Example 1 or Example 2, the specific process of preloading the target data into the multi-level cache of the computing module according to the access popularity based on the allocation result in step S6 is as follows: S61: Calculate the access heat value based on the historical access frequency, recent access time and related request heat of each data segment of the target data, and divide the target data into multiple data shards according to the heat value from high to low (the shard size is dynamically adjusted according to the storage capacity of each level of the multi-level cache, usually the L1 cache shard is the smallest and the L3 cache shard is the largest). S62: Based on the communication bandwidth determined in step S4, allocate a corresponding transmission bandwidth quota to each segment. The higher the popularity value of the segment, the higher the bandwidth quota allocated. S63: During preloading, the most popular fragments are transferred to the highest-level cache of the computing module first, the next most popular fragments are transferred to the next highest-level cache, and so on. During the preloading process, an intelligent gap-aware asynchronous loading strategy is adopted. By monitoring the CPU idle time slots and memory bandwidth utilization of the current request in real time, the preloading rate is dynamically adjusted to maximize the use of idle resources while avoiding blocking the current request.

[0046] The specific process of step S7 is as follows: S71: Collects specified data including load rate, memory fragmentation level, temperature data and hardware aging factor in real time at sampling intervals of 50-100ms. The load rate is the weighted average of CPU core utilization (active cores have a higher weight than idle cores). The memory fragmentation level is calculated by the standard deviation of the size distribution of idle memory blocks. The temperature data is the average of the CPU package temperature and the memory controller temperature. The hardware aging factor is determined based on the node runtime, the cumulative number of failures and the performance degradation curve (fitted through historical performance benchmark tests). S72: Standardize the collected data to the range [0, 1]. The higher the value, the better the status. Calculate the score through a dynamic health assessment model. The initial weight allocation of the model is as follows: load rate (0.3), memory fragmentation degree (0.2), temperature data (0.2), and hardware aging coefficient (0.3). For every 30 days of runtime, the weight of the hardware aging coefficient increases by 5% (maximum not exceeding 0.5), and the weights of other parameters decrease proportionally. The scoring formula is: Health score = 0.3 × (1 - load rate) + 0.2 × (1 - memory fragmentation level) + 0.2 × (1 - temperature / maximum allowable temperature) + 0.3 × (1 - hardware aging factor); S73: When the health score is lower than the preset threshold (default 0.6), the migration process is triggered. A node topology graph is constructed using a graph neural network (nodes are computing nodes, and edge weights are real-time communication delays). Combined with the real-time communication delay matrix (matrix elements are the average communication delay between any two nodes), the comprehensive cost of each candidate node is calculated. Overall cost = a × communication delay + b × (1 - candidate node health) + c × topology distance, where a, b, and c are weighting coefficients, and the sum is 1. The overall cost is dynamically adjusted according to the current system load, and the node with the lowest overall cost is selected as the optimal target node. S74: Call the bandwidth allocation resources in step S4 to reserve a dedicated bandwidth channel for the migration process, accounting for 10%-20% of the total bandwidth, and migrate interaction requests and associated data to the optimal target node.

[0047] Example 4 In this embodiment, taking an autonomous driving data processing scenario as an example, the computing and storage split-type server control method of the present invention for low-latency interaction is specifically described: In autonomous driving systems, massive amounts of data collected in real time from onboard sensors such as LiDAR and cameras need to be processed and interacted with servers to achieve functions such as environmental perception and path planning. The control method of this invention is implemented as follows: The architecture separates the computing and storage modules. The computing module contains five computing nodes responsible for real-time data processing and decision-making; the storage module contains three storage nodes used to store high-definition maps, historical sensor data, etc. The computing and storage modules are connected via a high-speed interconnect network integrating RDMA technology, with pre-defined initial communication protocols and resource scheduling interfaces to provide a basic environment for data interaction. See Figure 2 As shown, the computing nodes (C1-C5) of the computing module serve as execution units, responsible for performing high-intensity real-time data processing tasks, such as environmental perception calculations like image recognition, point cloud processing, multi-sensor fusion, and path planning in autonomous driving scenarios. The computing module controller (CM) is responsible for managing internal resources, monitoring the health status of all computing nodes, and responding to scheduling instructions from higher layers.

[0048] Storage nodes (S1-S3) of the storage module: Responsible for persistently storing different types of data. S1 is dedicated to storing high-definition maps, S2 is dedicated to storing historical data generated by sensors such as cameras and LiDAR, and S3 may be used to store logs, models, and other data. Storage module controller (SM): Responsible for managing the distribution, replication, migration, and strict data access permission control, ensuring that compute nodes can only access their authorized data.

[0049] A high-speed interconnect network connects computing and storage modules, incorporating Remote Direct Data Access (RDMA) technology. This allows the network interface card (NIC) of the compute node to directly read and write to the memory of the storage node without the intervention of the operating system kernel, thereby significantly reducing latency and CPU overhead. Support for dynamic bandwidth allocation means that different network bandwidths can be flexibly allocated to different data streams based on the real-time needs of computing tasks, ensuring the performance of critical tasks.

[0050] The foundational support layer includes the initial communication protocol, which is the upper-layer software protocol running on the high-speed network. It defines the format, commands, serialization, and error handling mechanisms for data transmission between computing and storage modules, ensuring that both parties can correctly understand the information sent by each other and achieving a unified data transmission format. The resource scheduling interface is a set of standardized APIs (Application Programming Interfaces) or services that provide a unified set of commands to the upper-layer control logic for invoking functions such as dynamic bandwidth allocation, data preloading, and node status query. It is a key channel for realizing automated operation and maintenance and advanced control.

[0051] Extract features and construct request feature vectors: Monitor the interaction requests between the computing module and the storage module in real time, and identify the corresponding business scenario tags of the requests through a deep learning model. For example, real-time point cloud data transmission requests from LiDAR can be identified as real-time perception requests, and historical path data backup requests can be identified as batch storage requests, and mapped to high-dimensional semantic vectors. The sliding time window length is set to 100ms, and the requested data volume is sampled in units of this window. The DBSCAN algorithm is used to divide geographically proximate compute and storage nodes into three spatial clusters. The average data volume, data volume variance, and interaction frequency within each cluster are calculated to form a two-dimensional distribution matrix. For example, if the normalized data volume of cluster 1 within a certain window is 0.8, it reflects frequent data interaction within this cluster. The priority attenuation coefficient is calculated by integrating SLA level, historical response time deviation, and system load. Assuming a real-time sensing request has a high SLA level, the corresponding SLA... w =0.9. The current request response time is 20ms faster than the historical average, Δt = -20ms, Δt max =50ms, system average load rate Load=0.6, maximum load rate Load max =1.0; α=0.4, β=0.3, γ=0.3, then: λ(t)=0.4×0.9+0.3×(1-(-20) / 50)+0.3×(1-0.6) =0.36 + 0.3 × 1.4 + 0.3 × 0.4 =0.36+0.42+0.12=0.9. The feature fusion network integrates the three types of features into a request feature vector with weights. The vector dimension is dynamically adjusted according to the request complexity. For example, the request vector dimension for real-time perception is 256, and that for batch storage is 128. Dynamically allocate basic bandwidth and elastic resources: Match the high-dimensional semantic vector of real-time perception requests with preset templates. The real-time interaction template with the highest matching degree corresponds to a basic weight range of 0.6-0.8, and the basic weight is determined to be 0.75 based on its timeout threshold (50ms). Batch storage requests are matched with batch processing templates, and the basic weight is determined to be 0.3. Time-series data of the spatial cluster to which the real-time sensing request belongs are extracted from the distribution matrix. The data volume growth rates of three consecutive windows are calculated to be 60%, 70%, and 55%, respectively. The standard deviation exceeds the preset fluctuation threshold, and the growth rate of adjacent windows all exceeds 50%, which is judged as pulse-like growth. The historical average data volume is 100MB, the current window is 160MB, K=(160-100) / 100=0.6, and temporary resources of 0.6×base bandwidth are allocated from the elastic bandwidth pool of 25% of the total bandwidth. Correcting the bandwidth allocation results: Calculate the priority attenuation coefficient difference Δλ = 0.9 - 0.8 = 0.1 between two consecutive windows, t2 - t1 = 100 ms, and the attenuation rate. v =0.1 / 0.1=1. The total bandwidth utilization rate is 60%, and k=0.3 is taken. The bandwidth ratio adjustment coefficient η=1-0.3×1=0.7. The initial bandwidth value is the sum of the base bandwidth and elastic resources. Multiplying this by 0.7 yields the corrected bandwidth allocation result, which is not lower than the minimum threshold of 0.2 × total bandwidth. Predicted data access probability: Count the number of times a certain high-definition map data has been accessed in the past 20 windows, and combine the higher weight of recent windows to calculate the weighted access frequency of 18 times. Constructing a request association graph reveals that the high-definition map data is associated with the route planning request. The associated request accounted for 30% of the total number of requests and 25% of the total data volume within the recent window, with a popularity value of 0.3 + 0.25 = 0.55.

[0052] The difference between the last access time and the current time of this data is ΔT=5s, μ=0.05, and the time decay factor f(ΔT)=e^(-0.05×5)=e^(-0.25)≈0.7788. After standardizing the above features, the input is fed into an LSTM model with a fusion attention mechanism. The predicted access probability within the next second is 0.85, exceeding the preset threshold of 0.7. Therefore, the target data is preloaded into a multi-level cache: based on historical access frequency, recent access time, and correlation popularity, the high-definition map data is divided into three segments with popularity values ​​of 0.9, 0.7, and 0.5, respectively, and segment sizes of 10MB, 20MB, and 30MB. Based on the allocation results, the bandwidth quota ratio allocated to the three segments is 5:3:2. The most frequently accessed shards are prioritized for loading into the L1 cache, followed by the L2 cache, and then the L3 cache. Loading speed is adjusted using CPU idle time slots to avoid impacting real-time data processing. Compute node health management and data migration: Compute node data is collected at 80ms intervals. A node's load rate is 0.8, memory fragmentation is 0.3, temperature is 65℃ (maximum allowable 70℃), and hardware aging factor is 0.2. Health score = 0.3 × (1 - 0.8) + 0.2 × (1 - 0.3) + 0.2 × (1 - 65 / 70) + 0.3 × (1 - 0.2) =0.06+0.14+0.2×(1-0.9286)+0.24 =0.06 + 0.14 + 0.2 × 0.0714 + 0.24 ≈0.06+0.14+0.014+0.24=0.454; If the value falls below the threshold of 0.6, migration is triggered. A topology graph is constructed using a graph neural network, the comprehensive cost of candidate nodes is calculated, and the node with the lowest cost is selected as the target node. 15% of the total bandwidth is reserved as a dedicated channel to migrate the real-time perception requests and associated data of the target node to it, ensuring that the migration process does not affect the real-time processing of autonomous driving data. Through the above implementation process, this method realizes dynamic allocation of communication bandwidth, intelligent preloading of data, and efficient migration of nodes in autonomous driving scenarios, thereby reducing data interaction latency, improving bandwidth utilization, and ensuring the real-time performance and stability of the autonomous driving system.

Claims

1. A control method for a separate computing and storage server for low-latency interaction, characterized in that, Includes the following steps: S1: Construct a split architecture that includes computing modules and storage modules. The computing modules and storage modules establish a communication connection through a high-speed interconnect network. This architecture pre-sets an initial communication protocol and resource scheduling interface. S2: Real-time monitoring of interaction requests between the computing module and the storage module; dynamically extracting the type characteristics, spatiotemporal distribution characteristics of data volume, and priority attenuation coefficient of interaction requests through the resource scheduling interface; constructing a request feature vector. S3: Determine the basic bandwidth allocation weight based on the type characteristics. The basic weight of real-time interactive requests is higher than that of batch processing requests. Combine the spatiotemporal distribution characteristics of data volume to calculate the bandwidth demand fluctuation coefficient. Allocate elastic bandwidth pool resources to requests whose data volume increases in a pulse manner over time. S4: Based on the priority decay coefficient, the allocation result is corrected. For requests whose priority decays over time, the bandwidth ratio is dynamically reduced according to the decay rate. The allocation result is synchronized to the interaction interface between the computing module and the storage module through the initial communication protocol. S5: Obtain historical access data from the server, extract multi-dimensional features including access frequency, related request popularity, and time decay factor, input them into an LSTM model with an attention mechanism to predict the access probability within a future preset time window, and execute step S6 when the access frequency is greater than the preset probability threshold. S6: Based on the allocation results corrected by S4, the target data is preloaded into the multi-level cache of the computing module by being sharded according to access popularity; S7: Real-time acquisition of specified data from each computing node in the computing module during the data processing process, and dynamic health score of each computing node based on the specified data. When the health score is lower than the preset threshold, the optimal target computing node is selected based on the real-time communication delay matrix to perform interactive requests and migration of related data.

2. The control method for a separate computing and storage server for low-latency interaction according to claim 1, characterized in that, In step S2, the specific process of dynamically extracting the type characteristics, spatiotemporal distribution characteristics of data volume, and priority attenuation coefficient of the interaction request through the resource scheduling interface to construct the request feature vector is as follows: S21: Automatically identify the business scenario tags corresponding to interactive requests through deep learning models and map them into high-dimensional semantic vectors; S22: Combining sliding time window and spatial clustering algorithm, the discrete request data volume is transformed into a distribution matrix with spatiotemporal correlation. The real-time request data volume is sliced ​​and sampled in units of windows, and the amount of interactive data between each computing node and storage node in each window is recorded. The DBSCAN spatial clustering algorithm is used to cluster the sampled data, dividing nodes that are geographically or topologically close into the same spatial cluster. The average data volume, data volume variance, and interaction frequency of node pairs within the cluster are calculated. Using the time window as the row index and the spatial cluster as the column index, the statistical data within the cluster are filled in chronological order to form a two-dimensional distribution matrix. The matrix element value is the standardized data volume of the corresponding spatial cluster within the time window. S23: Combine the SLA level of the request initiator, historical response time deviation, and current system load pressure to construct a dynamic attenuation function to calculate the priority attenuation coefficient; S24: The three types of features are adaptively weighted into a unified request feature vector through a feature fusion network.

3. The control method for a separate computing and storage server for low-latency interaction according to claim 2, characterized in that, The specific process of step S3 is as follows: S31: Based on the high-dimensional semantic vector of the type features in step S2, the pre-set business scenario template vector is matched with the cosine similarity. The scenario template with the highest matching degree corresponds to the pre-set basic weight range. The specific weight value is dynamically determined within the range in combination with the real-time parameters of the request. S32: Calculate the bandwidth demand fluctuation coefficient by combining the spatiotemporal distribution characteristics of data volume: Extract the time series data of the spatial cluster to which the target request belongs from the spatiotemporal distribution matrix of data volume, calculate the data volume growth rate of three consecutive sliding windows, and determine the pulse growth when the standard deviation of the growth rate exceeds the preset fluctuation threshold and there are two adjacent windows where the data volume increase exceeds 50%. Calculate the bandwidth demand fluctuation coefficient K = (current window data volume − historical average data volume) / historical average data volume, and allocate temporary resources of K × base bandwidth from the elastic bandwidth pool according to the K value. The total capacity of the elastic bandwidth pool is 20%-30% of the total bandwidth, and a token bucket mechanism is used for dynamic scheduling.

4. The control method for a separate computing and storage server for low-latency interaction according to claim 1, characterized in that, In step S4, the allocation result is corrected based on the priority attenuation coefficient. The specific process of dynamically reducing the bandwidth ratio for requests whose priority decays over time according to the attenuation rate is as follows: S41: Calculate the difference in priority decay coefficients for a request over two consecutive sliding time windows: Δλ = λ(t1) − λ(t2); Where λ(t1) is the priority decay coefficient of the sliding time window t1, λ(t2) is the priority decay coefficient of the sliding time window t2, t2>t1, and when Δλ>0, it is determined that the priority decays with time, and the decay rate is... v =Δλ·(t2−t1); S42: Determine the bandwidth proportion adjustment coefficient based on the attenuation rate. η =1− k · v , k This is a proportionality coefficient, ranging from 0.1 to 0.5, which is dynamically adjusted based on the total bandwidth utilization. S43: Multiply the initial bandwidth value after basic bandwidth allocation and elastic bandwidth pool resource allocation by the adjustment coefficient η to obtain the corrected bandwidth allocation result.

5. The control method for a separate computing and storage server for low-latency interaction according to claim 1, characterized in that, The specific process of extracting multidimensional features, including access frequency, popularity of associated requests, and time decay factor, in step S5 is as follows: S51: Count the total number of times the target data has been accessed in the past N sliding time windows. N is dynamically set according to the data type and the value range is 10-30. S52: By constructing a request association graph, identify other requests that have data dependencies or temporal relationships with the current storage access request, count the number of accesses and data volume of associated requests in the most recent sliding window, and calculate the heat value. Popularity score = Number of related requests / Total number of requests + Data volume of related requests / Total data volume; S53: Based on the difference ΔT between the last access time and the current time of the target data, the following calculation is performed using an exponential decay function: f (ΔT)=e −μ·ΔT , μ This is the attenuation coefficient, which is dynamically adjusted according to the data lifecycle, and its value ranges from 0.01 to 0.

1.

6. The control method for a separate computing and storage server for low-latency interaction according to claim 5, characterized in that, The specific process of the LSTM model using the multi-dimensional feature input fusion attention mechanism in step S5 to predict the access probability within the future preset time window is as follows: S54: The extracted access frequency, associated request popularity, and time decay factor are standardized. The feature values ​​are transformed into a distribution with a mean of 0 and a standard deviation of 1 through z-score normalization, forming a feature sequence with uniform dimensions. S55: The feature sequence is divided into fixed-length input windows in chronological order, with the window size consistent with the sliding time window, which serve as the input to the LSTM model; the LSTM model processes the input sequence, captures the long-term temporal dependencies of the features through a gating mechanism, and outputs the hidden state sequence; S56: Introduce a self-attention mechanism to weight the hidden state sequence, calculate the attention weight of the hidden state at each time step, the weight value is positively correlated with the contribution of that time step to the prediction result, after normalizing the weights through the softmax function, the hidden state and the corresponding weight are weighted and summed to obtain the context vector. S57: Input the context vector into the fully connected layer, and output the access probability within the future preset time window through the sigmoid activation function, with a value range of [0, 1].

7. The method for controlling a separate computing and storage server for low-latency interaction according to claim 1, characterized in that, The specific process of preloading the target data into the multi-level cache of the computing module based on the access popularity according to the allocation result in step S6 is as follows: S61: Calculate the access popularity value based on the historical access frequency, recent access time and related request popularity of each data segment of the target data, and divide the target data into multiple data fragments according to the popularity value from high to low. S62: Based on the communication bandwidth determined in step S4, allocate a corresponding transmission bandwidth quota to each segment. The higher the popularity value of the segment, the higher the bandwidth quota allocated. S63: During preloading, the most popular fragments are transferred to the highest-level cache of the computing module first, the next most popular fragments are transferred to the next highest-level cache, and so on. During the preloading process, an intelligent gap-aware asynchronous loading strategy is adopted. By monitoring the CPU idle time slots and memory bandwidth utilization of the current request in real time, the preloading rate is dynamically adjusted to maximize the use of idle resources while avoiding blocking the current request.

8. The control method for a separate computing and storage server for low-latency interaction according to claim 1, characterized in that, The specific process of step S7 is as follows: S71: Collects specified data including load rate, memory fragmentation level, temperature data and hardware aging factor in real time at sampling intervals of 50-100ms; S72: Standardize the collected data to the range [0, 1]. The higher the value, the better the status. Perform dynamic health scoring for each computing node based on the specified data. The initial weight allocation is: load rate 0.3, memory fragmentation degree 0.2, temperature data 0.2, hardware aging coefficient 0.

3. S73: When the health score is lower than the preset threshold, the migration process is triggered. A node topology graph is constructed through a graph neural network, and the comprehensive cost of each candidate node is calculated. S74: Call the bandwidth allocation resources in step S4 to reserve a dedicated bandwidth channel for the migration process, accounting for 10%-20% of the total bandwidth, and migrate interaction requests and associated data to the optimal target node.

Images