A distributed single-node scheduling method based on multi-layer perceptron decision making
By adopting a multi-layer perceptron decision-making scheduling method in distributed data centers and building a scheduling decision recommendation model, we can solve the high latency and low throughput problems of centralized scheduling architecture, achieve more efficient task scheduling and reduce the latency of long-tail tasks.
Patent Information
- Application Number
- CN202411173117.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2025-09-26
- Estimated Expiration
- 2044-08-26
AI Technical Summary
In modern large-scale data center scheduling systems, centralized scheduling architectures are unable to cope with the high throughput and scheduling latency requirements of large-scale scheduling and a large number of serverless feature services. Simple heuristic scheduling algorithms affect the throughput of the overall solution, resulting in reduced scheduling quality and an excessive number of long-tail tasks running in the cluster.
A distributed single-node scheduling method based on multi-layer perceptron (MLP) decision-making is adopted. By collecting and analyzing scheduling status information, a multi-layer perceptron decision network is constructed, and a scheduling decision recommendation model is trained. Scheduling decisions are optimized to improve cluster throughput and reduce long-tail task delays.
It improves the cluster throughput of the distributed scheduling system, reduces the delay of long-tail tasks, optimizes the scheduling effect, alleviates the pressure of centralized scheduling, and improves the overall scheduling quality.
Smart Images
Figure CN119071292B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of task scheduling, and relates to a distributed single-node scheduling method based on multi-layer perceptron decision-making in a distributed cluster environment. Background Art
[0002] Most modern large-scale data center scheduling systems use a centralized scheduling architecture, meaning all resource allocation and service orchestration within the cluster are calculated and decided by a central scheduling component. This faces two challenges: a single central scheduling component cannot handle large-scale scheduling and cannot meet the high throughput and scheduling latency requirements of numerous serverless services.
[0003] To address the above issues, a team proposed a high-throughput scheduling architecture based on distributed node independent decision-making based on detection queues, including: (1) a central controller that stores cluster status information (2) distributed nodes with independent decision-making modules and scheduling detection queues (3) a node scheduling basic logic module, including the following steps: S01: Task distribution phase: The central controller node will receive external task requests and distribute tasks to cluster nodes in a random distribution manner. At the same time, the central controller node will write task-related information into the database for subsequent task status tracking. S02: Node self-scheduling phase: An algorithm within the node determines whether the task is executed locally or forwarded to other nodes. S03: Task forwarding phase: If S02 decides to forward the task to certain nodes, these nodes will become receiving nodes, and the receiving nodes will put the forwarded tasks into their own detection queues for storage and processing. S04: Forwarding Task Scheduling Phase: The receiving node detects forwarding tasks in the queue and first checks for timeouts. If the task waits for a timeout, it returns a timeout error to the scheduling node. If not, the receiving node examines its own resources. If sufficient, it sends a confirmation request to the scheduling node. If insufficient, it adds the task to the delayed binding queue. S05: Task Feedback Processing Phase: After receiving feedback from the receiving node, the scheduling node processes the feedback. If the feedback is an error, it counts errors. When the error count reaches a set threshold, the task is retried, retrying steps S02-S05. If the feedback is a confirmation request, it checks whether the task has been preempted by another node. If not, it returns a confirmation message; if so, it returns a cancellation message. S06: Task Execution Recording Phase: After receiving the confirmation message, the receiving node executes the task. The scheduling node, based on a heartbeat mechanism, periodically reports confirmed task information and its own node resource information to the central controller node.
[0004] In the high-throughput scheduling architecture described above, step S02 uses the simplest heuristic scheduling algorithm. When a task arrives at a scheduling node, it checks whether local resources are sufficient to run the task. If so, it automatically schedules the task to the node; if not, it forwards the task. The task forwarding logic uses a two-step random selection strategy, selecting an appropriate forwarding node address from the cluster's neighboring node table. This simple heuristic scheduling algorithm affects the overall solution's throughput, resulting in poor scheduling quality and an excessive number of long-tail tasks running on the cluster. Summary of the Invention
[0005] In view of the above, the purpose of the present invention is to provide a distributed single-node scheduling method based on multi-layer perceptron (MLP) decision-making. This method finds relatively optimal decisions based on the status of the cluster and historical simulation results, further leverages the advantages of single-node parallel scheduling, improves the cluster throughput of the distributed scheduling system, reduces long-tail task delays, and optimizes the scheduling effect.
[0006] To achieve the above-mentioned object of the invention, an embodiment provides a distributed single-node scheduling method based on multi-layer perceptron decision making, comprising the following steps:
[0007] Collect the scheduling status information and scheduling decision information of each distributed node in the distributed scheduling system at each step, evaluate the scheduling results according to the scheduling log record information, and obtain the evaluation results including the best scheduling strategy or the worst scheduling strategy;
[0008] A scheduling decision recommendation model is constructed, which includes two multilayer perceptron decision networks corresponding to the optimal scheduling strategy and the inferior scheduling strategy, and a linear fusion layer. The scheduling status information is used as sample data and divided according to the evaluation results. The data is then input into the corresponding multilayer perceptron decision network for scheduling decision prediction. The scheduling decision information is used as supervision data and a loss function is constructed with the scheduling decision prediction results to train the multilayer perceptron decision network.
[0009] Each distributed node uses the trained scheduling policy recommendation model to make scheduling decision recommendations based on scheduling status information, and uses the recommended scheduling decisions for scheduling.
[0010] In the method of the present invention, the scheduling status information includes the current time T when the decision is made, the task submission time t c 、Resources required for the task res req , the remaining resources localres of this node when making a decision, the neighboring node information known to this node when making a decision, and the state update time of each neighboring node n , and the remaining resource information of the neighborhood when making decisions res n .
[0011] In the method of the present invention, the scheduling decision information includes three types of scheduling decisions, namely:
[0012] Local waiting: Add the scheduled task to the waiting queue of the local node and wait for the next cycle of scheduling;
[0013] Neighborhood forwarding: Select two neighboring nodes and send task detection information to them. If the neighboring node has surplus resources, it reserves the task and returns an acceptance message. If no neighboring node has surplus resources, it returns a rejection message. If both neighboring nodes have surplus resources, the task is sent to the neighboring node that sent the acceptance message first, and the reserved resources of the other neighboring node are released.
[0014] Neighborhood multi-hop: Select a neighboring node and send a multi-hop request to it to directly transfer the task ownership to the neighboring node. After the transfer is successful, the task will be scheduled by the new node to which it is transferred.
[0015] In the method of the present invention, the scheduling decision information needs to be encoded before participating in the multi-layer perceptron decision network training as supervisory data. Specifically, each node has n neighboring nodes, that is, there are n*2+1 scheduling decisions. Therefore, the encoding vector of the scheduling decision information of each node is an n*2+1 dimensional vector. In the n*2+1 dimensional vector, the vector value corresponding to the selected scheduling decision is 1, otherwise the vector value is 0.
[0016] In the method of the present invention, the scheduling result is evaluated according to the scheduling log record information to obtain an evaluation result including an excellent scheduling strategy or an inferior scheduling strategy, including:
[0017] If the scheduling result determined based on the scheduling log information meets one of the following conditions, it is evaluated as an excellent scheduling strategy; otherwise, it is evaluated as an inferior scheduling strategy:
[0018] Condition 1: The task is executed within two decision time units after the local wait action is selected;
[0019] Condition 2: The message is received by the node after the forwarding schedule is selected and the forwarding node is determined;
[0020] Condition 3: The task is executed within two decision time units after selecting the multi-hop action.
[0021] In the method of the present invention, the loss function of the training strategy network adopts the cross entropy loss function.
[0022] In the method of the present invention, each distributed node uses the trained scheduling strategy recommendation model to make scheduling decision recommendations based on scheduling status information, including:
[0023] The scheduling status information of each distributed node at each step is input into two decision networks respectively. After forward reasoning, two scheduling decision prediction probability vectors are obtained. The two scheduling decision prediction probability vectors are fused through a linear fusion layer to obtain a comprehensive scheduling decision probability vector. The comprehensive scheduling decision probability vector is activated by a softmax operation to select the recommended scheduling decision. The fusion coefficient in the linear fusion layer is set based on experience.
[0024] The method of the present invention also includes: regularly collecting scheduling status information and scheduling decision information when the scheduling strategy recommendation model makes scheduling decision recommendations and executes them, and using this information to retrain the scheduling decision recommendation model, and using the newly trained scheduling decision recommendation model to replace the original scheduling decision recommendation model, and executing subsequent scheduling decision recommendations until the long-tail task delay is reduced to meet the requirements.
[0025] Compared with the prior art, the present invention has the following beneficial effects:
[0026] Compared to centralized and other scheduling frameworks, the distributed scheduling architecture has an independent decision-making module for each node. This shifts scheduling decisions down to the cluster nodes, alleviating decision-making pressure and improving throughput. When nodes make decisions based on scheduling status information, using multi-layer perceptron-based scheduling can reduce task latency compared to naive heuristic algorithms, reducing the latency of long-tail tasks and effectively improving overall cluster throughput. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0028] Figure 1 is a flow chart of a distributed single-node scheduling method based on multi-layer perceptron decision-making provided by an embodiment;
[0029] Figure 2 is a schematic diagram of the structure of distributed nodes in a distributed scheduling system provided by an embodiment;
[0030] Figure 3 3 is a structural diagram of a scheduling decision recommendation model provided in an embodiment. DETAILED DESCRIPTION
[0031] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.
[0032] The inventive concept of the present invention is: the present invention provides a distributed single-node scheduling method based on multi-layer perceptron decision-making. The method is applied to a distributed scheduling system. Each distributed node abandons the naive heuristic algorithm and adopts a scheduling decision recommendation model constructed based on the multi-layer perceptron to make scheduling decisions. It can reduce task delays, reduce long-tail task delays, and effectively improve the overall cluster throughput.
[0033] like Figure 1 As shown, the distributed single-node scheduling method based on multi-layer perceptron decision-making provided in the embodiment includes the following steps:
[0034] S1, collects the scheduling status information and scheduling decision information of each distributed node in the distributed scheduling system at each step, and evaluates the scheduling results according to the scheduling log record information to obtain an evaluation result containing an optimal scheduling strategy or an inferior scheduling strategy.
[0035] like Figure 2 As shown, in a distributed node system, node A includes a task queue, neighborhood status information and a decision module. The scheduling method of the present invention is applied to the decision module in each distributed node in the distributed scheduling system, and the decision information of the present invention is based on the scheduling status information provided by the node. There are a fixed number of neighborhood nodes for each node in the architecture. The node can send information to the neighborhood node and record the neighborhood resource information within the node. The neighborhood nodes can be unidirectional neighbors or bidirectional neighbors to each other. The neighborhood network of the overall framework is a small-world network. The scheduling status information obtained is a record obtained when the scheduling node and the neighborhood node interact with each other, and is not the neighborhood real-time resource information. The neighborhood node is a node-schedulable interaction node. The scheduling node can choose to forward the task to any neighborhood node or perform multi-hop operations.
[0036] Specifically, the scheduling status information includes the current time T when the decision is made, the task submission time t c 、Resources required for the task res req , the remaining resources localres of this node when making a decision, the neighboring node information known to this node when making a decision, assuming there are n neighboring nodes (usually n=6 or 8), record the state update time of each neighboring node n , and the remaining resource information of the neighborhood when making decisions res n (Due to communication delays, this information is not accurate resource information of neighboring nodes).
[0037] At the same time, the scheduling decision information of each step must be collected. The scheduling decision information includes three types of scheduling decisions, namely:
[0038] Local waiting: Add the scheduled task to the waiting queue of the local node and wait for the next cycle of scheduling;
[0039] Neighborhood forwarding: Selects two neighboring nodes and sends task detection information to them. If the neighboring node has surplus resources, it reserves the task and returns an accept message. If no neighboring node has surplus resources, it returns a reject message. If both neighboring nodes have surplus resources, the task is sent to the neighboring node that sends the accept message first, and the reserved resources of the other neighboring node are released.
[0040] Neighborhood multi-hop: Select a neighboring node and send a multi-hop request to it to directly transfer the task ownership to the neighboring node. After the transfer is successful, the task will be scheduled by the new node to which it is transferred.
[0041] In the method of the present invention, the scheduling results are evaluated based on the scheduling log information to obtain an evaluation result including an optimal scheduling strategy or an inferior scheduling strategy. This evaluation result is used to divide the scheduling status information for input into different multilayer perceptron decision networks. Specifically, the evaluation process includes:
[0042] The scheduling result determined based on the scheduling log information is considered an optimal scheduling strategy if it meets one of the following conditions: i) the task is executed within two decision time units after the local wait action is selected; ii) the task is received by the node after the forwarding scheduling is selected and the forwarding node is determined; iii) the task is executed within two decision time units after the multi-hop action is selected. Otherwise, it is considered an inferior scheduling strategy.
[0043] S2. Construct a scheduling decision recommendation model that includes two multi-layer perceptron decision networks corresponding to the optimal scheduling strategy and the inferior scheduling strategy, and a linear fusion layer. Use the scheduling status information as sample data and divide it according to the evaluation results, then input it into the corresponding multi-layer perceptron decision network for scheduling decision prediction. Use the scheduling decision information as supervision data and construct a loss function with the scheduling decision prediction results to train the multi-layer perceptron decision network.
[0044] In the method of the present invention, two identical multilayer perceptron decision networks are constructed, corresponding to the optimal scheduling strategy and the inferior scheduling strategy respectively. The multilayer perceptron decision network comprises three layers, two hidden layers and one output layer, each of which is a fully connected layer. The input data first passes through the first fully connected layer fc1, corresponding to the following formula:
[0045] ;
[0046] Then, the activation function ReLU is applied to obtain the activation value of the first hidden layer:
[0047] ;
[0048] Next, the data passes through the dropout layer, randomly shutting down some neurons, and then passes through the second fully connected layer fc2, which corresponds to the following formula:
[0049] ;
[0050] Apply the activation function ReLU to get the activation value of the second hidden layer:
[0051] ;
[0052] The data passes through the dropout layer again, randomly shutting down some neurons, and then passes through three fully connected layers fc3, corresponding to the following formula:
[0053] ;
[0054] Get the output vector of the output layer .
[0055] The present invention constructs a scheduling decision recommendation model based on two constructed multi-layer perceptron decision networks, such as Figure 3 As shown, the scheduling decision recommendation model initially includes two multi-layer perceptron decision networks and a linear fusion layer. The fusion parameters of the linear fusion layer are set based on experience. The parameters of the two multi-layer perceptron decision networks need to be trained and optimized. Specifically, the scheduling status information is used as sample data and divided according to the evaluation results before being input into the corresponding multi-layer perceptron decision network for scheduling decision prediction. Specifically, the scheduling status information corresponding to the optimal scheduling strategy is input in batches into one multi-layer perceptron decision network corresponding to the optimal scheduling strategy, and the scheduling status information corresponding to the inferior scheduling strategy is input in batches into another multi-layer perceptron decision network corresponding to the inferior scheduling strategy. Through forward reasoning, the scheduling decision prediction results output by the two networks can be obtained.
[0056] When training two multi-layer perceptron decision networks, the scheduling decision information is also used as supervisory data. Before participating in the training of the multi-layer perceptron decision network, these supervisory data need to be encoded. Specifically, each node has n neighboring nodes, that is, there are n*2+1 scheduling decisions. Therefore, the encoding vector of the scheduling decision information of each node is an n*2+1-dimensional vector. In the n*2+1-dimensional vector, the vector value corresponding to the selected scheduling decision is 1, otherwise the vector value is 0.
[0057] For each multi-layer perceptron decision network, a loss function is constructed based on the encoding vector of the node's scheduling decision information and the scheduling decision prediction result. The loss function can adopt the cross-entropy loss function, and the network is trained by minimizing the cross-entropy loss so that the decision prediction values of the two networks are close to the optimal value and the worst value. After training, they can be used for subsequent scheduling decision predictions.
[0058] S3. Each distributed node uses the trained scheduling strategy recommendation model to make scheduling decision recommendations based on the scheduling status information, and uses the recommended scheduling decisions for scheduling.
[0059] In the method of the present invention, when the scheduling strategy recommendation model is used to make scheduling decision recommendations based on scheduling status information, the scheduling status information of each distributed node at each step is input into a multi-layer perceptron decision network corresponding to the optimal scheduling strategy and another multi-layer perceptron decision network corresponding to the inferior scheduling strategy, and two scheduling decision prediction probability vectors are obtained through forward reasoning. and , two scheduling decision prediction probability vectors and After fusion through the linear fusion layer, the comprehensive scheduling decision probability vector is obtained ,in, and is the fusion coefficient in the linear fusion layer. The fusion parameters corresponding to different tasks are different. For a large number of short-time tasks, the value of k1 / k2 is in the range of [-2, -1]. After the softmax activation operation, the recommended scheduling decision is selected.
[0060] In the method of the present invention, the scheduling strategy recommendation model is regularly collected for scheduling decision recommendation and scheduling status information and scheduling decision information during execution, and the scheduling decision recommendation model is retrained using this information. The newly trained scheduling decision recommendation model is used to replace the original scheduling decision recommendation model, and subsequent scheduling decision recommendations are executed until the long-tail task delay is reduced to meet the requirements. Such continuous iteration can optimize the model, improve the decision accuracy of the model, thereby reducing task delays, reducing long-tail task delays, and effectively improving the overall cluster throughput.
[0061] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A distributed single-node scheduling method based on multi-layer perceptron decision making, characterized in that: The following steps are involved: Collect the scheduling status information and scheduling decision information of each distributed node in the distributed scheduling system at each step, evaluate the scheduling results according to the scheduling log record information, and obtain the evaluation results including the best scheduling strategy or the worst scheduling strategy; A scheduling decision recommendation model is constructed, which includes two multilayer perceptron decision networks corresponding to the optimal scheduling strategy and the inferior scheduling strategy, and a linear fusion layer. The scheduling status information is used as sample data and divided according to the evaluation results. The data is then input into the corresponding multilayer perceptron decision network for scheduling decision prediction. The scheduling decision information is used as supervision data and a loss function is constructed with the scheduling decision prediction results to train the multilayer perceptron decision network. Each distributed node uses the trained scheduling policy recommendation model to make scheduling decision recommendations based on scheduling status information, and uses the recommended scheduling decisions for scheduling.
2. The distributed single-node scheduling method based on multi-layer perceptron decision-making according to claim 1 is characterized in that: Scheduling status information includes the current time T when the decision is made, the task submission time t c 、Resources required for the task res req , the remaining resources localres of this node when making a decision, the neighboring node information known to this node when making a decision, and the state update time of each neighboring node n , and the remaining resource information of the neighborhood when making decisions res n .
3. The distributed single-node scheduling method based on multi-layer perceptron decision-making according to claim 1 is characterized in that: Scheduling decision information includes three types of scheduling decisions: Local waiting: Add the scheduled task to the waiting queue of the local node and wait for the next cycle of scheduling; Neighborhood forwarding: Select two neighboring nodes and send task detection information to them. If the neighboring node has surplus resources, it reserves the task and returns an acceptance message. If no neighboring node has surplus resources, it returns a rejection message. If both neighboring nodes have surplus resources, the task is sent to the neighboring node that sent the acceptance message first, and the reserved resources of the other neighboring node are released. Neighborhood multi-hop: Select a neighboring node and send a multi-hop request to it to directly transfer the task ownership to the neighboring node. After the transfer is successful, the task will be scheduled by the new node to which it is transferred.
4. The distributed single-node scheduling method based on multi-layer perceptron decision making according to claim 3 is characterized in that: Before the scheduling decision information participates in the multi-layer perceptron decision network training as supervisory data, it needs to be encoded. Specifically, each node has n neighboring nodes, that is, there are n*2+1 scheduling decisions. Therefore, the encoding vector of the scheduling decision information of each node is an n*2+1 dimensional vector. In the n*2+1 dimensional vector, the vector value corresponding to the selected scheduling decision is 1, otherwise the vector value is 0.
5. The distributed single-node scheduling method based on multi-layer perceptron decision-making according to claim 1 is characterized in that: The scheduling results are evaluated based on the scheduling log information to obtain an evaluation result including an optimal scheduling strategy or an inferior scheduling strategy, including: If the scheduling result determined based on the scheduling log information meets one of the following conditions, it is evaluated as an excellent scheduling strategy; otherwise, it is evaluated as an inferior scheduling strategy: Condition 1: The task is executed within two decision time units after the local wait action is selected; Condition 2: The message is received by the node after the forwarding schedule is selected and the forwarding node is determined; Condition 3: The task is executed within two decision time units after selecting the multi-hop action.
6. The distributed single-node scheduling method based on multi-layer perceptron decision-making according to claim 1 is characterized in that: The loss function for training the policy network adopts the cross entropy loss function.
7. The distributed single-node scheduling method based on multi-layer perceptron decision-making according to claim 1 is characterized in that: Each distributed node uses the trained scheduling strategy recommendation model to make scheduling decision recommendations based on scheduling status information, including: The scheduling status information of each distributed node at each step is input into two decision networks respectively. After forward reasoning, two scheduling decision prediction probability vectors are obtained. The two scheduling decision prediction probability vectors are fused through a linear fusion layer to obtain a comprehensive scheduling decision probability vector. The comprehensive scheduling decision probability vector is activated by a softmax operation to select the recommended scheduling decision. The fusion coefficient in the linear fusion layer is set based on experience.
8. The distributed single-node scheduling method based on multi-layer perceptron decision-making according to claim 1 is characterized in that: Also includes: Regularly collect the scheduling status information and scheduling decision information during the scheduling decision recommendation and execution of the scheduling strategy recommendation model, and use this information to retrain the scheduling decision recommendation model, and use the newly trained scheduling decision recommendation model to replace the original scheduling decision recommendation model, and execute subsequent scheduling decision recommendations until the long-tail task delay is reduced to meet the requirements.