Data classification processing method and system for parallel computing of multi-core processor

Abstract

The invention discloses a data classification processing method and system for parallel computing of a multi-core processor, relates to the technical field of data classification processing, and effectively solves the problem of abnormal disturbance in multi-core parallel data classification through a confidence-driven heterogeneous pipeline architecture. According to dynamic grading processing based on data block confidence, high-confidence data is ensured to completely process a link, and meanwhile, rapid classification and cache isolation are performed on low-confidence data, so that an assembly line is prevented from being blocked by abnormal data; a self-adaptive window segmentation mechanism is linked with confidence coefficient fluctuation and calculation resource ratio, and inter-core load unbalance caused by sudden abnormity is relieved; a double-closed-loop control strategy combines a real-time queue state with a historical load mode, and the cache pollution risk caused by thread migration is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of data classification processing, and in particular to a data classification processing method and system for parallel computing on a multi-core processor. Background Art

[0002] Current multi-core processors need to cope with the continuous influx of data from high-throughput scenarios such as the Industrial Internet of Things and financial transactions in real-time streaming data classification tasks, while also handling abnormal data disturbances caused by factors such as sensor failures and network jitter. The mainstream solution uses a three-stage pipeline architecture to achieve parallelism between stages by binding a thread pool to a specific core group. However, as the frequency of data disturbances increases, such as sudden environmental noise in autonomous driving, traditional methods expose adaptability defects under the impact of sudden abnormal flows.

[0003] Currently, newer solutions such as the dynamic watermark control (DWC) solution can adjust the processing rhythm according to the queue depth. However, when encountering continuous abnormal data blocks, the global beat adjustment of DWC is prone to cause inter-stage resource mismatch. The preprocessing core will idle while waiting for the downstream to release the buffer, and the classification decision core will be idle due to insufficient input data. The elastic thread migration (ETM) technology proposed in recent years alleviates this problem by dynamically rebinding threads to idle cores. However, the cache pollution caused by frequent migration increases the L2 miss rate, resulting in unacceptable quality fluctuations in low-tolerance scenarios such as real-time classification of medical images.

[0004] The latest research direction is lightweight anti-disturbance mechanism, such as embedding soft switches based on data confidence between pipeline stages. When the confidence of a data block is lower than the threshold, the subsequent feature calculation stage is automatically skipped, and the simplified classifier is directly called to output the emergency result. However, the abnormal data features and normal data share computing resources, causing interference residues; some solutions introduce the concept of computing path isolation, retaining the complete pipeline for high-confidence data, while low-confidence data is transferred to a dedicated lightweight channel for processing. This design works well in video stream analysis, but the control of state synchronization overhead during channel switching is still not ideal, and it is difficult to adapt to microsecond decision-making requirements. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a data classification processing method and system for parallel computing on multi-core processors to solve the problems that existing solutions rely on fixed pipelines and global load balancing, are prone to inter-level resource mismatch and cache efficiency degradation when facing sudden abnormal data, and the dynamic adjustment mechanism lacks deep coupling with data characteristics.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] In a first aspect, an embodiment of the present invention provides a data classification processing method for parallel computing of a multi-core processor, which includes:

[0009] Step S1, dividing the input data stream into continuous data blocks and adding a timestamp and source identifier to each data block;

[0010] Step S2: Calculate the confidence index of each data block in real time, where the confidence index is generated based on a composite function of the feature distribution dispersion of the current data block and the historical classification accuracy;

[0011] Step S3, dynamically allocating computing resources based on the confidence index, includes:

[0012] Assigning data blocks with confidence levels higher than a first threshold to a complete classification pipeline consisting of a group of primary computing cores, wherein the complete classification pipeline includes a serial processing chain of feature extraction, dimensionality reduction, and classification decision making;

[0013] Assigning data blocks with a confidence score lower than a first threshold but higher than a second threshold to a lightweight computing core group to perform a fast classification process, wherein the fast classification process skips the dimensionality reduction stage and adopts a simplified classification model;

[0014] For data blocks with a confidence level lower than a second threshold, triggering an exception processing channel to perform cache isolation and feature marking;

[0015] Periodically collect the cache hit rate and queue depth indicators of each core group, and dynamically adjust the core ratio between the main computing core group and the lightweight computing core group.

[0016] As a preferred solution of the data classification processing method for parallel computing of a multi-core processor described in the present invention, the data block segmentation in step S1 adopts an adaptive window mechanism:

[0017] When it is detected that the confidence standard deviation of N consecutive data blocks exceeds the preset threshold, the segmentation window size is automatically reduced;

[0018] When the queue depth is lower than the critical value and lasts for T seconds, the split window size is gradually expanded to the initial value.

[0019] As a preferred solution of the data classification processing method for parallel computing of multi-core processors described in the present invention, in step S1, the confidence sequence of the latest N data blocks is continuously tracked during the process of automatically reducing the segmentation window size. When the fluctuation is severe, the window size is adaptively reduced. The process includes:

[0020] Calculate the standard deviation of the confidence sequence to quantify the degree of fluctuation. If the standard deviation exceeds the preset threshold, the segmentation window size is dynamically adjusted according to the fluctuation amplitude. The formula is:

[0021]

[0022] Where N represents the number of consecutive data blocks used for statistics, c i represents the confidence index of the i-th data block,

[0023] represents the arithmetic mean of the latest N confidences, t represents the time sequence number of the current data block, σ C represents the calculated confidence standard deviation;

[0024] When σ C >T σ When , update the split window size at the next moment as follows:

[0025] α=exp(-γ(σ C -T σ )),

[0026]

[0027] Among them, T σ represents the preset threshold of the confidence standard deviation, γ represents the adaptive scaling rate coefficient, α represents the scaling factor, and W t Indicates the current split window size, W t+1 Indicates the updated split window size, W min Indicates the minimum allowed size of the split window, Indicates floor operation.

[0028] As a preferred solution of the data classification processing method for parallel computing of a multi-core processor described in the present invention, the calculation of the confidence index in step S2 includes:

[0029] Extract the feature distribution histogram from the head area of the data block and calculate its KL divergence with the historical normal data distribution;

[0030] Get the most recent K classification results for the same source identifier and calculate the sliding weighted average of their accuracy;

[0031] The KL divergence and accuracy are weighted and fused, where the weight coefficient increases dynamically with the freshness of the data block timestamp.

[0032] As a preferred solution of the data classification processing method for parallel computing of multi-core processors described in the present invention, in step S2, a composite confidence is calculated based on the feature distribution dispersion and the historical classification accuracy, and a confidence index c is generated for each data block in real time. t , calculated and integrated in three stages:

[0033] Calculate the feature distribution dispersion: extract the feature distribution histogram from the head area of the t-th data block And with the historical normal data distribution histogram Calculate KL divergence:

[0034]

[0035] Among them, M represents the total number of histogram buckets, H t (m) represents the frequency of the mth feature bucket of the current data block, H hist (m) represents the frequency of the mth feature bucket of historical normal data, D t Indicates the discreteness of the characteristic distribution of the current data block, H hist In the middle, hist indicates the historical normal data distribution;

[0036] Calculate historical classification accuracy: Get the most recent K classification result sequences {r t-K+1 ,…,r t}, where r j ∈{0,1}, and use exponential decay weight ρ∈(0,1) for sliding weighted average, the formula is:

[0037]

[0038] Among them, K represents the number of historical classification results, ρ represents the sliding weighted attenuation factor, and r t-K+j Indicates the t-K+j classification result, A t Represents the sliding weighted average of historical classification accuracy;

[0039] Dynamic fusion generates confidence: define the freshness coefficient φ t With timestamp τ t Incremental, expressed as:

[0040] φ t =1-exp[-λ(τ t -τ ref )],

[0041] Among them, τ t Indicates the timestamp of the current data block, τ ref represents the reference start time, λ represents the freshness growth rate, φ t Represents the freshness weight coefficient;

[0042] The discreteness and accuracy are fused according to the weights, and the discreteness is normalized to take the maximum possible divergence D max , the formula is:

[0043]

[0044] Among them, D max represents the maximum value of the divergence normalization, c t Represents the generated confidence metric.

[0045] As a preferred solution of the data classification processing method for parallel computing of a multi-core processor described in the present invention, the exception processing channel in step S3 includes:

[0046] Create an isolated storage area in the last-level cache to store low-confidence data blocks and their feature tags;

[0047] Start an independent monitoring thread to periodically scan the isolated storage area and trigger recalculation when one of the following conditions is met:

[0048] The confidence level of the newly arrived data block with the same source identifier is restored to above the first threshold;

[0049] The overall system load rate is lower than the set threshold and lasts longer than the preset period.

[0050] As a preferred solution of the data classification processing method for parallel computing of a multi-core processor described in the present invention, the core ratio adjustment in step S3 adopts a double closed-loop control strategy:

[0051] The inner loop adjusts the frequency of thread migration between core groups based on the real-time queue depth, and the outer loop predicts the core group size requirements in future time windows based on historical load patterns;

[0052] When the average processing delay of the lightweight computing core group exceeds β times the delay of the main computing core group, a specified number of core resources are forcibly reclaimed and allocated to the main computing core group.

[0053] As a preferred solution of the data classification processing method for parallel computing of multi-core processors described in the present invention, in step S3, when the average processing delay L of the lightweight computing core group is t Exceeds the main computing core group delay M t When the latency is β times greater than the initial latency, the following resource recovery and allocation mechanism is triggered: by calculating the latency ratio and comparing it with the threshold, the number of cores to be recovered is dynamically determined. Then, while ensuring the availability of the lightweight group, resources are preferentially allocated to the main computing core group to restore overall performance.

[0054] If R t >1, then:

[0055]

[0056] N L,t+1 =N L,t -ΔN,N M,t+1 =N M,t +ΔN,

[0057] Among them, L t represents the average processing delay of the lightweight computing core group at time t, M t represents the average processing delay of the main computing core group at time t, β represents the delay multiplier threshold, R t represents the delay ratio, N L,t Indicates the current number of cores in the lightweight computing core group, N M,t Indicates the current number of cores in the main computing core group, N L,min Indicates the minimum number of cores allowed in the lightweight group, κ indicates the recovery coefficient, which is used to adjust the recovery intensity, L indicates the lightweight core group, M indicates the host core group, and ΔN indicates the number of core resources recovered this time. Indicates rounding up, and min(·) indicates taking the minimum value of the two.

[0058] In a second aspect, the present invention provides a data classification processing system for parallel computing on a multi-core processor, comprising: a data intake module configured to segment an input data stream into consecutive data blocks and add a timestamp and a source identifier to each data block;

[0059] a confidence prediction module, connected to the data ingestion module and configured to generate a confidence indicator for a data block;

[0060] an elastic computing module, comprising a main computing core group, a lightweight computing core group, and an exception handling channel, wherein an input end of the elastic computing module is connected to the confidence prediction module;

[0061] a dynamic scheduling module configured to collect operating indicators of each core group and adjust the core number ratio, with its control end connected to the elastic computing module;

[0062] The main computing core group includes a cascaded feature extraction unit, a dimensionality reduction unit, and a classification decision unit, and the lightweight computing core group bypasses the dimensionality reduction unit and is directly connected to the classification decision unit.

[0063] As a preferred solution of the data classification processing system for parallel computing of multi-core processors described in the present invention, the exception processing channel includes:

[0064] a cache isolation controller configured to dynamically partition isolated storage areas in the last-level cache; the cache isolation controller implements the partitioning of isolated areas by modifying a bit mask of a CPU cache allocation table CAT, where the bit mask value increases in units of 256KB based on the current number of isolated data blocks;

[0065] A recalculation trigger is connected to the cache isolation controller and the dynamic scheduling module, and is configured to activate the data reprocessing process when a preset condition is met.

[0066] The beneficial effects of the present invention are as follows: the present invention effectively copes with the problem of abnormal disturbance in multi-core parallel data classification through a confidence-driven heterogeneous pipeline architecture; based on the dynamic hierarchical processing of data block confidence, while ensuring the complete processing link of high-confidence data, it implements rapid classification and cache isolation for low-confidence data to avoid abnormal data blocking the pipeline; the adaptive window segmentation mechanism links confidence fluctuations with computing resource ratios to alleviate the inter-core load imbalance caused by sudden anomalies; the dual closed-loop control strategy combines real-time queue status with historical load patterns to reduce the cache pollution risk caused by thread migration; the exception handling channel uses hardware-level cache isolation and conditional triggering recalculation to maintain the mainstream pipeline throughput while ensuring delayed repair of abnormal data. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0068] Figure 1 This is a flow chart of the data classification processing method for parallel computing on multi-core processors in Example 1.

[0069] Figure 2 This is a schematic diagram of the framework of the data classification processing system for parallel computing on multi-core processors in Example 1. DETAILED DESCRIPTION

[0070] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0071] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0072] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0073] Example 1, with reference to Figure 1 and Figure 2This embodiment provides a data classification processing method for parallel computing on a multi-core processor, comprising the following steps:

[0074] Step S1, dividing the input data stream into continuous data blocks and adding a timestamp and source identifier to each data block;

[0075] The data block segmentation in step S1 adopts the adaptive window mechanism:

[0076] When it is detected that the confidence standard deviation of N consecutive data blocks exceeds the preset threshold, the segmentation window size is automatically reduced;

[0077] When the queue depth is lower than the critical value and lasts for T seconds, the split window size is gradually expanded to the initial value;

[0078] In step S1, while automatically reducing the segmentation window size, the confidence sequence of the latest N data blocks is continuously tracked. When the fluctuation is severe, the window size is adaptively reduced. The process includes:

[0079] Calculate the standard deviation of the confidence sequence to quantify the degree of fluctuation. If the standard deviation exceeds the preset threshold, the segmentation window size is dynamically adjusted according to the fluctuation amplitude. The formula is:

[0080]

[0081] Where N represents the number of consecutive data blocks used for statistics, c i represents the confidence index of the i-th data block,

[0082] represents the arithmetic mean of the latest N confidences, t represents the time sequence number of the current data block, σ C represents the calculated confidence standard deviation;

[0083] When σ C >T σ When , update the split window size at the next moment as follows:

[0084] α=exp(-γ(σ C -T σ )),

[0085]

[0086] Among them, T σ represents the preset threshold of the confidence standard deviation, γ represents the adaptive scaling rate coefficient, α represents the scaling factor, and W t Indicates the current split window size, W t+1 Indicates the updated split window size, W minIndicates the minimum allowed size of the split window, Indicates floor operation;

[0087] Specifically, in high-volatility scenarios, the confidence level changes are quickly responded to, and shrinking the window can increase the sensitivity to the burst mode and avoid the information ambiguity caused by the large window. The dynamic factor α is scaled step by step according to the fluctuation amplitude to ensure that the contraction is neither excessive nor slow. The lower limit W min The settings ensure that the window will not shrink infinitely, thus maintaining system stability;

[0088] Step S2: Calculate the confidence index of each data block in real time. The confidence index is generated based on a composite function of the feature distribution dispersion of the current data block and the historical classification accuracy.

[0089] The calculation of the confidence index in step S2 includes:

[0090] Extract the feature distribution histogram from the head area of the data block and calculate its KL divergence with the historical normal data distribution;

[0091] Get the most recent K classification results for the same source identifier and calculate the sliding weighted average of their accuracy;

[0092] The KL divergence and accuracy are weighted and integrated, where the weight coefficient increases dynamically with the freshness of the data block timestamp;

[0093] In step S2, the composite confidence is calculated based on the feature distribution dispersion and the historical classification accuracy, and the confidence index c is generated in real time for each data block. t , calculated and integrated in three stages:

[0094] Calculate the feature distribution dispersion: extract the feature distribution histogram from the head area of the t-th data block And with the historical normal data distribution histogram Calculate KL divergence:

[0095]

[0096] Among them, M represents the total number of histogram buckets, H t (m) represents the frequency of the mth feature bucket of the current data block, H hist (m) represents the frequency of the mth feature bucket of historical normal data, D t Indicates the discreteness of the characteristic distribution of the current data block, H hist In the middle, hist indicates the historical normal data distribution;

[0097] Calculate historical classification accuracy: Get the most recent K classification result sequences {r t-K+1 ,…,r t}, where r j ∈{0,1}, and use exponential decay weight ρ∈(0,1) for sliding weighted average, the formula is:

[0098]

[0099] Among them, K represents the number of historical classification results, ρ represents the sliding weighted attenuation factor, and r t-K+j Indicates the t-K+j classification result, A t Represents the sliding weighted average of historical classification accuracy;

[0100] Dynamic fusion generates confidence: define the freshness coefficient φ t With timestamp τ t Incremental, expressed as:

[0101] φ t =1-exp[-λ(τ t -τ ref )],

[0102] Among them, τ t Indicates the timestamp of the current data block, τ ref represents the reference start time, λ represents the freshness growth rate, φ t Represents the freshness weight coefficient;

[0103] The discreteness and accuracy are fused according to the weights, and the discreteness is normalized to take the maximum possible divergence D max , the formula is:

[0104]

[0105] Among them, D max represents the maximum value of the divergence normalization, c t represents the generated confidence indicator;

[0106] Specifically, this composite function takes into account both current features and historical performance, organically integrating distribution dispersion and classification accuracy. It also dynamically adjusts weight distribution through the freshness coefficient, allowing the confidence indicator to promptly reflect sudden offsets while maintaining model stability. The distribution dispersion term detects deviations between data block characteristics and the normal distribution, while the accuracy term reflects the model's recent performance. The two complement each other, and the freshness mechanism ensures greater sensitivity to the latest data, preventing judgment lags caused by outdated historical information. This method improves the robustness and responsiveness of confidence assessment and is suitable for real-time data classification in multi-core parallel environments.

[0107] Step S3, dynamically allocating computing resources based on the confidence index, includes:

[0108] Assigning data blocks with confidence levels higher than a first threshold to a complete classification pipeline consisting of a main computing core group, where the complete classification pipeline includes a serial processing chain of feature extraction, dimensionality reduction, and classification decision making;

[0109] Data blocks with confidence levels lower than the first threshold but higher than the second threshold are assigned to the lightweight computing core group to perform a fast classification process that skips the dimensionality reduction phase and uses a simplified classification model.

[0110] For data blocks with a confidence level lower than a second threshold, triggering an exception processing channel to perform cache isolation and feature marking;

[0111] Periodically collect cache hit rates and queue depth metrics for each core group, and dynamically adjust the core count ratio between the main computing core group and the lightweight computing core group.

[0112] The exception handling channel in step S3 includes:

[0113] Create an isolated storage area in the last-level cache to store low-confidence data blocks and their feature tags;

[0114] Start an independent monitoring thread to periodically scan the isolated storage area and trigger recalculation when one of the following conditions is met:

[0115] The confidence level of the newly arrived data block with the same source identifier is restored to above the first threshold;

[0116] The overall system load rate is lower than the set threshold and lasts longer than the preset period;

[0117] In step S3, the core ratio adjustment adopts a double closed-loop control strategy:

[0118] The inner loop adjusts the frequency of thread migration between core groups based on the real-time queue depth, and the outer loop predicts the core group size requirements in future time windows based on historical load patterns;

[0119] When the average processing delay of the lightweight computing core group exceeds β times the delay of the main computing core group, a specified number of core resources are forcibly reclaimed and allocated to the main computing core group;

[0120] In step S3, when the average processing delay L of the lightweight computing core group is t Exceeds the main computing core group delay M t When the latency is β times greater than the initial latency, the following resource recovery and allocation mechanism is triggered: by calculating the latency ratio and comparing it with the threshold, the number of cores to be recovered is dynamically determined. Then, while ensuring the availability of the lightweight group, resources are preferentially allocated to the main computing core group to restore overall performance.

[0121] If R t >1, then:

[0122]

[0123] N L,t+1 =N L,t -ΔN,N M,t+1 =N M,t +ΔN,

[0124] Among them, L t represents the average processing delay of the lightweight computing core group at time t, M t represents the average processing delay of the main computing core group at time t, β represents the delay multiplier threshold, R t represents the delay ratio, N L,t Indicates the current number of cores in the lightweight computing core group, N M,t Indicates the current number of cores in the main computing core group, N L,min Indicates the minimum number of cores allowed in the lightweight group, κ indicates the recovery coefficient, which is used to adjust the recovery intensity, L indicates the lightweight core group, M indicates the host core group, and ΔN indicates the number of core resources recovered this time. Indicates rounding up operation, min(·) means taking the minimum value of the two;

[0125] Specifically, this mechanism can quickly identify and proportionally recycle core resources when the processing delay of the lightweight group increases abnormally, and allocate resources to the main computing core group first to restore the performance of key classification tasks. The recovery coefficient κ can be tuned offline according to the system tolerance, so that the core resource recovery is neither too drastic nor too conservative, and the lower limit N L,min Ensure that the lightweight group always has enough cores to avoid complete failure.

[0126] This embodiment further provides a data classification processing system for parallel computing on a multi-core processor, comprising:

[0127] a data ingestion module configured to segment the input data stream into consecutive data chunks and add a timestamp and a source identifier to each data chunk;

[0128] a confidence prediction module, connected to the data ingestion module and configured to generate confidence indicators for data blocks;

[0129] The elastic computing module includes a main computing core group, a lightweight computing core group, and an exception handling channel, and its input end is connected to the confidence prediction module;

[0130] The dynamic scheduling module is configured to collect the operating indicators of each core group and adjust the core ratio. Its control end is connected to the elastic computing module;

[0131] The main computing core group includes a cascaded feature extraction unit, a dimensionality reduction unit, and a classification decision unit. The lightweight computing core group bypasses the dimensionality reduction unit and is directly connected to the classification decision unit.

[0132] The exception handling channel includes:

[0133] a cache isolation controller configured to dynamically partition isolated storage areas in the last-level cache; the cache isolation controller implements the partitioning of isolated areas by modifying a bit mask in the CPU cache allocation table CAT, where the bit mask value increases in 256KB increments based on the number of currently isolated data blocks;

[0134] The recomputation trigger is connected to the cache isolation controller and the dynamic scheduling module, and is configured to activate the data reprocessing process when a preset condition is met.

[0135] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A data classification processing method for parallel computing of multi-core processors, characterized in that: include, Step S1, dividing the input data stream into continuous data blocks and adding a timestamp and source identifier to each data block; Step S2: Calculate the confidence index of each data block in real time, where the confidence index is generated based on a composite function of the feature distribution dispersion of the current data block and the historical classification accuracy; Step S3, dynamically allocating computing resources based on the confidence index, includes: Assigning data blocks with confidence levels higher than a first threshold to a complete classification pipeline consisting of a group of primary computing cores, wherein the complete classification pipeline includes a serial processing chain of feature extraction, dimensionality reduction, and classification decision making; Assigning data blocks with a confidence score lower than a first threshold but higher than a second threshold to a lightweight computing core group to perform a fast classification process, wherein the fast classification process skips the dimensionality reduction stage and adopts a simplified classification model; For data blocks with a confidence level lower than a second threshold, triggering an exception processing channel to perform cache isolation and feature marking; Periodically collect the cache hit rate and queue depth indicators of each core group, and dynamically adjust the core ratio between the main computing core group and the lightweight computing core group.

2. The data classification processing method for parallel computing of a multi-core processor according to claim 1, characterized in that: The data blocks in step S1 are segmented using an adaptive window mechanism: When it is detected that the confidence standard deviation of N consecutive data blocks exceeds the preset threshold, the segmentation window size is automatically reduced; When the queue depth is lower than the critical value and lasts for T seconds, the split window size is gradually expanded to the initial value.

3. The data classification processing method for parallel computing of a multi-core processor according to claim 2, characterized in that: In step S1, while automatically reducing the segmentation window size, the confidence sequence of the latest N data blocks is continuously tracked. When the fluctuation is severe, the window size is adaptively reduced. The process includes: Calculate the standard deviation of the confidence sequence to quantify the degree of fluctuation. If the standard deviation exceeds the preset threshold, the segmentation window size is dynamically adjusted according to the fluctuation amplitude. The formula is: Where N represents the number of consecutive data blocks used for statistics, c i represents the confidence index of the i-th data block, represents the arithmetic mean of the latest N confidences, t represents the time sequence number of the current data block, σ C represents the calculated confidence standard deviation; When σ C >T σ When , update the split window size at the next moment as follows: α=exp(-γ(σ C -T σ )), Among them, T σ represents the preset threshold of the confidence standard deviation, γ represents the adaptive scaling rate coefficient, α represents the scaling factor, and W t Indicates the current split window size, W t+1 Indicates the updated split window size, W min Indicates the minimum allowed size of the split window, Indicates floor operation.

4. The data classification processing method for parallel computing of a multi-core processor according to claim 1, characterized in that: The calculation of the confidence index in step S2 includes: Extract the feature distribution histogram from the head area of the data block and calculate its KL divergence with the historical normal data distribution; Get the most recent K classification results for the same source identifier and calculate the sliding weighted average of their accuracy; The KL divergence and accuracy are weighted and fused, where the weight coefficient increases dynamically with the freshness of the data block timestamp.

5. The data classification processing method for parallel computing of a multi-core processor according to claim 4, characterized in that: In step S2, the composite confidence is calculated based on the feature distribution dispersion and the historical classification accuracy, and the confidence index c is generated for each data block in real time. t , calculated and integrated in three stages: Calculate the feature distribution dispersion: extract the feature distribution histogram from the head area of the t-th data block And with the historical normal data distribution histogram Calculate KL divergence: Among them, M represents the total number of histogram buckets, H t (m) represents the frequency of the mth feature bucket of the current data block, H hist (m) represents the frequency of the mth feature bucket of historical normal data, D t Indicates the discreteness of the characteristic distribution of the current data block, H hist In the middle, hist indicates the historical normal data distribution; Calculate historical classification accuracy: Get the most recent K classification result sequences {r t-K+1 ,…,r t }, where r j ∈{0,1}, and use exponential decay weight ρ∈(0,1) for sliding weighted average, the formula is: Among them, K represents the number of historical classification results, ρ represents the sliding weighted attenuation factor, and r t-K+j Indicates the t-K+j classification result, A t Represents the sliding weighted average of historical classification accuracy; Dynamic fusion generates confidence: define the freshness coefficient φ t With timestamp τ t Incremental, expressed as: f t =1-exp[-λ(τ t -t ref )], Among them, τ t Indicates the timestamp of the current data block, τ ref represents the reference start time, λ represents the freshness growth rate, φ t Represents the freshness weight coefficient; The discreteness and accuracy are fused according to the weights, and the discreteness is normalized to take the maximum possible divergence D max , the formula is: Among them, D max represents the maximum value of the divergence normalization, c t Represents the generated confidence metric.

6. The data classification processing method for parallel computing of a multi-core processor according to claim 1, characterized in that: The exception handling channel in step S3 includes: Create an isolated storage area in the last-level cache to store low-confidence data blocks and their feature tags; Start an independent monitoring thread to periodically scan the isolated storage area and trigger recalculation when one of the following conditions is met: The confidence level of the newly arrived data block with the same source identifier is restored to above the first threshold; The overall system load rate is lower than the set threshold and lasts longer than the preset period.

7. The data classification processing method for parallel computing of a multi-core processor according to claim 6, characterized in that: The core ratio adjustment in step S3 adopts a double closed-loop control strategy: The inner loop adjusts the frequency of thread migration between core groups based on the real-time queue depth, and the outer loop predicts the core group size requirements in future time windows based on historical load patterns; When the average processing delay of the lightweight computing core group exceeds β times the delay of the main computing core group, a specified number of core resources are forcibly reclaimed and allocated to the main computing core group.

8. The data classification processing method for parallel computing of a multi-core processor according to claim 7, characterized in that: In step S3, when the average processing delay L of the lightweight computing core group is t Exceeds the main computing core group delay M t When the latency is β times greater than the initial latency, the following resource recovery and allocation mechanism is triggered: by calculating the latency ratio and comparing it with the threshold, the number of cores to be recovered is dynamically determined. Then, while ensuring the availability of the lightweight group, resources are preferentially allocated to the main computing core group to restore overall performance. If R t >1, then: N L,t+1 =N L,t -ΔN,N M,t+1 =N M,t +ΔN, Among them, L t represents the average processing delay of the lightweight computing core group at time t, M t represents the average processing delay of the main computing core group at time t, β represents the delay multiplier threshold, R t represents the delay ratio, N L,t Indicates the current number of cores in the lightweight computing core group, N M,t Indicates the current number of cores in the main computing core group, N L,min Indicates the minimum number of cores allowed in the lightweight group, κ indicates the recovery coefficient, which is used to adjust the recovery intensity, L indicates the lightweight core group, M indicates the host core group, and ΔN indicates the number of core resources recovered this time. Indicates rounding up, and min(·) indicates taking the minimum value of the two.

9. A data classification processing system for parallel computing on multi-core processors, based on a data classification processing method for parallel computing on multi-core processors according to any one of claims 1 to 8, characterized in that: include: a data ingestion module configured to segment the input data stream into consecutive data chunks and add a timestamp and a source identifier to each data chunk; a confidence prediction module, connected to the data ingestion module and configured to generate a confidence indicator for a data block; an elastic computing module, comprising a main computing core group, a lightweight computing core group, and an exception handling channel, wherein an input end of the elastic computing module is connected to the confidence prediction module; a dynamic scheduling module configured to collect operating indicators of each core group and adjust the core number ratio, with its control end connected to the elastic computing module; The main computing core group includes a cascaded feature extraction unit, a dimensionality reduction unit, and a classification decision unit, and the lightweight computing core group bypasses the dimensionality reduction unit and is directly connected to the classification decision unit.

10. The data classification processing system for parallel computing of multi-core processors according to claim 9, characterized in that: The exception handling channel includes: a cache isolation controller configured to dynamically partition isolated storage regions in the last level cache; A recalculation trigger is connected to the cache isolation controller and the dynamic scheduling module, and is configured to activate the data reprocessing process when a preset condition is met.

Images