A distributed database-based intelligent search system for massive amounts of information

By using a semantically aware storage control layer and a multi-tenant joint scheduling engine, the problems of prefetching mechanism failure and insufficient resource isolation in distributed databases are solved, achieving efficient data prefetching and resource isolation, and improving semantic retrieval performance and service quality in a multi-tenant environment.

CN123115241APending Publication Date: 2026-11-17GUIZHOU GUIJING TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611336306.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-31
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The existing prefetching mechanism of the distributed database storage control layer schedules based on the continuity of logical block addresses or historical access frequency, which cannot adapt to the highly discrete access patterns in semantic retrieval scenarios. This leads to prefetching mechanism failure, decreased cache hit rate, and severe read/write amplification. The hot and cold data tiering mechanism cannot predict semantic hotspot changes, resulting in long-tail query latency. In a multi-tenant environment, the computation pushdown scheduling strategy does not perform resource isolation, and high-priority tenant requests are blocked by low-priority tenants.

Method used

A semantically aware storage control layer is adopted. By querying the semantic distance between the semantic vector and the semantic centroid vector, the cluster of data blocks to be prefetched is selected. Cold and hot storage media are migrated in advance based on the semantic heat value. Combined with the multi-tenant joint scheduling engine, resources are allocated according to the priority of the tenant service level agreement to achieve resource isolation.

Benefits of technology

It improves the prefetch hit rate in semantic retrieval scenarios, reduces random read/write amplification, reduces long-tail query latency, and enhances the service quality assurance capability in multi-tenant scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115241A_ABST
    Figure CN123115241A_ABST
Patent Text Reader

Abstract

This invention discloses a massive information intelligent search system for distributed databases. The invention relates to the fields of distributed databases and intelligent storage technology. The system includes an application retrieval layer, a distributed metadata management layer, a semantic-aware storage control layer, and a physical storage medium layer. The semantic-aware storage control layer is equipped with a semantic mapping prefetching engine, a semantic heat level scheduling engine, and a multi-tenant joint scheduling engine. The former selects clusters of data blocks to be prefetched based on the semantic distance between the query semantic vector and the semantic centroid vector; the latter determines the current heat value of the data block based on the heat decay value and semantic heat increment to drive the migration of hot and cold media; and the latter allocates bandwidth and computing time slices according to the tenant service level agreement weight. This invention improves the semantic retrieval prefetching hit rate, achieves seamless migration before hot spot changes, and ensures multi-tenant resource isolation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of distributed database and intelligent storage technology, specifically to a massive information intelligent search system for distributed databases. Background Technology

[0002] When distributed databases handle massive data storage and retrieval, they typically adopt a layered architecture: the computation layer executes indexing algorithms such as vector similarity retrieval, the distributed file system layer is responsible for data sharding and metadata maintenance, and the storage control layer maps logical data blocks to physical storage media and performs data prefetching, cache replacement, and hot and cold data layered migration.

[0003] Existing underlying storage prefetching mechanisms mainly rely on the principles of spatial locality and temporal locality. This means that data blocks with logically adjacent addresses are likely to be accessed consecutively in the near future, or high-frequency data is kept in the cache based on historical access frequency and time. Existing hot and cold data tiering mechanisms are usually based on statistical rules such as data creation time, recent access time, or access count thresholds. Cross-node data retrieval is only initiated when cold data is detected to be accessed. In addition, existing computable storage devices push down computational tasks such as filtering and aggregation to the local computing unit of the storage device for execution, but their scheduling strategies are mostly first-come, first-served or simple round-robin, without resource isolation for service level differences in multi-tenant scenarios.

[0004] The closest existing technical solution to this invention is: the distributed database storage engine determines whether to trigger prefetching based on whether the logical block addresses are consecutive. When the addresses are consecutive, the subsequent logical blocks are prefetched. When the addresses are not consecutive, synchronous read requests are initiated block by block. The cache is maintained based on the least recently used queue. When cold data is detected to be hit, cross-node data retrieval is initiated synchronously.

[0005] The above schemes rely on physical address locality or historical access statistics for scheduling, which is difficult to adapt to the characteristics of highly discrete addresses, frequent changes in popularity, and multi-tenant resource contention in semantic retrieval scenarios. They also suffer from problems such as prefetching failure, migration lag, and insufficient resource isolation. Summary of the Invention

[0006] The technical problem to be solved by this invention is that the prefetching mechanism of the existing distributed database storage control layer mainly schedules based on the continuity of logical block addresses or historical access frequency, while semantic retrieval requests present highly discrete access patterns on the physical storage medium, resulting in the failure of the prefetching mechanism, a decrease in cache hit rate, and severe read / write amplification.

[0007] The existing hot and cold data stratification mechanism makes migration decisions based solely on historical access statistics, which is unrelated to the hot data mutation patterns determined by the evolution of user intent in semantic retrieval scenarios. It cannot complete data migration before semantic hotspots undergo mutations, resulting in long-tail query delays caused by synchronous cross-node fetching.

[0008] Existing computational pushdown scheduling strategies do not isolate resources based on the service level differences among multiple tenants, resulting in high-priority tenant requests being blocked by dense requests from low-priority tenants. This invention aims to provide a distributed database-based intelligent search system for massive information to solve the above-mentioned technical problems.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] A massive information intelligent search system for distributed databases includes an application retrieval layer, a distributed metadata management layer, a semantic-aware storage control layer, and a physical storage medium layer. The application retrieval layer receives retrieval requests and outputs a list of logical addresses of candidate data blocks and a query semantic vector corresponding to the retrieval request. The distributed metadata management layer maintains the logical addresses, semantic centroid vector labels, and physical location mapping relationships of physical data blocks. The semantic-aware storage control layer is connected to the distributed metadata management layer and the physical storage medium layer, and includes a semantic mapping prefetching engine, a semantic heat-based hierarchical scheduling engine, and a multi-tenant joint scheduling engine. The semantic mapping prefetching engine is used to retrieve data from multiple data blocks corresponding to the multiple semantic centroid vectors based on the semantic distance between the query semantic vector and multiple pre-obtained semantic centroid vectors. The system selects a preset number of data block clusters with the closest semantic distance from the data block clusters, prefetches the physical data blocks corresponding to the selected data block clusters from the physical storage medium layer to the storage-level cache, and uses the semantic heat level scheduling engine to obtain the current semantic heat value of the data block located in the cold storage medium in the physical storage medium layer based on the historical semantic heat value and the semantic heat increment calculated from the query semantic vector sequence. When the current semantic heat value reaches a preset migration threshold, the system migrates the data block from the cold storage medium to the hot storage medium in the physical storage medium layer. The multi-tenant joint scheduling engine allocates flash channel bandwidth time slices and computing unit computing power time slices to the multiple tenants according to the service level protocol priority weights of the multiple tenants accessing the system, according to the proportion of the priority weights.

[0011] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0012] 1. This invention selects clusters of data blocks to be prefetched based on the semantic distance between the query semantic vector and the semantic centroid vector, so that the prefetch range is associated with the semantic similarity of the data rather than the physical address continuity, thereby improving the prefetch hit rate in semantic retrieval scenarios and reducing read / write amplification caused by random read / write.

[0013] 2. This invention reduces the long-tail latency of queries caused by synchronous cross-node fetching by predicting and triggering data migration between cold and hot storage media in advance based on the superposition result of historical heat decay value and semantic heat increment before the data block is hit by the real semantic hot spot, and realizes application-layer-agnostic hierarchical scheduling.

[0014] 3. This invention achieves isolation of underlying storage hardware resources in a multi-tenant environment by jointly allocating flash channel bandwidth time slices and computing unit computing power time slices based on the priority weight of the tenant service level agreement. This reduces the interference of low-priority tenant requests on high-priority tenant requests and improves the service quality assurance capability in multi-tenant scenarios. Attached Figure Description

[0015] Figure 1 This is a diagram illustrating the overall architecture of the intelligent search system for massive information in this embodiment of the invention.

[0016] Figure 2 This is a flowchart illustrating the semantic mapping prefetching engine's execution of semantic awareness prefetching in an embodiment of the present invention.

[0017] Figure 3 This is a flowchart of the multi-tenant time-slice scheduling process of the multi-tenant joint scheduling engine in an embodiment of the present invention. Detailed Implementation

[0018] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0019] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be noted that the embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0020] The distributed database intelligent search system for massive information provided by this invention is divided into an application retrieval layer, a distributed metadata management layer, a semantic-aware storage control layer, and a physical storage medium layer in its overall architecture. The semantic-aware storage control layer is the core innovation of this invention. It has a semantic mapping prefetching engine, a semantic heat level scheduling engine, and a multi-tenant joint scheduling engine set up in parallel. The three of them exchange query semantic vectors and heat status information through a shared semantic feature bus. The physical storage medium layer includes hot storage media and cold storage media, which are connected through a remote direct memory access channel and support asynchronous data block migration.

[0021] The application retrieval layer includes a vector retrieval index module, which receives retrieval requests from users, executes retrieval algorithms such as graph index traversal, and outputs a list of logical addresses of candidate data blocks and the query semantic vector corresponding to this retrieval request. The distributed metadata management layer includes metadata service nodes, which maintain information such as data block logical addresses, semantic centroid vector labels, and physical location mapping tables, and distribute semantic labels to the semantic-aware storage control layer through the storage interface. The calling relationship between the layers is as follows: the application retrieval layer initiates retrieval requests with query semantic vectors downwards, the semantic-aware storage control layer returns prefetched data blocks upwards, and the asynchronous migration between cold storage media and hot storage media is driven by a semantic heat-based hierarchical scheduling engine.

[0022] I. Specific Implementation Methods of the Semantic Mapping Prefetch Engine

[0023] The core principle of the semantic mapping prefetch engine is to attach a semantic centroid vector label of its cluster to the metadata of the distributed metadata management layer for each physical data block.

[0024] When the application retrieval layer initiates a retrieval request and provides the query semantic vector, the semantic mapping prefetch engine does not determine the prefetch range based on the continuity of logical block addresses. Instead, it calculates the semantic distance between the query semantic vector and each semantic centroid vector and prefetches the physical data blocks corresponding to the clusters with the closest semantic distance to the storage-level cache.

[0025] Semantic distance is calculated using the normalized vector dot product, as shown in the following formula:

[0026] ;

[0027] in, Indicates the query semantic vector and the first The semantic distance between semantic centroid vectors ranges from zero to two, with smaller values ​​indicating higher semantic similarity.

[0028] This represents the query semantic vector of the current retrieval request;

[0029] Indicates the first The semantic centroid vector of a cluster of data blocks;

[0030] Representing vectors with vector The dot product;

[0031] and Representing vectors respectively with vector The Euclidean norm.

[0032] The specific process of semantic mapping prefetching engine performing semantic-aware prefetching is as follows: In the offline stage, the distributed metadata management layer performs clustering on the physical data blocks that have been written to the physical storage medium layer according to their content feature vectors, resulting in multiple data block clusters, and calculates the semantic centroid vector of each data block cluster. The centroid vector and its logical address list for the physical data block are recorded in the semantic centroid index library. Simultaneously, the cluster number of each physical data block is permanently stored in the storage mapping table entry for that data block. During the online phase, when the application retrieval layer initiates a retrieval request, the vector retrieval index module outputs the query semantic vector for this retrieval while completing the graph index traversal. ;

[0033] The semantic mapping prefetch engine receives query semantic vectors. Then, the semantic centroid index is traversed, and the calculation is performed according to the aforementioned formula. With each semantic centroid vector semantic distance The clusters are sorted by semantic distance from smallest to largest, and a predetermined number of clusters at the top of the sorting are selected as the candidate cluster set for this pre-selection.

[0034] Then, the logical address list of all physical data blocks in the candidate cluster set is obtained from the distributed metadata management layer, a batch prefetch request is generated, and these physical data blocks are read from the physical storage medium layer to the storage-level cache through the asynchronous input / output interface;

[0035] Finally, the list of data block addresses that the application retrieval layer actually needs to access is compared with the list of addresses in the prefetch cache. Data blocks that match the match are returned directly from the cache, while data blocks that do not match trigger synchronous reading, and the miss information is fed back to the semantic centroid index library for iterative correction of the corresponding semantic centroid vector.

[0036] For example, let's say the query semantic vector for a certain retrieval semantic centroid vectors of the three candidate clusters , , The semantic distances calculated between them are respectively , , If the preset prefetch width parameter is two, the semantic mapping prefetch engine selects the physical data blocks corresponding to cluster one and cluster two with the smallest semantic distance as the prefetch objects for this time, while the data blocks corresponding to cluster three do not participate in this prefetch. This achieves the association between the prefetch range and semantic similarity, rather than the association with physical address continuity.

[0037] II. Specific Implementation Methods of the Semantic Popularity-Based Hierarchical Scheduling Engine

[0038] The core principle of the semantic popularity hierarchical scheduling engine is to maintain a semantic popularity value for each data block that decays over time. This popularity value is driven by a decay term driven by historical access statistics and a popularity increment calculated by the retrieval query semantic features obtained from bypass monitoring.

[0039] When the semantic heat value of a data block in a cold storage medium reaches the preset migration threshold, the block is migrated to a hot storage medium in advance through a remote direct memory access channel, and the media migration is completed before the actual query arrives.

[0040] The semantic popularity is updated using the following formula:

[0041] ;

[0042] in, Represents data blocks At any moment The semantic popularity value;

[0043] This indicates the semantic heat value of the block at the previous time step;

[0044] This represents the preset heat attenuation coefficient. A larger value indicates a faster decline in historical popularity;

[0045] Indicates the time interval between two popularity updates;

[0046] Indicates the time interval The semantic heat increment obtained by internal calculation.

[0047] Increase in popularity The formula is as follows: The similarity between the query semantic vector sequence and the block semantic centroid vector is weighted and summed.

[0048] ;

[0049] in, Indicates the time interval The total number of query requests detected by the internal bypass;

[0050] Indicates the first The semantic vector of a query request;

[0051] Indicates the first The weight of each query request in the direction of query sequence intent evolution is determined based on the proximity of the query semantic vector to the direction of intent evolution; the closer the proximity, the larger the weight value.

[0052] The result of the aforementioned semantic distance formula is as follows: This indicates that the query request is related to the block. Semantic similarity.

[0053] The specific process of the semantic popularity hierarchical scheduling engine to execute semantic popularity decay and mutation prediction hierarchical scheduling is as follows: continuously collect the query semantic vector sequence issued by the application retrieval layer in the side-channel listening module and record the corresponding timestamp;

[0054] Based on the evolution trajectory of several recent query semantic vectors, sequence similarity analysis is used to calculate the weight of each query request in the direction of intent evolution. ;

[0055] For each data block located on the cold storage medium Calculated based on the aforementioned formula for heat increment ;

[0056] Based on the aforementioned heat update formula, the blocks are... The current semantic popularity value is obtained by exponentially decaying the historical semantic popularity value and then adding the current popularity increment. ;

[0057] Will The data is compared with a preset migration threshold; if the threshold is reached, the block is... Mark the block as awaiting migration and add it to the migration queue;

[0058] The migration scheduling unit retrieves blocks to be migrated from the migration queue in descending order of popularity value, initiates a data block copy from cold storage medium to hot storage medium through the remote direct memory access channel, takes physical data blocks as the atomic operation unit, and synchronously updates the physical location mapping and version number of the block in the distributed metadata management layer.

[0059] After the migration is completed, if the semantic popularity value of the block continues to fall below the product of the preset migration threshold and the preset fall coefficient in the subsequent time window, a reverse migration is triggered, and the block is downgraded from hot storage medium to cold storage medium, releasing hot storage space and forming a dynamic cyclic scheduling. The preset fall coefficient is greater than zero and less than one.

[0060] For example, suppose the semantic heat value of a cold-stored data block at the previous time step is... Heat decay coefficient for The time interval for this update For ten minutes, the historical heat value after decay is approximately ;

[0061] Five query requests were detected during this time interval, and the increase in popularity was calculated by weighted summation. The current semantic popularity value is approximately ;

[0062] If the preset migration threshold is If the block is marked as a block to be migrated, it will be migrated from the cold storage medium to the hot storage medium before the arrival of the real semantic hotspot, thereby realizing advance scheduling without the application layer's awareness.

[0063] III. Specific Implementation Methods of the Multi-Tenant Joint Scheduling Engine

[0064] The core principle of the multi-tenant joint scheduling engine is: within the storage controller that supports computation pushdown, based on the service level agreement priority weights of different tenants, the flash channel bandwidth time slice and the computing unit computing power time slice are dynamically allocated, so that the vector distance calculation and data filtering requests of high-priority tenants can obtain a higher proportion of underlying resource quotas, thus eliminating resource contention when multiple tenants share the underlying storage hardware.

[0065] Time slice allocation uses the following formula:

[0066] ;

[0067] in, Indicates assignment to tenant Time slice length;

[0068] Indicates the total length of time slices that can be allocated within a scheduling cycle;

[0069] Indicates tenant The service level agreement priority weight is calculated based on tenant level, historical latency default count, and current queue backlog length.

[0070] This indicates the total number of currently active tenants;

[0071] This represents the sum of the weights of all active tenants.

[0072] The specific process of multi-tenant joint scheduling engine executing multi-tenant joint time slice scheduling is as follows: maintain an independent input / output and computation request queue for each accessing tenant, and record the service level protocol level, historical latency default count and current queue backlog length for each tenant;

[0073] At the beginning of each scheduling cycle, the priority weight of each tenant is calculated according to the service level agreement level, the number of defaults, and the backlog length using a preset weighting function. ;

[0074] Based on the aforementioned time slice allocation formula, calculate the flash channel bandwidth time slice and computing unit computing power time slice that each tenant should obtain in this scheduling cycle;

[0075] Within this cycle, according to the time slice length obtained by each tenant, the right to use the flash channel and the right to use the computing unit are sequentially allocated to the pending requests of the corresponding tenant. Each tenant's request exclusively executes pushdown tasks such as vector distance calculation or data filtering within its allocated time slice.

[0076] If a tenant fails to process all requests in the queue within its allocated time slice, the unprocessed requests remain in the queue and are included in the tenant's backlog length in the next scheduling cycle for recalculation of weights.

[0077] If a tenant's actual response latency exceeds the threshold agreed upon in its service level agreement (SLA) within multiple consecutive scheduling cycles, the tenant will be given a compensatory weight boost when calculating the weight, and will be allocated more time slices until the tenant's response latency returns to the range agreed upon in the SLA.

[0078] For example, suppose there are three active tenants in a certain scheduling period, with priority weights as follows: , , If the total time slice length that can be allocated within the scheduling period is 100 milliseconds, then according to the aforementioned formula, the three tenants will obtain time slices of approximately 60 milliseconds, 30 milliseconds, and 10 milliseconds respectively, thereby enabling the retrieval requests of high-priority tenants to obtain a higher proportion of flash channel bandwidth and computing resources, and avoid being blocked by the dense requests of low-priority tenants.

[0079] IV. Specific Implementation Methods of Semantic-Aware Metadata Structure

[0080] To support the coordinated operation of the above three functional modules, the distributed metadata management layer has been expanded on the basis of the original metadata structure. Each metadata entry includes the logical address field of the physical data block, the semantic cluster number field, the current semantic heat value field, the tenant identifier field and the corresponding service level protocol level field, the current storage medium type field, and the most recent migration timestamp field. Among them, the logical address field and the semantic cluster number field are used for the semantic mapping prefetch engine to read and write.

[0081] The current semantic popularity value field and storage medium type field are available for reading and writing by the semantic popularity hierarchical scheduling engine;

[0082] The tenant identifier field and the corresponding service level agreement level field are available for the multi-tenant joint scheduling engine to read and write. The three functional modules share the same metadata base and work together to form a system-level closed loop, avoiding information inconsistency caused by decentralized maintenance of metadata.

[0083] V. Other Implementation Methods

[0084] In an alternative implementation, the generation and updating of semantic centroid vectors can be achieved not through offline clustering, but through an online incremental clustering algorithm. The semantic centroid index library is updated in real time during the data writing process, enabling the cluster centroids to dynamically adapt to the drift of data distribution. This implementation has better adaptability in scenarios where the data distribution changes rapidly over time.

[0085] In another alternative implementation, the hot and cold data migration triggering mechanism does not rely entirely on the heat increment calculated from the query semantic vector sequence. Instead, it adopts an association rule mining method based on historical query logs to pre-count offline the set of semantically related data blocks that are frequently retrieved. When any block in the set is accessed and hit, other blocks in the association set that are in cold storage media are pre-fetched. This implementation has a lower implementation cost in deployment scenarios where computing resources are limited and it is difficult to run intent prediction in real time.

[0086] The above embodiments are only some embodiments of the present invention. Any technical solutions obtained by equivalent substitution or simple modification under the technical concept of the present invention should fall within the protection scope of the present invention.

[0087] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A massive information intelligent search system for a distributed database, comprising an application retrieval layer, a distributed metadata management layer, a semantic-aware storage control layer, and a physical storage medium layer; The application retrieval layer is used to receive retrieval requests, output a list of logical addresses of candidate data blocks and a query semantic vector corresponding to the retrieval request; The distributed metadata management layer is used to maintain the logical address, semantic centroid vector label, and physical location mapping relationship of physical data blocks; The semantic-aware storage control layer is communicatively connected to the distributed metadata management layer and the physical storage medium layer, characterized in that... The semantic-aware storage control layer includes a semantic mapping prefetch engine, a semantic heat-based hierarchical scheduling engine, and a multi-tenant joint scheduling engine. The semantic mapping prefetch engine is used to select a preset number of data block clusters with the closest semantic distance from the multiple data block clusters corresponding to the multiple semantic centroid vectors based on the semantic distance between the query semantic vector and the multiple pre-obtained semantic centroid vectors, and prefetch the physical data blocks corresponding to the selected data block clusters from the physical storage medium layer to the storage-level cache. The semantic heat level scheduling engine is used to obtain the current semantic heat value of the data block located in the cold storage medium in the physical storage medium layer based on the historical semantic heat value and the semantic heat increment calculated from the query semantic vector sequence, and when the current semantic heat value reaches the preset migration threshold, the data block is migrated from the cold storage medium to the hot storage medium in the physical storage medium layer. The multi-tenant joint scheduling engine is used to allocate flash channel bandwidth time slices and computing unit computing power time slices to the multiple tenants according to the service level agreement priority weights of the multiple tenants accessing the system, and according to the proportion of the priority weights.

2. The system of claim 1, wherein, The semantic mapping prefetching engine is also used to: cluster physical data blocks that have been written to the physical storage medium layer according to content feature vectors to obtain multiple data block clusters, calculate the semantic centroid vector of each data block cluster, and record the semantic centroid vector and the logical address list of the physical data blocks contained in the data block cluster in the semantic centroid index library. The normalized vector dot product between the query semantic vector and each semantic centroid vector recorded in the semantic centroid index is used as a similarity measure. Based on this, the semantic distance is determined, and the data blocks are sorted from smallest to largest according to the semantic distance. A preset number of data block clusters at the top of the sorted list are selected as clusters to be prefetched.

3. The system of claim 2, wherein, The semantic mapping prefetching engine is also used to: compare the logical address list of candidate data blocks output by the application retrieval layer with the logical address list of physical data blocks prefetched into the storage-level cache; return the matching physical data blocks from the storage-level cache; initiate synchronous reading of the matching physical data blocks; and feed back the missing information to the semantic centroid index library for correcting the corresponding semantic centroid vector.

4. The system of claim 1, wherein, The semantic heat level scheduling engine is also used to: exponentially decay the historical semantic heat value of the data block according to a preset decay coefficient to obtain the decayed historical semantic heat value. The system listens to the query semantic vector sequence emitted by the application retrieval layer in the bypass mode, determines the intent weight based on the proximity of each query semantic vector in the sequence to the intent evolution direction, and obtains the semantic heat increment by weighted summation based on the intent weight and the similarity between each query semantic vector and the semantic centroid vector of the data block. The current semantic heat value of the data block is obtained by adding the attenuated historical semantic heat value to the semantic heat increment.

5. The system of claim 1, wherein, The semantic heat level scheduling engine is also used to: continuously monitor the semantic heat value of the data block after the data block is migrated from the cold storage medium to the hot storage medium; if the semantic heat value of the data block falls below the product of the preset migration threshold and the preset fall coefficient, then the data block is migrated from the hot storage medium back to the cold storage medium, wherein the preset fall coefficient is greater than zero and less than one.

6. The system of claim 1, wherein, The multi-tenant joint scheduling engine is also used to: maintain an independent input / output and computation request queue for each tenant accessing the system, and record the service level agreement level, historical latency default count and current queue backlog length of each tenant; The priority weight of each tenant is determined based on the service level agreement level, the number of historical delay defaults, and the current queue backlog length. Within each scheduling cycle, the flash channel bandwidth time slice and the computing unit computing power time slice are calculated and allocated according to the proportion of each tenant's priority weight to the sum of the priority weights of all active tenants.

7. The system of claim 6, wherein, The multi-tenant joint scheduling engine is also used to: if the actual response latency of a tenant exceeds the threshold agreed upon in the service level agreement for the tenant in multiple consecutive scheduling cycles, then give a compensatory weight increase when calculating the priority weight of the tenant, until the actual response latency of the tenant recovers to the range agreed upon in the service level agreement.

8. The system according to claim 1, characterized in that, The metadata entries maintained by the distributed metadata management layer include the logical address field of the physical data block, the semantic cluster number field, the current semantic heat value field, the tenant identifier field and the corresponding service level protocol level field, the current storage medium type field, and the most recent migration timestamp field. The metadata entries are available for joint reading, writing, and calling by the semantic mapping prefetch engine, the semantic heat hierarchical scheduling engine, and the multi-tenant joint scheduling engine.

9. A method for intelligent search of massive information in a distributed database, applied to a distributed database system including an application retrieval layer, a distributed metadata management layer, a semantically aware storage control layer, and a physical storage medium layer, characterized in that, The method includes: Receive a retrieval request and output a list of logical addresses of candidate data blocks and a query semantic vector corresponding to the retrieval request; Based on the semantic distance between the query semantic vector and the multiple pre-obtained semantic centroid vectors, a preset number of data block clusters with the closest semantic distance are selected from the multiple data block clusters corresponding to the multiple semantic centroid vectors, and the physical data blocks corresponding to the selected data block clusters are prefetched from the physical storage medium layer to the storage-level cache. Based on the historical semantic heat value and the semantic heat increment calculated from the query semantic vector sequence, the current semantic heat value of the data block located in the cold storage medium in the physical storage medium layer is obtained, and when the current semantic heat value reaches the preset migration threshold, the data block is migrated from the cold storage medium to the hot storage medium in the physical storage medium layer. Based on the service level agreement priority weights of the multiple tenants, and in proportion to the priority weights, flash channel bandwidth time slices and computing unit computing power time slices are allocated to the multiple tenants respectively.

10. The method of claim 9, wherein, Before the step of selecting a preset number of data block clusters with the closest semantic distance from the multiple data block clusters corresponding to the multiple semantic centroid vectors based on the semantic distance between the query semantic vector and the multiple pre-obtained semantic centroid vectors, the method further includes: clustering the physical data blocks that have been written to the physical storage medium layer according to the content feature vector to obtain the multiple data block clusters, and calculating the semantic centroid vector of each data block cluster.