A service arrangement method under limited computing power and computer-readable storage medium

Abstract

The present invention relates to a service arrangement method under limited computing power and a computer-readable storage medium, wherein the service arrangement method comprises: obtaining N A computing power-constrained cluster is composed of server devices, and the resource information of each server device and the resource configuration and usage information of each service in each server device are collected from the computing power-constrained cluster, and the above information and the resource demand information of each service in the service queue to be arranged are combined into a training set; a service orchestration model is constructed, and the constructed service orchestration model is trained using the training set to obtain the trained service orchestration model; the trained service orchestration model is migrated to the scheduling center; a service queue to be allocated is constructed, and the constructed service queue to be allocated is input into the scheduling center to obtain the service orchestration result. This method can reflect the utilization of resources by services; and ensures the balance of service scheduling and allocation, and improves the utilization rate of cluster server devices.

Description

Technical Field

[0001] The present invention relates to the technical field of resource allocation, and in particular to a service orchestration method under limited computing power and a computer-readable storage medium. Background Art

[0002] In the era of intelligent big data, with the continuous development of technology, data is growing rapidly, which means that the demand for data processing is also increasing. However, due to the limitations of CPU manufacturing technology, there are still great technical obstacles to achieving breakthrough progress in computing power on smaller chips. Due to the limitation of physical resources, how to schedule and use more services under limited resources has become a difficult problem faced by many companies.

[0003] To orchestrate resources, increase the number of services available at the same time and the utilization rate of physical resources. For example, the Chinese invention patent with patent number ZL202110313443.6 (authorization announcement number CN113051053B) discloses a heterogeneous resource scheduling method, device, equipment and computer-readable storage medium. The scheduling method uses a unified interface of a deep learning framework to receive overall general tasks, and uses a unified application interface for distribution, so as to achieve the purpose of breaking through the isolation of heterogeneous resources and balancing the heterogeneous resources of the AI ​​platform. To a certain extent, it solves the current problems of isolation and versatility between heterogeneous resources. However, for the service scheduling problem under limited resources, the above scheme still has the following two problems: first, the above method cannot reasonably adjust the service according to the existing cluster resource configuration and usage load; second, the above method cannot measure and evaluate the computing power and service load. To this end, it is necessary to further improve the existing technology. Summary of the invention

[0004] The first technical problem to be solved by the present invention is to provide a service scheduling method under limited computing power that can ensure the balance of service scheduling and allocation in response to the above-mentioned prior art.

[0005] The second technical problem to be solved by the present invention is to provide a computer-readable storage medium for applying the above-mentioned service orchestration method under limited computing power in response to the above-mentioned prior art.

[0006] The technical solution adopted by the present invention to solve the above-mentioned first technical problem is: a service orchestration method under limited computing power, characterized by comprising the following steps:

[0007] Step 1: Get N A computing power-constrained cluster consisting of 100 server devices. Nis a positive integer, and collects resource information of each server device and resource configuration and usage information of each service in each server device from the computing power-constrained cluster, and combines the resource information of each server device, the resource configuration and usage information of each service in each server device, and the resource demand information of each service in the service queue to be arranged into a training set; the service queue to be arranged is composed of the services to be arranged in the computing power-constrained cluster;

[0008] Step 2: Build a service orchestration model, and use the training set in step 1 to train the built service orchestration model to obtain a trained service orchestration model;

[0009] The process of training the constructed service orchestration model is as follows:

[0010] Step 2-1: Build a DQN reinforcement learning network, and use the constructed service orchestration model as the evaluation network and target network of the DQN reinforcement learning network;

[0011] Step 2-2, normalize the training set, input the normalized resource information of each server device, the normalized service resource configuration and usage information in each server device, and the normalized resource demand information of each service in the service queue to be orchestrated into the DQN reinforcement learning network, calculate the reward function r, and use the reward function r to update the network parameters of the DQN reinforcement learning network to obtain the target network after training. The target network after training is the service orchestration model after the training.

[0012] Step 3: Migrate the trained service orchestration model to the scheduling center;

[0013] Step 4: Construct a service queue to be allocated, and input the constructed service queue to be allocated to the dispatch center, that is, the dispatch center collects the resource information of each server device and the resource configuration and usage information of each service in each server device from the cluster, and inputs the resource information of each server device, the resource configuration and usage information of each service in each server device, and the resource demand information of each service in the service queue to be allocated into the trained service orchestration model to obtain the service orchestration result.

[0014] Preferably, the resource information of each server device in step 1 includes CPU default frequency, number of CPU cores, total memory, GPU total video memory, number of GPUs, GPU rated power, intranet bandwidth, and utilization of the service under the configured resources in an ideal state.

[0015] Preferably, the service resource configuration and usage information in each server device in step 1 includes the allocated CPU operating frequency, the number of CPU cores, the allocated memory size, the GPU memory occupancy rate, the GPU power consumption, the utilization rate of the allocated bandwidth and the utilization rate of the service under the configured resources.

[0016] In the above scenario, the service type Utilization of configured resources The calculation formula is:

[0017]

[0018] in, To obtain the average value, is the service type existing on server device 1 The utilization rate under the configured resources, is the service type existing on server device 2 The utilization rate under the configured resources, For the server device Service Types The utilization rate under the configured resources, The calculation formula is:

[0019]

[0020] For server equipment Service Type The number of CPU cores in the resource configuration usage information, For server equipment Service Type The CPU operating frequency in the resource configuration usage information, For server equipment The default CPU frequency on For server equipment The number of GPUs on For server equipment Service Type GPU power consumption in resource configuration usage information, For server equipment The GPU power rating on For server equipment Service Type The utilization of the allocated bandwidth in the resource configuration usage information.

[0021] Furthermore, the ideal state is the utilization rate of the service under the configured resources. The calculation formula is:

[0022]

[0023] In ideal state , , For server equipment The number of CPU cores on the , .

[0024] Preferably, the reward function in step 2-2 is The calculation formula is:

[0025] in, is the first reward function, The calculation formula is:

[0026]

[0027] For server equipment The number of services enabled on To increase the service types that all servers can carry after adding similar services The upper limit of The calculation formula is: , The service type within the interval to be counted The historical probability value of occurrence, The calculation formula is: , The service type within the interval to be counted The number of occurrences, is the total number of services in the interval to be counted, is the number of service types within the interval to be counted; The service type within the interval to be counted Number of consecutive occurrences; Server equipment before dispatching actions for services Next services and service types The utilization rate under the configured resources, Server equipment after dispatching actions for services Next services and service types Utilization rate under configured resources;

[0028] is the second reward function, The calculation formula is:

[0029] ;

[0030] is the third reward function, The calculation formula is:

[0031] ,

[0032] For server equipment The first service under and service type The utilization rate under the configured resources, For server equipment The next services and service types Utilization of configured resources.

[0033] The above service orchestration model can use a neural network. In order to be able to carry out the optimized scheduling of thousands of server devices with low latency, preferably, the service orchestration model includes a convolution block, a linear block, a fusion module, two MobileOne network blocks, a residual connection layer, an upsampling layer and an activation function layer; the specific process of the service orchestration model is:

[0034] (1) The normalized resource information of each server device and the normalized service resource configuration and usage information of each server device are combined into a first input matrix , the first input matrix Input into the convolution block to obtain the first feature map; in addition, the normalized resource demand information of each service in the service queue to be arranged constitutes the second input matrix , the second input matrix Input into the linear block to obtain the second feature map;

[0035] (2) inputting the first feature map and the second feature map into a fusion module, where the fusion module is used to perform a dot product calculation on the first feature map and the second feature map to obtain a third feature map;

[0036] (3) After the third feature map passes through the two connected MobileOne network blocks, the residual connection layer, the upsampling layer and the activation function layer in sequence, the service scheduling action is output.

[0037] The technical solution adopted by the present invention to solve the above-mentioned second technical problem is: a computer-readable storage medium, which stores a computer program and can be read and executed by a processor, and is characterized in that: when the computer program is executed by the processor, it implements the service orchestration method under limited computing power as mentioned above.

[0038] Compared with the prior art, the advantages of the present invention are: by collecting the resource information of each server device and the resource configuration and usage information of each service in each server device from the computing power-constrained cluster, forming a training set, and training the constructed service orchestration model. Therefore, on the one hand, the resource information of each server device and the resource configuration and usage information of each service in each server device can reflect the resource utilization of the service to a certain extent, and play a positive role in the rationalization of resource allocation; on the other hand, the combination of DQN reinforcement learning and reward function ensures the balance of service scheduling and allocation, and improves the utilization rate of cluster server devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of a service orchestration method under limited computing power in an embodiment of the present invention;

[0040] Figure 2 for Figure 1 The use architecture of the service orchestration method in

[0041] Figure 3 for Figure 1 Block diagram of the service orchestration model in . DETAILED DESCRIPTION

[0042] The present invention is further described in detail below with reference to the accompanying drawings.

[0043] like Figure 1 As shown, the service orchestration method under limited computing power in this embodiment includes the following steps:

[0044] Step 1: Get N A computing power-constrained cluster consisting of 100 server devices. N is a positive integer, and collects resource information of each server device and resource configuration and usage information of each service in each server device from the computing power-constrained cluster, and combines the resource information of each server device, the resource configuration and usage information of each service in each server device, and the resource demand information of each service in the service queue to be arranged into a training set; the service queue to be arranged is composed of the services to be arranged in the computing power-constrained cluster;

[0045] like Figure 2 As shown in the figure, the total number of server devices on the computing power-constrained cluster is N, and the maximum number of services is H. , h (1)、 h (2)… are respectively the number of services enabled on server device 1, the number of services enabled on server device 2, ... the number of services enabled on server device n;

[0046] In this embodiment, the resource information of each server device includes the CPU default frequency, the number of CPU cores, the total memory, the GPU total memory, the number of GPUs, the GPU rated power, the intranet bandwidth and the utilization rate of the service under the configured resources in the ideal state; the resource configuration and usage information of each service in each server device includes the allocated CPU operating frequency, the number of CPU cores, the allocated memory size, the GPU memory occupancy rate, the GPU power consumption, the utilization rate of the allocated bandwidth and the utilization rate of the service under the configured resources; the utilization rate of the above services under the configured resources reflects the load status of the service;

[0047] Among them, service type Utilization of configured resources The calculation formula is:

[0048]

[0049] in, To obtain the average value, is the service type existing on server device 1 The utilization rate under the configured resources, is the service type existing on server device 2 The utilization rate under the configured resources, For the server device Service Types The utilization rate under the configured resources, The calculation formula is:

[0050]

[0051] For server equipment Service Type The number of CPU cores in the resource configuration usage information, For server equipment Service Type The CPU operating frequency in the resource configuration usage information, For server equipment The default CPU frequency on For server equipment The number of GPUs on For server equipment Service Type GPU power consumption in resource configuration usage information, For server equipment The GPU power rating on For server equipment Service Type The utilization rate of the allocated bandwidth in the resource configuration and usage information;

[0052] Ideal utilization of services under configured resources The calculation formula is:

[0053]

[0054] In ideal state The above 1.9 times is the ratio between the CPU's maximum turbo frequency and the default frequency. Due to the CPU's large and small core settings, it is impossible for all cores to reach the CPU's maximum turbo frequency at the same time, so this value is a theoretical value. , For server equipment The number of CPU cores on the , ;

[0055] The resource information of each server device and the configuration and usage information of each service resource in each server device are collected by polling, with a time span of two to three days and a collection interval of 10 to 15 minutes;

[0056] Step 2: Build a service orchestration model, and use the training set in step 1 to train the built service orchestration model to obtain a trained service orchestration model;

[0057] The process of training the constructed service orchestration model is as follows:

[0058] Step 2-1: Build a DQN reinforcement learning network, and use the constructed service orchestration model as the evaluation network and target network of the DQN reinforcement learning network;

[0059] Step 2-2, normalize the training set, input the normalized resource information of each server device, the normalized service resource configuration and usage information in each server device, and the normalized resource demand information of each service in the service queue to be orchestrated into the DQN reinforcement learning network, calculate the reward function r, and use the reward function r to update the network parameters of the DQN reinforcement learning network to obtain the target network after training. The target network after training is the service orchestration model after the training.

[0060] In this embodiment, the DQN reinforcement learning network is a prior art, and its specific training process can refer to the content disclosed in https: / / blog.csdn.net / qq_37266917 / article / details / 109151368. In the learning process, the target network is used to evaluate the reward obtained by the self-benefit, and it is used as the learning goal. In the process of weight updating, only the weight of the evaluation network is updated, and the weight of the target network is not updated. In this way, the target targeted when updating the weight will not change in each iteration and is a fixed target. After completing a certain number of updates, the weight value of the evaluation network is assigned to the target network, and then the next batch of updates is performed. In this way, the target network can also be updated. Since the estimate of the reward is relatively fixed during a period of time when the target network does not change, the introduction of the target network increases the stability of learning;

[0061] In this embodiment, the reward function The calculation formula is:

[0062] in, is the first reward function, The calculation formula is:

[0063]

[0064] For server equipment The number of services enabled on To increase the service types that all servers can carry after adding similar services The upper limit of The calculation formula is: , The service type within the interval to be counted The historical probability value of occurrence, The calculation formula is: , The service type within the interval to be counted The number of occurrences, is the total number of services in the interval to be counted, is the number of service types within the interval to be counted; The service type within the interval to be counted Number of consecutive occurrences; Server equipment before dispatching actions for services Next services and service types The utilization rate under the configured resources, Server equipment after dispatching actions for services Next services and service types Utilization rate under configured resources;

[0065] In this embodiment, The value is obtained by obtaining the most recent The service is calculated based on The preferred setting is 100;

[0066] is the second reward function, The calculation formula is:

[0067] ;

[0068] is the third reward function, The calculation formula is:

[0069] ,

[0070] For server equipment The first service under and service type The utilization rate under the configured resources, For server equipment The next services and service types Utilization rate under configured resources;

[0071] like Figure 3 As shown, the above service orchestration model includes a convolution block, a linear block, a fusion module, two MobileOne network blocks, a residual connection layer, an upsampling layer, and an activation function layer; the number of output channels of the convolution block and the linear block in this embodiment is 64, the convolution block is composed of a convolution with a convolution kernel size of 3 and a step size of 1, LayerNorm (normalization), and a ReLu activation function connected in sequence, and the linear block is composed of three linear layers connected in sequence. The CH in the figure represents the output feature map size, and the feature map is first enlarged to twice the number of output channels and then shrunk to the size of the output channels; the fusion module is a dot multiplication calculation; the MobileOne network block is an existing technology, and its main feature is that it uses multi-branch 3×3 convolution to extract diversified features during training, and simplifies branches to reduce latency through parameter reconstruction during inference; the residual connection layer further aggregates information of various types of data in the feature map through 5×5 deep convolution, 1×1 convolution, and residual connection. The upsampling layer completes the aggregation of the final semantic information through the upsampling module and two consecutive convolution blocks. Preferably, except that the output feature map of the last convolution block in the upsampling layer is set to 4, the number of output feature maps of the remaining convolution blocks is set to 64; the activation function layer adopts the Sigmoid activation function.

[0072] The specific process of the above service orchestration model is as follows:

[0073] (1) The normalized resource information of each server device and the normalized service resource configuration and usage information of each server device form a first input matrix , the first input matrix Input into the convolution block to obtain the first feature map; in addition, the normalized resource demand information of each service in the service queue to be arranged constitutes the second input matrix , the second input matrix Input into the linear block to obtain the second feature map;

[0074] (2) inputting the first feature map and the second feature map into a fusion module, where the fusion module is used to perform a dot product calculation on the first feature map and the second feature map to obtain a third feature map;

[0075] (3) After the third feature map passes through the two connected MobileOne network blocks, the residual connection layer, the upsampling layer and the activation function layer in sequence, the service scheduling action is output;

[0076] Step 3: Migrate the trained service orchestration model to the scheduling center;

[0077] After the dispatch center completes each dispatch action, it will collect resource information of each server device, configuration and usage information of each service resource in each server device, and the service queue to be orchestrated from the cluster, and then build a new training set again. The service orchestration model will be trained again using the constructed new training set. The training process refers to the content shown in step 2. Finally, the weight of the evaluation network after training is used to replace the weight of the service orchestration model saved by the dispatch center, and the weight of the service orchestration model of the dispatch center is updated, so that the dispatch center can continuously optimize the dispatch strategy during actual use.

[0078] Step 4: Construct a service queue to be allocated, and input the constructed service queue to be allocated to the dispatch center, that is, the dispatch center collects the resource information of each server device and the resource configuration and usage information of each service in each server device from the cluster, and inputs the resource information of each server device, the resource configuration and usage information of each service in each server device, and the resource demand information of each service in the service queue to be allocated into the trained service orchestration model to obtain the service orchestration result.

[0079] like Figure 2As shown, in this embodiment, the user submits a task at the platform layer, and the platform layer parses the task submitted by the user into data and services, which flow into the scheduling center; after receiving the request, the scheduling center submits the parsed data to the data sharing center, and the scheduling center inserts the parsed service into the service queue to be allocated. The scheduling center analyzes the status of each server device in the current computing power-constrained cluster based on the service orchestration model, and generates a scheduling action until all services in the service queue to be allocated are scheduled. Figure 2 The nodes in the system refer to the server devices deployed in the system, the services refer to the deep learning algorithms that provide image recognition capabilities, the cluster refers to the whole of all the servers that the system relies on, and the dispatch center refers to the functional module in the system that deploys deep learning algorithms to servers based on the status of all servers and the assigned tasks.

[0080] The computer-readable storage medium in this embodiment stores a computer program and can be read and executed by a processor. When the computer program is executed by the processor, the above-mentioned service orchestration method under limited computing power is implemented.

[0081] The utilization rate of the service type under the configured resources, the utilization rate of the service under the configured resources in the ideal state and the probability value of the service history calculated in this embodiment can reflect the utilization of resources by the service to a certain extent, which has a positive significance for the rational allocation of resources; and the proposed service scheduling model is based on the convolutional neural network (CNN) architecture, and the fully convolutional structure enables the model to carry out optimized scheduling of thousands of server devices with low latency; in addition, the reward function in this embodiment solves the problem of low resource utilization caused by service exclusivity under resource constraints through the calculation of three parts.

[0082] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A service orchestration method under limited computing power, characterized in that The steps include: Step 1: Get N A computing power-constrained cluster consisting of 100 server devices. N is a positive integer, and collects resource information of each server device and resource configuration and usage information of each service in each server device from the computing power-constrained cluster, and the resource information of each server device, the resource configuration and usage information of each service in each server device and the resource demand information of each service in the queue to be arranged form a training set; the queue to be arranged is composed of the services to be arranged in the computing power-constrained cluster; The resource information of each server device includes the CPU default frequency, number of CPU cores, total memory, total GPU memory, number of GPUs, GPU rated power, intranet bandwidth, and the utilization rate of the service under the configured resources in the ideal state; The configuration and usage information of each service resource in each server device includes the allocated CPU operating frequency, number of CPU cores, allocated memory size, GPU memory occupancy, GPU power consumption, allocated bandwidth utilization, and service utilization under the configured resources; Step 2: Build a service orchestration model, and use the training set in step 1 to train the built service orchestration model to obtain a trained service orchestration model; The service orchestration model includes a convolution block, a linear block, a fusion module, two MobileOne network blocks, a residual connection layer, an upsampling layer and an activation function layer; the specific process of the service orchestration model is: (1) The normalized resource information of each server device and the normalized service resource configuration and usage information of each server device are combined into a first input matrix , the first input matrix Input into the convolution block to obtain the first feature map; in addition, the normalized resource demand information of each service in the service queue to be arranged constitutes the second input matrix , the second input matrix Input into the linear block to obtain the second feature map; (2) inputting the first feature map and the second feature map into a fusion module, where the fusion module is used to perform a dot product calculation on the first feature map and the second feature map to obtain a third feature map; (3) After the third feature map passes through the two connected MobileOne network blocks, the residual connection layer, the upsampling layer and the activation function layer in sequence, the service scheduling action is output; The process of training the constructed service orchestration model is as follows: Step 2-1: Build a DQN reinforcement learning network, and use the constructed service orchestration model as the evaluation network and target network of the DQN reinforcement learning network; Step 2-2, normalize the training set, input the normalized resource information of each server device, the normalized service resource configuration and usage information in each server device, and the normalized resource demand information of each service in the service queue to be orchestrated into the DQN reinforcement learning network, calculate the reward function r, and use the reward function r to update the network parameters of the DQN reinforcement learning network to obtain the target network after training. The target network after training is the service orchestration model after the training. Step 3: Migrate the trained service orchestration model to the scheduling center; Step 4: Construct a service queue to be allocated, and input the constructed service queue to be allocated to the dispatch center, that is: the dispatch center collects the resource information of each server device and the resource configuration and usage information of each service in each server device from the computing power-constrained cluster, and inputs the resource information of each server device, the resource configuration and usage information of each service in each server device, and the resource demand information of each service in the service queue to be allocated into the trained service orchestration model to obtain the service orchestration result.

2. The service arrangement method according to claim 1, characterized in that: Service Type Utilization of configured resources The calculation formula is: in, To obtain the average value, is the service type existing on server device 1 The utilization rate under the configured resources, is the service type existing on server device 2 The utilization rate under the configured resources, For the server device Service Types The utilization rate under the configured resources, The calculation formula is: For server equipment Service Type The number of CPU cores in the resource configuration usage information, For server equipment Service Type The CPU operating frequency in the resource configuration usage information, For server equipment The default CPU frequency on For server equipment The number of GPUs on For server equipment Service Type GPU power consumption in resource configuration usage information, For server equipment The GPU power rating on For server equipment Service Type The utilization of the allocated bandwidth in the resource configuration usage information.

3. The service arrangement method according to claim 2, characterized in that: The ideal state is the utilization rate of the service under the configured resources The calculation formula is: In ideal state , , For server equipment The number of CPU cores on the , .

4. The service arrangement method according to claim 3, characterized in that: The reward function in step 2-2 The calculation formula is: in, is the first reward function, The calculation formula is: For server equipment The number of services enabled on To increase the service types that all servers can carry after adding similar services The upper limit of The calculation formula is: , The service type within the interval to be counted The historical probability value of occurrence, The calculation formula is: , The service type within the interval to be counted The number of occurrences, is the total number of services in the interval to be counted, is the number of service types within the interval to be counted; The service type within the interval to be counted Number of consecutive occurrences; Server equipment before dispatching actions for services Next services and service types The utilization rate under the configured resources, Server equipment after dispatching actions for services Next services and service types Utilization rate under configured resources; is the second reward function, The calculation formula is: ； is the third reward function, The calculation formula is: ； For server equipment The first service under and service type The utilization rate under the configured resources, For server equipment The next services and service types The utilization rate under the configured resources, H is the maximum number of services.

5. A computer-readable storage medium storing a computer program and capable of being read and executed by a processor, characterized in that: When the computer program is executed by a processor, the service orchestration method under computing power limitation as claimed in any one of claims 1 to 4 is implemented.

Images