Cross-machine-room scheduling optimization method based on dynamic water level value

Abstract

The invention discloses a cross-machine-room scheduling optimization method based on a dynamic water level value, which is suitable for a framework with a CPU (Central Processing Unit) as a main node and a GPU (Graphics Processing Unit) as a sub-node, and comprises the following steps of: 1, estimating the load of each sub-node in real time by adopting a weighted moving average method; step 2, based on the real-time load data of the weighted moving average in the step 1, calculating a corresponding task allocation adjustment amount through a PID controller so as to realize dynamic task allocation for each child node; step 3, using the communication frameworks of Netty, transmitting real-time load information to the main node by the child node, and performing global scheduling decision by the main node according to the collected information; and step 4, the main node adjusts a task allocation strategy in real time according to the collected load condition of each child node. According to the cross-machine-room scheduling optimization method based on the dynamic water level value, the load of each child node is estimated in real time by adopting a weighted moving average method, so that the hysteresis of load information acquisition is avoided.

Description

Technical Field

[0001] The present invention relates to a scheduling optimization method, and more particularly to a cross-machine room scheduling optimization method based on dynamic water level values. Background Art

[0002] In the current data center and cloud computing environment, scheduling computing power across computer rooms faces many challenges. Traditional resource scheduling methods may lead to low resource utilization, uneven request distribution, and even risks to data security and integrity. In addition, with the rapid development of cloud computing and big data applications, more effective ways are needed to manage and schedule resources distributed in different geographical locations to meet the needs of high-performance computing and real-time data processing.

[0003] There is a patent number CN202010271217.1 in the prior art, and the name is a method and device for cross-computer room task scheduling and task processing. The invention patent discloses that before the job object starts processing, the management node determines several target computing nodes based on the collected load information of each computing node, and completes the segmentation of the job object, and then sends the segmented job data and task scheduling information to the target computing node. However, the scheduling and task processing method in the above manner adopts a real-time collection method for the load information of the computing node, so this method has a certain lag in the collection of the load information of the computing node. Therefore, the above scheduling method has a certain lag and cannot meet the needs of high-performance computing and real-time data processing. Summary of the invention

[0004] In view of the deficiencies in the prior art, the object of the present invention is to provide a cross-machine room scheduling optimization method based on dynamic water level values ​​that can effectively meet the needs of high-performance computing and real-time data processing.

[0005] To achieve the above object, the present invention provides the following technical solution: a cross-machine room scheduling optimization method based on dynamic water level value, suitable for the architecture with CPU as the main node and GPU as the sub-node, characterized in that it includes the following steps:

[0006] Step 1: Use the weighted moving average method to estimate the load of each child node in real time to obtain an accurate reflection of the current load status;

[0007] Step 2: Based on the real-time load data of the weighted moving average in step 1, the corresponding task allocation adjustment amount is calculated through the PID controller to realize dynamic task allocation for each sub-node;

[0008] Step 3: Using Netty and other communication frameworks, the child nodes transmit real-time load information to the master node, and the master node makes global scheduling decisions based on the collected information;

[0009] Step 4: The master node adjusts the task allocation strategy in real time according to the collected load conditions of each sub-node to ensure load balance among nodes and optimize the overall resource utilization of the system.

[0010] As a further improvement of the present invention, the formula of the weighted moving average method in step 1 is as follows:

[0011] in

[0012] ·WMALoad i (t) is the weighted moving average load for i at time point t.

[0013] N is the size of the moving window, i.e. the number of data points in the time period considered.

[0014] ·ω j is the weight of the jth data point

[0015] Load i (j) is the load on i at time point j

[0016] The summation symbol ∑ represents the accumulation of all values ​​of j from t-N+1 to t. As a further improvement of the present invention, the formula for calculating the corresponding task allocation adjustment amount by the PID controller in step 2 is as follows:

[0017]

[0018] Where, (μ(t)) is the current error, K p ), (K i ), (K d ) are the proportional, integral, and derivative coefficients, respectively. This formula shows how the PID controller combines the current error (proportional P), the historical accumulation of the error (integral I), and the rate of change of the error (derivative D) to generate a control signal μ(t), which is used to adjust the behavior of the system to achieve the desired goal.

[0019] As a further improvement of the present invention, there is an initialization and parameter setting step before step one to set the window size and weight of the weighted moving average, and the parameters of the PID controller.

[0020] As a further improvement of the present invention, by periodically executing step one, each GPU node periodically calculates its own weighted moving average load, and transmits the real-time load information to the CPU master node through Netty communication.

[0021] As a further improvement of the present invention, the specific method of dynamic task allocation of each sub-node in step 2 is to allocate by sending the scheduling decision back to each GPU node.

[0022] The beneficial effect of the present invention is that through the setting of step one, the weighted moving average method can be used to estimate the load of each sub-node in real time. Compared with the real-time collection method in the prior art, this has the effect of pre-estimation and effectively avoids the lag caused by real-time collection. In addition, this method manages each sub-computer room according to the master-slave node architecture, and Netty is used as the communication framework of the master-slave node of the communication framework. Scheduling is based on the dynamic water level of the preloaded queue. The system can intelligently allocate tasks according to the actual processing capacity and load of each computer room to avoid the situation where some computer rooms are overloaded while other computer rooms are lightly loaded. This dynamic scheduling strategy helps to improve the overall computing power utilization and the balance of request distribution. Therefore, it has the following advantages:

[0023] Real-time and stability: Combining weighted moving average and PID controller, it can respond quickly to load changes while maintaining system stability.

[0024] Resource optimization: By dynamically adjusting task allocation, the computing resources of each node can be maximized to improve the overall processing capacity and efficiency of the system.

[0025] Communication efficiency: Utilize an efficient network communication framework to achieve real-time data transmission and information reporting between nodes, ensuring the timeliness and accuracy of scheduling decisions. DETAILED DESCRIPTION

[0026] The present invention will be further described in detail with reference to the given embodiments below.

[0027] The cross-machine room scheduling optimization method based on dynamic water level value of this embodiment realizes effective scheduling and management of heterogeneous resources through the architecture of CPU as the main node and GPU as the sub-node. The main features include:

[0028] Load estimation by weighted moving average method (WMA):

[0029] The weighted moving average method is used to estimate the load of each child node (GPU node) in real time to obtain an accurate reflection of the current load status.

[0030]

[0031] Among them, WMALoad i (t) represents the load of the (i)th GPU node at time (t), (ω j ) is the weighting coefficient, and (N) is the weighting window size.

[0032] PID controller task allocation adjustment:

[0033] Based on the real-time load data of weighted moving average, the corresponding task allocation adjustment is calculated by PID controller to realize dynamic task allocation to each sub-node.

[0034]

[0035] Among them, (e(t)) is the current error, K p , K i , K d are the proportional, integral and differential coefficients respectively.

[0036] Network communication and information reporting:

[0037] Using communication frameworks such as Netty, child nodes transmit real-time load information to the master node, and the master node makes global scheduling decisions based on the collected information.

[0038] Dynamic scheduling strategy:

[0039] The master node adjusts the task allocation strategy in real time based on the collected load conditions of each sub-node to ensure load balance among each node and optimize the overall resource utilization of the system.

[0040] The scheduling optimization method using this embodiment is as follows:

[0041] Initialization and parameter setting: At system startup, set the window size and weight of the weighted moving average, as well as the parameters of the PID controller.

[0042] Load estimation and information reporting: Each GPU node periodically calculates its own weighted moving average load and transmits real-time load information to the CPU master node through Netty communication.

[0043] Master node scheduling decision: The master node receives and analyzes the collected load information, uses the PID controller to calculate the corresponding task allocation adjustment, and sends the scheduling decision back to each GPU node.

[0044] Task allocation adjustment: The GPU node adjusts its own task allocation according to the scheduling instructions of the master node to ensure the stable operation and resource optimization of the overall system under dynamic load.

[0045] Through the above steps, the present invention effectively solves the problems of heterogeneous resource management and task allocation optimization in a multi-computer room resource scheduling system, and improves the overall performance and efficiency of the system.

[0046] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A cross-computer room scheduling optimization method based on dynamic water level value, applicable to the architecture with CPU as the main node and GPU as the sub-node, characterized by: The steps include: Step 1: Use the weighted moving average method to estimate the load of each child node in real time to obtain an accurate reflection of the current load status; Step 2: Based on the real-time load data of the weighted moving average in step 1, the corresponding task allocation adjustment amount is calculated through the PID controller to realize dynamic task allocation for each sub-node; Step 3: Using Netty and other communication frameworks, the child nodes transmit real-time load information to the master node, and the master node makes global scheduling decisions based on the collected information; Step 4: The master node adjusts the task allocation strategy in real time according to the collected load conditions of each sub-node to ensure load balance among nodes and optimize the overall resource utilization of the system.

2. The cross-machine room scheduling optimization method based on dynamic water level value according to claim 1 is characterized in that: The formula of the weighted moving average method in step 1 is as follows: Among them, (WMALoad i ( t ) ) represents the load of the (i)th GPU node at time (t), (ω j ) is the weighting coefficient, and (N) is the weighting window size.

3. The cross-machine room scheduling optimization method based on dynamic water level value according to claim 1 or 2 is characterized in that: The formula for calculating the corresponding task allocation adjustment amount by the PID controller in step 2 is as follows: Among them, (μ ( t ) ) is the current error, (K p ), (K i ), (K d ) are the proportional, integral and differential coefficients respectively.

4. The cross-machine room scheduling optimization method based on dynamic water level value according to claim 1 or 2 is characterized in that: The method also includes an initialization and parameter setting step before step 1 to set the window size and weight of the weighted moving average and the parameters of the PID controller.

5. The cross-machine room scheduling optimization method based on dynamic water level value according to claim 1 or 2 is characterized in that: By executing step 1 regularly, each GPU node periodically calculates its own weighted moving average load, and transmits the real-time load information to the CPU master node through Netty communication.

6. The cross-machine room scheduling optimization method based on dynamic water level value according to claim 1 or 2 is characterized in that: The specific method of dynamic task allocation of each sub-node in step 2 is to allocate by sending the scheduling decision back to each GPU node.

Images