A multi-level heterogeneous video computing resource catalog-based video computing resource management method

Abstract

The application discloses a multi-level heterogeneous video computing power management method based on a video computing power resource directory. The application comprises the following steps: constructing a multi-level directory architecture comprising a top directory node, a middle directory node, a lower directory node and an edge computing node according to the video computing power resource directory; obtaining a video analysis task, and the top directory node queries an optimal node for processing the video analysis task in a global resource view; the top directory node distributes task data of the video analysis task to the optimal node, and the optimal node processes the video analysis task and returns the processed video result to the top directory node. The application realizes effective organization, accurate monitoring, safe access and efficient scheduling of multi-level heterogeneous video computing power resources, significantly improves resource positioning accuracy, cross-domain collaborative scheduling efficiency and data security guarantee capability, and meets the diversified computing power demand of high-concurrency tasks such as video target identification.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of video processing and computing resource management, and particularly relates to a multi-level heterogeneous video computing resource management method based on a video computing resource directory. BACKGROUND

[0002] With the rapid development of video monitoring, intelligent analysis and other technologies, video data is growing explosively, and the computing power demand for video target recognition, tracking, abnormal behavior detection and other tasks is increasing. In practical applications, computing resources are characterized by multi-level heterogeneity, including different types of computing devices (such as CPU, GPU, FPGA, etc.) and resources distributed in different hierarchical management domains (upper, middle, lower, edge).

[0003] Currently, there is a lack of unified management and scheduling mechanism for these multi-level heterogeneous video computing resources, and there are problems such as inaccurate resource positioning, low matching efficiency, and difficult security protection. In terms of resource identification, there is no unified standard, resulting in fuzzy positioning of the management domain to which the resource belongs, which is not convenient for hierarchical management and scheduling; in terms of security, there is a lack of effective real-time monitoring and control of the security state of computing nodes, and there is a risk of data leakage and illegal access; in terms of scheduling, it is difficult to efficiently schedule across domains according to task requirements and resource load conditions, resulting in resource waste or task execution delay.

[0004] Therefore, there is an urgent need for a method that can effectively manage and schedule multi-level heterogeneous video computing resources. SUMMARY

[0005] In order to solve the problems and needs in the background art, the application proposes a multi-level heterogeneous video computing resource management method based on a video computing resource directory, which improves the utilization efficiency of computing resources, ensures data security, meets the needs of various video processing tasks, and reasonably schedules computing resources.

[0006] The technical solution of the application is as follows:

[0007] I. A multi-level heterogeneous video computing resource management method based on a video computing resource directory

[0008] According to the video computing resource directory, a multi-level directory architecture including a top-level directory node, a middle-level directory node, a lower-level directory node and an edge computing node is constructed; the relationship between the nodes, the communication mode and the resource information format comply with a preset standard. A video analysis task is obtained, the top-level directory node queries an optimal node for processing the video analysis task in a global resource view; the top-level directory node distributes task data of the video analysis task to the optimal node, and the optimal node processes the video analysis task and returns the processed video result to the top-level directory node.

[0009] The construction process of the global resource view is as follows:

[0010] The computing power resource identifier corresponding to all nodes is generated and submitted to the immediate superior directory node; each middle directory node in the multi-level directory architecture monitors the real-time load state of all subordinate lower directory nodes and edge computing power nodes, and forms a resource pool of the middle directory node in accordance with the metadata format, and the resource pool of the middle directory node is uploaded to the top directory node, and the metadata format is JSON format; the top directory node generates a global resource view according to the resource pool of each middle directory node, that is, the top directory node aggregates the resource information uploaded by each middle directory node to generate a global resource view; the top directory node broadcasts the generated global resource view to all other nodes.

[0011] The computing power resource identifier of each node includes the management domain code, multi-level application scenario label and device fingerprint of the node, and the format is “[6-bit management domain code]_[4-level application scenario label]_[16-bit device fingerprint]”.

[0012] The device fingerprint is generated by encrypting the hardware features of each node and then attaching an enclave signature.

[0013] When a new node is added at a lower directory node or an edge computing power node, the new node performs bidirectional certificate authentication with the authentication server at the lower directory node or the edge computing power node using the SM2 algorithm. After authentication, the new node submits registration information to the lower directory node or the edge computing power node, and the registration information includes the computing power resource identifier of the new node.

[0014] Each middle directory node uses a resource state monitoring mechanism to monitor the real-time load state of the lower directory nodes and edge computing power nodes, generates a resource change set ΔD (including change type, timestamp, resource ID and attribute set) and uploads it to the top directory node. Optionally, the resource state monitoring mechanism is an anti-entropy algorithm based on incremental update mechanism, and the synchronization period is ≤ 30 seconds, only the newly added, attribute modified, state offline and CPU / memory / bandwidth load change information are synchronized.

[0015] The top directory node aggregates the information in the resource pool of each middle directory node according to the global resource view construction rule. Optionally, the global resource view construction rule is a timestamp arbitration mechanism to solve the cross-domain conflict in the resource pool of each middle directory node.

[0016] Every interval, each middle directory node uses a trusted execution environment technology to perform security verification on its lower directory nodes and edge computing power nodes, and obtains a dynamic trust measurement value corresponding to the nodes.

[0017] The top-level directory node queries an optimal node for processing the video analysis task in a global resource view, including:

[0018] According to the video analysis task, the top-level directory node searches in the global resource view by using a task scheduling strategy to obtain the optimal node. Optionally, the task scheduling strategy is an intelligent scheduling method based on task priority and resource load balancing.

[0019] The top-level directory node distributes task data of the video analysis task to the optimal node, including:

[0020] The top-level directory node distributes the task data of the video analysis task to the optimal node according to a data distribution security strategy. Optionally, the data distribution security strategy includes data desensitization according to 0-3 levels of data desensitization levels and encrypted transmission according to 0-4 levels of task security levels. That is, the top-level directory node desensitizes the task data of the video analysis task according to the data desensitization level, and then encrypts the desensitized task data according to the task security level before distributing the desensitized task data to the optimal node.

[0021] After the optimal node processes the video analysis task, the processed video result is returned to the top-level directory node according to a result return mechanism, and the result return mechanism includes distributed storage of the execution result by using a distributed hash table (DHT) technology and aggregation of the result by using a hash algorithm.

[0022] II. A computer device

[0023] The device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the multi-level heterogeneous video computing resource directory-based multi-level heterogeneous video computing resource management method when executing the computer program.

[0024] The beneficial effects of the present application are:

[0025] The present application realizes effective organization, accurate monitoring, safe access and efficient scheduling of multi-level heterogeneous video computing resources, significantly improves resource positioning accuracy, cross-domain collaborative scheduling efficiency and data security protection capability, and meets the diversified computing power needs of high-concurrency tasks such as video target recognition. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 It is a multi-level directory architecture diagram of the present application.

[0027] Figure 2 It is a method flowchart of the present application.

[0028] Figure 3 It is a construction flowchart of the global resource view. DETAILED DESCRIPTION

[0029] In order to make the objects, technical solutions and advantages of the embodiments of the present disclosure clearer, the following will be combined with the drawings of the embodiments of the present disclosure to describe the technical solutions of the embodiments of the present disclosure clearly and completely. Obviously, the described embodiments are part of the embodiments of the present disclosure, rather than all the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the protection scope of the present disclosure.

[0030] Unless otherwise defined, technical terms or scientific terms used in the present disclosure shall have the ordinary meaning understood by a person of ordinary skill in the art to which the present disclosure pertains. The terms “first”, “second” and similar terms used in the present disclosure do not indicate any order, number or importance, but are only used to distinguish different components. Similarly, the terms “one”, “an” or “the” and similar terms do not indicate a number limitation, but indicate that there is at least one. The terms “include” or “contain” and similar terms mean that the elements or objects before the terms cover the elements or objects listed after the terms and their equivalents, and do not exclude other elements or objects. The terms “connect” or “connected” and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms “up”, “down”, “left”, “right” and the like only represent relative positional relationships, which may change accordingly when the absolute position of the described object changes.

[0031] As shown in Figure 2 The multi-level heterogeneous video computing resource management method based on a video computing resource directory proposed by the present application specifically includes the following steps:

[0032] According to the video computing resource directory, a multi-level directory architecture including a top-level directory node, a middle-level directory node, a lower-level directory node and an edge computing node is constructed, as shown in Figure 1 The relationship between the nodes, the communication mode and the resource information format comply with a preset standard. A video analysis task is obtained, the top-level directory node queries an optimal node for processing the video analysis task in a global resource view, the top-level directory node distributes task data of the video analysis task to the optimal node, the optimal node processes the video analysis task and returns the processed video result to the top-level directory node. After the optimal node completes the video task, the execution result is stored in the middle-level, lower-level and edge nodes by using a distributed hash table (DHT) technology. The top-level directory node quickly locates the result position by using a hash algorithm, efficiently aggregates the execution result and improves the result collection efficiency.

[0033] In one feasible implementation, for a province, its video computing resource catalog includes a top-level catalog node, several mid-level catalog nodes, lower-level catalog nodes, and edge computing nodes. Each node corresponds to a management domain. The mid-level catalog nodes can be first-level mid-level catalog nodes or composed of multiple levels of subordinate mid-level catalog nodes.

[0034] In the flattened model of the video computing resource directory proposed in this invention, middle-level directory nodes establish stable and continuous connections with their directly subordinate lower-level directory nodes, and lower-level directory nodes establish stable and continuous connections with their directly subordinate edge directory nodes through long-lived connection technology (such as the WebSocket protocol). Through this long-lived connection, computing nodes report their own status information, including online status, resource usage, and task execution progress, to their directly subordinate upper-level directory nodes in real time. Once a node's status changes, such as due to excessive resource load or device failure, a registration information update process is immediately triggered. The node encapsulates the changed status information according to the metadata standard format and quickly sends it to its directly subordinate directory nodes, ensuring that the directory nodes can accurately and in real time grasp the latest resource status, providing a reliable basis for resource scheduling and management decisions. Optionally, the metadata format is JSON.

[0035] In one feasible implementation, top-level directory nodes and mid-level directory nodes are added to the access permission whitelist. The access permission whitelist uses a two-factor authentication method of IP address and digital certificate. Only nodes whose IP addresses are on the whitelist and who hold valid digital certificates can access the corresponding resources, effectively preventing unauthorized access and data leakage.

[0036] The access permission whitelist is maintained by the administrator through a dedicated security management console. This console has a user-friendly interface and supports batch import (such as importing a large number of IP addresses and digital certificate information via Excel spreadsheet), export (the current whitelist information can be exported for backup), and dynamic update (adding, deleting, or modifying records in the whitelist in real time), allowing administrators to flexibly manage access permissions.

[0037] In one feasible implementation, the video analysis task can be sent by a node in the management domain of a middle-level directory node or a lower-level directory node, or it can be sent by another top-level directory node to the current top-level directory node.

[0038] like Figure 3 As shown, the process of building the global resource view is as follows:

[0039] The computing power resource identifier corresponding to all nodes (the identifier of each node is unique) is generated and submitted to its immediate superior directory node; each middle-level directory node in the multi-level directory architecture monitors the real-time load state of all subordinate lower-level directory nodes and edge computing power nodes, and uses an efficient merging algorithm to quickly integrate all resource change sets according to the metadata format to form a resource pool and / or a resource change pool of the middle-level directory node, and the resource pool of the middle-level directory node is uploaded to the top-level directory node; the top-level directory node generates a global resource view according to the resource pool of each middle-level directory node, that is, the top-level directory node aggregates the resource information uploaded by each middle-level directory node to generate a global resource view; the top-level directory node broadcasts the generated global resource view to all other nodes (i.e., all nodes in the video computing power resource directory except the top-level directory node) through the UDP broadcast protocol, ensuring that the resource information of each directory node is real-time consistent. The UDP broadcast protocol is a low-latency and high-reliability broadcast method.

[0040] In a feasible implementation, the global resource view VG is stored in a distributed key-value pair storage system such as RedisCluster. The computing power resource identifier is used as the key, and the resource detailed attributes (including all metadata information such as the basic field composed of the management domain code, the application scenario multi-level label, the device fingerprint, the resource capability description field, the security control field (i.e., the data desensitization level), etc.) are used as the value for storage. This storage method supports efficient key-value pair query and update operations, and can quickly respond to resource retrieval and scheduling requests. Among them, the resource capability description is used to clearly define the computing power type, the video stream concurrent number, the hash calculation method, etc. The computing power type includes CPU general computing, GPU accelerated computing and FPGA customized computing. For example, CPU general computing is suitable for general video data processing; GPU accelerated computing is suitable for high-concurrency and high-computing tasks such as video image analysis; FPGA customized computing can meet the efficient hardware acceleration needs of specific identification algorithms. For the video stream concurrent number, according to the hardware performance and network bandwidth, the number of video streams that can be processed by the node resource at the same time is quantified to ensure that the resource capability is clear and searchable. SHA-256 algorithm is used to calculate the hash of the algorithm container image to generate a unique hash value. By comparing the hash values, the integrity and uniqueness of the algorithm container image in the storage and transmission process are guaranteed, and the image is prevented from being tampered with. For example, a node is a GPU accelerated computing node suitable for video image analysis; the video stream concurrent number is limited to 32; the SHA-256 hash calculation is performed on the algorithm container image to generate a unique hash value "a1b2c3d4e5f6g7h8i9j0k1l2m3n4o5p6q7r8s9t0u1v2w3x4y5z6".

[0041] The top-level directory node sends a heartbeat packet to the middle-level and edge directory nodes periodically (e.g., every 1 minute) through a heartbeat mechanism, and the lower-level directory nodes reply to the confirmation information immediately after receiving the heartbeat packet. By monitoring the reply, the top-level directory node can detect the synchronization state of the lower-level directory nodes in the VG in a timely manner. If it is found that a lower-level directory node has not replied for a long time or the synchronization state is abnormal, the top-level directory node will take measures such as re-sending the global resource view VG and re-establishing the connection to ensure the consistency of the global resource view VG of the directory nodes at all levels. Figure 1

[0042] The computing power resource identifier of each node includes the management domain code, multi-level application scenario label and device fingerprint of the node, and its format is "[6-bit management domain code]_[4-level application scenario label]_[16-bit device fingerprint]". Specifically, a 6-bit standard management domain code is used, which is accurate to the management domain of the lower-level directory node, realizes the positioning of the management domain to which the resource belongs, and facilitates the hierarchical management and scheduling of computing power resources. For example, "100000" represents the management domain of a top-level directory node, "100100" represents the management domain of a middle-level directory node under the management domain of the top-level directory node, and "100101" represents the management domain of a lower-level directory node under the management domain of the middle-level directory node. A 4-bit pre-defined application scenario classification label is set, which is subdivided into categories such as target tracking, abnormal behavior detection and traffic statistics from the macro application dimension, and defines the application scenario of resource adaptation to provide support for accurate matching of tasks and computing power resources. For example, a three-level retrieval mode of upper layer-middle layer-specific scenario is set, the upper layer label is from the macro global application strategy perspective, the middle layer label is refined to the application demand of each regional feature, and the specific scenario label is accurately corresponding to the specific identification task. For example, vehicle tracking in traffic monitoring and personnel gathering monitoring in park management. For vehicle tracking in traffic monitoring, the upper layer label is "city traffic global monitoring", the middle layer label is "certain city traffic monitoring", and the specific scenario label is "intersection vehicle tracking". Through this multi-level label system, the retrieval efficiency and matching accuracy of video computing power resources in different levels and different scenarios are greatly improved. For example, the management domain code of a certain node is "100101", the multi-level application scenario label is "upper domain_middle domain_edge domain_CJ004", and the device fingerprint is "F1A2B3C4D5E6G7H8". The unique computing power resource identifier of the node is "100101_upper domain_middle domain_edge domain_CJ004_F1A2B3C4D5E6G7H8".

[0043] The device fingerprint is generated by encrypting the hardware features (such as CPU serial number, motherboard unique identifier, etc.) of each node and then attaching an enclave signature, ensuring the uniqueness and unforgeability of the device fingerprint. For example,

[0044] ​In a feasible implementation, when generating the hash value of the device fingerprint, at least three hardware features, such as the CPU serial number, the motherboard MAC address "BOARD789012", and the hard disk unique identifier, are selected, spliced in a specific order, and then subjected to SHA-256 hash operation to generate a 16-bit hash value. In the generation process, the SGX enclave secure execution environment is used for signing to ensure the uniqueness, integrity and non-tamperability of the device fingerprint. For example, the hardware features include the CPU serial number "CPU123456" and the motherboard unique identifier "BOARD789012", and a 16-bit hash value "F1A2B3C4D5E6G7H8" is generated through an encryption algorithm.

[0045] When a new node is added at a certain lower directory node or edge computing node, in the special network, the new node performs bidirectional certificate authentication with the authentication server at the lower directory node or edge computing node by using the national standard SM2 algorithm (i.e., the elliptic curve public key cryptography algorithm). In the authentication process, the node and the authentication server verify the legality and authenticity of each other's certificate. After the authentication is passed, the new node submits registration information to the lower directory node or edge computing node. The registration information includes the computing resource identifier of the new node, effectively reduces the intermediate level transmission, and significantly improves the registration efficiency and information accuracy.

[0046] Each middle directory node monitors the real-time load state of the nodes of the lower directory nodes and edge computing nodes by using the anti-entropy algorithm based on the incremental update mechanism, generates a resource change set ΔD, and uploads it to the top directory node. The top directory node updates the global resource view according to the resource change set ΔD and broadcasts it to other nodes. Specifically, the middle directory node synchronizes the change information of the lower directory nodes and edge computing nodes, such as addition, attribute modification, state offline, CPU usage, memory occupancy, network bandwidth utilization, every ≤30 seconds, to provide key basis for resource scheduling and load balancing of the upper directory node.

[0047] In a feasible implementation, the resource change set ΔD is stored in the JSON (JavaScript Object Notation) format. This format has the characteristics of simplicity, readability, and ease of parsing and processing. The resource change set ΔD includes the change type (explicitly marked as addition, modification, deletion), the change timestamp (accurately records the time when the change occurs, accurate to the millisecond level, used for timestamp arbitration mechanism and state tracing), the unique identifier of the changed resource (through which the changed resource can be accurately located), and the detailed attribute information after the change (comprehensively and accurately reflects the state after the resource change). Through the standardized JSON format storage, the upper, middle and edge directory nodes can quickly parse the content of the resource change set ΔD and efficiently complete the resource change synchronization operation.

[0048] The top-level directory node adopts a timestamp arbitration mechanism to solve the cross-domain conflict in the resource pool or resource change set ΔD of each middle-level directory node. That is, when different regional directory nodes make conflicting changes to the same resource, the latest change in terms of timestamp is adopted.

[0049] Every interval of a preset time (such as 5 minutes), each middle-level directory node uses a trusted execution environment (TEE) technology to perform security verification on the lower-level directory nodes and edge computing nodes thereunder, and obtains a dynamic trust metric value of the corresponding nodes, that is, the software integrity and running environment security of the nodes are periodically verified to reflect the real-time security state of the nodes. When the dynamic trust metric value is lower than a pre-set trust metric threshold T, the resource identifier of the node is automatically invalidated, and any access operation is prohibited, thereby ensuring the security and trustworthiness of the accessed computing resources. When the dynamic trust metric value of the node recovers to above the pre-set trust metric threshold T, the node is added to the global resource view again. For example, T is 80 (full score is 100). When the dynamic trust metric value of a node is 75, which is lower than T, the resource identifier of the node is automatically invalidated, and other nodes are prohibited from accessing the node.

[0050] The top-level directory node queries the optimal node for processing the video analysis task in the global resource view, including:

[0051] According to the video analysis task, the top-level directory node adopts an intelligent scheduling method based on task priority and resource load balancing to search in the global resource view and obtain the optimal node. Specifically, in terms of task priority, the priority is divided according to the urgency and importance of the video analysis task; in terms of resource load balancing, the real-time load state of each node (such as CPU usage, memory occupancy, etc.) is comprehensively considered to search and schedule the subordinate middle-level, lower-level and edge computing nodes, thereby ensuring efficient execution of the task.

[0052] The top-level directory node distributes the task data of the video analysis task to the optimal node, including:

[0053] The top-level directory node performs data desensitization on the task data of the video analysis task according to the data desensitization level, and distributes the desensitized task data to the optimal node according to the task security level. The dynamic trust metric value of the optimal node is not lower than the pre-set trust metric threshold T. The task data is transmitted through an SSL / TLS encrypted tunnel for encryption transmission, thereby ensuring the security of the data in the transmission process.

[0054] In a feasible implementation, a 0-3 level data desensitization level is provided. The data of the 0th level data desensitization level is data not involving sensitive information, and no desensitization processing is required. The data of the 1st level data desensitization level is non-key sensitive information, and then character replacement, randomization and other algorithms are used to perform fuzzy processing on the non-key sensitive information. For example, “13800138000” is processed into “138****8000”. The data of the 2nd level data desensitization level is the main sensitive information such as faces and license plates in a video stream, and image recognition and shielding algorithms are used to perform local fuzzy shielding on the main sensitive information such as faces and license plates in the video stream. The data of the 3rd level data desensitization level is key sensitive information, and advanced generative adversarial network (GAN) technology is used to perform deep fake replacement on the key sensitive information and generate substitute data similar to the original information but not identifiable, to ensure that the data meets the business requirements under the premise of security and usability.

[0055] In a feasible implementation, the security level is divided into 0-4 levels; from low to high, they are 0, 1, 2, 3, and 4. For each security level, detailed and differentiated protection measures are developed. For example, 0-level resources can be transmitted through a normal network, 1-level resources need to be transmitted on a dedicated network, and 2-level and above resources enable encrypted transmission channels and are accompanied by strict access permission control. 4-level resources can also enable quantum key distribution (QKD) + SM4 dual-mode encryption channels. Combined with the absolute security of quantum encryption and the efficiency of SM4 algorithm, the absolute security of data transmission is ensured.

[0056] In an embodiment, the top-level directory node receives an urgent vehicle tracking task (high priority), queries the global resource view VG, and finds that a GPU computing power node under the management domain of a certain middle-level directory node has a low load (CPU usage rate 30%, memory occupancy rate 25%) and a dynamic trust metric value of 90 is greater than the threshold T (80). The task data is distributed to the node through an SSL / TLS encrypted tunnel. After the node executes the task, the processed video result is returned to the top-level directory node according to the result return mechanism. The result return mechanism includes using distributed hash table (DHT) technology. The execution result is stored in the local and related edge nodes by using the distributed hash table (DHT) technology, and the storage location is located by using a hash algorithm and the execution result is efficiently aggregated, and finally the execution result is returned to the task initiator.

[0057] The application provides a multi-level heterogeneous video computing power management system based on a video computing power resource directory, which comprises:

[0058] Catalog node cluster: composed of upper, middle and lower catalog nodes in the video computing resource catalog and edge computing nodes. The nodes adopt a distributed cluster architecture. The nodes are connected through a high-speed low-latency network, such as using Gigabit Ethernet technology, to ensure efficient data transmission. Distributed transaction processing technology is used, such as based on two-phase commit protocol (2PC) or three-phase commit protocol (3PC), to ensure the atomicity and consistency of operations such as resource registration, catalog synchronization and global resource view generation. That is, these operations are either all successfully executed or all rolled back, preventing data inconsistency. The cluster has load balancing and fault tolerance capabilities. Through load balancing algorithms such as round-robin algorithm, weighted round-robin algorithm, and minimum connection number algorithm, requests are evenly distributed to each node. When a node fails, the system can automatically detect and switch tasks to other normal nodes to ensure continuous and stable operation of the system.

[0059] Computing agent gateway: The computing agent gateway interfaces with GPUs, FPGAs, edge computing devices and other types of heterogeneous devices. Hardware feature extraction chips and specially developed drivers are used to obtain device hardware features in real time, such as reading CPU serial numbers and motherboard hardware information through specific chips. A protocol conversion engine is built in, which can recognize the resource attribute format output by heterogeneous devices and convert it to VCRDP standard format, realizing seamless access of heterogeneous devices to the system. It supports hot plug function. When new heterogeneous devices are added, the system can automatically identify the device type, load the corresponding driver, complete the device initialization and registration process, and realize plug-and-play. It has resource monitoring function. Through sensors, monitoring software and other means, it collects device computing usage (such as CPU usage, GPU memory occupancy), temperature, power consumption and other indicators in real time, and reports these indicators to the upper, middle or edge catalog node cluster, providing data support for resource scheduling and device management.

[0060] In a feasible implementation, the computing agent gateway uses specially designed hardware feature extraction chips in hardware feature extraction, which have high-speed data acquisition and processing capabilities and can quickly and accurately obtain device hardware features. In protocol conversion, multiple protocol conversion rules and algorithms are built in for different types of heterogeneous devices to ensure accurate conversion of various device resource attributes to VCRDP standard format. The catalog node cluster adopts a hierarchical storage strategy. The middle and edge catalog nodes store the complete information of the subordinate computing nodes, including detailed resource attributes, historical usage records, etc. At the same time, a global resource view simplified index based on Bloom filter is constructed. Through the efficient search characteristics of Bloom filter, the query efficiency is greatly improved and the storage overhead is reduced. When processing query requests, first determine whether the resource exists through Bloom filter, and if it exists, further query the detailed information, effectively reducing disk I / O operations and improving system response speed.

[0061] Security authentication module: Based on a trusted network environment, a security authentication system is constructed using national cryptographic algorithms such as SM2, SM3, and SM4. Combined with trusted execution environment (TEE) technology, such as ARM TrustZone or Intel SGX technology, the security of the certificate storage and verification process is ensured. In terms of certificate storage, digital certificates are stored in the secure area provided by TEE to prevent unauthorized reading and tampering. During the verification process, the secure execution environment of TEE verifies the identity information and legality of the certificate of the computing power node, effectively verifying the registration authority of the computing power node. The TEE remote verification agent is integrated to communicate with external TEE verification services, obtain dynamic trust measurement values, and provide them to other system modules for decision-making.

[0062] Synchronization engine: Using multi-threading and asynchronous processing technology, the resource change set ΔD is quickly pushed between the middle layer and the upper layer, and between the edge and the middle layer directory nodes. Through multi-threading technology, multiple resource change set pushing tasks can be processed simultaneously, improving the pushing efficiency. Using asynchronous processing mechanism, other business operations are not blocked during the pushing process, ensuring the overall performance of the system. Using multicast technology, the global resource view is efficiently broadcasted by sending the global resource view to a specific multicast address, reducing network bandwidth consumption and improving broadcast efficiency. The load-sensitive adapter is built-in, which dynamically adjusts the resource change set ΔD pushing frequency according to the CPU / memory multi-level thresholds defined by the multi-element heterogeneous system of China Unicom. When the system load is low, the pushing frequency is appropriately increased to speed up the resource synchronization speed; when the system load is too high, the pushing frequency is reduced to avoid excessive pressure on the system performance.

[0063] Cross-domain collaboration module: Based on the global resource view, intelligent scheduling algorithms such as task priority queue and resource load balancing algorithm are used to achieve precise scheduling of upper, middle, lower, and edge computing power nodes across management domains. Virtual private network (VPN) technology is used to build a secure transmission channel, such as IPsec VPN or SSL VPN, to encrypt the transmission of task data and ensure data transmission security. Distributed message queues (such as Kafka) are used to aggregate task execution results, and the high throughput and low delay characteristics of message queues are used to send task execution results to the message queue, and then relevant modules obtain the results from the queue for processing, improving the efficiency of result processing.

Claims

1. A multi-level heterogeneous video computing resource catalog-based management method, characterized in that, The method comprises the following steps: According to the video computing resource directory, a multi-level directory architecture including a top-level directory node, a middle-level directory node, a lower-level directory node, and an edge computing node is constructed; A video analysis task is obtained, and the top-level directory node queries an optimal node for processing the video analysis task in a global resource view; The top-level directory node distributes task data of the video analysis task to the optimal node, and the optimal node processes the video analysis task and returns the processed video result to the top-level directory node; The construction process of the global resource view is as follows: An algorithm resource identifier corresponding to each node is generated and submitted to the immediate superior directory node of the node; each middle-level directory node in the multi-level directory architecture monitors the real-time load state of all lower-level directory nodes and edge computing nodes under the middle-level directory node and forms a resource pool of the middle-level directory node; The top-level directory node generates a global resource view according to the resource pool of each middle-level directory node; The top-level directory node broadcasts the generated global resource view to all other nodes; Every interval, each middle-level directory node uses a trusted execution environment technology to perform security verification on the lower-level directory nodes and edge computing nodes under the middle-level directory node, and obtains a dynamic trust metric value of the corresponding nodes; When the dynamic trust metric value is lower than a preset trust metric threshold T, the resource identifier of the node is automatically invalidated, and any access operation is prohibited, thereby ensuring the security and trustworthiness of the accessed algorithm resource; when the dynamic trust metric value of the node recovers to above the preset trust metric threshold T, the node is added to the global resource view again; The top-level directory node distributes task data of the video analysis task to the optimal node, comprising: The top-level directory node distributes task data of the video analysis task to the optimal node according to a data distribution security strategy, specifically comprising: The top-level directory node performs data desensitization on the task data of the video analysis task according to a data desensitization level, and then encrypts the desensitized task data according to a task security level before distributing the task data to the optimal node.

2. The multi-level heterogeneous video computing resource catalog-based management method according to claim 1, characterized in that, The algorithm resource identifier of each node includes a management domain code, a multi-level label of an application scenario, and a device fingerprint of the node.

3. The multi-level heterogeneous video computing resource catalog-based management method according to claim 1, characterized in that, When a new node is added at a lower-level directory node or an edge computing node, the new node performs bidirectional certificate authentication with an authentication server at the lower-level directory node or the edge computing node by using a national cryptographic SM2 algorithm; after the authentication is passed, the new node submits registration information to the lower-level directory node or the edge computing node.

4. The multi-level heterogeneous video computing resource catalog-based management method according to claim 1, characterized in that, Each middle-level directory node uses a resource state monitoring mechanism to monitor the real-time load state of the lower-level directory nodes and edge computing nodes under the middle-level directory node, generates a resource change set ΔD, and uploads the resource change set ΔD to the top-level directory node.

5. The multi-level heterogeneous video computing resource catalog-based management method according to claim 1, characterized in that, The top-level directory node aggregates information in the resource pool of each middle-level directory node according to a global resource view construction rule.

6. The multi-level heterogeneous video computing resource catalog-based management method according to claim 1, characterized in that, The top-level directory node queries an optimal node for processing the video analysis task in a global resource view, comprising: According to the video analysis task, the top-level directory node uses a task scheduling strategy to search in the global resource view, and obtains the optimal node. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor implements the steps of the multi-level heterogeneous video computing resource management method based on the video computing resource directory according to any one of claims 1-6 when executing the computer program.

Images