Task scheduling method and device in heterogeneous multi-core environment based on deep learning

Abstract

The invention discloses a task scheduling method in a heterogeneous multi-core environment based on deep learning. The method comprises the following steps: forming a DAG node graph according to a task flow; a key feature matrix is extracted according to the DAG node graph in combination with a GCN of a transform module; performing training learning optimization on a node network through the key feature matrix, and selecting proper nodes; scheduling the appropriate node to the most appropriate processor by adopting a heuristic scheduling algorithm; and after the strategy is optimized through the PPO algorithm, the above process is repeated until the optimal node distribution mode optimized by the algorithm is output to the target processor.

Description

Technical Field

[0001] This invention belongs to the fields of high-performance computing (HPC), cloud computing and data centers, communication networks, Internet of Things (IoT) and edge computing, artificial intelligence and deep learning, and specifically relates to a task scheduling method and apparatus based on deep learning in a heterogeneous multi-core environment. Background Technology

[0002] In modern multi-core heterogeneous systems, processors not only include multiple general-purpose CPU cores, but also dedicated hardware accelerators such as GPUs, DSPs, and FPGAs. These different types of processor cores have their own advantages in performance, energy consumption, and computing power. Therefore, in such heterogeneous computing environments, how to efficiently schedule tasks to optimize system performance and resource utilization has become a key research topic.

[0003] Traditional task scheduling algorithms, such as static scheduling or heuristic methods, typically require prior knowledge of task characteristics and system resource status. This approach has limited effectiveness in handling the complex and ever-changing scenarios of heterogeneous multi-core systems. Especially when tasks arrive dynamically, their complexity changes, and system resource status fluctuates, traditional scheduling methods struggle to make optimal decisions in real-time and flexibly. Furthermore, multi-core heterogeneous systems have different core types and vastly different load characteristics. Task scheduling requires not only matching computing resources but also balancing factors such as energy consumption, response time, and task dependencies, further increasing the complexity of task scheduling.

[0004] In this context, Deep Reinforcement Learning (DRL) has emerged as a potentially effective method for solving task scheduling problems in multi-core heterogeneous systems due to its ability to learn optimal policies in complex and dynamic environments. DRL continuously learns the mapping relationship between the current system state (including task load, core availability, energy consumption, etc.) and scheduling decisions through the interaction between the agent and the environment, thereby optimizing the overall performance of the system over long-term operation.

[0005] However, because deep reinforcement learning increases the computational complexity of data and takes a relatively long time, a method that combines traditional heuristic algorithms with deep reinforcement learning to dynamically adjust strategies is used to achieve task scheduling. Summary of the Invention

[0006] In view of this, the main objective of the present invention is to provide a task scheduling method, apparatus, and medium based on deep learning in a heterogeneous multi-core environment.

[0007] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0008] This invention provides a task scheduling method based on deep learning in a heterogeneous multi-core environment. The method is as follows:

[0009] A DAG node graph is constructed based on the task flow;

[0010] The key feature matrix is ​​extracted based on the DAG node graph and the GCN of the transformer module.

[0011] The node network is trained and optimized using the key feature matrix to select suitable nodes.

[0012] The suitable nodes are scheduled to the most suitable processor using a heuristic scheduling algorithm.

[0013] After optimizing the strategy using the PPO algorithm, repeat the above process until the optimal node allocation method optimized by the algorithm is output to the target processor.

[0014] In the above scheme, the step of extracting the feature values ​​of the DAG task graph based on the DAG node graph and the GCN of the transformer module specifically involves: for the input sequence H = [h1, h2, ... hn], mapping it to the query space Q, the key space K, and the value space V, with the relevant calculation formulas as follows:

[0015] The formula for calculating the i-th attention head is expressed as:

[0016] By concatenating the outputs of all attention heads, we obtain the output of the multi-head attention mechanism: MultiHead(H) = Concat(head1, head2, ..., head...). x W o .

[0017] In the above scheme, the step of training and optimizing the node network using the feature values ​​of the DAG task graph to select suitable nodes specifically involves: using three feature matrices to represent environmental state information in real time, inputting these three feature matrices into the next-level node network for learning and node selection, and finally using a non-linear activation function to select any node n. i The score is normalized to the probability value P. ni A suitable node is selected based on the probability value.

[0018] In the above scheme, the score of any node ni is normalized to a probability value using a nonlinear activation function. Specifically: According to Determine the probability value P ni .

[0019] In the above scheme, the heuristic scheduling algorithm is used to schedule the suitable node to the most suitable processor. Specifically, the unscheduled task nodes in the sliding window are copied. The earliest start time and the earliest finish time are calculated by combining the copied node with the predecessor node and placing it into each processor. These values ​​are then put into the copy table as a local search solution. Finally, the optimal solution is output based on the minimum value among these solutions.

[0020] In the above scheme, the strategy is optimized using the PPO algorithm, and the above process is repeated until the optimal node allocation method optimized by the algorithm is output to the target processor. Specifically, this includes:

[0021] Define the reward function;

[0022] The cropping range is adjusted using linear and logarithmic decay methods during training.

[0023] During training, the old strategy π is used. old Update the data to obtain a new strategy π θ ;

[0024] In an epoch, the data is divided into multiple mini-batches. Each training session uses mini-batch samples. During each training session, the model parameters are updated multiple times, and the pruning factor also decays.

[0025] Finally, the optimal strategy is obtained through the objective optimization function of the critic network.

[0026] The reward function is:

[0027] The pruning range applied during training using linear and logarithmic decay methods is specifically defined as follows: the decay function is linear decay. Nonlinear decay

[0028] The optimal strategy obtained through the objective optimization function of the critic network is specifically as follows:

[0029] Assuming there are L training steps, each step generates a dataset containing actions, states, and rewards in the training environment. The advantage function is determined based on this dataset of actions, states, and rewards.

[0030] Then, combining this with the PPO algorithm based on the policy objective function, the objective function for policy network updates is obtained as follows:

[0031]

[0032] Finally, the objective optimization function of the commentator network is applied.

[0033] This invention also provides a task scheduling device based on deep reinforcement learning in a heterogeneous multi-core environment, the device comprising:

[0034] The building module is used to construct a DAG node graph based on the task flow;

[0035] The transformer module is used to extract the key feature matrix based on the DAG node graph and the GCN of the transformer module;

[0036] The training module is used to train and optimize the node network using the key feature matrix to select suitable nodes.

[0037] The scheduling module is used to schedule the suitable nodes to the most suitable processor using a heuristic scheduling algorithm; it is also used to repeatedly train the strategy after optimization using the PPO algorithm until the optimal node allocation method optimized by the algorithm is output to the target processor.

[0038] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0039] A DAG node graph is constructed based on the task flow;

[0040] The key feature matrix is ​​extracted based on the DAG node graph and the GCN of the transformer module.

[0041] The node network is trained and optimized using the key feature matrix to select suitable nodes.

[0042] The suitable nodes are scheduled to the most suitable processor using a heuristic scheduling algorithm.

[0043] After optimizing the strategy using the PPO algorithm, repeat the above process until the optimal node allocation method optimized by the algorithm is output to the target processor.

[0044] This invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0045] A DAG node graph is constructed based on the task flow;

[0046] The key feature matrix is ​​extracted based on the DAG node graph and the GCN of the transformer module.

[0047] The node network is trained and optimized using the key feature matrix to select suitable nodes.

[0048] The suitable nodes are scheduled to the most suitable processor using a heuristic scheduling algorithm.

[0049] After optimizing the strategy using the PPO algorithm, repeat the above process until the optimal node allocation method optimized by the algorithm is output to the target processor.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] This invention finds the optimal scheduling strategy through algorithm training and optimization in a heterogeneous multi-core environment. Furthermore, it can control the parameter λ in the reward function according to the actual needs of the algorithm user, thereby controlling power consumption and time consumption. This expands the application scenarios of the algorithm and solves the problem that power consumption is not taken into consideration in traditional algorithms. Attached Figure Description

[0052] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and, together with their descriptions, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart illustrating a task scheduling method based on deep learning in a heterogeneous multi-core environment, as provided in an embodiment of the present invention.

[0054] Figure 2 This invention provides a system architecture for a task scheduling method based on deep learning in a heterogeneous multi-core environment.

[0055] Figure 3 This is a schematic diagram of the transformer combined with attention mechanism module structure in a task scheduling method based on deep learning in a heterogeneous multi-core environment provided by an embodiment of the present invention;

[0056] Figure 4 This is the node selection module structure in this embodiment of the invention;

[0057] Figure 5 The simulation results of the algorithm's reward under different learning rates are shown in the figure.

[0058] Figure 6 This is a simulation result of the average loss of the present invention under different numbers of tasks and different parameters λ. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0060] This invention provides a task scheduling method based on deep learning in a heterogeneous multi-core environment, such as... Figure 1 As shown, the method is as follows:

[0061] Step 101: Construct a DAG node graph based on the task flow;

[0062] Step 102: Extract the key feature matrix based on the DAG node graph and the GCN of the transformer module;

[0063] Specifically, for an input sequence H = [h1, h2, ... hn], it is mapped to a query space Q, a key space K, and a value space V, where the W value corresponding to the three spaces is a weight matrix, which is represented as matrix W in the following formulas. q W k and W v To distinguish them, these are parameters learned during model training. The relevant calculation formulas for these three spatial values ​​are as follows:

[0064] The formula for calculating the i-th attention head is expressed as: Q i K i and V i These are the values ​​of Q, K, and V in space at the i-th attention head, respectively. In the formula, Ki is calculated in its transpose form, and dk is the dimension of the corresponding attention head.

[0065] Each attention head learns the correlations between different parts of the input sequence, thereby capturing different features. Finally, the outputs of all attention heads are concatenated into vectors to achieve three feature matrices, representing the node-level feature matrix, the DAG-level feature matrix, and the global-level feature matrix, respectively. Therefore, the output of the multi-head attention mechanism can be expressed as: MultiHead(H) = Concat(head1, head2, ..., head...). x W o Among them, head x Let be the calculated value of the x-th attention head, and Wo be the linear transformation mapping matrix.

[0066] Step 103: Train and optimize the node network using the key feature matrix to select suitable nodes;

[0067] Specifically, environmental state information is represented in real time using three feature matrices. These three feature matrices are then input into the next-level node network for learning and node selection. Finally, a non-linear activation function is used to activate any node n. i The score is normalized to the probability value P. niBased on the principle of maximizing probability, select a P from among them. ni The node with the maximum value.

[0068] according to Figure 4 The structure diagram shows that three feature matrices are input to the node network as the information to be extracted for each node, and the network is trained and learned together. Finally, N output values ​​are obtained, and a corresponding probability value is obtained from these N output values.

[0069] according to Determine the probability value Where e ni represents the neural network training values ​​for node ni.

[0070] Step 104: Use a heuristic scheduling algorithm to schedule the suitable node to the most suitable processor;

[0071] Specifically, unscheduled task nodes in the sliding window are copied. The earliest start time and earliest finish time are calculated by combining the copied node with the predecessor node and placing it into each processor. These values ​​are then put into the copy table as a local search solution. Finally, the optimal solution is output based on the minimum value among these solutions.

[0072] Step 105: After optimizing the strategy using the PPO algorithm, repeat the above process until the optimal node allocation method optimized by the algorithm is output to the target processor.

[0073] Specifically, define the reward function;

[0074] The cropping range is adjusted using linear and logarithmic decay methods during training.

[0075] During training, the old strategy π is used. old Update the data to obtain a new strategy π θ ;

[0076] New Strategy π θ r, as the objective function of the subsequent PPO algorithm t The numerator of (θ) is used to limit the policy update magnitude, improve sample efficiency, and realize functions such as importance sampling. It is an indispensable parameter in the PPO algorithm.

[0077] In an epoch, the data is divided into multiple mini-batches. Each training session uses mini-batch samples. During each training session, the model parameters are updated multiple times, and the pruning factor also decays.

[0078] Finally, the optimal strategy is obtained through the objective optimization function of the critic network.

[0079] The reward function is: Where J(k) is the step tn to t n+1 The nodes that have completed scheduling execution between TD comm and TD proc These are the communication time and computation time of these nodes, E. comm and E proc These represent the communication and computational energy consumption of these nodes, respectively, with λ representing the weighting factor between latency and power consumption.

[0080] In this invention, a three-layer fully connected neural network is used for node selection. The input layer receives three feature matrices output from the previous stage, the hidden layer is a single fully connected neural network consisting of 128 units, and the output layer contains 16 units.

[0081] The pruning range applied during training using linear and logarithmic decay methods specifically refers to a decay function that is non-linear. Where T is the total number of iterations in the algorithm training process, t is the current iteration number, ε0 is the initial pruning factor value, and ξ is the decay factor.

[0082] The optimal strategy obtained through the objective optimization function of the critic network is specifically as follows:

[0083] Assuming there are T training steps, each step generates a dataset containing actions, states, and rewards in the training environment. The advantage function is determined based on this dataset of actions, states, and rewards. In the formula, G t For the reward function G t =R t+1 +γR t+2 +…+γ n-1 R t+n +γ n V(s t+n );

[0084] Where V(s) t ,a t ;π k ) is π k State-action function s under policy t Let a be the state when the step is t. t Let be the action value at step t, and Rt+1 be the reward at step t+1, γ be the discount factor, and V(st+n) be the state function value at the final state st+n. Then, combining this with the PPO algorithm based on the policy objective function, the objective function for updating the policy network is obtained as follows: Referent r t (θ) is represented as: E represents the expected value, and τ represents the strategy π. k The trajectory is given below, where θ is the hyperparameter value of the current policy function (the hyperparameter that needs to be optimized).old Here are the hyperparameter values ​​for the old policy network, rt(θ) is the policy ratio (probability ratio between the old and new policies), clip() is the clipping function, and πθ(a t |s t When in state 'st', action a is executed. t The current policy probability value, and π θold (a t |s t ) represents the probability value of the old strategy.

[0085] Finally, the objective optimization function of the commentator network is applied. Where φ is a hyperparameter of the critic network, and the other parameters have been explained above. The subscript φ indicates that it is the value of this parameter in the critic network.

[0086] This invention references shared memory-based communication protocols and the PCIe (Peripheral Component Interconnect Express) 4.0 protocol to study data exchange in heterogeneous multi-core systems. Through algorithm training and optimization, the optimal scheduling strategy is found to reduce scheduling time. Furthermore, based on the actual needs of algorithm users, the parameter λ in the reward function is controlled to achieve the purpose of controlling power consumption and time consumption, thereby increasing the application scenarios of the algorithm.

[0087] This invention improves the heuristic task scheduling algorithm in heterogeneous multi-core environments by combining it with deep reinforcement learning algorithms, thereby reducing the computational complexity of the task scheduling process. In the optimization of the PPO algorithm, the algorithm is improved by dynamically adjusting the pruning factor, which can effectively avoid prematurely reducing the exploration variance and prevent slow progress and getting stuck in local optima. Furthermore, the power consumption factor is considered in the reward function, filling the gap in the task scheduling algorithm's incomplete consideration of power consumption.

[0088] like Figure 5 As shown in the figure, when the learning rate α is 0.00001, the move reward increases with the increase of training batches. The optimization strategy increases the move reward, but the efficiency is very low. When the learning rate α is 0.0001, the trend is the same, but the growth rate is faster. When the training batch reaches 450, the growth rate basically stops and reaches the maximum value. When the learning rate α is 0.001, the growth rate starts to be very fast, but due to the excessive learning rate, the model may cross the optimal solution when updating parameters. The value of the move reward diverges, and the final performance is poor.

[0089] like Figure 5As shown, the weight of power consumption and time consumption is controlled by the value of parameter λ. When λ is 0, only power consumption loss is considered, and the average loss is the largest as the number of heterogeneous tasks increases. When λ is 1, only time consumption loss is considered, showing the same trend, but the average loss is slightly less. When λ is 0.6, time consumption is given more weight, and the average loss is the best. It can be seen that time consumption and power consumption are closely related, and the reduction of time consumption during task execution has a greater impact on the reduction of average loss.

[0090] In one embodiment, a task scheduling device based on deep reinforcement learning in a heterogeneous multi-core environment is proposed, the device comprising:

[0091] The building module is used to construct a DAG node graph based on the task flow;

[0092] The transformer module is used to extract the key feature matrix based on the DAG node graph and the GCN of the transformer module;

[0093] The training module is used to train and optimize the node network using the key feature matrix to select suitable nodes.

[0094] The scheduling module is used to schedule the suitable nodes to the most suitable processor using a heuristic scheduling algorithm; it is also used to repeatedly train the strategy after optimization using the PPO algorithm until the optimal node allocation method optimized by the algorithm is output to the target processor.

[0095] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0096] Step 101: Construct a DAG node graph based on the task flow;

[0097] Step 102: Extract the key feature matrix based on the DAG node graph and the GCN of the transformer module;

[0098] Step 103: Train and optimize the node network using the key feature matrix to select suitable nodes;

[0099] Step 104: Use a heuristic scheduling algorithm to schedule the suitable node to the most suitable processor;

[0100] Step 105: After optimizing the strategy using the PPO algorithm, repeat the above process until the optimal node allocation method optimized by the algorithm is output to the target processor.

[0101] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps:

[0102] Step 101: Construct a DAG node graph based on the task flow;

[0103] Step 102: Extract the key feature matrix based on the DAG node graph and the GCN of the transformer module;

[0104] Step 103: Train and optimize the node network using the key feature matrix to select suitable nodes;

[0105] Step 104: Use a heuristic scheduling algorithm to schedule the suitable node to the most suitable processor;

[0106] Step 105: After optimizing the strategy using the PPO algorithm, repeat the above process until the optimal node allocation method optimized by the algorithm is output to the target processor.

[0107] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0108] Simulation data

[0109] The embodiments of this invention are based on simulations of this invention combined with other algorithms in a Python environment, and the current algorithm structure is designed by comparing the advantages and disadvantages of other algorithms.

[0110] Table 1 Hardware and Software Environment

[0111]

[0112] The paper (Song Y, Li C, Tian L, et al. A reinforcement learning based jobscheduling algorithm for heterogeneous computing environment[J]. Computers and Electrical Engineering, 2023, 107: 108653.) also studies task scheduling in heterogeneous environments, combining traditional heuristic scheduling with deep reinforcement learning to reduce the computational complexity of the system. In the heuristic scheduling algorithm, an optimistic cost table and parent node replication are used to balance performance instability by estimating the future cost of the current allocation. In deep reinforcement learning, the PPO algorithm is used to directly evaluate the allocated nodes, continuously optimizing the allocation method and node selection order. This invention improves upon PPO, effectively avoiding premature reduction of exploration variance, minimizing slow progress and preventing getting trapped in local optima. It uses a dynamically adjusted pruning factor to gradually reduce exploration variance during training, avoiding local optima. Furthermore, considering that task scheduling algorithms in real-world environments also need to constrain power consumption, power consumption is incorporated into the reward function design to address different environmental requirements.

[0113] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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.

[0114] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A task scheduling method based on deep learning in a heterogeneous multi-core environment, characterized in that, The method is as follows: A DAG node graph is constructed based on the task flow; The key feature matrix is ​​extracted based on the DAG node graph and the GCN of the transformer module. The node network is trained and optimized using the key feature matrix to select suitable nodes. The suitable nodes are scheduled to the most suitable processor using a heuristic scheduling algorithm. After optimizing the strategy using the PPO algorithm, repeat the above process until the optimal node allocation method optimized by the algorithm is output to the target processor.

2. The task scheduling method based on deep learning in a heterogeneous multi-core environment according to claim 1, characterized in that, The step of extracting the feature values ​​of the DAG task graph based on the DAG node graph and the GCN of the transformer module is as follows: For the input sequence H = [h1, h2, ... hn], it is mapped to the query space Q, the key space K, and the value space V. The relevant calculation formulas are as follows: The formula for calculating the i-th attention head is expressed as: By concatenating the outputs of all attention heads, we obtain the output of the multi-head attention mechanism: MultiHead(H) = Concat(head1, head2, ..., head...). x W o .

3. The task scheduling method based on deep learning in a heterogeneous multi-core environment according to claim 1 or 2, characterized in that, The process of training and optimizing the node network using the feature values ​​of the DAG task graph to select suitable nodes specifically involves: using three feature matrices to represent environmental state information in real time; inputting these three feature matrices into the next-level node network for learning and node selection; and finally, using a non-linear activation function to optimize any node n. i The scores are normalized to probability values. Select a suitable node based on the probability value.

4. The task scheduling method based on deep learning in a heterogeneous multi-core environment according to claim 3, characterized in that, The score of any node ni is normalized to a probability value using a nonlinear activation function. Specifically: According to Determine the probability value 5. The task scheduling method based on deep learning in a heterogeneous multi-core environment according to claim 4, characterized in that, The heuristic scheduling algorithm is used to schedule the suitable nodes to the most suitable processor. Specifically, the unscheduled task nodes in the sliding window are copied. The earliest start time and the earliest finish time are calculated by combining the copied nodes with the predecessor nodes and placing them into each processor. These values ​​are then put into the copy table as local search solutions. Finally, the optimal solution is output based on the minimum value among these solutions.

6. The task scheduling method based on deep learning in a heterogeneous multi-core environment according to claim 5, characterized in that, After optimizing the strategy using the PPO algorithm, the above process is repeated until the optimal node allocation method determined by the algorithm is output to the target processor. Specifically, this includes: Define the reward function; The cropping range is adjusted using linear and logarithmic decay methods during training. During training, the old strategy π is used. old Update the data to obtain a new strategy π θ ; In an epoch, the data is divided into multiple mini-batches. Each training session uses mini-batch samples. During each training session, the model parameters are updated multiple times, and the pruning factor also decays. Finally, the optimal strategy is obtained through the objective optimization function of the critic network. The reward function is: The pruning range applied during training using linear and logarithmic decay methods is specifically defined as follows: the decay function is linear decay. Nonlinear decay The optimal strategy obtained through the objective optimization function of the critic network is specifically as follows: Assuming there are L training steps, each step generates a dataset containing actions, states, and rewards in the training environment. The advantage function is determined based on this dataset of actions, states, and rewards. Then, combining this with the PPO algorithm based on the policy objective function, the objective function for policy network updates is obtained as follows: Finally, the objective optimization function of the commentator network is applied.

7. A task scheduling device based on deep reinforcement learning in a heterogeneous multi-core environment, characterized in that, The device includes: The building module is used to construct a DAG node graph based on the task flow; The transformer module is used to extract the key feature matrix based on the DAG node graph and the GCN of the transformer module; The training module is used to train and optimize the node network using the key feature matrix to select suitable nodes. The scheduling module is used to schedule the suitable nodes to the most suitable processor using a heuristic scheduling algorithm; it is also used to repeatedly train the strategy after optimization using the PPO algorithm until the optimal node allocation method optimized by the algorithm is output to the target processor.

8. A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method as claimed in any one of claims 1 to 6.

9. A computer device comprising a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as claimed in any one of claims 1 to 6.

Images