Computing power resource scheduling method and system for model training

Abstract

The invention relates to the field of big data, and discloses a computing power resource scheduling method and system for model training, and the method comprises the steps: receiving computing power resource request information through a cloud computing power scheduling server, judging the legality of a computing power resource request according to computing power resource authorization information in the computing power resource request information, and if the computing power resource request is a legal request, sending the computing power resource request to the cloud computing power scheduling server; and obtaining model training task information in the computing power resource request information, obtaining training features in the model training task information, and performing feature matching in the cloud computing power scheduling server to complete a model training task. Through dynamic analysis and intelligent scheduling, cloud computing power resources can be utilized more effectively, idle or overload of the resources is avoided, and the resource utilization rate is remarkably increased.

Description

Technical Field

[0001] The present invention relates to the field of big data, and in particular to a computing resource scheduling method and system for model training. Background Art

[0002] With the rapid development of artificial intelligence technology, deep learning models have been widely used in various fields, from image recognition, speech recognition to natural language processing, all of which demonstrate their powerful data processing and pattern recognition capabilities. However, the training process of deep learning models is often accompanied by a huge demand for high-performance computing resources, especially when processing large-scale data sets or complex models. A single computing node is often difficult to meet the computing power required for training, which requires efficient computing resource scheduling methods to optimize resource allocation and improve training efficiency.

[0003] Traditional computing resource scheduling methods are mostly based on static allocation strategies, that is, fixed computing resources are allocated to different training tasks according to preset rules or strategies. This method is not flexible enough when facing changing task requirements and dynamic resource status, and it is easy to cause unbalanced resource utilization, with some resources being overloaded while others are idle, which seriously affects the overall training efficiency and resource utilization. In addition, with the popularization of cloud computing technology, cloud computing resources have become increasingly rich and diverse. How to efficiently and dynamically schedule these distributed computing resources to meet the needs of different training tasks has become an urgent problem to be solved. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a computing resource scheduling method for model training, comprising the following steps:

[0005] Step 1: The cloud computing resource scheduling server receives the computing resource request information, and determines the legitimacy of the computing resource request based on the computing resource authorization information in the computing resource request information. If it is a legitimate request, it proceeds to step 2; otherwise, it rejects the computing resource request and issues a warning message;

[0006] Step 2: Obtain the model training task information in the computing resource request information, obtain the training features in the model training task information, and perform feature matching in the cloud computing scheduling server. If a training container corresponding to the training feature is matched, the model training task is a repeated training task type, and proceed to step 3; otherwise, it is a second training task type, and proceed to step 4;

[0007] Step 3: According to the matched training containers of each corresponding training feature, a training container sequence of the corresponding training feature is obtained, and the training container sequence of the corresponding training feature is matched according to the computing power requirement information in the model training task information. If the training container corresponding to the computing power requirement information is matched, the model training task is sent to the task execution list of the training container corresponding to the matching computing power requirement information. Otherwise, according to the computing power requirement information, the training container with the smallest computing power difference with the computing power requirement is selected in the training container sequence corresponding to the training feature, and the model training task is sent to the merged training container for model training, and then the model training is performed, and step 7 is entered;

[0008] Step 4: The cloud computing power scheduling server establishes a training container corresponding to the second training feature according to the second training feature in the computing power resource request information, and matches the training container corresponding to the second computing power demand information in the cloud computing power scheduling server according to the second computing power demand information in the computing power resource request information. If the training container corresponding to the second computing power demand information is matched, proceed to step 5; otherwise, proceed to step 6.

[0009] Step 5: obtain a training container sequence corresponding to the second computing power requirement information, and respectively obtain the call features of each training container in the training container sequence corresponding to the second computing power requirement information, obtain a scheduling training container according to the call features, connect the scheduling training container to the training container corresponding to the second training feature, perform model training, and proceed to step 7;

[0010] Step 6: The cloud computing power scheduling server generates a computing power container corresponding to the second computing power demand according to the second computing power demand information, and connects it to the training container corresponding to the second training feature to perform model training, and then proceeds to step 7;

[0011] Step 7: Complete the model training task.

[0012] Furthermore, judging the legitimacy of the computing resource request according to the computing resource authorization information in the computing resource request information includes:

[0013] The computing power resource authorization information includes authorization instructions and authorized computing power usage scope; if the authorization instruction authorization verification is passed and the authorized computing power usage scope is within the set single training task computing power usage scope, then the computing power resource request is legal, otherwise, it is illegal.

[0014] Furthermore, the obtaining of the model training task information in the computing resource request information, obtaining the training features in the model training task information, and performing feature matching in the cloud computing resource scheduling server include:

[0015] The training characteristics are one or more of the data set size, training rounds and model scale; in the cloud computing power scheduling server, the training container includes the computing power size and training characteristics, and different training containers include different computing power sizes and training characteristics; the corresponding training container is matched according to the training characteristics in the training task information.

[0016] Further, the training container sequence corresponding to the training feature is obtained according to the matched training containers of each corresponding training feature, and the training container sequence corresponding to the training feature is matched according to the computing power requirement information in the model training task information, including:

[0017] The remaining time for completing the model training task of each training container corresponding to the training feature is obtained respectively, and the training container sequence corresponding to the training feature is obtained from short to long according to the length of the remaining time for completing the model training task, and the computing power difference between the computing power requirement in the model training task information and the computing power of each training container in the training container sequence corresponding to the training feature is obtained respectively, and the training container matching degree is obtained according to the computing power difference and the remaining time for completing the model training task, and the training container corresponding to the minimum training container matching degree is the matching training container, wherein the training container matching degree is:

[0018] Training container matching degree = duration weight × remaining time to complete the model training task + computing power difference × computing power weight.

[0019] Furthermore, the matching of the training container corresponding to the second computing power requirement information in the cloud computing power scheduling server according to the second computing power requirement information in the computing power resource request information includes:

[0020] A training container whose computing power is greater than the second computing power requirement information is obtained from the cloud computing power scheduling server, and a training container matching the second computing power requirement information is obtained.

[0021] Furthermore, the obtaining of the training container sequence corresponding to the second computing power requirement information and respectively obtaining the calling features of each training container in the training container sequence corresponding to the second computing power requirement information include:

[0022] The calling feature is the network transmission delay of the training container, wherein the training container with the smallest network delay is the scheduling training container.

[0023] A computing resource scheduling system for model training, applying the computing resource scheduling method for model training, including a cloud computing resource scheduling server, a communication module, a computing resource request module, a data processing module, an early warning module and a display module;

[0024] The computing power request module, communication module, warning module and display module are respectively connected to the data processing module; the cloud computing power scheduling server is communicatively connected to the communication module.

[0025] The beneficial effects of the present invention are: improving resource utilization: through dynamic analysis and intelligent scheduling, the present invention can more effectively utilize cloud computing resources, avoid resource idleness or overload, and significantly improve resource utilization.

[0026] Shorten model training time: By accurately matching training features and computing power requirements, and optimizing network latency, the present invention can speed up model training and shorten the training cycle.

[0027] Enhanced system flexibility: For special or emerging training tasks, the system can dynamically generate new computing power containers, enhancing the flexibility and scalability of the system.

[0028] Ensure resource security: Through a strict legitimacy verification mechanism, the present invention ensures the safe use of computing resources and prevents illegal or unauthorized access to resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A flowchart of a computing resource scheduling method for model training;

[0030] Figure 2 This is a schematic diagram of the principles of the computing resource scheduling system used for model training. DETAILED DESCRIPTION

[0031] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0032] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0033] like Figure 1 As shown in FIG. 1 , the computing resource scheduling method for model training includes the following steps:

[0034] Step 1: The cloud computing resource scheduling server receives the computing resource request information, and determines the legitimacy of the computing resource request based on the computing resource authorization information in the computing resource request information. If it is a legitimate request, it proceeds to step 2; otherwise, it rejects the computing resource request and issues a warning message;

[0035] Step 2: Obtain the model training task information in the computing resource request information, obtain the training features in the model training task information, and perform feature matching in the cloud computing scheduling server. If a training container corresponding to the training feature is matched, the model training task is a repeated training task type, and proceed to step 3; otherwise, it is a second training task type, and proceed to step 4;

[0036] Step 3: According to the matched training containers of each corresponding training feature, a training container sequence of the corresponding training feature is obtained, and the training container sequence of the corresponding training feature is matched according to the computing power requirement information in the model training task information. If the training container corresponding to the computing power requirement information is matched, the model training task is sent to the task execution list of the training container corresponding to the matching computing power requirement information. Otherwise, according to the computing power requirement information, the training container with the smallest computing power difference with the computing power requirement is selected in the training container sequence corresponding to the training feature, and the model training task is sent to the merged training container for model training, and then the model training is performed, and step 7 is entered;

[0037] Step 4: The cloud computing power scheduling server establishes a training container corresponding to the second training feature according to the second training feature in the computing power resource request information, and matches the training container corresponding to the second computing power demand information in the cloud computing power scheduling server according to the second computing power demand information in the computing power resource request information. If the training container corresponding to the second computing power demand information is matched, proceed to step 5; otherwise, proceed to step 6.

[0038] Step 5: obtain a training container sequence corresponding to the second computing power requirement information, and respectively obtain the call features of each training container in the training container sequence corresponding to the second computing power requirement information, obtain a scheduling training container according to the call features, connect the scheduling training container to the training container corresponding to the second training feature, perform model training, and proceed to step 7;

[0039] Step 6: The cloud computing power scheduling server generates a computing power container corresponding to the second computing power demand according to the second computing power demand information, and connects it to the training container corresponding to the second training feature to perform model training, and then proceeds to step 7;

[0040] Step 7: Complete the model training task.

[0041] The determining the legitimacy of the computing resource request according to the computing resource authorization information in the computing resource request information includes:

[0042] The computing power resource authorization information includes authorization instructions and authorized computing power usage scope; if the authorization instruction authorization verification is passed and the authorized computing power usage scope is within the set single training task computing power usage scope, then the computing power resource request is legal, otherwise, it is illegal.

[0043] The step of obtaining the model training task information in the computing resource request information, obtaining the training features in the model training task information, and performing feature matching in the cloud computing resource scheduling server includes:

[0044] The training characteristics are one or more of the data set size, training rounds and model scale; in the cloud computing power scheduling server, the training container includes the computing power size and training characteristics, and different training containers include different computing power sizes and training characteristics; the corresponding training container is matched according to the training characteristics in the training task information.

[0045] The method of obtaining a training container sequence corresponding to the training feature according to the matched training containers of each corresponding training feature, and matching the training container sequence corresponding to the training feature according to the computing power requirement information in the model training task information, includes:

[0046] The remaining time for completing the model training task of each training container corresponding to the training feature is obtained respectively, and the training container sequence corresponding to the training feature is obtained from short to long according to the length of the remaining time for completing the model training task, and the computing power difference between the computing power requirement in the model training task information and the computing power of each training container in the training container sequence corresponding to the training feature is obtained respectively, and the training container matching degree is obtained according to the computing power difference and the remaining time for completing the model training task, and the training container corresponding to the minimum training container matching degree is the matching training container, wherein the training container matching degree is:

[0047] Training container matching degree = duration weight × remaining time to complete the model training task + computing power difference × computing power weight.

[0048] The method of matching the training container corresponding to the second computing power requirement information in the cloud computing power scheduling server according to the second computing power requirement information in the computing power resource request information includes:

[0049] A training container whose computing power is greater than the second computing power requirement information is obtained from the cloud computing power scheduling server, and a training container matching the second computing power requirement information is obtained.

[0050] The obtaining of the training container sequence corresponding to the second computing power requirement information, and respectively obtaining the calling features of each training container in the training container sequence corresponding to the second computing power requirement information, includes:

[0051] The calling feature is the network transmission delay of the training container, wherein the training container with the smallest network delay is the scheduling training container.

[0052] like Figure 2As shown, a computing resource scheduling system for model training, which applies the computing resource scheduling method for model training, includes a cloud computing resource scheduling server, a communication module, a computing resource request module, a data processing module, an early warning module and a display module;

[0053] The computing power request module, communication module, warning module and display module are respectively connected to the data processing module; the cloud computing power scheduling server is communicatively connected to the communication module.

[0054] Specifically, the present invention provides a computing resource scheduling method for model training, which includes the following key steps:

[0055] Step 1: Legality Verification

[0056] The cloud computing power scheduling server first receives computing power resource request information from the user, which contains computing power resource authorization information to verify the legitimacy of the request. The authorization information specifically includes the authorization instruction and the authorized computing power usage scope. The server determines whether the request is legal by verifying the validity of the authorization instruction and ensuring that the requested computing power usage scope does not exceed the preset upper limit of computing power usage for a single training task. If the request is legal, it proceeds to the next step; otherwise, the request is rejected and a warning message is issued to prevent illegal or unauthorized resource use.

[0057] Step 2: Identify the training task type

[0058] After confirming that the request is legitimate, the server extracts the model training task information, especially the training features, such as data set size, training rounds, model scale, etc., from the computing resource request information. The server maintains a training container library, and each training container predefines the computing power size and training features it supports. Through feature matching, the server can identify whether the current task is a repeated training task (that is, a task with the same or similar features has been processed before). If it is a repeated task, proceed to step three; otherwise, it is regarded as the second training task type and proceed to step four.

[0059] Step 3: Repeat training task scheduling

[0060] For repeated training tasks, the server will search for all training containers that match the current training features, and generate a training container sequence based on the remaining time and computing power of the model training tasks of these containers. Then, according to the computing power requirements of the current task, the most suitable training container is found in the sequence. If a fully matched container is found, the task is directly added to the task execution list of the container. If there is no full match, the container with the smallest computing power difference and the shortest remaining time to complete the task is selected and the task is merged into it to maximize resource utilization.

[0061] Step 4: Scheduling the second training task type

[0062] For the second training task type, the server needs to create a corresponding training container based on its unique training characteristics. Next, the server searches for a training container that can meet the computing power requirements of the task in the existing resources. If found, it will be scheduled according to step five; if not, it will proceed to step six to dynamically generate a new computing power container.

[0063] Step 5: Schedule training container connection

[0064] After finding the training container that matches the second computing power requirement information, the server will further consider the calling characteristics of the container, such as network transmission delay, select the container with the smallest delay as the scheduling training container, and connect it to the newly created training container to ensure efficient data transmission and model training.

[0065] Step 6: Generate a new computing container

[0066] If the existing resources cannot meet the computing power requirements of the second training task, the server will dynamically generate a new computing power container based on the computing power requirement information of the task and connect it to the corresponding training container to meet the special needs of the task.

[0067] Step 7: Complete the model training task

[0068] Whether it is a repeated task or a second training task, the model training will eventually be completed in the selected training container (or container combination), and the server will monitor the training process until the task is completed.

[0069] The present invention ensures the legality and security of resource use through double verification (authorization instructions + computing power usage scope).

[0070] Use training features for precise matching, combine computing power requirements and the remaining time for task completion, and implement intelligent container selection and task scheduling. For computing power requirements that cannot be met, new computing power containers can be dynamically generated, which improves the flexibility and adaptability of the system. When selecting and scheduling training containers, consider network transmission delays to reduce data transmission time and improve training efficiency.

[0071] Example 1: Efficient Scheduling of Repeated Training Tasks

[0072] Assume that there is a deep learning research team that often uses the same model architecture and data set to perform different rounds of model training to optimize model performance. They submit a computing resource request through the computing resource scheduling system of the present invention, hoping to use cloud computing power for model training.

[0073] Legality verification: The research team submitted computing resource request information including authorization instructions and authorized computing power usage range through the computing power request module. After receiving the request, the cloud computing power scheduling server first verifies the validity of the authorization instructions and checks whether the requested computing power usage range is within the preset upper limit of computing power usage for a single training task. After the verification is passed, the request is confirmed to be legal and proceeds to the next step.

[0074] Identification of training task type: The server extracts model training task information from the request information, including training features such as dataset size, training rounds, and model scale. By matching features with containers in the training container library, the server identifies that this is a repeated training task because a task with the same features has been processed before.

[0075] Repeated training task scheduling: The server searches for all training containers that match the current training features, and generates a training container sequence based on the remaining time to complete the model training tasks and the computing power of these containers. Suppose a container A is found with computing power close to the current task requirements and a shorter remaining time to complete the task, but there is no exact matching container. The server selects container A and merges the current task into the task execution list of container A to maximize resource utilization.

[0076] Completing the model training task: The server monitors the model training process in container A until the task is completed. The research team uses the display module to view the training progress and resource usage in real time and is satisfied with the scheduling efficiency.

[0077] Through the computing power resource scheduling method of the present invention, the repetitive training tasks of the research team are efficiently scheduled, which not only shortens the model training time but also improves the utilization rate of cloud computing power resources.

[0078] Example 2: Dynamic resource scheduling for the second training task type

[0079] An AI startup is developing a new deep learning model that has unique training characteristics and high computing power requirements. They decided to use the computing power resource scheduling system of the present invention to apply for cloud computing power resources.

[0080] Legality verification: AI Company submitted computing resource request information including authorization instructions and authorized computing power usage scope through the computing power request module. The cloud computing power scheduling server verifies the legality of the request and proceeds to the next step after confirming that the request is legal.

[0081] Training task type identification: The server extracts the model training task information from the request information, including unique training features such as dataset size, training rounds, and model scale. By matching features with containers in the training container library, the server identifies that this is a second training task type because no tasks with the same features have been processed before.

[0082] Second training task type scheduling: The server creates a corresponding training container B based on the unique training characteristics of the task. Next, the server searches for a training container in the existing resources that can meet the computing power requirements of the task. Since the computing power requirements of the task are high, none of the existing containers can meet them. Therefore, the server enters step 6, dynamically generates a new computing power container C, and connects it to the newly created training container B.

[0083] Scheduling training container connections: After generating a new computing container C, the server considers the calling characteristics of the container, such as network transmission delay, and selects the path with the least delay to connect container C to training container B.

[0084] Completing the model training task: The server monitors the model training process in training container B and computing container C to ensure efficient data transmission and model training. AI companies view the training progress and resource usage in real time through the display module, and highly appreciate the system's ability to dynamically generate new computing containers to meet special needs.

[0085] Through the computing power resource scheduling method of the present invention, the second training task of the AI ​​company was successfully scheduled and executed, which not only met the high computing power requirements, but also demonstrated the flexibility and adaptability of the system when facing new tasks.

Claims

1. A computing resource scheduling method for model training, characterized in that: The steps include: Step 1: The cloud computing resource scheduling server receives the computing resource request information, and determines the legitimacy of the computing resource request based on the computing resource authorization information in the computing resource request information. If it is a legitimate request, it proceeds to step 2; otherwise, it rejects the computing resource request and issues a warning message; Step 2: Obtain the model training task information in the computing resource request information, obtain the training features in the model training task information, and perform feature matching in the cloud computing scheduling server. If a training container corresponding to the training feature is matched, the model training task is a repeated training task type, and proceed to step 3; otherwise, it is a second training task type, and proceed to step 4; Step 3: According to the matched training containers of each corresponding training feature, a training container sequence of the corresponding training feature is obtained, and the training container sequence of the corresponding training feature is matched according to the computing power requirement information in the model training task information. If the training container corresponding to the computing power requirement information is matched, the model training task is sent to the task execution list of the training container corresponding to the matching computing power requirement information. Otherwise, according to the computing power requirement information, the training container with the smallest computing power difference with the computing power requirement is selected in the training container sequence corresponding to the training feature, and the model training task is sent to the merged training container for model training, and then the model training is performed, and step 7 is entered; Step 4: The cloud computing power scheduling server establishes a training container corresponding to the second training feature according to the second training feature in the computing power resource request information, and matches the training container corresponding to the second computing power demand information in the cloud computing power scheduling server according to the second computing power demand information in the computing power resource request information. If the training container corresponding to the second computing power demand information is matched, proceed to step 5; otherwise, proceed to step 6. Step 5: obtain a training container sequence corresponding to the second computing power requirement information, and respectively obtain the call features of each training container in the training container sequence corresponding to the second computing power requirement information, obtain a scheduling training container according to the call features, connect the scheduling training container to the training container corresponding to the second training feature, perform model training, and proceed to step 7; Step 6: The cloud computing power scheduling server generates a computing power container corresponding to the second computing power demand according to the second computing power demand information, and connects it to the training container corresponding to the second training feature to perform model training, and then proceeds to step 7; Step 7: Complete the model training task.

2. The computing resource scheduling method for model training according to claim 1, characterized in that: The determining the legitimacy of the computing resource request according to the computing resource authorization information in the computing resource request information includes: The computing power resource authorization information includes authorization instructions and authorized computing power usage scope; if the authorization instruction authorization verification is passed and the authorized computing power usage scope is within the set single training task computing power usage scope, then the computing power resource request is legal, otherwise, it is illegal.

3. The computing resource scheduling method for model training according to claim 2 is characterized in that: The step of obtaining the model training task information in the computing resource request information, obtaining the training features in the model training task information, and performing feature matching in the cloud computing resource scheduling server includes: The training features are one or more of the data set size, training rounds, and model scale; in the cloud computing power scheduling server, the training container includes the computing power size and the training features, and different training containers include different computing power sizes and training features; the corresponding training container is matched according to the training features in the training task information.

4. The computing resource scheduling method for model training according to claim 3 is characterized in that: The method of obtaining a training container sequence corresponding to the training feature according to the matched training containers of each corresponding training feature, and matching the training container sequence corresponding to the training feature according to the computing power requirement information in the model training task information, includes: The remaining time for completing the model training task of each training container corresponding to the training feature is obtained respectively, and the training container sequence corresponding to the training feature is obtained from short to long according to the length of the remaining time for completing the model training task, and the computing power difference between the computing power requirement in the model training task information and the computing power of each training container in the training container sequence corresponding to the training feature is obtained respectively, and the training container matching degree is obtained according to the computing power difference and the remaining time for completing the model training task, and the training container corresponding to the minimum training container matching degree is the matching training container, wherein the training container matching degree is: Training container matching degree = duration weight × remaining time to complete the model training task + computing power difference × computing power weight.

5. The computing resource scheduling method for model training according to claim 3 is characterized in that: The method of matching the training container corresponding to the second computing power requirement information in the cloud computing power scheduling server according to the second computing power requirement information in the computing power resource request information includes: A training container whose computing power is greater than the second computing power requirement information is obtained from the cloud computing power scheduling server, and a training container matching the second computing power requirement information is obtained.

6. The computing resource scheduling method for model training according to claim 5, characterized in that: The obtaining of the training container sequence corresponding to the second computing power requirement information, and respectively obtaining the calling features of each training container in the training container sequence corresponding to the second computing power requirement information, includes: The calling feature is the network transmission delay of the training container, wherein the training container with the smallest network delay is the scheduling training container.

7. A computing resource scheduling system for model training, characterized in that: A computing resource scheduling method for model training according to any one of claims 1 to 6 is applied, comprising a cloud computing resource scheduling server, a communication module, a computing resource request module, a data processing module, an early warning module and a display module; The computing power request module, communication module, warning module and display module are respectively connected to the data processing module; the cloud computing power scheduling server is communicatively connected to the communication module.

Images