Network scheduling method and device, computer equipment, storage medium and program product

Abstract

The embodiment of the invention provides a network scheduling method and device, computer equipment, a storage medium and a program product, relates to the technical field of network scheduling, and can realize topology awareness of distributed training, so that a traffic transmission path can be adjusted in time, the communication bandwidth utilization rate in network communication can be improved, and the user experience is improved. The network communication quality is improved, and the efficiency of distributed training can be further improved. The method comprises the following steps: acquiring network demand information of a target tenant server and topological information of a network architecture to which the target tenant server belongs; determining a target communication path of the target tenant server according to the network demand information and the topological information; according to the target communication path, video card configuration information corresponding to a target cloud host server is determined; and accessing the target tenant server to the target cloud host server according to the graphics card configuration information.

Description

Technical Field

[0001] The present application relates to the technical field of network scheduling, and in particular to a network scheduling method, apparatus, computer equipment, storage medium and program product. Background Art

[0002] In an era when large model applications are prevalent, as the complexity of deep learning models increases, the number of model parameters is also growing rapidly. In order to provide more accurate predictions, these models usually need to be trained on a large amount of data, which requires powerful computing power to support the training of these large model applications. However, due to the limitations of manufacturing processes, the development speed of artificial intelligence chips lags far behind the computing power required by the models. In order to solve the computing power challenges faced by large models, people use distributed training, that is, collaborative training between multiple machines and calculations, thereby improving computing power and training speed.

[0003] However, in distributed training tasks, each graphics card calculation needs to frequently synchronize the parameters and gradients of model training to ensure the ultimate consistency and accuracy of the model. Therefore, as the scale of the distributed training system increases, the communication overhead accounts for an increasingly higher proportion of the entire training overhead, thus becoming a performance bottleneck for large-scale distributed training.

[0004] However, in traditional network communications, the underlying collective communication library is usually unable to directly perceive the actual physical networking topology in the network architecture to which it belongs, making it difficult to adjust the traffic transmission path in a timely manner, which can easily lead to a reduction in communication bandwidth utilization, thereby affecting the efficiency of distributed training. Summary of the invention

[0005] Based on this, it is necessary to provide a network scheduling method, device, computer equipment, storage medium and program product to address the above-mentioned technical problems, which can realize topology awareness of distributed training, so that the traffic transmission path can be adjusted in time, thereby improving the communication bandwidth utilization in network communication, improving the quality of network communication, and further improving the efficiency of distributed training.

[0006] According to a first aspect of an embodiment of the present application, a network scheduling method is provided, including:

[0007] Obtaining network requirement information of a target tenant server and topology information of a network architecture to which the target tenant server belongs;

[0008] Determining a target communication path of the target tenant server according to the network demand information and the topology information;

[0009] Determine the graphics card configuration information corresponding to the target cloud host server according to the target communication path;

[0010] According to the graphics card configuration information, the target tenant server is connected to the target cloud host server.

[0011] In some implementations, the network requirement information includes at least one of the number of graphics cards required by the target tenant server and a target communication bandwidth required by the target tenant server.

[0012] In some implementations, determining the graphics card configuration information corresponding to the target cloud host server according to the target communication path includes:

[0013] Determining the number of the target cloud host servers according to the target communication path;

[0014] Determine the graphics card configuration information corresponding to the target cloud host server according to the number of the target cloud host servers;

[0015] Among them, the graphics card configuration information includes single-machine graphics card configuration information and multi-machine graphics card configuration information. The single-machine graphics card configuration information includes a target configuration graphics card in the target cloud host server that is communicatively connected to the target tenant server, and the multi-machine graphics card configuration information includes at least two target configuration graphics cards in the target cloud host server that are communicatively connected to the target tenant server, a target connection graphics card used for communication connection between different target cloud host servers, and a connection relationship between each of the target connection graphics cards.

[0016] In some implementations, connecting the target tenant server to the target cloud host server according to the graphics card configuration information includes:

[0017] In the case where the target cloud host server includes one, connecting the target tenant server to the target cloud host server through the target configured graphics card according to the single-machine graphics card configuration information;

[0018] In the case where the target cloud host servers include at least two, according to the multi-machine graphics card configuration information, at least two of the target cloud host servers are communicatively connected through the target connection graphics card according to the connection relationship, and the target tenant server is connected to different target cloud host servers through the target configuration graphics card.

[0019] In some implementations, determining a target communication path of the target tenant server according to the network demand information and the topology information includes:

[0020] Determining at least one initial communication path according to the network demand information and the topology information;

[0021] Obtaining the switch hop count and network communication bandwidth of the initial communication path;

[0022] The target communication path is determined according to at least one of the switch hop count and the network communication bandwidth.

[0023] In some implementations, before the step of obtaining the network requirement information of the target tenant server and the topology information of the network architecture to which the target tenant server belongs, the method further includes:

[0024] Obtaining original topology information of the network architecture of the target tenant server;

[0025] Monitoring the operating status of the network architecture;

[0026] According to the running state, the original topology information is modified to obtain the topology information of the network architecture to which the target tenant server belongs;

[0027] The network scheduling method further includes:

[0028] When there is an abnormality in the operation status of the network architecture, an alarm message is generated according to the location of the network link where the abnormality occurs and the abnormal status.

[0029] According to a second aspect of an embodiment of the present application, a network scheduling device is provided, including:

[0030] An information acquisition module, used to acquire network demand information of a target tenant server and topology information of a network architecture to which the target tenant server belongs;

[0031] A first determining module, configured to determine a target communication path of the target tenant server according to the network demand information and the topology information;

[0032] A second determination module is used to determine the graphics card configuration information corresponding to the target cloud host server according to the target communication path;

[0033] The access control module is used to connect the target tenant server to the target cloud host server according to the graphics card configuration information.

[0034] According to a third aspect of an embodiment of the present application, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0035] Obtaining network requirement information of a target tenant server and topology information of a network architecture to which the target tenant server belongs;

[0036] Determining a target communication path of the target tenant server according to the network demand information and the topology information;

[0037] Determine the graphics card configuration information corresponding to the target cloud host server according to the target communication path;

[0038] According to the graphics card configuration information, the target tenant server is connected to the target cloud host server.

[0039] According to a fourth aspect of the embodiments of the present application, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0040] Obtaining network requirement information of a target tenant server and topology information of a network architecture to which the target tenant server belongs;

[0041] Determining a target communication path of the target tenant server according to the network demand information and the topology information;

[0042] Determine the graphics card configuration information corresponding to the target cloud host server according to the target communication path;

[0043] According to the graphics card configuration information, the target tenant server is connected to the target cloud host server.

[0044] According to a fifth aspect of the embodiments of the present application, a computer program product is provided, including a computer program, which, when executed by a processor, implements the following steps:

[0045] Obtaining network requirement information of a target tenant server and topology information of a network architecture to which the target tenant server belongs;

[0046] Determining a target communication path of the target tenant server according to the network demand information and the topology information;

[0047] Determine the graphics card configuration information corresponding to the target cloud host server according to the target communication path;

[0048] According to the graphics card configuration information, the target tenant server is connected to the target cloud host server.

[0049] The network scheduling method, device, computer equipment, storage medium and program product provided in the embodiments of the present application can obtain the network demand information of the target tenant server and the topology information of the network architecture to which the target tenant server belongs, and can further determine the target communication path for the target tenant server to access the communication network according to the network demand information and the topology information of the network architecture, and can first determine the target communication path of the target tenant server, determine the target cloud host server that is connected to the target tenant server in the target communication path, and further determine the graphics card configuration information in each target cloud host server to achieve the decomposition of the target communication path, so that the graphics card configuration information can be sent to the corresponding target cloud host server, so that the target cloud host server can directly optimize the access quality of the target tenant server according to the graphics card configuration information, and further achieve the topology perception of the network architecture by the target host server. Therefore, the network scheduling method provided in the embodiments of the present application can achieve the topology perception of distributed training, so that the traffic transmission path can be adjusted in time, and then the communication bandwidth utilization rate in network communication can be improved, the network communication quality can be improved, and the efficiency of distributed training can be further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] FIG1 is a schematic flow chart of a network scheduling method provided in an embodiment of the present application;

[0052] FIG2 is a schematic structural diagram of an original communication path of a network scheduling method provided in an embodiment of the present application;

[0053] FIG3 is a schematic structural diagram of a target communication path of a network scheduling method provided in an embodiment of the present application;

[0054] FIG4 is a schematic structural diagram of a target configuration graphics card of a network scheduling method provided in an embodiment of the present application;

[0055] FIG5 is a schematic diagram of target cloud host server connection of a network scheduling method provided in an embodiment of the present application;

[0056] FIG6 is a schematic diagram of connecting a target cloud host server to another network scheduling method provided by an embodiment of the present application;

[0057] FIG7 is a schematic diagram of connecting a target cloud host server of another network scheduling method provided in an embodiment of the present application;

[0058] FIG8 is a schematic diagram of connecting a target cloud host server of another network scheduling method provided in an embodiment of the present application;

[0059] FIG9 is a schematic structural diagram of a network scheduling device provided in an embodiment of the present application;

[0060] FIG10 is a schematic structural diagram of a computer device provided in an embodiment of the present application;

[0061] FIG11 is a schematic structural diagram of a computer-readable storage medium provided in an embodiment of the present application;

[0062] FIG12 is a schematic structural diagram of a computer program product provided in an embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0064] In an era where large model applications are prevalent, as the complexity of deep learning models increases, the number of model parameters is also growing rapidly. In addition, in order to provide more accurate predictions, these models need to be trained on a large amount of data, which requires powerful computing power to support the training of these large model applications. Taking OpenAI's (Open AI Research Center) GPT (Generative Pre-trained Transformer) as an example, the GPT-3 model has 175 billion parameters, which is 116 times the number of parameters of GPT-2. GPT-3 training takes several weeks using a thousand NVIDIA A100 graphics cards. The iteration of GPT-4 has increased the parameters by another 10 times.

[0065] However, due to the limitation of manufacturing process, the development speed of AI chips lags far behind the computing power required by the model. In order to solve the computing power challenge faced by large models, people use distributed training, that is, using multiple machines and GPGPU (General-Purpose on Graphics Computing Unit) for collaborative training, so as to improve computing power and training speed. Among them, the commonly used distributed training frameworks are PyTorch and TensorFlow. The underlying layer calls the collective communication library, such as NVIDIA's NCCL, to realize the communication operation of multiple computing GPGPUs between single or multiple nodes. In the distributed training task, each graphics card computing GPGPU needs to frequently synchronize the parameters and gradients of the model training to ensure the final consistency and accuracy of the model. Therefore, as the scale of the distributed training system increases, the communication overhead accounts for an increasing proportion of the entire training overhead, thus becoming a performance bottleneck for large-scale distributed training. In the intelligent computing scenario, cloud computing is the provider of computing power, and excessive communication consumption will also cause computing power blocking and waste of computing resources. The underlying collective communication library cannot perceive the actual physical networking topology. The networking topology with too many hops will greatly reduce the network bandwidth and training efficiency. The distributed training method using the NCCL collective communication library as an example has the following shortcomings in the intelligent computing scenario: First, NCCL itself does not perceive the physical networking topology, that is, in the intelligent computing network, cross-core switch communication scenarios may occur, affecting the utilization of network bandwidth. Second, NCCL uses the NBT (Double Binary Tree) structure in multi-machine scenarios, but only maintains the local topology information of the graphics card in the node (such as the parent node and child node of the GPGPU in the binary tree), and only has node-level topology information between nodes, lacks global topology information at the GPGPU granularity, and cannot perceive the actual networking information and links.

[0066] In view of this, proposing a topology-aware network scheduling method that can realize distributed training of the network is a technical problem that needs to be solved urgently.

[0067] As shown in FIG. 1 , a first aspect of an embodiment of the present application provides a network scheduling method, including:

[0068] Step S110: Acquire network requirement information of a target tenant server and topology information of a network architecture to which the target tenant server belongs.

[0069] Exemplarily, the topology information of the network architecture to which the target tenant server belongs can be determined based on the initial deployment of the network architecture. The topology information can be the physical networking information of the network architecture, which can specifically include the location and connection relationship of the access switch and the core switch in the network architecture, the physical location of each cloud host server, and the connection relationship between the cloud host server and the core switch.

[0070] Step S120: Determine a target communication path of a target tenant server according to network demand information and topology information.

[0071] Exemplarily, based on network demand information and topology information, the combination of all cloud host servers that meet the requirements of the target tenant server can be determined, and further based on the combination of cloud host servers, the access method of the switch between the target tenant server and the cloud host server can be determined, so that the target communication path can be determined.

[0072] Step S130: Determine the graphics card configuration information corresponding to the target cloud host server according to the target communication path.

[0073] As shown in FIG. 2 and FIG. 3 , it can be seen that the complexity and path length of the target communication path provided by FIG. 3 are significantly smaller than those of the initial communication path provided by FIG. 2 .

[0074] For example, the connection status of each graphics card in the target cloud host server can be determined according to the target communication path. After the target communication path is determined, the corresponding NodeRank sequence number can be assigned to the target cloud host server in the target communication path to facilitate identification and scheduling in the subsequent distributed training process.

[0075] Step S140: Connect the target tenant server to the target cloud host server according to the graphics card configuration information.

[0076] The network scheduling method, device, computer equipment, storage medium and program product provided in the embodiments of the present application can obtain the network demand information of the target tenant server and the topology information of the network architecture to which the target tenant server belongs, and can further determine the target communication path for the target tenant server to access the communication network according to the network demand information and the topology information of the network architecture, and can first determine the target communication path of the target tenant server, determine the target cloud host server that is connected to the target tenant server in the target communication path, and further determine the graphics card configuration information in each target cloud host server to achieve the decomposition of the target communication path, so that the graphics card configuration information can be sent to the corresponding target cloud host server, so that the target cloud host server can directly optimize the access quality of the target tenant server according to the graphics card configuration information, and further achieve the topology perception of the network architecture by the target host server. Therefore, the network scheduling method provided in the embodiments of the present application can achieve the topology perception of distributed training, so that the traffic transmission path can be adjusted in time, and then the communication bandwidth utilization rate in network communication can be improved, the network communication quality can be improved, and the efficiency of distributed training can be further improved.

[0077] In some feasible implementations, the network requirement information includes at least one of the number of graphics cards required by the target tenant server and the target communication bandwidth required by the target tenant server.

[0078] Exemplarily, at least one cloud host server having a number of graphics cards that is greater than or equal to the number of graphics cards required by the target tenant server can be determined based on the number of graphics cards required by the target tenant server.

[0079] Exemplarily, at least one cloud host server having a graphics card bandwidth greater than or equal to the target communication bandwidth required by the target tenant server may be determined according to the target communication bandwidth required by the target tenant server.

[0080] The network scheduling method provided in the embodiment of the present application determines the target communication path by the number of graphics cards required by the target tenant server and the required target communication bandwidth, which can further improve the network access quality of the target tenant server, improve the network communication quality, ensure the efficiency and stability of network communication, and further improve the efficiency of distributed training.

[0081] In some feasible implementations, determining graphics card configuration information corresponding to a target cloud host server according to a target communication path includes: determining the number of target cloud host servers according to the target communication path; determining graphics card configuration information corresponding to the target cloud host servers according to the number of target cloud host servers; wherein the graphics card configuration information includes single-machine graphics card configuration information and multi-machine graphics card configuration information, the single-machine graphics card configuration information includes a target configuration graphics card in a target cloud host server that is communicatively connected to a target tenant server, the multi-machine graphics card configuration information includes a target configuration graphics card in at least two target cloud host servers that is communicatively connected to a target tenant server, a target connection graphics card used for communication connection between different target cloud host servers, and a connection relationship between each target connection graphics card.

[0082] Exemplarily, the target transmission path can be transmitted in the form of parameters or configuration files, and can be transmitted to the MPI framework and PyTorch framework, etc. Among them, for the parallel MPI framework, the serial number of the target cloud host server can be bound to the corresponding node IP based on the rankfile rule as a configuration file, and the distributed command mpirun can be run on the Master node to control the order of the host nodes. For the training framework TensorFlow, the number of nodes nnodes of the entire distributed training task is defined, and the information of the Master node is filled in the master_addr and master_port parameters. Based on the node order planned in the early stage, it is distributed to each node to generate the corresponding node_rank, and the training script of the node is executed. Further, the corresponding Local Rank and Rank serial numbers can be assigned to each graphics card according to the number of graphics cards in each computing node and the node order, which respectively represent the serial number of the graphics card in this computing node and the global serial number in the entire distributed training task.

[0083] The network scheduling method provided in the embodiment of the present application can clarify the connection method of the graphics cards in each target cloud host server through the single-machine graphics card configuration information and the multi-machine graphics card configuration information. At the same time, when the target cloud host server includes multiple ones, the connection method between different target cloud host servers can be determined, thereby realizing topological awareness of distributed training, so as to facilitate timely adjustment of the traffic transmission path, further improving the quality of network communication and improving the efficiency of distributed training.

[0084] In some feasible implementations, the target tenant server is connected to the target cloud host server according to the graphics card configuration information, including: when there is one target cloud host server, the target tenant server is connected to the target cloud host server through the target configuration graphics card according to the single-machine graphics card configuration information; when there are at least two target cloud host servers, according to the multi-machine graphics card configuration information, at least two target cloud host servers are communicatively connected through the target connection graphics card according to the connection relationship, and the target tenant server is connected to different target cloud host servers through the target configuration graphics card.

[0085] As shown in FIG4, when the target cloud host servers include four, the target configuration graphics cards in each target cloud host server are connected inside the target cloud host server. As shown in FIG5, when the target cloud host servers include four, the target cloud host servers are connected through the first connection method. As shown in FIG6, in the first connection method shown in FIG5, all graphics cards in each target cloud host server establish a communication connection. As shown in FIG7, when the target cloud host servers include four, the target cloud host servers are connected through the second connection method. As shown in FIG8, in the second connection method shown in FIG7, all graphics cards in each target cloud host server establish a communication connection.

[0086] For example, the collective communication library can establish the topology of the PCIe system in each target cloud host server based on the target cloud host server provided by the upper layer and the number of target configured graphics cards corresponding to it, bind the rank relationship to each graphics card, search for the optimal path and bandwidth between GPGPU nodes through the topology, and maintain the graphics card topology within a single node. In a multi-machine scenario, the collective communication library establishes a dual binary tree structure based on the number of host nodes, and further connects the graphics cards in each node responsible for inter-host communication, thus forming a multi-machine multi-card topology.

[0087] The network scheduling method provided in the embodiment of the present application can clarify the connection method of the graphics cards in each target cloud host server through the single-machine graphics card configuration information and the multi-machine graphics card configuration information. At the same time, when there are multiple target cloud host servers, the connection method between different target cloud host servers can be determined, thereby improving the convenience of network access, realizing the topology awareness of the underlying communication library, and then realizing the topology awareness of distributed training, so as to facilitate timely adjustment of the traffic transmission path, further improving the quality of network communication, and improving the efficiency of distributed training.

[0088] In some feasible implementations, a target communication path of a target tenant server is determined based on network demand information and topology information, including: determining at least one initial communication path based on the network demand information and topology information; obtaining the number of switch hops and the network communication bandwidth of the initial communication path; and determining the target communication path based on at least one of the number of switch hops and the network communication bandwidth.

[0089] Exemplarily, in the process of determining the target communication path, the initial communication path that does not involve the core switch can be determined as the target communication path. It should be noted that in the spine-leaf network architecture, if the network communication can be completed between the access layer switches, the number of switch hops under the communication path will be minimized, thereby improving the communication speed and further improving the efficiency of network communication. However, in the case where network communication must be established through a core switch, for example, when the number of target cloud host servers determined by the number of graphics cards required by the target tenant server exceeds the number of cloud host servers that can be connected to an access switch, or when the distance between different graphics cards connected to the target tenant server is far and a large amount of cross-node communication is required, it usually results in the target communication path having to pass through the core switch.

[0090] For example, when determining the target communication path according to the switch hop count, the initial communication path with the smallest switch hop count is usually selected as the target communication path. When determining the target communication path according to the network communication bandwidth, the initial communication path with the largest network communication bandwidth is usually selected as the target communication path.

[0091] It should be noted that when the target communication path is determined by the topology structure, there may be a possibility that the target cloud host server fails. A failed cloud host server may cause the target tenant server to access a poor network environment, or even cause the target tenant server to be unable to access the network environment. Therefore, nodes that may fail can be removed from the topology information.

[0092] The network scheduling method provided in the embodiment of the present application determines the target communication path through the number of switch hops and the network communication bandwidth, which can further shorten the network transmission time for the target tenant server to access, improve the network communication efficiency and the stability of network communication, and can also increase the network bandwidth for the target tenant server to access, thereby ensuring the access quality of the target tenant server.

[0093] In some feasible implementations, before the step of obtaining the network demand information of the target tenant server and the topology information of the network architecture to which the target tenant server belongs, it also includes: obtaining the original topology information of the network architecture of the target tenant server; monitoring the operating status of the network architecture; and correcting the original topology information according to the operating status to obtain the topology information of the network architecture to which the target tenant server belongs.

[0094] For example, taking NCCL as an example, the graphics card tree structure within a single node (presented in a linked list structure) and the node tree structure information between multiple nodes (presented in a double binary tree structure) are integrated to build a graphics card tree structure between nodes. Each tree node represents the graphics card node that communicates with other host nodes and contains all the graphics card topology information within the node. The actual networking information is further persisted in YAML format for feedback to further verify the feasibility of the optimized networking.

[0095] The network scheduling method provided in the embodiment of the present application can facilitate the correction of topology information by reconstructing a global topology network at the granularity of a graphics card, thereby further improving the accuracy of network scheduling and further improving the network access quality of the target tenant server.

[0096] In some feasible implementations, the network scheduling method further includes: when there is an abnormality in the operating state of the network architecture, generating alarm information according to the location and abnormal state of the network link where the abnormality occurs.

[0097] Exemplarily, the corrected actual networking information can be pushed to the Pushgateway component, and the data can be collected and stored through the Prometheus ecosystem to achieve graphics card granularity monitoring on the entire communication link, providing data support for fault diagnosis and rapid recovery.

[0098] The network scheduling method provided in the embodiment of the present application can facilitate the maintenance of network access, improve the stability and robustness of distributed training tasks, and further improve the security of network access by monitoring the operating status of the network architecture and generating alarm information in a timely manner based on network anomalies.

[0099] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0100] Based on the same inventive concept, the embodiment of the present application also provides a memory management device for implementing the memory management method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more memory management device embodiments provided below can refer to the limitations on the memory management method above, and will not be repeated here.

[0101] As shown in FIG. 9 , a second aspect of an embodiment of the present application provides a network scheduling device, including:

[0102] The information acquisition module 100 is used to acquire the network demand information of the target tenant server and the topology information of the network architecture to which the target tenant server belongs;

[0103] A first determination module 200 is used to determine a target communication path of a target tenant server according to network demand information and topology information;

[0104] The second determination module 300 is used to determine the graphics card configuration information corresponding to the target cloud host server according to the target communication path;

[0105] The access control module 400 is used to connect the target tenant server to the target cloud host server according to the graphics card configuration information.

[0106] Each module in the above memory management device can be implemented in whole or in part by software, hardware or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0107] As shown in FIG10, the third aspect of the embodiment of the present application provides a computer device, which may be a device integrating multiple target applications, and the specific implementation process of the memory monitoring module in the target application may be performed by the computer device; wherein, the computer device may be a server, and its internal structure diagram may be shown in FIG10. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Wherein, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Wherein, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store memory data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a memory management method is implemented.

[0108] Those skilled in the art will appreciate that the structure shown in FIG. 1 is merely a block diagram of a portion of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0109] As shown in FIG. 11 , in a fourth aspect of an embodiment of the present application, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0110] Obtaining network requirement information of the target tenant server and topology information of the network architecture to which the target tenant server belongs;

[0111] Determine the target communication path of the target tenant server based on network demand information and topology information;

[0112] According to the target communication path, determine the graphics card configuration information corresponding to the target cloud host server;

[0113] Connect the target tenant server to the target cloud host server based on the graphics card configuration information.

[0114] As shown in FIG. 12 , in a fifth aspect of an embodiment of the present application, a computer program product is provided, including a computer program, which implements the following steps when executed by a processor:

[0115] Obtaining network requirement information of the target tenant server and topology information of the network architecture to which the target tenant server belongs;

[0116] Determine the target communication path of the target tenant server based on network demand information and topology information;

[0117] According to the target communication path, determine the graphics card configuration information corresponding to the target cloud host server;

[0118] Connect the target tenant server to the target cloud host server based on the graphics card configuration information.

[0119] It should be noted that the information (including but not limited to application information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0120] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0121] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0122] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A network scheduling method, characterized in that: include: Obtaining network requirement information of a target tenant server and topology information of a network architecture to which the target tenant server belongs; Determining a target communication path of the target tenant server according to the network demand information and the topology information; Determine the graphics card configuration information corresponding to the target cloud host server according to the target communication path; According to the graphics card configuration information, the target tenant server is connected to the target cloud host server.

2. The network scheduling method according to claim 1, characterized in that: The network requirement information includes at least one of the number of graphics cards required by the target tenant server and the target communication bandwidth required by the target tenant server.

3. The network scheduling method according to claim 1, characterized in that: The step of determining the graphics card configuration information corresponding to the target cloud host server according to the target communication path includes: Determining the number of the target cloud host servers according to the target communication path; Determine the graphics card configuration information corresponding to the target cloud host server according to the number of the target cloud host servers; Among them, the graphics card configuration information includes single-machine graphics card configuration information and multi-machine graphics card configuration information. The single-machine graphics card configuration information includes a target configuration graphics card in the target cloud host server that is communicatively connected to the target tenant server, and the multi-machine graphics card configuration information includes at least two target configuration graphics cards in the target cloud host server that are communicatively connected to the target tenant server, a target connection graphics card used for communication connection between different target cloud host servers, and a connection relationship between each of the target connection graphics cards.

4. The network scheduling method according to claim 3, characterized in that: The step of connecting the target tenant server to the target cloud host server according to the graphics card configuration information includes: In the case where the target cloud host server includes one, connecting the target tenant server to the target cloud host server through the target configured graphics card according to the single-machine graphics card configuration information; In the case where the target cloud host servers include at least two, according to the multi-machine graphics card configuration information, at least two of the target cloud host servers are communicatively connected through the target connection graphics card according to the connection relationship, and the target tenant server is connected to different target cloud host servers through the target configuration graphics card.

5. The network scheduling method according to claim 1, characterized in that: The determining, according to the network requirement information and the topology information, a target communication path of the target tenant server includes: Determining at least one initial communication path according to the network demand information and the topology information; Obtaining the switch hop count and network communication bandwidth of the initial communication path; The target communication path is determined according to at least one of the switch hop count and the network communication bandwidth.

6. The network scheduling method according to claim 1, characterized in that: Before the step of obtaining the network requirement information of the target tenant server and the topology information of the network architecture to which the target tenant server belongs, the step further includes: Obtaining original topology information of the network architecture of the target tenant server; Monitoring the operating status of the network architecture; According to the running state, the original topology information is modified to obtain the topology information of the network architecture to which the target tenant server belongs; The network scheduling method further includes: When there is an abnormality in the operation status of the network architecture, an alarm message is generated according to the location of the network link where the abnormality occurs and the abnormal status.

7. A network scheduling device, characterized in that: include: An information acquisition module, used to acquire network demand information of a target tenant server and topology information of a network architecture to which the target tenant server belongs; A first determining module, configured to determine a target communication path of the target tenant server according to the network demand information and the topology information; A second determination module is used to determine the graphics card configuration information corresponding to the target cloud host server according to the target communication path; The access control module is used to connect the target tenant server to the target cloud host server according to the graphics card configuration information.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the network scheduling method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the network scheduling method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the network scheduling method described in any one of claims 1 to 6 are implemented.

Images