Multi-level resource isolation and concurrency management method for autonomous driving model evaluation

Abstract

The present invention discloses a multi-level resource isolation and concurrent management method for autonomous driving model evaluation, and the present invention relates to the field of autonomous driving technology. The multi-level resource isolation and concurrent management method for autonomous driving model evaluation, by classifying the collected data, constructs an architecture based on the hybrid deployment of virtual machines and containers, and then extracts various evaluation tasks for the autonomous driving model, and determines the parallel processing operations on the evaluation tasks by combining the priority setting of the evaluation tasks and the resource quota algorithm, which can provide exclusive computing resources for tasks requiring high performance and high isolation, ensure the reliability and accuracy of the evaluation results, and significantly improve the concurrent execution efficiency of the evaluation tasks by optimizing resource allocation and task scheduling strategies, and maximize the use of existing computing resources.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and specifically to a multi-level resource isolation and concurrency management method for autonomous driving model evaluation. Background Art

[0002] The rapid development of autonomous driving technology has put forward higher requirements for model evaluation. Traditional evaluation methods often face problems such as resource conflicts and low evaluation efficiency, and cannot meet the evaluation needs of autonomous driving models in complex scenarios.

[0003] The reference patent name is: A method, device and equipment for evaluating an autonomous driving model (patent publication number: CN117235030A, patent publication date: 2023-12-15), including: obtaining a pre-processing file, and processing the pre-processing file based on a preset first autonomous driving model to obtain a first output file; based on the network file system NFS sharing service, sending the pre-processing file to at least one domain controller; based on the NFS sharing service, receiving a second output file sent by at least one domain controller; determining the evaluation results based on the first output file and the second output file sent by at least one domain controller, which can reduce the model evaluation gap between different types of device ends and obtain more accurate evaluation results, so as to improve the efficiency of autonomous driving model evaluation.

[0004] Based on the description in the above-mentioned documents, in the prior art, for tasks that require high-performance resource isolation, the existing containerization technology cannot provide sufficient resource exclusivity, which easily causes resource competition and affects the authenticity and reliability of the evaluation results. In addition, for virtual machine tasks that require a higher degree of isolation, the scheduling support is insufficient, resulting in the inability to efficiently and uniformly manage on the same platform. For this reason, the present invention provides a multi-level resource isolation and concurrency management method for autonomous driving model evaluation. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a multi-level resource isolation and concurrency management method for autonomous driving model evaluation, which solves the problem that for tasks requiring high-performance resource isolation, the existing containerization technology cannot provide sufficient resource exclusivity, which easily causes resource competition and affects the authenticity and reliability of the evaluation results. In addition, for virtual machine tasks that require a higher degree of isolation, there is insufficient scheduling support, resulting in the inability to efficiently and uniformly manage on the same platform.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multi-level resource isolation and concurrency management method for autonomous driving model evaluation, specifically comprising the following steps:

[0007] S1. Collect parameter data involved in the evaluation of the autonomous driving model and implement data storage and transmission operations;

[0008] S2. Classify the collected data, build an architecture based on the hybrid deployment of virtual machines and containers, and then extract various evaluation tasks for the autonomous driving model. Determine the parallel processing operations of the evaluation tasks by combining the priority setting of the evaluation tasks and the resource quota algorithm;

[0009] S3. Display the generated results through a visual display interface.

[0010] Preferably, the operation of classifying the collected data in S2 is:

[0011] A1. Set a classification template, where the category items of the classification template are resource data category and evaluation task data category, and multiple single categories with different category names are set under the resource data category and the evaluation task data category;

[0012] A2. Content matching is performed in the collected data based on the individual categories with different category names as the search content, and the subsequent result data corresponding to the matching content is filled into the classification template until the individual categories with multiple different category names are filled with results;

[0013] A3. After the results are filled, they are classified according to the resource data category and the evaluation task data category to form a resource data set and an evaluation task data set.

[0014] Preferably, the operation of constructing a hybrid deployment architecture based on virtual machines and containers in S2 is:

[0015] B1. Create a virtual machine and set it to achieve isolation and resource management of the complete evaluation task;

[0016] By submitting task information to k8s, k8s detects the virtual machine task and submits the task to the VM control plane. The VM control plane detects the task, assigns the task to the node that meets the requirements, and starts to create a virtual machine. Multiple containers will be created in the virtual machine. The evaluation program is used to pull data for processing and evaluate after obtaining the results.

[0017] B2. Create a container and set it to achieve isolation and resource management of a single evaluation task;

[0018] By submitting the task information to k8s, k8s detects the container task and assigns it to the node that meets the requirements, and starts to create the task pod. The task pod contains multiple containers. The evaluation container is used to pull data for processing and evaluate the results after obtaining them.

[0019] B3. After completing the evaluation task, the container and virtual machine are released and the resources are recycled.

[0020] Preferably, the operation of creating a virtual machine in B1 is:

[0021] b11. After the evaluation task data set is introduced, the resource requirements for the complete evaluation task are determined based on the priority of the evaluation task, and parameter configuration is defined according to the resource requirements;

[0022] b12. Then, the virtualization software provides API or management tools to implement the creation, start, stop, migration and deletion of virtual machines.

[0023] Preferably, the priority setting operation of the evaluation task in S2 is:

[0024] C1. First, divide each evaluation task according to the difference between containers and virtual machines, forming single evaluation tasks and complete evaluation tasks;

[0025] C2. Then, the urgency of time processing of single evaluation tasks and complete evaluation tasks is divided, and sorting operations are first performed according to the time nodes required to complete the processing of the tasks, and then the classification cycle is set to form a multi-level evaluation task summary, and the multi-level categories of single evaluation tasks and complete evaluation tasks should be expressed as P n and Q n , and n∈[1, 2, 3, ...], and the smaller the n number, the higher the priority of the multi-level category;

[0026] C3. Sort the single-level evaluation tasks again, compare the values ​​required for each resource of the evaluation tasks in the single-level, and determine the priority of the evaluation tasks.

[0027] Preferably, the specific operation of determining the priority of the evaluation task in C3 is:

[0028] c31. Select the single evaluation task of level category P1 and the complete evaluation task of level category Q1, and extract the numerical results of the corresponding single categories for comparison;

[0029] c32, set the numerical results corresponding to each category in the single evaluation task of level category P1 to M xy , and x∈[1, 2, 3, ...], y∈[1, 2, 3, ...], and x represents the corresponding single evaluation task, y represents the single category under the corresponding evaluation task, and the numerical results corresponding to each category in the complete evaluation task with the hierarchical category Q1 are N uy , and u∈[1, 2, 3, …], where u represents the corresponding number of complete evaluation tasks, y represents the single category under the corresponding evaluation task, and the total resource value of each single category is extracted and marked as R y ;

[0030] c33. Compare the single category values ​​in single evaluation tasks and complete evaluation tasks with the corresponding total resource values, retain the evaluation tasks that meet the requirements, and then continue to compare the values ​​of other single category values ​​until the priority sorting of the evaluation tasks is completed.

[0031] Preferably, the comparison operation of the single category value and the corresponding total resource value in the single evaluation task and the complete evaluation task in c33 is:

[0032] c331. Select the complete evaluation task N with the hierarchical category Q1 and the first ranking. 1y , and the first single category value N of the complete evaluation task 11 As a first reference value;

[0033] c332, the total resource value R1 of the corresponding single category minus the first reference value N 11 Obtaining a first threshold, and then comparing the first threshold with the value of the single category in the corresponding single evaluation task, and retaining the single evaluation task that meets the requirements;

[0034] c333, and then through the complete evaluation task of the second single category value N 12 As the second reference value, the total resource value R2 of the corresponding single category minus the second reference value N 12 Obtain a second threshold, and then compare the second threshold with the corresponding single category value in the single evaluation task retained in step c332, and retain the single evaluation task that meets the requirements;

[0035] c334. By analogy, realize the other single category values ​​required for the subsequent complete evaluation tasks, and compare the remaining value of the total resource value with the corresponding single category value in the single evaluation task, and then obtain the final retained single evaluation task, and retain the single category value result with the largest result at the end of the retained single evaluation task, and take the final retained single evaluation task and the first complete evaluation task as the first submitted task, and realize the priority sorting of the subsequent submitted tasks through steps c331 to c334.

[0036] Preferably, the comparison method involved in step c331 to step c334 is:

[0037] The calculation formula of the first threshold is: T1=R1-N 11 ;

[0038] T1 and the value M of a single category in a single evaluation task x1 Compare and retain the single evaluation task M whose single category value is less than or equal to T1 x1 ;

[0039] The calculation formula of the second threshold is: T2=R2-N12 ;

[0040] Combine T2 with the single category value M in the retained single evaluation task x1 Compare and retain the single evaluation task M whose single category value is less than or equal to T2 x1 ;

[0041] Repeat the operation until all individual category comparisons are completed.

[0042] Preferably, the operation of the resource quota algorithm in S2 is:

[0043] D1. Based on the determined evaluation task priority, various resources are allocated according to the evaluation task priority, thereby realizing the independent use of resources for containers and virtual machines;

[0044] D2. After the container or virtual machine completes the operation, the evaluation task is progressively implemented. After a certain resource is released, it is combined with the existing resources to achieve secondary allocation, and the resource allocation operation that can be processed in the progressive evaluation task is completed.

[0045] Preferably, the steps of the parallel processing operation in S2 are:

[0046] E1. Submit the sorted single evaluation tasks and complete evaluation tasks, and then implement synchronous parallel processing operations through containers and virtual machines;

[0047] E2. Then, the remaining resources are compared with the remaining evaluation tasks, and the evaluation tasks that meet the remaining resources are introduced into the synchronized evaluation tasks, and the resources between the evaluation tasks are independent.

[0048] The present invention provides a multi-level resource isolation and concurrency management method for autonomous driving model evaluation. Compared with the prior art, it has the following beneficial effects:

[0049] 1. This multi-level resource isolation and concurrency management method for autonomous driving model evaluation builds an architecture based on the hybrid deployment of virtual machines and containers by classifying the collected data, and then extracts various evaluation tasks for the autonomous driving model. By combining the priority setting of the evaluation task and the resource quota algorithm, it determines the parallel processing operations of the evaluation task. It can provide exclusive computing resources for tasks that require high performance and high isolation, ensure the reliability and accuracy of the evaluation results, and significantly improve the concurrent execution efficiency of the evaluation tasks by optimizing resource allocation and task scheduling strategies, thereby maximizing the use of existing computing resources.

[0050] 2. This multi-level resource isolation and concurrency management method for autonomous driving model evaluation submits task information to k8s. After k8s detects the virtual machine task or container task, it submits the task, assigns the task to the node that meets the requirements, starts to create the virtual machine or container, and implements the data pulling and processing operation. It expands the capabilities of the k8s platform, enabling it to support the scheduling of container and virtual machine tasks at the same time, and realizes the unified management of tasks with different isolation levels on the same platform, which greatly improves the flexibility and efficiency of task scheduling.

[0051] 3. This multi-level resource isolation and concurrency management method for autonomous driving model evaluation determines the parallel processing operations of evaluation tasks by combining the priority setting of evaluation tasks and the resource quota algorithm, realizes the processing sorting of submitted tasks and the quota processing of corresponding resource usage, and realizes resource isolation and scheduling optimization of different types of tasks under a unified architecture, and ensures that computing resources between tasks are not shared, thereby avoiding resource contention and performance interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is an operation flow chart of the resource isolation and concurrency management method of the present invention;

[0053] Figure 2 A flowchart of the operation of setting the priority of the evaluation task of the present invention;

[0054] Figure 3 A logical flow chart for evaluating the tasks of the present invention. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] See also Figure 1-Figure 3 , the present invention provides two technical solutions:

[0057] Embodiment 1: A multi-level resource isolation and concurrency management method for autonomous driving model evaluation, specifically comprising the following steps:

[0058] S1. Collect parameter data involved in the evaluation of the autonomous driving model and implement data storage and transmission operations;

[0059] S2. Classify the collected data, build an architecture based on the hybrid deployment of virtual machines and containers, and then extract various evaluation tasks for the autonomous driving model. Determine the parallel processing operations of the evaluation tasks by combining the priority setting of the evaluation tasks and the resource quota algorithm;

[0060] S3. Display the generated results through a visual display interface.

[0061] Among them, by classifying the collected data, building an architecture based on the hybrid deployment of virtual machines and containers, and then extracting various evaluation tasks for the autonomous driving model, and determining the parallel processing operations of the evaluation tasks by combining the priority setting of the evaluation tasks and the resource quota algorithm, it can provide exclusive computing resources for tasks that require high performance and high isolation, ensuring the reliability and accuracy of the evaluation results, and by optimizing resource allocation and task scheduling strategies, significantly improving the concurrent execution efficiency of evaluation tasks, and maximizing the use of existing computing resources.

[0062] In the embodiment of the present invention, the operation of classifying the collected data in S2 is:

[0063] A1. Set a classification template, where the category items of the classification template are resource data category and evaluation task data category, and multiple single categories with different category names are set under the resource data category and the evaluation task data category;

[0064] A2. Content matching is performed in the collected data based on the individual categories with different category names as the search content, and the subsequent result data corresponding to the matching content is filled into the classification template until the individual categories with multiple different category names are filled with results;

[0065] A3. After the results are filled, they are classified according to the resource data category and the evaluation task data category to form a resource data set and an evaluation task data set.

[0066] In the embodiment of the present invention, the operation of constructing a hybrid deployment architecture based on virtual machines and containers in S2 is:

[0067] B1. Create a virtual machine and set it to achieve isolation and resource management of the complete evaluation task;

[0068] By submitting task information to k8s, k8s detects the virtual machine task and submits the task to the VM control plane. The VM control plane detects the task, assigns the task to the node that meets the requirements, and starts to create a virtual machine. Multiple containers will be created in the virtual machine. The evaluation program is used to pull data for processing and evaluate after obtaining the results.

[0069] B2. Create a container and set it to achieve isolation and resource management of a single evaluation task;

[0070] By submitting the task information to k8s, k8s detects the container task and assigns it to the node that meets the requirements, and starts to create the task pod. The task pod contains multiple containers. The evaluation container is used to pull data for processing and evaluate the results after obtaining them.

[0071] B3. After completing the evaluation task, the container and virtual machine are released and the resources are recycled.

[0072] Among them, k8s is kubernetes, and the native components include: kube-apiserver, kubelet, etc.

[0073] vm-related components include:

[0074] Control plane: vm-controller: manages and monitors vm objects and their associated pods, and manages their status; vm-api: a unified vm management interface layer that connects to k8s and is a unified entry point for managing and operating vm. Both components of the control plane run on k8s in the form of pods.

[0075] Data plane: vm-manager: runs in pod mode, one for each virtual machine, used to manage virtual machines, including issuing actions such as start, stop, and release, and monitoring the status of virtual machines and reporting to the control plane. qemu: the main component for managing virtual machines.

[0076] By submitting task information to k8s, k8s submits the task after detecting the virtual machine task or container task, assigns the task to the node that meets the requirements, starts creating the virtual machine or container, and implements the data pulling and processing operation, which expands the capabilities of the k8s platform, enabling it to support the scheduling of container and virtual machine tasks at the same time, and realize the unified management of tasks with different isolation levels on the same platform, greatly improving the flexibility and efficiency of task scheduling.

[0077] In the embodiment of the present invention, the operation of creating a virtual machine in B1 is:

[0078] b11. After the evaluation task data set is introduced, the resource requirements for the complete evaluation task are determined based on the priority of the evaluation task, and parameter configuration is defined according to the resource requirements;

[0079] b12. Then, the virtualization software provides API or management tools to implement the creation, start, stop, migration and deletion of virtual machines.

[0080] Among them, API (Application Programming Interface) refers to a set of methods for interaction between two different software applications and is an existing mature tool.

[0081] In the embodiment of the present invention, the priority setting operation of the evaluation task in S2 is:

[0082] C1. First, divide each evaluation task according to the difference between containers and virtual machines, forming single evaluation tasks and complete evaluation tasks;

[0083] C2. Then, the urgency of time processing of single evaluation tasks and complete evaluation tasks is divided, and sorting operations are first performed according to the time nodes required to complete the processing of the tasks, and then the classification cycle is set to form a multi-level evaluation task summary, and the multi-level categories of single evaluation tasks and complete evaluation tasks should be expressed as P n and Q n , and n∈[1, 2, 3, ...], and the smaller the n number, the higher the priority of the multi-level category;

[0084] C3. Sort the single-level evaluation tasks again, compare the values ​​required for each resource of the evaluation tasks in the single-level, and determine the priority of the evaluation tasks.

[0085] Preferably, the specific operation of determining the priority of the evaluation task in C3 is:

[0086] c31. Select the single evaluation task of level category P1 and the complete evaluation task of level category Q1, and extract the numerical results of the corresponding single categories for comparison;

[0087] c32, set the numerical results corresponding to each category in the single evaluation task of level category P1 to M xy , and x∈[1, 2, 3, ...], y∈[1, 2, 3, ...], and x represents the corresponding single evaluation task, y represents the single category under the corresponding evaluation task, and the numerical results corresponding to each category in the complete evaluation task with the hierarchical category Q1 are N uy , and u∈[1, 2, 3, …], where u represents the corresponding number of complete evaluation tasks, y represents the single category under the corresponding evaluation task, and the total resource value of each single category is extracted and marked as R y ;

[0088] c33. Compare the single category values ​​in single evaluation tasks and complete evaluation tasks with the corresponding total resource values, retain the evaluation tasks that meet the requirements, and then continue to compare the values ​​of other single category values ​​until the priority sorting of the evaluation tasks is completed.

[0089] In the embodiment of the present invention, the comparison operation of the single category value and the corresponding total resource value in the single evaluation task and the complete evaluation task in c33 is:

[0090] c331. Select the complete evaluation task N with the hierarchical category Q1 and the first ranking. 1y , and the first single category value N of the complete evaluation task 11 As a first reference value;

[0091] c332, the total resource value R1 of the corresponding single category minus the first reference value N 11 Obtaining a first threshold, and then comparing the first threshold with the value of the single category in the corresponding single evaluation task, and retaining the single evaluation task that meets the requirements;

[0092] c333, and then through the complete evaluation task of the second single category value N 12 As the second reference value, the total resource value R2 of the corresponding single category minus the second reference value N 12 Obtain a second threshold, and then compare the second threshold with the corresponding single category value in the single evaluation task retained in step c332, and retain the single evaluation task that meets the requirements;

[0093] c334. By analogy, realize the other single category values ​​required for the subsequent complete evaluation tasks, and compare the remaining value of the total resource value with the corresponding single category value in the single evaluation task, and then obtain the final retained single evaluation task, and retain the single category value result with the largest result at the end of the retained single evaluation task, and take the final retained single evaluation task and the first complete evaluation task as the first submitted task, and realize the priority sorting of the subsequent submitted tasks through steps c331 to c334.

[0094] In the embodiment of the present invention, the comparison method involved in step c331 to step c334 is:

[0095] The calculation formula of the first threshold is: T1=R1-N 11 ;

[0096] T1 and the value M of a single category in a single evaluation task x1 Compare and retain the single evaluation task M whose single category value is less than or equal to T1 x1 ;

[0097] The calculation formula of the second threshold is: T2=R2-N 12 ;

[0098] Combine T2 with the single category value M in the retained single evaluation task x1 Compare and retain the single evaluation task M whose single category value is less than or equal to T2x1 ;

[0099] Repeat the operation until all individual category comparisons are completed.

[0100] In the embodiment of the present invention, the operation of the resource quota algorithm in S2 is:

[0101] D1. Based on the determined evaluation task priority, various resources are allocated according to the evaluation task priority, thereby realizing the independent use of resources for containers and virtual machines;

[0102] D2. After the container or virtual machine completes the operation, the evaluation task is progressively implemented. After a certain resource is released, it is combined with the existing resources to achieve secondary allocation, and the resource allocation operation that can be processed in the progressive evaluation task is completed.

[0103] By combining the priority setting of the evaluation task and the resource quota algorithm to determine the parallel processing operations on the evaluation task, the processing sorting of the submitted tasks and the quota processing of the corresponding resource usage are realized. Under a unified architecture, resource isolation and scheduling optimization of different types of tasks are achieved, and it is ensured that the computing resources between tasks are not shared, thereby avoiding resource contention and performance interference.

[0104] In the embodiment of the present invention, the steps of the parallel processing operation in S2 are:

[0105] E1. Submit the sorted single evaluation tasks and complete evaluation tasks, and then implement synchronous parallel processing operations through containers and virtual machines;

[0106] E2. Then, the remaining resources are compared with the remaining evaluation tasks, and the evaluation tasks that meet the remaining resources are introduced into the synchronized evaluation tasks, and the resources between the evaluation tasks are independent.

[0107] The difference between the second embodiment and the first embodiment is that after the task is submitted, the specific operations of the container and the virtual machine are as follows:

[0108] Submit the visual lane line task:

[0109] 1. If the task isolation level is detected as container, the task information (including CPU, memory, GPU, etc.) is directly submitted to k8s;

[0110] 2. K8s detects the container task, assigns it to the node that meets the requirements, and starts creating the task pod;

[0111] 3. Execute the task. The task pod contains multiple containers, which are divided into visual lane line model container and evaluation container. The visual lane line model container is mainly used to run the artificial intelligence model that can identify lane lines; the evaluation container is used to pull data, broadcast packets, and evaluate after obtaining the results;

[0112] 4. After the evaluation is completed, the container is released;

[0113] Submit an end-to-end evaluation task:

[0114] 1. Detect that the task isolation level is a virtual machine, and submit the task information (including CPU, memory, GPU, disk IO, etc.) to k8s;

[0115] 2. K8s detects the virtual machine task and submits the task to the VM control plane;

[0116] 3. The VM control plane detects the task, assigns the task to the node that meets the requirements, and starts creating the VM;

[0117] 4. Execute tasks. Multiple containers will be created in the virtual machine to run models and evaluation programs such as vision, perception, fusion, prediction, and regulation. Start each module, simulate the running scenario of the vehicle, and the evaluation program will broadcast the package. After obtaining the results, the evaluation will be carried out;

[0118] 5. After the evaluation is completed, the virtual machine is released.

[0119] Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0120] It should be noted that, in this article, relational terms such as first and second, etc. 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 these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0121] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-level resource isolation and concurrency management method for autonomous driving model evaluation, characterized by: The specific steps include: S1. Collect parameter data involved in the evaluation of the autonomous driving model and implement data storage and transmission operations; S2. Classify the collected data, build an architecture based on the hybrid deployment of virtual machines and containers, and then extract various evaluation tasks for the autonomous driving model. Determine the parallel processing operations of the evaluation tasks by combining the priority setting of the evaluation tasks and the resource quota algorithm; S3, displaying the generated results through a visual display interface; The priority setting operation of the evaluation task in S2 is: C1. First, divide each evaluation task according to the difference between containers and virtual machines, forming single evaluation tasks and complete evaluation tasks; C2. Then, the urgency of time processing of single evaluation tasks and complete evaluation tasks is divided, and sorting operations are first performed according to the time nodes required to complete the processing of the tasks, and then the classification cycle is set to form a multi-level evaluation task summary, and the multi-level categories of single evaluation tasks and complete evaluation tasks should be expressed as P n and Q n , and n∈[1, 2, 3, ...], and the smaller the n number, the higher the priority of the multi-level category; C3. Sort the single-level evaluation tasks, compare the values ​​required for each resource of the evaluation tasks in the single-level, and determine the priority of the evaluation tasks; The operation of the resource quota algorithm in S2 is: D1. Based on the determined evaluation task priority, various resources are allocated according to the evaluation task priority, thereby realizing the independent use of resources for containers and virtual machines; D2. After the container or virtual machine completes the operation, the evaluation task is progressively implemented. After a certain resource is released, it is combined with the existing resources to achieve secondary allocation, and the resource allocation operation that can be processed in the progressive evaluation task is completed.

2. The multi-level resource isolation and concurrency management method for autonomous driving model evaluation according to claim 1, characterized in that: The operation of classifying the collected data in S2 is: A1. Set a classification template, where the category items of the classification template are resource data category and evaluation task data category, and multiple single categories with different category names are set under the resource data category and the evaluation task data category; A2. Content matching is performed in the collected data based on the individual categories with different category names as the search content, and the subsequent result data corresponding to the matching content is filled into the classification template until the individual categories with multiple different category names are filled with results; A3. After the results are filled, they are classified according to the resource data category and the evaluation task data category to form a resource data set and an evaluation task data set.

3. The multi-level resource isolation and concurrency management method for autonomous driving model evaluation according to claim 1, characterized in that: The operations of constructing a hybrid deployment architecture based on virtual machines and containers in S2 are: B1. Create a virtual machine and set it to achieve isolation and resource management of the complete evaluation task; By submitting task information to k8s, K8s detects the virtual machine task and submits the task to the vm control plane. The vm control plane detects the task, assigns the task to the node that meets the requirements, and starts to create a virtual machine. Multiple containers will be created in the virtual machine. The evaluation program is used to pull data for processing and evaluate after obtaining the results. B2. Create a container and set it to achieve isolation and resource management of a single evaluation task; By submitting the task information to k8s, k8s detects the container task and assigns it to the node that meets the requirements, and starts to create the task pod. The task pod contains multiple containers. The evaluation container is used to pull data for processing and evaluate the results after obtaining them. B3. After completing the evaluation task, the container and virtual machine are released and the resources are recycled.

4. The multi-level resource isolation and concurrent management method for autonomous driving model evaluation according to claim 3, characterized in that: The operation of creating a virtual machine in B1 is: b11. After the evaluation task data set is introduced, the resource requirements for the complete evaluation task are determined based on the priority of the evaluation task, and parameter configuration is defined according to the resource requirements; b12. Then, the virtualization software provides API or management tools to implement the creation, start, stop, migration and deletion of virtual machines.

5. The multi-level resource isolation and concurrency management method for autonomous driving model evaluation according to claim 1, characterized in that: The specific operation of determining the priority of the evaluation task in C3 is: c31. Select the single evaluation task of level category P1 and the complete evaluation task of level category Q1, and extract the numerical results of the corresponding single categories for comparison; c32, set the numerical results corresponding to each category in the single evaluation task of level category P1 to M xy , and x∈[1, 2, 3, ...], y∈[1, 2, 3, ...], and x represents the corresponding single evaluation task, y represents the single category under the corresponding evaluation task, and the numerical results corresponding to each category in the complete evaluation task with the hierarchical category Q1 are N uy , and u∈[1, 2, 3, …], where u represents the corresponding number of complete evaluation tasks, y represents the single category under the corresponding evaluation task, and the total resource value of each single category is extracted and marked as R y ; c33. Compare the single category values ​​in single evaluation tasks and complete evaluation tasks with the corresponding total resource values, retain the evaluation tasks that meet the requirements, and then continue to compare the values ​​of other single category values ​​until the priority sorting of the evaluation tasks is completed.

6. The multi-level resource isolation and concurrent management method for autonomous driving model evaluation according to claim 5, characterized in that: The comparison operation between the single category value and the corresponding total resource value in the single evaluation task and the complete evaluation task in c33 is: c331. Select the complete evaluation task N with the hierarchical category Q1 and the first ranking. 1y , and the first single category value N of the complete evaluation task 11 As a first reference value; c332, the total resource value R1 of the corresponding single category minus the first reference value N 11 Obtaining a first threshold, and then comparing the first threshold with the value of the single category in the corresponding single evaluation task, and retaining the single evaluation task that meets the requirements; c333, and then through the complete evaluation task of the second single category value N 12 As the second reference value, the total resource value R2 of the corresponding single category minus the second reference value N 12 Obtain a second threshold, and then compare the second threshold with the corresponding single category value in the single evaluation task retained in step c332, and retain the single evaluation task that meets the requirements; c334. By analogy, realize the other single category values ​​required for the subsequent complete evaluation tasks, and compare the remaining value of the total resource value with the corresponding single category value in the single evaluation task, and then obtain the final retained single evaluation task, and retain the single category value result with the largest result at the end of the retained single evaluation task, and take the final retained single evaluation task and the first complete evaluation task as the first submitted task, and realize the priority sorting of the subsequent submitted tasks through steps c331 to c334.

7. The multi-level resource isolation and concurrency management method for autonomous driving model evaluation according to claim 6, characterized in that: The comparison method involved in step c331 to step c334 is: The calculation formula of the first threshold is: T1=R1-N 11 ; T1 and the value M of a single category in a single evaluation task x1 Compare and retain the single evaluation task M whose single category value is less than or equal to T1 x1 ; The calculation formula of the second threshold is: T2=R2-N 12 ; Combine T2 with the single category value M in the retained single evaluation task x1 Compare and retain the single evaluation task M whose single category value is less than or equal to T2 x1 ; Repeat the operation until all individual category comparisons are completed.

8. The multi-level resource isolation and concurrency management method for autonomous driving model evaluation according to claim 1, characterized in that: The steps of the parallel processing operation in S2 are: E1. Submit the sorted single evaluation tasks and complete evaluation tasks, and then implement synchronous parallel processing operations through containers and virtual machines; E2. Then, the remaining resources are compared with the remaining evaluation tasks, and the evaluation tasks that meet the remaining resources are introduced into the synchronized evaluation tasks, and the resources between the evaluation tasks are independent.

Images