A heterogeneous computing platform and model training method based on hybrid cloud resources
By using a heterogeneous computing platform based on hybrid cloud resources, the problems of inconsistent resource management and insufficient computing power in the hybrid cloud management system are solved, realizing unified management and self-service operation and maintenance of multiple clusters, and improving the efficiency and flexibility of model training.
Patent Information
- Application Number
- CN202110049064.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-01-14
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-01-14
AI Technical Summary
Existing hybrid cloud management systems suffer from lengthy and error-prone manual operations, leading users to frequently switch between different management interfaces. They also fail to achieve unified monitoring and management at the private cloud level and have issues such as limited computing power, single AI chip compatibility, and fixed frameworks.
This paper provides a heterogeneous computing platform based on hybrid cloud resources, including a basic component layer, a computing framework layer, and a resource management layer. It supports deep learning and reinforcement learning frameworks, implements resource scheduling through Kubernetes and Docker modules, integrates heterogeneous computing power resources, network resources, and storage resources, and supports unified management of multiple clusters and large-scale multi-user synchronous use.
It has enabled a transformation from an administrator-centric operation and maintenance approach to a decentralized self-service operation and maintenance approach, improving the efficiency of managing and using heterogeneous resources, solving the problems of inconsistent resource management and insufficient computing power in existing cloud platforms, and making the model training process fast and efficient.
Smart Images

Figure CN112667594B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of cloud technology, and in particular to a heterogeneous computing platform and a model training method based on hybrid cloud resources. Background Art
[0002] The existing computing, storage, and network resources are isolated in different virtualization platforms, making it impossible to achieve unified monitoring and management at the private cloud level. With the development of cloud computing technology, in order to enable management users to frequently switch between different management interfaces and master the different management logics and virtualization models of multiple platforms, enterprises need to hire or train managers who are familiar with specific virtualization platforms to manage them separately.
[0003] A hybrid cloud is a solution that combines a private cloud with one or more public cloud services. It not only provides a private and secure data storage and computing environment, but also provides more flexible and lower-cost computing, storage, and network resources.
[0004] Currently, most hybrid cloud management systems manage multi-cloud systems based on cloud management platforms (CMPs). However, cloud management platforms generally have problems such as lengthy processes and error-prone manual operations, which prevents users from applying for resources and improving self-service capabilities in a unified manner. Summary of the invention
[0005] In order to solve the problems existing in the prior art, the present invention provides the following technical solutions.
[0006] One aspect of the present invention relates to a heterogeneous computing platform based on hybrid cloud resources, comprising:
[0007] The basic component layer is used to provide an interface for user operations, including setting model training tasks;
[0008] The computing framework layer is used to provide a learning framework used by the model training task;
[0009] A resource management layer, used to allocate and schedule the hybrid cloud resources in the infrastructure layer to perform the model training task;
[0010] The infrastructure layer is used to provide hybrid cloud resources, including heterogeneous computing resources, network resources, and storage resources.
[0011] Furthermore, the learning framework includes a deep learning framework and a reinforcement learning framework.
[0012] Furthermore, the resource management layer includes a resource management module, a Kubernetes module and a Docker module. The resource management module implements the scheduling of heterogeneous computing resources, network resources and storage resources in the infrastructure layer through the Kubernetes module and the Docker module.
[0013] Furthermore, the heterogeneous computing resources include distributed CPU, GPU, and ASIC processor resources, the network resources include RDMA networks, and the storage resources include distributed storage systems HDFS, Ceph, and / or ClusterFS.
[0014] Furthermore, the user operation also includes uploading a data set and / or uploading an algorithm.
[0015] Furthermore, the computing framework layer also includes a big data engine for managing the uploaded data sets.
[0016] Another aspect of the present invention relates to a model training method, which is implemented using the above-mentioned heterogeneous computing platform based on hybrid cloud resources, comprising:
[0017] The user sets a model training task through the basic component layer and starts the task, wherein the setting of the model training task includes selecting a model, a data set, a learning framework and / or computing resources;
[0018] The computing framework layer provides the selected learning framework;
[0019] The resource management layer allocates and calls the computing resources, network resources and storage resources of the infrastructure layer for the model training task according to the setting of the model training task to perform model training.
[0020] Preferably, the resource management layer allocates and calls the computing resources, network resources and storage resources of the infrastructure layer for the model training task according to the setting of the model training task, including:
[0021] The resource management layer allocates computing resources, network resources and storage resources to the model training task according to the setting of the model training task, and calls the Kubernetes module and the Docker module to establish a container for the model training task, wherein the container includes the images of the allocated computing resources, network resources and storage resources.
[0022] Furthermore, the resource management layer allocates computing resources to the model training task according to the setting of the model training task, including:
[0023] Get the currently available computing resources;
[0024] If the setting of the model training task includes the selection of computing resources, the corresponding computing resources are allocated based on the selection;
[0025] Otherwise, identify the type of the model training task, and determine the type and size of the required computing resources according to the type;
[0026] Allocate the required computing resources from the currently available computing resources according to their type and size.
[0027] Furthermore, the resource management layer records the resources used by each model training task in real time, and dynamically adjusts the allocated computing resources, network resources and storage resources during the model training process.
[0028] The beneficial effects of the present invention are as follows: the present invention provides a heterogeneous computing platform and model training method based on hybrid cloud resources, which transforms from an administrator-centered operation and maintenance mode to a decentralized self-service operation and maintenance mode, and from a one-way supply operation mode to a transparent and autonomous operation mode, so that the work efficiency of managing and using heterogeneous resources is improved. In addition, it can realize unified management of multiple clusters and large-scale multi-user synchronous use. The heterogeneous computing platform can visualize the entire machine learning modeling process by supporting multiple reinforcement learning architectures and ultra-large-scale distributed training, and at the same time solves the problems of limited computing power, single AI chip adaptation, and fixed framework that are common in existing cloud platforms, making the model training process convenient, fast, and efficient. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 This is a schematic diagram of the structure of a heterogeneous computing platform based on hybrid cloud resources according to the present invention;
[0030] Figure 2 This is a flow chart of the model training method implemented by the heterogeneous computing platform based on hybrid cloud resources described in the present invention. DETAILED DESCRIPTION
[0031] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0032] Embodiment 1
[0033] like Figure 1 As shown, an embodiment of the present invention provides a heterogeneous computing platform based on hybrid cloud resources, including:
[0034] The basic component layer 11 is used to provide an interface for user operations, wherein the user operations include setting model training tasks;
[0035] The computing framework layer 12 is used to provide a learning framework used by the model training task;
[0036] The resource management layer 13 is used to allocate and schedule the hybrid cloud resources in the infrastructure layer to perform the model training task;
[0037] The infrastructure layer 14 is used to provide hybrid cloud resources, including heterogeneous computing resources, network resources, and storage resources.
[0038] Among them, the basic component layer 11 includes a data management module 111, an algorithm development module 112 and a model training module 113. Users upload data sets, delete, modify and export data sets through the data management module 111. Users upload algorithms, modify and delete algorithms through the algorithm development module 112. Users set model training tasks through the model training module 113, including setting the algorithms, data sets and / or learning frameworks used for model training. Optionally, the basic component layer 11 also includes a customized orchestration module 114, which is used to customize the resources used for model training, including processor type, number of processors, etc.
[0039] The learning framework 121 provided by the computing framework layer 12 includes a deep learning framework and a reinforcement learning framework. The deep learning framework includes international mainstream deep learning frameworks such as TensorFlow, mxnet, Caffe, PyTorch, and domestic frameworks such as OneFlow, MegEngine, PaddlePaddle, and MindSpore. The reinforcement learning framework includes the multi-tenant reinforcement learning framework Ray. The above learning frameworks are preset in the platform.
[0040] During use, the user can specify the computing framework used for model training through the model training module 113 of the basic component layer 11. During model training, the platform directly calls the specified computing framework from the computing framework layer 12, which is convenient and fast, greatly simplifies the deployment process and improves operating efficiency.
[0041] The computing framework layer 12 also includes a big data engine 122 for operating the uploaded data set, including storage, computing, mining, management, etc. The big data engine includes multiple data engines, such as SPARK, HADOOP, STORM, HIVE, FLINK, KAFKA, etc., so as to achieve full data interoperability and zero configuration use, and create a unified and rich ecological data collection body.
[0042] The resource management layer 13 includes a resource management module 131, a Kubernetes module 132 and a Docker module 133. The resource management module 131 implements the allocation and scheduling of heterogeneous computing resources, network resources and storage resources in the infrastructure layer through the Kubernetes module 132 and the Docker module 133. Specifically, the resource management module 131 allocates computing resources, network resources and storage resources of the infrastructure layer 14 to the model training task according to the setting of the model training task, and then calls the Kubernetes module 132 to establish a container for the model training task. The container includes the image of the allocated computing resources, network resources and storage resources, and the container is stored in the Docker module 133, so that resource scheduling is performed in units of Docker containers.
[0043] The infrastructure layer 14 includes a private cloud module 141 and a public cloud module 142. The private cloud module 141 is used to provide private cloud resources, and the public cloud module 142 is used to provide public cloud resources 142. The private cloud resources include heterogeneous computing resources, network resources and storage resources. Heterogeneous computing resources include various types of processors, such as distributed CPUs, GPUs, ASICs, and processor series from different manufacturers, such as Cambrian, Huawei Ascend, Baidu Kunlun, etc., so as to meet various computing needs of users; the network resources include RDMA networks, which avoid the copy overhead from user space to system space and improve the CPU utilization efficiency of remote servers; the storage resources include distributed storage systems HDFS, Ceph and / or ClusterFS, so that users can more easily access shared files distributed on the network. The public cloud resources include Huawei Cloud, Alibaba Cloud, Kingsoft Cloud, etc.
[0044] The present invention provides a heterogeneous computing platform based on hybrid cloud resources, which integrates heterogeneous computing resources, multiple computing frameworks and big data engines to achieve unified management of multiple clusters and large-scale multi-user synchronous use, supports multiple reinforcement learning architectures and ultra-large-scale distributed training, and can make the machine learning modeling fully visualized. At the same time, it solves the problems commonly existing in existing cloud platforms, such as limited computing power, single AI chip adaptation, fixed framework, etc., making the model training process convenient, fast and efficient.
[0045] Embodiment 2
[0046] like Figure 2 As shown, this embodiment provides a model training method, which is implemented using the heterogeneous computing platform based on hybrid cloud resources described in Example 1, including:
[0047] S101, the user sets a model training task through the basic component layer and starts the task, wherein setting the model training task includes selecting an algorithm, data set, learning framework and / or computing resources to be used for training;
[0048] S102, the computing framework layer provides the selected learning framework;
[0049] S103, the resource management layer allocates and calls the computing resources, network resources and storage resources of the infrastructure layer for the model training task according to the setting of the model training task to perform model training.
[0050] Specifically, the user sets the model training task through the model training module 113, including setting the algorithm, data set and / or learning framework used for model training. The algorithm used can be uploaded by the user through the algorithm development module 112, and the data set used can be uploaded by the user through the data management module 111. The algorithm and data set can also be uploaded in advance by the administrator or other users. When the administrator or user uploads the algorithm or data set, he can choose whether to make it public. If it is public, all users of the platform can choose to use it. The user customizes the resource usage plan for model training through the customized orchestration module 114, including selecting a public cloud or a private cloud, and selecting computing power resources in the private cloud, such as processor type, number of processors, processor series, etc. As a result, the user can flexibly set training resources based on their own needs. For example, if the user wants to increase the training speed, a larger number of processors can be selected; if the user has requirements for the processor type, a GPU or CPU can be selected; if the user wants to test the processor of a specific manufacturer, the processor series of the manufacturer can be selected, such as Cambrian. As a result, a flexible and unified resource usage method can be provided to users, which can meet the personalized resource needs of users.
[0051] After the settings are completed, the user starts the model training task. The platform calls the selected learning framework from the learning framework 121 of the computing framework layer 12 according to the user's settings, extracts the data set as training data, extracts the algorithm code and executes it;
[0052] At the same time, the resource management layer 13 allocates computing resources, network resources and storage resources according to the customized resource usage plan. For the case where no resources are customized or only some resources are customized, the allocation is based on the model training settings and the current usage of the uncustomized resources. For example, if the customized resource usage plan only limits the type and number of processors, the resource management layer 13 allocates idle network resources and storage resources according to the model training settings (such as the size of the data set used for model training). If the currently available resources are less than the customized resource usage plan, they are allocated according to the currently available resources, and the model training task is recorded. When there are new idle resources, they are allocated to the task first until the customized resource usage plan is reached.
[0053] Then, the platform calls the Kubernetes module to create a Docker container for the model training task, which is stored in the Docker module, and packages the allocated computing resources, network resources, and storage resources into an image and places it in the created container. Therefore, when there are multiple model training tasks executed in parallel, each task has a corresponding Docker container, and the platform can call the Kubernetes model to uniformly manage each Docker container, such as recording the resources used by each model training task in real time, and dynamically adjusting the image of the computing resources, network resources, and storage resources contained in each container.
[0054] The heterogeneous computing platform based on hybrid cloud resources and the model training method provided by the embodiments of the present invention can be well applied to scenarios involving various artificial intelligence fields, such as machine translation, face recognition, AI medical treatment, brain-like computing, intelligent simulation, etc.
[0055] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.
Claims
1. A model training method, characterized in that: Implemented using a heterogeneous computing platform based on hybrid cloud resources; The heterogeneous computing platform based on hybrid cloud resources includes: a basic component layer, which is used to provide an interface for user operations, and the user operations include setting model training tasks; a computing framework layer, which is used to provide a learning framework used by the model training tasks; a resource management layer, which is used to allocate and schedule hybrid cloud resources in the infrastructure layer to execute the model training tasks; an infrastructure layer, which is used to provide hybrid cloud resources, including heterogeneous computing resources, network resources and storage resources; the learning framework includes a deep learning framework and a reinforcement learning framework; the resource management layer includes a resource management module, a Kubernetes module and a Docker module, and the resource management module implements the scheduling of heterogeneous computing resources, network resources and storage resources in the infrastructure layer through the Kubernetes module and the Docker module; the heterogeneous computing resources include distributed CPU, GPU, ASIC processor resources, the network resources include RDMA networks, and the storage resources include distributed storage systems HDFS, Ceph and / or ClusterFS; the user operations also include uploading data sets and / or uploading algorithms; the computing framework layer also includes a big data engine, which is used to manage the uploaded data sets; The model training method comprises: The user sets the model training task through the basic component layer and starts the task. The setting of the model training task includes selecting a model, a data set, a learning framework and / or computing resources; the computing framework layer provides the selected learning framework; the resource management layer allocates and calls the computing resources, network resources and storage resources of the infrastructure layer for the model training task according to the setting of the model training task to perform model training; the resource management layer allocates and calls the computing resources, network resources and storage resources of the infrastructure layer for the model training task according to the setting of the model training task, including: the resource management layer allocates computing resources, network resources and storage resources for the model training task according to the setting of the model training task, calls the Kubernetes module and the Dock The er module establishes a container for the model training task, and the container includes a mirror of the allocated computing power resources, network resources and storage resources; the resource management layer allocates computing power resources to the model training task according to the setting of the model training task, including: obtaining currently available computing power resources; if the setting of the model training task includes the selection of computing power resources, then allocating corresponding computing power resources based on the selection; otherwise, identifying the type of the model training task, and determining the type and size of the required computing power resources according to the type; allocating from the currently available computing power resources according to the type and size of the required computing power resources; the resource management layer records the resource status of each model training task in real time, and dynamically adjusts the allocated computing power resources, network resources and storage resources during the model training process.
