An intelligent hierarchical caching method and system based on hyper-converged architecture

Abstract

The application discloses an intelligent hierarchical caching method and system based on a super-fusion architecture, which comprises fusion architecture collection, initial caching hierarchy, three-way load prediction, caching resource scheduling and intelligent hierarchical caching. The application relates to the technical field of data caching intelligent hierarchy, and particularly discloses an intelligent hierarchical caching method and system based on a super-fusion architecture. The initial caching hierarchy method based on heat score calculation is adopted in the scheme, a more dynamic, fine-grained and controllable three-layer caching initialization structure is realized, and the cold and hot data misplacement rate and initial scheduling cost are effectively reduced. A time sequence alignment enhanced adaptive load three-channel prediction model is adopted, a time sequence alignment enhancement and neighborhood enhancement mechanism are introduced, a double-channel gate fusion backbone network of a time channel and a resource channel is combined, and three-way load prediction is performed. A reinforcement learning method based on risk-sensitive reward improvement is adopted for caching resource scheduling.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of data cache intelligent grading, and particularly relates to an intelligent grading cache method and system based on a hyper-converged architecture. BACKGROUND

[0002] The intelligent grading cache method and system based on the hyper-converged architecture is a distributed technology integrating computing, storage and network resources, which dynamically grades data and intelligently schedules cache strategies through a machine learning algorithm, so that high-frequency access data is preferentially stored in a high-performance storage layer, and low-frequency data is sunk to a low-cost storage layer. The method and system significantly reduce I / O delay and improve resource utilization, and rely on the horizontal expansion capability of the hyper-converged architecture to guarantee system high availability, and are suitable for cloud computing, edge computing and other scenarios requiring real-time response and massive data processing.

[0003] However, the existing intelligent method has the problems of design fragmentation at the overall architecture level, disconnection between prediction and scheduling in each link, inability of the cache grading strategy to dynamically optimize with the system running state, and difficulty in balancing long-term performance and medium life.

[0004] In the existing initial cache grading setting process, the existing initial cache grading setting depends on static medium characteristics or simple access frequency, lacks comprehensive modeling of access timing, read-write characteristics and request size, and causes the technical problem of disconnection between the initial grading result and actual business load.

[0005] In the existing load prediction method, the traditional load prediction method has the technical problems of coarse prediction granularity in cache management, insufficient capture of bursty access patterns and neighborhood correlation, and lack of constraint fusion between multi-channel prediction results, which easily leads to unbalanced scheduling strategies.

[0006] In the existing cache resource scheduling method, the simple reward function design only considers performance improvement or migration cost, and does not uniformly model physical constraints such as capacity, bandwidth, temperature and life and delay risks, which easily leads to violation of hard constraints or life overdraft under extreme load. SUMMARY

[0007] In view of the above, in order to overcome the defects of the prior art, the present application provides an intelligent hierarchical caching method and system based on a hyper-converged architecture. In the existing data hierarchical caching method based on the hyper-converged architecture, there are problems such as design fragmentation in the overall architecture level, disconnection between prediction and scheduling in each link of the existing intelligent method, resulting in that the caching hierarchical strategy cannot be dynamically optimized according to the system running state, and the long-term performance and medium life are difficult to balance. The present scheme starts from the top-level design of the system, and constructs a progressive integrated technology link composed of "initial hierarchical - multi-dimensional load prediction - constraint-aware scheduling - intelligent cache execution". Under the unified architecture, the whole process linkage of resource state awareness, load trend prediction, risk constraint optimization and hierarchical execution is realized, which significantly improves the dynamic adaptability and global optimization ability of cache management, and solves the structural bottleneck of the traditional method "static initial setting, step-by-step independent, lack of feedback". In the existing initial cache hierarchical setting process, there are technical problems such as that the existing initial cache hierarchical setting depends on static medium characteristics or simple access frequency, lacks comprehensive modeling of access timing, read-write characteristics and request size, and the initial hierarchical result is disconnected with the actual business load. The present scheme creatively adopts an initial cache hierarchical method based on heat score calculation, introduces weight indexes such as access frequency, recency, read-write ratio and request size, and combines hysteresis interval control and use condition restriction, to realize a more dynamic, fine-grained and controllable three-layer cache initialization structure, effectively reducing the cold and hot data misplacement rate and initial scheduling cost. In the existing load prediction method, there are technical problems such as that the traditional load prediction method has coarse prediction granularity in cache management, is insufficient in capturing bursty access patterns and neighborhood correlation, and lacks constraint fusion between multi-channel prediction results, which easily leads to unbalanced scheduling strategy. The present scheme creatively adopts a time alignment enhanced adaptive load three-channel prediction model, performs three-way load prediction, introduces time alignment enhancement and neighborhood enhancement mechanisms, combines double-channel gating fusion backbone network of time channel and resource channel, and embeds prototype constraint, cross-time domain consistency constraint and monotonicity constraint in access type, access heat and high delay risk three-way prediction, to realize significant improvement of the prediction result in time consistency, category stability and resource availability, and provide a more robust multi-dimensional prediction basis for cache scheduling. In the existing cache resource scheduling method, there are technical problems such as that the simple reward function design only considers performance improvement or migration cost, without unified modeling of physical constraints such as capacity, bandwidth, temperature and life and delay risk, which easily leads to violation of hard constraints or causes life overdraft under extreme load. The present scheme creatively adopts a reinforcement learning method based on risk-sensitive reward improvement, performs cache resource scheduling, so that the scheduling decision can dynamically balance between performance and resource safety, and significantly reduces the risk of cache system instability or medium premature aging under high load.

[0008] The technical scheme adopted by the present application is as follows: the present application provides an intelligent hierarchical caching method based on an ultra-converged architecture, which comprises the following steps:

[0009] Step S1: converging architecture collection;

[0010] Step S2: initial cache hierarchical arrangement;

[0011] Step S3: three-way load prediction;

[0012] Step S4: cache resource scheduling;

[0013] Step S5: intelligent hierarchical caching.

[0014] Further, in step S1, the converging architecture collection is used to collect and organize the running conditions of computing, storage and network devices and data access condition data, specifically, the data resource is collected, integrated and stored through the construction of an ultra-converged architecture to obtain system running original data;

[0015] The specific content of the system running original data includes device running state data, storage medium attribute data and data access record data.

[0016] Further, in step S2, the initial cache hierarchical arrangement is used to divide the initial level for the cache space according to the speed, capacity and cost of different storage media, and set the use conditions and restrictions of each layer, specifically, according to the system running original data, the initial cache hierarchical arrangement is performed through the construction of a three-layer structure to obtain a cache level initialization structure, comprising the following steps:

[0017] Step S21: cache level division, specifically, the cache space is divided into a fast layer, a high-speed layer and a capacity layer, the fast layer uses a medium with the lowest delay and the smallest capacity in the three layers, the high-speed layer uses a medium with a medium delay and a medium capacity in the three layers, and the capacity layer uses a medium with the highest delay and the highest capacity in the three layers;

[0018] Step S22: data block management, specifically, a fixed-size data block is used as a management unit, and a hotness score is calculated according to the historical access record, the weighted access frequency factor, the access recency factor, the read-write ratio factor and the request size factor of each data block to obtain the hotness score;

[0019] Step S23: file writing strategy setting, specifically, the file writing in the storage system is set with a hierarchical strategy, including: temporarily storing the newly written data in the write buffer of the fast layer, and subsequently migrating the data to other storage layers according to the access characteristic judgment result in priority order;

[0020] The sequential read data detected as large volume is preferentially written to the high-speed layer to reduce the occupancy of the fast layer and optimize the subsequent access performance;

[0021] Step S24: hysteresis interval control, used as a scheduling screening condition of the cache resource scheduling in step S4, specifically setting a double-threshold hysteresis control mechanism, setting independent hotness score thresholds for the up-regulation and down-regulation between the fast layer, the high-speed layer and the capacity layer, and setting the data block regulation mechanism between the layers;

[0022] Step S25: usage condition limitation, specifically setting unified usage conditions and limitations for different layers, including limiting the capacity usage rate of each layer to be not higher than a preset capacity threshold; when the device temperature reaches a preset temperature threshold, reducing the data migration rate and preferentially sinking the data to the low-speed layer; setting a write budget for the solid-state medium according to a period and limiting the write proportion to be not more than the budget; for data with a single write exceeding a preset object capacity threshold, defaulting to be placed in the high-speed layer or the capacity layer; and reserving a minimum access bandwidth for foreground services in various operations to guarantee service continuity;

[0023] Step S26: initial cache grading, specifically generating a cache level initialization structure, including the medium type and capacity upper limit, the migration rate upper limit, the temperature and lifetime threshold, the hysteresis interval, the preheating list and the fixed list of the three layers, and the initial level mapping table of each data block;

[0024] The cache level initialization structure is used as the input of the subsequent steps, for analysis by the three-way load prediction in step S3, and for planning and boundary control by the cache resource scheduling in step S4.

[0025] Further, in step S3, the three-way load prediction is used to predict the service type, access hotness and high-delay risk of each type of data, specifically using a time sequence alignment enhanced improved adaptive load three-channel prediction model to perform three-way load prediction according to the system running original data and the cache level initialization structure, to obtain load comprehensive prediction reference data, including the following steps:

[0026] Step S31: time-aware alignment, specifically extracting the access time sequence from the system running original data, and calculating the time-aware alignment factor of each data item to obtain time-aligned access feature data;

[0027] Step S32: access field enhancement, specifically constructing the access hotness feature between adjacent data items by calculating the neighborhood enhancement weight between adjacent data items to obtain neighborhood-enhanced access feature data;

[0028] Step S33: adaptive backbone network construction, specifically constructing a dual-channel gated fusion network as an adaptive backbone network for data feature extraction, including a time channel subnet, a resource channel subnet and a feature fusion subblock, and performing feature extraction using the adaptive backbone network to obtain fusion feature data;

[0029] The time channel subnet specifically performs time sequence convolution processing on the time feature vector and extracts access form feature; the resource channel subnet specifically constructs a multilayer perceptron and extracts resource state feature from the cache resource state parameter; the feature fusion subblock specifically performs gated fusion of the access form feature and the resource state feature by constructing a gated fusion mechanism to obtain fusion feature data;

[0030] Step S34: three-way prediction channel construction, specifically constructing an access type prediction channel, an access heat prediction channel and a high delay risk prediction channel based on the adaptive backbone network to obtain three-way prediction channels, and performing data prediction for access type prediction tasks, access heat prediction tasks and high delay risk prediction tasks to obtain task branch prediction data;

[0031] Step S35: soft budget constraint loss improvement, specifically introducing a soft budget constraint loss, calculating the data occupancy capacity exceeding the threshold under the predicted heat in the next 10 minutes, comparing the data occupancy capacity exceeding the threshold with the available budget capacity, and adding the exceeding part to the optimization target as a square loss, limiting the prediction result within the system capacity and resource budget, and integrating the soft budget constraint loss with the original prototype constraint loss, the cross-time domain consistency loss, the monotonicity constraint loss and the task prediction loss to obtain a joint optimization loss function;

[0032] Step S36: three-way load prediction, specifically constructing the adaptive backbone network and the three-way prediction channel, combining the joint optimization loss function, and performing model training to obtain a three-way load prediction model, and using the three-way load prediction model to obtain load comprehensive prediction reference data;

[0033] The load comprehensive prediction reference data specifically includes data identification, access prediction type, predicted access heat, predicted access delay and high delay risk prediction probability.

[0034] Further, in step S4, the cache resource scheduling is used to formulate an optimal placement and migration scheme of data among cache layers under the constraints of capacity, bandwidth, temperature and lifetime, specifically, based on the system running original data, the cache level initialization structure and the load comprehensive prediction reference data, a reinforcement learning method based on risk-sensitive reward improvement is used for cache resource scheduling to obtain cache resource scheduling decision data, including the following steps:

[0035] Step S41: state constraint construction, based on the system running original data, the cache level initialization structure and the load comprehensive prediction reference data, a scheduling state vector and a constraint boundary are constructed, the state vector includes the heat, delay risk, current level and recent migration record of each data object, and the constraint boundary includes the hierarchical capacity, available bandwidth, temperature and lifetime threshold and its budget parameter;

[0036] Step S42: risk-sensitive reward improvement, used to construct a reinforcement learning problem with budget and bandwidth constraints, specifically, a reward function is constructed by introducing a performance benefit item, a migration cost penalty and a risk-sensitive constraint penalty to obtain a risk-sensitive reward improvement reward function;

[0037] Step S43: reinforcement learning training, specifically, through the state constraint construction and the risk-sensitive reward improvement, an action vector is defined, and reinforcement learning training is performed to obtain a cache resource scheduling model;

[0038] The action vector includes a level migration action, a migration target level and a migration speed limit;

[0039] The level migration action specifically includes a keep level, an up level and a down level operation;

[0040] Step S44: cache resource scheduling, specifically, by using the cache resource scheduling model, based on the system running original data, the cache level initialization structure and the load comprehensive prediction reference data, cache resource scheduling is performed to obtain cache resource scheduling decision data;

[0041] The cache resource scheduling decision data specifically includes a cache resource scheduling migration object identifier, a cache resource scheduling operation type, a cache resource source level, a cache resource target level, a migration speed limit and a decision time.

[0042] Further, in step S5, the intelligent hierarchical cache is used to perform migration and adjustment of data between different cache layers, specifically, according to the cache resource scheduling decision data, the migration and hierarchical adjustment of data between different cache layers are controlled, under the condition of meeting the operation constraint and the hotness score threshold, the target data is moved up, sunk down or adjusted across layers according to a preset migration strategy, and the distribution of data in each layer is dynamically updated in combination with real-time business priority and access hotness, so as to form a cache scheduling structure meeting the performance and life balance, and obtain an intelligent cache scheduling structure.

[0043] The application provides an intelligent hierarchical cache system based on a hyper-converged architecture, which comprises a data collection module, an initial hierarchical module, a load prediction module, a resource scheduling module and an intelligent hierarchical module.

[0044] The data collection module is used for hyper-converged architecture collection, and the system operation original data is obtained through the hyper-converged architecture collection, and the system operation original data is sent to the initial hierarchical module, the load prediction module and the resource scheduling module.

[0045] The initial hierarchical module is used for initial cache hierarchical, and the cache hierarchical initialization structure is obtained through the initial cache hierarchical, and the cache hierarchical initialization structure is sent to the load prediction module and the resource scheduling module.

[0046] The load prediction module is used for three-way load prediction, and the load comprehensive prediction reference data is obtained through the three-way load prediction, and the load comprehensive prediction reference data is sent to the resource scheduling module.

[0047] The resource scheduling module is used for cache resource scheduling, and the cache resource scheduling decision data is obtained through the cache resource scheduling, and the cache resource scheduling decision data is sent to the intelligent hierarchical module.

[0048] The intelligent hierarchical module is used for intelligent hierarchical cache, and the intelligent cache scheduling structure is obtained through the intelligent hierarchical cache.

[0049] The application has the following beneficial effects by adopting the above scheme:

[0050] (1) In view of the problem that in the existing data hierarchical caching method based on hyper-converged architecture, the existing intelligent method has design fragmentation at the overall architecture level, the prediction and scheduling between the links are relatively disconnected, the caching hierarchical strategy cannot be dynamically optimized with the system running state, and the long-term performance and medium life are difficult to balance, the scheme starts from the top-level design of the system, and constructs a progressive integrated technology link composed of "initial hierarchical - multi-dimensional load prediction - constraint aware scheduling - intelligent cache execution", realizes the whole-process linkage of resource state awareness, load trend prediction, risk constraint optimization and hierarchical execution under the unified architecture, significantly improves the dynamic adaptability and global optimization ability of cache management, and solves the structural bottleneck of traditional method "static initial setting, step-by-step independent, lack of feedback";

[0051] (2) In view of the problem that in the existing initial cache hierarchical setting process, the existing initial cache hierarchical setting depends on static medium characteristics or simple access frequency, lacks comprehensive modeling of access timing, read-write characteristics and request size, and the initial hierarchical result is disconnected with the actual business load, the scheme creatively adopts an initial cache hierarchical method based on heat score calculation, introduces weight indexes such as access frequency, recency, read-write ratio and request size, and combines hysteresis interval control and use condition restriction, realizes a more dynamic, fine-grained and controllable three-layer cache initialization structure, effectively reduces the cold and hot data misplacement rate and initial scheduling cost;

[0052] (3) In view of the problem that in the existing load prediction method, the traditional load prediction method has coarse prediction granularity in cache management, and lacks capture of bursty access mode and neighborhood correlation, and the multi-channel prediction results lack constraint fusion, which easily leads to unbalanced scheduling strategy, the scheme creatively adopts a time alignment enhanced adaptive load three-channel prediction model, performs three-way load prediction, introduces time alignment enhancement and neighborhood enhancement mechanism, combines double-channel gating fusion backbone network of time channel and resource channel, and embeds prototype constraint, cross-time domain consistency constraint and monotonicity constraint in access type, access heat and high delay risk three-way prediction, realizes significant improvement of prediction results in time consistency, category stability and resource availability, and provides more robust multi-dimensional prediction basis for cache scheduling;

[0053] (4) In the existing cache resource scheduling method, there is a simple reward function design that only considers performance improvement or migration cost, and does not uniformly model the physical constraints of capacity, bandwidth, temperature and life span and delay risk, resulting in the technical problem that the scheduling strategy is easy to violate the hard limit or cause life overdraft under extreme load. The scheme creatively uses a reinforcement learning method improved based on a risk-sensitive reward to schedule the cache resources, so that the scheduling decision can dynamically balance between performance and resource safety, significantly reducing the risk of cache system instability or medium premature aging under high load. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A flowchart of an intelligent hierarchical cache method based on a hyper-converged architecture provided by the present application is shown in the figure.

[0055] Figure 2 A schematic diagram of an intelligent hierarchical cache system based on a hyper-converged architecture provided by the present application is shown in the figure.

[0056] Figure 3 A flowchart of the initial cache hierarchy of step S2 is shown in the figure.

[0057] Figure 4 A flowchart of the three-way load prediction of step S3 is shown in the figure.

[0058] Figure 5 A flowchart of the cache resource scheduling of step S4 is shown in the figure.

[0059] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0061] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0062] Embodiment one, refer to Figure 1The application provides an intelligent hierarchical caching method based on a hyper-converged architecture, which comprises the following steps:

[0063] Step S1: hyper-converged architecture collection;

[0064] Step S2: initial caching hierarchy;

[0065] Step S3: three-way load prediction;

[0066] Step S4: caching resource scheduling;

[0067] Step S5: intelligent hierarchical caching.

[0068] By performing the above operation, for the existing data hierarchical caching method based on the hyper-converged architecture, the existing intelligent method has the problems of design fragmentation at the overall architecture level, disconnection between prediction and scheduling in each link, leading to the fact that the caching hierarchical strategy cannot be dynamically optimized according to the system running state, and the long-term performance and medium life are difficult to balance. The present application starts from the top-level design of the system, constructs a progressive integrated technology link composed of "initial hierarchical - multi-dimensional load prediction - constraint-aware scheduling - intelligent caching execution", realizes the whole-process linkage of resource state awareness, load trend prediction, risk constraint optimization and hierarchical execution under the unified architecture, significantly improves the dynamic adaptability and global optimization ability of the cache management, and solves the structural bottleneck of the traditional method "static initial setting, step-by-step independent, lack of feedback".

[0069] Embodiment two, refer to Figure 1 and Figure 2 This embodiment is based on the above-mentioned embodiment, in step S1, the hyper-converged architecture collection is used to collect and organize the running conditions and data access condition data of the computing, storage and network devices, specifically, the data resource is collected, integrated and stored through the construction of the hyper-converged architecture, and the system running original data is obtained;

[0070] The specific content of the system running original data includes device running state data, storage medium attribute data and data access record data;

[0071] The device running state data specifically includes processor utilization, memory occupation, storage read-write speed, access queue duration, network transceiving rate and round-trip delay;

[0072] The storage medium attribute data specifically includes storage medium working temperature, available capacity and predicted life;

[0073] The data access record data specifically includes access object identifier, access type, access time, access size, access interval, return delay, hit cache record and timeout record;

[0074] Preferably, the super-converged architecture is built, specifically including hardware resource deployment, network configuration management, data collection program deployment, data integration storage and expansion design;

[0075] In a preferred embodiment, the hardware resource deployment specifically deploys at least 3 physical nodes, each node being configured with an x86 or ARM architecture server containing a processor, memory, high-speed solid-state storage medium, large-capacity mechanical storage medium and high-speed network interface; computing virtualization is achieved through server virtualization software (such as KVM or VMware ESXi), and physical CPU and memory are abstracted into a virtual resource pool; distributed storage software is used to aggregate the storage media of each node into a unified storage pool to support block, file and object storage services;

[0076] The network configuration management specifically achieves logical isolation and traffic scheduling between virtual machines by building a virtualized logical network, and allocates high-priority bandwidth for storage traffic by configuring quality of service control strategies; network function elastic expansion is achieved by embedding a virtual firewall and load balancer in the node;

[0077] The data collection program deployment specifically deploys a lightweight data collection program on each node to obtain the running information of the node, including computing resource status, storage resource status, network resource status and data access record; the data collection program samples the running information at a predetermined time interval or trigger condition and transmits it to the centralized processing area through an internal data channel;

[0078] The data integration storage specifically receives the running information by building a centralized processing area and storage component, and aggregates, classifies and standardizes the collected information from each node to obtain system running raw data in a unified format and store it in the storage component;

[0079] The expansion design specifically refers to the redundant deployment of multiple copies of the data collection program and the storage component, and through consistent hash, the data distribution is balanced when a new node is added.

[0080] Embodiment three, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above-mentioned embodiments, in step S2, the initial cache grading is used to divide the initial level for the cache space according to the speed, capacity and cost of different storage media, and set the use conditions and restrictions of each layer, specifically, according to the system running raw data, the initial cache grading is performed through the cache level initialization three-layer structure construction to obtain the cache level initialization structure, including the following steps:

[0081] Step S21: cache level division, specifically, dividing the cache space into a fast layer, a high-speed layer, and a capacity layer, the fast layer adopts a medium with the lowest delay and the smallest capacity among the three layers, the high-speed layer adopts a medium with a median delay and a median capacity among the three layers, and the capacity layer adopts a medium with the highest delay and the highest capacity among the three layers;

[0082] Preferably, the initial capacity ratio of the three layers is 10%:30%:60%, and 10% of safety margin is reserved for each layer;

[0083] Step S22: data block management, specifically, taking a fixed-size data block as a management unit, and according to historical access records, a hotness score is calculated by synthesizing a weighted access frequency factor, an access recency factor, a read-write ratio factor, and a request size factor for each data block, to obtain the hotness score;

[0084] The calculation formula of the hotness score is:

[0085] ;

[0086] In the formula, H is the hotness score, w1 is the access frequency weight, F is the access frequency factor, w2 is the access recency weight, R is the access recency factor, w3 is the read-write ratio weight, W is the read-write ratio factor, w4 is the request size weight, and S is the request size factor;

[0087] Preferably, the default recommended weight of the access frequency weight is set to 0.4, the default recommended weight of the access recency weight is set to 0.3, the default recommended weight of the read-write ratio weight is set to 0.2, and the default recommended weight of the request size weight is set to 0.1;

[0088] The specific calculation formula of the access frequency factor is:

[0089] ;

[0090] In the formula, F is the access frequency factor, ln is the natural logarithm function, count is the total number of accesses of the data block, and C max is the configurable upper limit of the number of accesses, and the default value is 10000;

[0091] The specific calculation formula of the access recency factor is:

[0092] ;

[0093] In the formula, R is the access recency factor, exp is the natural base function, is the time elapsed since the last access in hours, is the time decay factor, which is set to 6 by default, and the unit is hour;

[0094] The specific calculation formula of the read-write ratio factor is:

[0095]

[0096] In the formula, W is the read-write ratio factor, read ps is the read frequency, and write ps is the write frequency.

[0097] The specific calculation formula of the request size factor is:

[0098]

[0099] In the formula, S is the request size factor, min is the minimum value function, AvgReqSize is the average request size, S ref is the reference size threshold, which is set to 256 by default, and the unit is KB.

[0100] Preferably, the size of the management unit is specifically set to 1 MB or 4 MB; when the hotness score is ≥80, the data block is initially placed in the fast layer, when the hotness score is 50-79, the data block is initially placed in the high-speed layer, and when the hotness score is <50, the data block is initially placed in the capacity layer.

[0101] Step S23: file write strategy setting, specifically, the file write in the storage system is set with a hierarchical strategy, including: temporarily storing the newly written data in the write buffer of the fast layer, and subsequently migrating the data to other storage layers according to the access characteristic judgment result in order of priority;

[0102] The sequential read data detected as large volume is preferentially written into the high-speed layer to reduce the occupation of the fast layer and optimize the subsequent access performance.

[0103] Preferably, the large volume is specifically judged according to the relative capacity threshold, and when the size of a single file or data block is ≥5% of the total capacity of the fast layer, it is identified as large volume data.

[0104] Step S24: back interval control, used as a scheduling screening condition for the cache resource scheduling in step S4, specifically, a double-threshold back control mechanism is set, and independent hotness score thresholds are set for the up-regulation and down-regulation between the fast layer, the high-speed layer and the capacity layer, to set the data block regulation mechanism between the layers.

[0105] Preferably, the migration direction and the hotness score threshold of the three-layer cache are set as follows:

[0106] ​​When the data block is promoted from the capacity layer to the high-speed layer, the hotness score thereof is required to be not less than 60 points; and when the data block is demoted from the high-speed layer to the capacity layer, the hotness score thereof only needs to be less than 50 points to trigger the migration.

[0107] When the data block is promoted from the high-speed layer to the fast layer, the hotness score thereof is required to be not less than 80 points; and when the data block is demoted from the fast layer to the high-speed layer, the hotness score thereof only needs to be less than 70 points to trigger the migration.

[0108] When the data block is directly promoted from the capacity layer to the fast layer, the hotness score thereof is required to be not less than 90 points; and when the data block is directly demoted from the fast layer to the capacity layer, the hotness score thereof only needs to be less than 60 points to trigger the migration.

[0109] Step S25: usage condition restriction, specifically, a unified usage condition and restriction is set for different levels, including limiting the capacity usage rate of each layer to be not higher than a preset capacity threshold; when the device temperature reaches a preset temperature threshold, the data migration rate is reduced and the data is preferentially sunk to a low-speed layer; the write budget of the solid-state medium is set according to a period and the write proportion is limited to not exceed the budget; the data whose single write exceeds a preset object capacity threshold is defaulted to be placed in the high-speed layer or the capacity layer; and the minimum access bandwidth is reserved for foreground services in various operations to guarantee service continuity;

[0110] Preferably, the preset capacity threshold is defaulted to be 90% of the capacity.

[0111] Step S26: initial cache grading, specifically, a cache level initialization structure is generated, including the medium type and capacity upper limit of three layers, the migration rate upper limit, the temperature and life threshold, the hysteresis interval, the preheating list and the fixed list, and the initial level mapping table of each data block.

[0112] The cache level initialization structure is used as the input of the subsequent steps, is used for analysis by the three-way load prediction of step S3, and is used for planning and boundary control by the cache resource scheduling of step S4.

[0113] By performing the above operations, in the process of the existing initial cache grading setting, there is a technical problem that the initial grading result is disconnected from the actual business load due to the lack of comprehensive modeling of multiple factors such as access timing, read-write characteristics and request size, and the initial grading setting in the existing initial cache grading setting depends on static medium characteristics or simple access frequency. The present scheme creatively adopts an initial cache grading method based on hotness score calculation, introduces weighted indexes such as access frequency, recency, read-write ratio and request size, and combines hysteresis interval control and usage condition restriction, to realize a more dynamic, fine-grained and controllable three-layer cache initialization structure, thereby effectively reducing the cold and hot data misplacement rate and the initial scheduling cost.

[0114] Embodiment four, refer toFigure 1 、 Figure 2 and Figure 4 , the embodiment is based on the above-mentioned embodiment, in step S3, the three-way load prediction is used to predict the service type, access heat and high delay risk of each type of data, specifically, according to the system running original data and the cache level initialization structure, a time alignment enhanced improved adaptive load three-channel prediction model is used for three-way load prediction, and load comprehensive prediction reference data is obtained, including the following steps:

[0115] Step S31: Time-aware alignment, specifically, the access time sequence is extracted from the system running original data, and the time-aware alignment factor of each data item is calculated to obtain time alignment access feature data, and the calculation formula is:

[0116] ;

[0117] In the formula, F i is the time alignment access feature of the i-th data item in the system running original data, i is the data item index, is the access time index, which is used as the time index of the access time sequence, is the access time weight, count i is the access frequency, j is the neighborhood data item index, which is used as the normalization index, count j is the normalized access frequency;

[0118] Preferably, the access time weight is calculated by introducing a time decay coefficient and a data level change time node, and the calculation formula is:

[0119] ;

[0120] In the formula, t is the current time index, is the time decay factor, the value range is [0.05, 0.2], is the data level migration compensation parameter, the value range is [0.2, 0.6], is a conditional value function, which takes the value 1 when the condition is true, and takes the value 0 when the condition is false, wherein t m is the data level migration time, is the migration process time parameter, the value range is [5, 15], the unit is minute;

[0121] Step S32: Access field enhancement, specifically, the neighborhood enhancement weight between adjacent data items is calculated to construct the access heat feature between adjacent data items, and the neighborhood enhancement access feature data is obtained, and the calculation formula is:

[0122] ;

[0123] In the formula, is the neighborhood enhanced access feature data of the i-th data item, j is the neighborhood data item index, i is the data item index, L ij is the neighborhood enhanced weight, is the heat value calculated by step S22 corresponding to the j-th neighborhood data item, is the heat value decay factor, the value range is [0.5, 1.5], and d(i, j) is the distance between adjacent data items;

[0124] The calculation formula of the neighborhood enhanced weight is:

[0125] ;

[0126] In the formula, L ij is the neighborhood enhanced weight, Coaccess(i, j) is the simultaneous access times, and access is the access times of the data item;

[0127] Step S33: adaptive backbone network construction, specifically constructing a double-channel gated fusion network as an adaptive backbone network for data feature extraction, including a time channel subnet, a resource channel subnet and a feature fusion subblock, and performing feature extraction by using the adaptive backbone network to obtain fusion feature data;

[0128] The time channel subnet specifically performs time sequence convolution processing on the time feature vector and extracts access pattern features; the resource channel subnet specifically constructs a multilayer perceptron and extracts resource state features from the cache resource state parameters; and the feature fusion subblock specifically performs gated fusion on the access pattern features and the resource state features by constructing a gated fusion mechanism to obtain fusion feature data.

[0129] Preferably, the time channel subnet specifically uses a group of 1D convolution layers (convolution kernel size of 3, step size of 1, and channel number of 64) combined with a time attention mechanism to capture short-term fluctuations and long-term trends of the access pattern, and outputs the access pattern features; and the resource channel subnet specifically adopts two layers of multilayer perceptrons (hidden layer dimensions of 128 and 64, and activation function of ReLU) to perform nonlinear mapping on the resource state to obtain a resource state feature vector.

[0130] Step S34: three-way prediction channel construction, specifically constructing an access type prediction channel, an access heat prediction channel and a high delay risk prediction channel based on the adaptive backbone network to obtain three-way prediction channels, and performing data prediction for the access type prediction task, the access heat prediction task and the high delay risk prediction task to obtain task branch prediction data.

[0131] Preferably, the access type prediction channel uses a 256-unit fully connected layer for feature compression and a ReLU activation function, followed by a Softmax output layer with an output dimension equal to the number of business type categories. During training, a prototype constraint loss is introduced, which dynamically maintains the feature center of each class and restricts the Euclidean distance between the current batch features and the center of that class. The formula for calculating the prototype constraint loss is:

[0132] ;

[0133] In the formula, This is the prototype constraint loss, where i is the index of the data item. It is a fusion of feature data, x i It is the original data input corresponding to the i-th data item. It is the yth i The feature centers of each predicted category, y i It is the predicted category corresponding to the i-th data item. This is the feature center stability weight, with a default value of 0.1. c is the class index. It is the feature center of the c-th access category. It is the feature center of the c-th visited class in the previous iteration during model training, and ||·||2 is the L2 norm operator;

[0134] The access popularity prediction channel uses a shared two-layer 128-unit fully connected network to extract time-invariant features. At the output, it splits into three independent linear regression heads, corresponding to prediction scales of 1 minute, 10 minutes, and 1 hour, respectively. Each regression head consists of a 64-unit ReLU fully connected layer and a linear output layer. A cross-temporal consistency loss is introduced, utilizing exponential moving average to constrain the prediction smoothness across different time scales. The formula for calculating the cross-temporal consistency loss is as follows:

[0135] ;

[0136] In the formula, It is a cross-temporal consistency loss. It's a predicted visit popularity on a 10-minute scale. It is the exponential moving average operator. It is the first smoothing coefficient. It is the second smoothing coefficient. It predicts visit popularity on a 1-minute scale. It is a predicted visit volume on an hourly scale;

[0137] The high delay risk prediction channel first passes through a full connection layer of 128 units to extract risk-related features, uses a Sigmoid activation function to limit the value range to (0, 1), then generates a delay prediction value through a linear layer, and introduces a combination term of load rate, concurrency number and session cost parameters when calculating the delay value, and ensures that the delay prediction value does not decrease with the increase of the load rate through the monotonicity constraint loss, and the calculation formula of the monotonicity constraint loss is:

[0138] ;

[0139] In the formula, is the monotonicity constraint loss, is the comprehensive prediction score of the i th data item, used to measure the comprehensive quality of the prediction result, wherein p is a delay percentile index, specifically taking a value range of {95, 99}, used to represent p95 delay and p99 delay, is the control variable proportion of the i th data item, taking a value range of (0, 1);

[0140] The calculation formula of the comprehensive prediction score is:

[0141] ;

[0142] In the formula, is the comprehensive prediction score, is the prediction result of the input data, specifically used to represent the prediction output of the high delay risk prediction channel, is a scaling coefficient, is a control variable proportion, taking a value range of (0, 1), c s is a scene weight coefficient, used to represent the evaluation of service time dispersion, and s is a medium service time reference parameter;

[0143] Step S35: soft budget constraint loss improvement, specifically introducing a soft budget constraint loss, calculating the data occupancy capacity exceeding the threshold under the prediction heat in the next 10 minutes, comparing the data occupancy capacity exceeding the threshold with the available budget capacity, and adding the exceeding part to the optimization target according to the square loss, limiting the prediction result within the range of system capacity and resource budget, and integrating the soft budget constraint loss with the original constraint loss, the cross-time domain consistency loss, the monotonicity constraint loss and the task prediction loss to obtain a joint optimization loss function;

[0144] The calculation formula of the soft budget constraint loss is:

[0145] ;

[0146] In the formula, is the soft budget constraint loss, sigmoid function, is a fast layer hotness threshold, is a smoothing parameter, C L1 is the maximum budget capacity corresponding to the fast layer, reserve is a budget reservation ratio parameter, is a square operation that takes a positive value, used to represent only the part that exceeds the budget is punished;

[0147] The calculation formula of the joint optimization loss function is:

[0148]

[0149] In the formula, is a joint optimization loss function, is a task prediction loss, used to represent the average weighting of the prediction loss of the access type prediction, the access hotness prediction and the high delay risk prediction, is a prototype constraint loss, is a cross-time domain consistency loss, is a monotonicity constraint loss, is a soft budget constraint loss, is a prototype constraint weight, and the default value is 0.5, is a cross-time domain consistency weight, and the default value is 0.3, is a monotonicity constraint loss, and the default value is 0.2, is a soft budget constraint weight, and the default value is 0.4;

[0150] Step S36: three-way load prediction, specifically, through the adaptive backbone network construction and the three-way prediction channel construction, combined with the joint optimization loss function, model training is performed to obtain a three-way load prediction model, and through the use of the three-way load prediction model, load comprehensive prediction reference data is obtained.

[0151] The load comprehensive prediction reference data specifically includes data identification, access prediction type, predicted access hotness, predicted access delay and high delay risk prediction probability.

[0152] ​By performing the above operation, in order to solve the technical problems that in the existing load prediction method, the traditional load prediction method has the problems of coarse prediction granularity in cache management, insufficient capture of burst access mode and neighborhood correlation, and lack of constraint fusion among multi-channel prediction results, which easily leads to imbalance of scheduling strategy, the scheme creatively adopts time sequence alignment to enhance the improved adaptive load three-channel prediction model, performs three-way load prediction, introduces time sequence alignment enhancement and neighborhood enhancement mechanism, combines the double-channel gate fusion backbone network of time channel and resource channel, and embeds prototype constraint, cross-time domain consistency constraint and monotonicity constraint in access type, access heat and high delay risk three-way prediction, which realizes significant improvement of prediction results in time consistency, category stability and resource availability, and provides more robust multi-dimensional prediction basis for cache scheduling.

[0153] In an embodiment five, refer to Figure 1 、 Figure 2 and Figure 5 , the embodiment is based on the above-mentioned embodiment, in step S4, the cache resource scheduling is used to formulate the optimal placement and migration scheme of data among cache layers under the limitation conditions of capacity, bandwidth, temperature and life, specifically, based on the system running original data, cache level initialization structure and load comprehensive prediction reference data, a reinforcement learning method improved based on risk sensitive reward is used for cache resource scheduling to obtain cache resource scheduling decision data, including the following steps:

[0154] Step S41: state constraint construction, according to the system running original data, cache level initialization structure and load comprehensive prediction reference data, a scheduling state vector and a constraint boundary are constructed, the state vector includes the heat, delay risk, current level and recent migration record of each data object, and the constraint boundary includes the layered capacity, available bandwidth, temperature and life threshold and its budget parameter;

[0155] Step S42: risk sensitive reward improvement, used for constructing a reinforcement learning problem with budget and bandwidth constraints, specifically, an reward function is constructed by introducing a performance benefit item, a migration cost penalty and a risk sensitive constraint penalty, a risk sensitive reward improvement reward function is obtained, and the calculation formula is:

[0156] ;

[0157] In the formula, R is the risk sensitive reward improvement reward function, i is the data item index, is the predicted access heat, is the predicted access delay of the i th data item in the current cache level, is the predicted access delay of the i th data item in the target cache level, C migis a migration cost penalty term, specifically calculated by dividing the data volume by the migration limit speed, U is the high latency risk prediction probability, is a capacity constraint violation amount, is a bandwidth constraint violation amount, is a temperature constraint violation amount, is a lifetime constraint violation amount;

[0158] Step S43: reinforcement learning training, specifically reinforcement learning training is performed by defining an action vector through the state constraint construction and the risk-sensitive reward improvement, to obtain a cache resource scheduling model;

[0159] The action vector includes a hierarchical migration action, a migration target hierarchy, and a migration limit speed.

[0160] The hierarchical migration action specifically includes a keep hierarchy, a raise hierarchy, and a sink hierarchy operation.

[0161] Step S44: cache resource scheduling, specifically cache resource scheduling is performed by using the cache resource scheduling model, according to the system running original data, the cache hierarchy initialization structure, and the load comprehensive prediction reference data, to obtain cache resource scheduling decision data.

[0162] The cache resource scheduling decision data specifically includes a cache resource scheduling migration object identifier, a cache resource scheduling operation type, a cache resource source hierarchy, a cache resource target hierarchy, a migration speed limit, and a decision time.

[0163] By performing the above operations, for the technical problem that in the existing cache resource scheduling method, a simple reward function design only considers performance improvement or migration cost, without unified modeling of physical constraints such as capacity, bandwidth, temperature, and lifetime with delay risk, resulting in a scheduling strategy that is easy to violate hard limits or cause lifetime overdraft under extreme load, the present scheme creatively uses a reinforcement learning method based on risk-sensitive reward improvement to perform cache resource scheduling, so that the scheduling decision can dynamically balance between performance and resource safety, significantly reducing the risk of cache system instability or medium premature aging under high load.

[0164] Embodiment six, refer to Figure 1 and Figure 2The embodiment is based on the above-mentioned embodiment, and in step S5, the intelligent hierarchical cache is used to perform migration and adjustment of data between different cache layers. Specifically, according to the cache resource scheduling decision data, the migration and hierarchical adjustment of data between different cache layers are controlled. Under the condition of meeting the operation constraint and the hotness score threshold, the target data is moved up, down or adjusted across layers according to a preset migration strategy. In combination with the real-time business priority and access hotness, the data distribution of each layer is dynamically updated, so as to form a cache scheduling structure meeting the performance and life balance, and obtain an intelligent cache scheduling structure.

[0165] Embodiment seven, refer to Figure 1 and Figure 2 The embodiment is based on the above-mentioned embodiment, and the application provides an intelligent hierarchical cache system based on a hyper-converged architecture, which comprises a data acquisition module, an initial hierarchical module, a load prediction module, a resource scheduling module and an intelligent hierarchical module.

[0166] The data acquisition module is used for hyper-converged architecture acquisition. Through hyper-converged architecture acquisition, system operation original data is obtained, and the system operation original data is sent to the initial hierarchical module, the load prediction module and the resource scheduling module.

[0167] The initial hierarchical module is used for initial cache hierarchical. Through initial cache hierarchical, a cache hierarchical initialization structure is obtained, and the cache hierarchical initialization structure is sent to the load prediction module and the resource scheduling module.

[0168] The load prediction module is used for three-way load prediction. Through three-way load prediction, load comprehensive prediction reference data is obtained, and the load comprehensive prediction reference data is sent to the resource scheduling module.

[0169] The resource scheduling module is used for cache resource scheduling. Through cache resource scheduling, cache resource scheduling decision data is obtained, and the cache resource scheduling decision data is sent to the intelligent hierarchical module.

[0170] The intelligent hierarchical module is used for intelligent hierarchical cache. Through intelligent hierarchical cache, an intelligent cache scheduling structure is obtained.

[0171] It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0172] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary and that changes can be made in the embodiments without departing from the spirit and scope of the application.

[0173] The above description of the application and its embodiments is not restrictive, and the embodiments shown in the drawings are only one of the embodiments of the application, and the actual structure is not limited thereto. In general, if a person skilled in the art is inspired by it, without departing from the purpose of the application, without creative design, similar structure and embodiments of the technical solution should belong to the protection scope of the application.

Claims

1. A smart hierarchical caching method based on a hyperconverged architecture, characterized in that: The method includes the following steps: Step S1: Data collection via converged architecture. Data resources are collected, integrated, and stored by building a hyperconverged architecture to obtain the raw data of system operation. Step S2: Initial cache hierarchy. Based on the original data of the system operation, the initial cache hierarchy is performed by constructing a three-layer cache hierarchy structure to obtain the cache hierarchy initialization structure. Step S3: Three-way load prediction. Based on the original system operation data and the cache level initialization structure, an adaptive three-channel load prediction model with time-series alignment enhancement is used to perform three-way load prediction, obtaining comprehensive load prediction reference data. This includes the following steps: Step S31: Time-aware alignment. Specifically, the access time series is extracted from the original system operation data, and the time-aware alignment factor for each data item is calculated to obtain time-aligned access feature data; Step S32: Access neighborhood enhancement. Specifically, by calculating the neighborhood enhancement weights between adjacent data items, access heat features between adjacent data items are constructed to obtain... Step S33: Construct an adaptive backbone network, specifically by constructing a dual-channel gated fusion network, including a time channel subnet, a resource channel subnet, and a feature fusion sub-block, to obtain fused feature data; Step S34: Construct three prediction channels, specifically by constructing an access type prediction channel, an access popularity prediction channel, and a high latency risk prediction channel based on the adaptive backbone network, to obtain task branch prediction data; Step S35: Improve soft budget constraint loss, specifically by introducing soft budget constraint loss and integrating it to obtain a joint optimized loss function; Step S36: Three-way load prediction; The soft budget constraint loss is calculated by taking into account the data usage capacity exceeding a threshold under the predicted heat in the next 10 minutes, comparing the data usage capacity exceeding the threshold with the available budget capacity, and including the excess portion as a squared loss in the optimization objective, thereby limiting the prediction results to the range of system capacity and resource budget. Step S4: Cache resource scheduling. Based on the system's original operating data, the cache hierarchy initialization structure, and the comprehensive load prediction reference data, a reinforcement learning method based on risk-sensitive reward improvement is used to perform cache resource scheduling decision data. The risk-sensitive reward improvement involves constructing a reward function that incorporates performance gains, migration cost penalties, and risk-sensitive constraint penalties. The calculation formula is as follows: ； In the formula, R is the risk-sensitive reward improvement function, and i is the index of the data item. It predicts visit popularity. It is the predicted access latency of the i-th data item at the current cache level. C is the predicted access latency of the i-th data item at the target cache level. mig This is the migration cost penalty, specifically calculated by dividing the data volume by the migration rate limit. U is the predicted probability of high latency risk. It is a capacity constraint violation. It is a bandwidth constraint violation. It is a violation of temperature constraints. It is a violation of the lifespan constraint; Step S5: Intelligent hierarchical caching, resulting in an intelligent cache scheduling structure.

2. The intelligent hierarchical caching method based on a hyperconverged architecture according to claim 1, characterized in that: In step S1, the converged architecture acquisition is used to collect and organize the operating status and data access status data of computing, storage and network devices. Specifically, it involves collecting, integrating and storing data resources by building a hyperconverged architecture to obtain the raw system operation data. The specific content of the raw system operation data includes device operating status data, storage medium attribute data and data access record data.

3. The intelligent hierarchical caching method based on a hyperconverged architecture according to claim 2, characterized in that: In step S2, the initial cache hierarchy is used to divide the cache space into initial levels according to the speed, capacity, and cost of different storage media, and to set the usage conditions and limitations of each level. Specifically, based on the original system operation data, the initial cache hierarchy is constructed through the initialization of the three-layer cache structure to obtain the cache hierarchy initialization structure, including the following steps: Step S21: Cache layer division, specifically dividing the cache space into a fast layer, a high-speed layer, and a capacity layer. The fast layer uses a medium with the lowest latency and the smallest capacity among the three layers. The high-speed layer uses a medium with intermediate latency and a medium capacity among the three layers. The capacity layer uses a medium with the highest latency and the highest capacity among the three layers. Step S22: Data block management, specifically, using fixed-size data blocks as the management unit, and calculating a popularity score for each data block based on historical access records by comprehensively weighting access frequency factors, access recentness factors, read / write ratio factors, and request size factors. The formula for calculating the heat score is: ； In the formula, H is the popularity score, w1 is the access frequency weight, F is the access frequency factor, w2 is the access recentity weight, R is the access recentity factor, w3 is the read / write ratio weight, W is the read / write ratio factor, w4 is the request size weight, and S is the request size factor. The specific formula for calculating the access frequency factor is as follows: ； In the formula, F is the access frequency factor, ln is the natural logarithm function, count is the total number of accesses to the data block, and C... max This is a configurable limit on the number of accesses; The specific formula for calculating the access proximity factor is as follows: ； In the formula, R is the proximity factor, and exp is the natural base function. It represents the time elapsed since the last access, in hours. It is the time decay factor, in hours; The specific formula for calculating the read / write ratio factor is as follows: ； In the formula, W is the read / write ratio factor, read ps It refers to the number of reads, write. ps It represents the number of writes; The specific formula for calculating the request size factor is as follows: ； In the formula, S is the request size factor, min is the function to find the minimum value, AvgReqSize is the average request size, and S ref This is a reference size threshold, in KB. Step S23: File write strategy setting, specifically, setting a hierarchical strategy for file writing in the storage system, including: temporarily storing newly written data in the write buffer of the fast layer, and subsequently migrating the data to other storage layers according to priority based on the access feature determination results; Data that is detected as large in volume and read sequentially is written to the high-speed layer first, so as to reduce the occupation of the fast layer and optimize the performance of subsequent accesses. Step S24: Hysteresis interval control, used as the scheduling filtering condition for cache resource scheduling in step S4, specifically sets a dual threshold hysteresis control mechanism, sets independent heat score thresholds for the upward and downward adjustments between the fast layer, high-speed layer and capacity layer, and sets the data block adjustment mechanism between layers. Step S25: Usage restrictions, specifically setting uniform usage conditions and restrictions for different levels; Step S26: Initial cache hierarchy, specifically generating the cache hierarchy initialization structure.

4. The intelligent hierarchical caching method based on a hyperconverged architecture according to claim 3, characterized in that: In step S2, the cache hierarchy initialization structure includes three layers of media type and capacity limit, migration rate limit, temperature and lifetime threshold, hysteresis interval, preheating list and fixed list, and an initial hierarchy mapping table for each data block; The cache hierarchy initialization structure serves as input for subsequent steps, providing data for the three-way load prediction in step S3 for analysis, and for the cache resource scheduling in step S4 for planning and boundary control.

5. The intelligent hierarchical caching method based on a hyperconverged architecture according to claim 4, characterized in that: In step S3, the three-way load prediction is used to predict the service type, access popularity, and high latency risk of each type of data. In step S31, the calculation of the time-aware alignment factor for each data item yields the following formula for calculating the time-aligned access feature data: ； In the formula, F i It is the time-aligned access characteristic of the i-th data item in the system's raw data, where i is the data item index. It is an access time index, used as a time index for access time series. It is the access time weight, count i Here, j is the access count, j is the neighborhood data item index, used as a normalized index, and count is the access count. j This is the normalized number of visits; In step S32, the formula for calculating the neighborhood enhancement weight is: ； In the formula, L ij It is the neighborhood enhancement weight, Coaccess(i,j) is the number of times it is accessed at the same time, and access is the number of times the data item is accessed; Step S33: Adaptive backbone network construction, specifically, constructing a dual-channel gated fusion network as an adaptive backbone network for data feature extraction, including a time channel subnet, a resource channel subnet, and a feature fusion sub-block, and performing feature extraction using the adaptive backbone network to obtain fused feature data; The time channel subnet specifically performs temporal convolution processing on the time feature vector and extracts access pattern features; the resource channel subnet specifically constructs a multilayer perceptron and extracts resource status features from the cached resource status parameters; the feature fusion sub-block specifically performs gated fusion of the access pattern features and the resource status features by constructing a gated fusion mechanism to obtain fused feature data. Step S34: Constructing three prediction channels. Specifically, based on the adaptive backbone network, an access type prediction channel, an access popularity prediction channel, and a high latency risk prediction channel are constructed respectively to obtain three prediction channels. Data prediction for the access type prediction task, access popularity prediction task, and high latency risk prediction task are performed respectively to obtain task branch prediction data. Step S35: Improvement of soft budget constraint loss, specifically by introducing soft budget constraint loss and integrating the soft budget constraint loss with prototype constraint loss, cross-temporal consistency loss, monotonicity constraint loss and task prediction loss to obtain a joint optimization loss function; The formula for calculating the prototype constraint loss is as follows: ； In the formula, This is the prototype constraint loss, where i is the index of the data item. It is a fusion of feature data, x i It is the original data input corresponding to the i-th data item. It is the yth i The feature centers of each predicted category, y i It is the predicted category corresponding to the i-th data item. Here, c represents the feature center stability weight, and c is the class index. It is the feature center of the c-th access category. It is the feature center of the c-th visited class in the previous iteration during model training, and ||·||2 is the L2 norm operator; The formula for calculating the cross-temporal consistency loss is as follows: ； In the formula, It is a cross-temporal consistency loss. It's a predicted visit popularity on a 10-minute scale. It is the exponential moving average operator. It is the first smoothing coefficient. It is the second smoothing coefficient. It predicts visit popularity on a 1-minute scale. It is a predicted visit volume on an hourly scale; The formula for calculating the monotonicity constraint loss is as follows: ； In the formula, It is a monotonicity-constrained loss. This is the comprehensive prediction score for the i-th data item, used to measure the overall quality of the prediction results, where p is the lag percentile index, ranging from {95, 99}, used to represent p95 lag and p99 lag. It is the proportion of the control variable for the i-th data item; The formula for calculating the soft budget constraint loss is as follows: ； In the formula, It is a soft budget constraint loss. It is the sigmoid function. It is the rapid layer heat threshold. It is the smoothing parameter, C L1 This is the maximum budget capacity corresponding to the fast layer, and reserve is the budget retention ratio parameter. It is a square operation for positive values, used to indicate that only the portion exceeding the budget is penalized; Step S36: Three-way load prediction, specifically, by constructing the adaptive backbone network and the three prediction channels, and combining the joint optimization loss function, the model is trained to obtain the three-way load prediction model, and by using the three-way load prediction model, the comprehensive load prediction reference data is obtained.

6. The intelligent hierarchical caching method based on a hyper-converged architecture according to claim 5, characterized in that: In step S3, the load comprehensive prediction reference data specifically includes data identifier, access prediction type, predicted access popularity, predicted access latency, and high latency risk prediction probability.

7. The intelligent hierarchical caching method based on a hyperconverged architecture according to claim 6, characterized in that: In step S4, the cache resource scheduling is used to formulate the optimal placement and migration scheme of data among various cache layers under the constraints of capacity, bandwidth, temperature, and lifetime. Specifically, based on the original system operation data, the cache layer initialization structure, and the comprehensive load prediction reference data, a reinforcement learning method based on risk-sensitive reward improvement is used to perform cache resource scheduling to obtain cache resource scheduling decision data, including the following steps: Step S41: State constraint construction. Based on the original system operation data, cache hierarchy initialization structure and load comprehensive prediction reference data, construct the scheduling state vector and constraint boundary. The state vector includes the popularity, latency risk, current level and recent migration record of each data object. The constraint boundary includes the layer capacity, available bandwidth, temperature and lifetime threshold and their budget parameters. Step S42: Risk-sensitive reward improvement, used to construct a reinforcement learning problem with budget and bandwidth constraints. Specifically, it involves constructing a reward function that introduces a performance benefit term, a migration cost penalty, and a risk-sensitive constraint penalty, resulting in the improved risk-sensitive reward function. Step S43: Reinforcement learning training, specifically, through the state constraint construction and risk-sensitive reward improvement, by defining action vectors, reinforcement learning training is performed to obtain the cache resource scheduling model; The action vector includes the hierarchical migration action, the migration target level, and the migration speed limit; The hierarchical migration action specifically includes operations such as maintaining the hierarchy, raising the hierarchy, and lowering the hierarchy. Step S44: Cache resource scheduling, specifically, by using the cache resource scheduling model, based on the original system operation data, the cache hierarchy initialization structure, and the load comprehensive prediction reference data, cache resource scheduling is performed to obtain cache resource scheduling decision data; The cache resource scheduling decision data specifically includes the cache resource scheduling migration object identifier, cache resource scheduling operation type, cache resource source level, cache resource target performance, migration speed limit, and decision time.

8. The intelligent hierarchical caching method based on a hyperconverged architecture according to claim 7, characterized in that: In step S5, the intelligent hierarchical cache is used to perform data migration and adjustment between different cache layers. Specifically, based on the cache resource scheduling decision data, it controls the data migration and hierarchical adjustment between different cache layers. Under the condition of meeting the operating constraints and the popularity score threshold, the target data is moved up, down, or adjusted across layers according to the preset migration strategy. The data distribution of each layer is dynamically updated in combination with real-time business priority and access popularity to form a cache scheduling structure that meets the balance between performance and lifespan, thus obtaining the intelligent cache scheduling structure.

9. A smart hierarchical caching system based on a hyperconverged architecture, used to implement the smart hierarchical caching method based on a hyperconverged architecture as described in any one of claims 1-8, characterized in that: It includes a data acquisition module, an initial classification module, a load prediction module, a resource scheduling module, and an intelligent classification module.

10. The intelligent hierarchical caching system based on a hyper-converged architecture according to claim 9, characterized in that: The data acquisition module is used for data acquisition through the fusion architecture. Through data acquisition through the fusion architecture, the raw data of system operation is obtained and sent to the initial classification module, the load prediction module and the resource scheduling module. The initial grading module is used to initialize cache grading, obtain cache grading initialization structure through initial cache grading, and send the cache grading initialization structure to the load prediction module and the resource scheduling module. The load prediction module is used for three-way load prediction. Through the three-way load prediction, it obtains comprehensive load prediction reference data and sends the comprehensive load prediction reference data to the resource scheduling module. The resource scheduling module is used to cache resource scheduling, obtain cache resource scheduling decision data through cache resource scheduling, and send the cache resource scheduling decision data to the intelligent hierarchical module. The intelligent hierarchical module is used for intelligent hierarchical caching, and through intelligent hierarchical caching, an intelligent cache scheduling structure is obtained.

Images