Distributed big data processing task scheduling method for non-peer communication

Abstract

The application discloses a distributed large data processing task scheduling method for non-peer communication. At the beginning of a processing round, a control node distributes the latest to-be-processed data and a model to the required computing nodes according to the current environmental conditions. After receiving the latest to-be-processed data and the model, the computing nodes execute multiple rounds of data processing according to the processing round information of the control node and send the processing results to the control node. The control node monitors the negative influence of the received processing results, discards the processing results of the computing nodes if there is a negative influence, and determines the aggregation weights of the processing results of the computing nodes which are not discarded. The control node aggregates the results processed by the different computing nodes according to the aggregation weights of the nodes, and then performs the next round of processing according to the current environmental conditions. The application can realize efficient distributed large data processing in a network environment with a large difference between uplink and downlink bandwidths, and improve the overall processing efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data processing and parallel computing, and particularly to a distributed big data processing task scheduling method for non-equivalent communication. BACKGROUND

[0002] Distributed big data processing refers to distributing data to multiple computing nodes or servers and using parallel processing technology to improve the efficiency and speed of data processing. With the rapid development of the Internet, the Internet of Things, social media and mobile devices, the global data volume is growing explosively. Traditional centralized data processing methods have been difficult to cope with such a large amount of data due to the limited computing power and storage capacity of a single server. In this case, distributed big data processing has emerged.

[0003] The basic idea of distributed processing is to decompose a big data task into several smaller tasks. By parallelizing these tasks, the processing time can be significantly shortened, and the storage and computing resources of multiple machines can be effectively utilized. For example, in order to meet the inference-type requests of computing users, a certain number of instances of each version of the inference-type model need to be deployed in each computing node with limited resources. These model instances have consistent functions and can concurrently process the requests of computing users. During data processing, the control node will continuously update the instances and parameters based on the processing results of each computing node and send the updated instances and parameters to these computing nodes. In order to reduce the delay between the control node and each computing node, it is more appropriate to deploy the control node near a large number of computing nodes than in a data center.

[0004] However, for a control node that needs to communicate with a large number of computing nodes simultaneously, the current symmetric communication method will limit the communication upper limit to a small uplink bandwidth, resulting in a small number of computing nodes that the control node can communicate with to send instances and requirements in each round, which not only wastes a large amount of computing node computing resource, but also wastes a large amount of downlink bandwidth.

[0005] Therefore, there is an urgent need for a new strategy to solve this problem and improve the utilization of control node downlink bandwidth and computing node computing resource, and improve the efficiency of distributed big data processing. SUMMARY

[0006] The present application aims to provide a distributed big data processing task scheduling method for non-equivalent communication, which combines asymmetric communication strategies and distributed big data processing methods to dynamically adjust task scheduling according to task requirements, control node conditions and network bandwidth conditions, and solves the problem of waste of control node downlink bandwidth and computing node computing resource in the current symmetric communication distributed big data processing task method.

[0007] In order to solve the above technical problems, the present application provides the following technical solutions: a distributed large data processing task scheduling method for non-peer communication, the scheduling method steps include:

[0008] At the beginning of the processing round, the control node distributes the latest to-be-processed data and model to the computing nodes required for the current processing round according to the current environmental conditions, so as to perform local calculation and model updating; the environmental conditions include but are not limited to the current to-be-processed demand, computing node condition and network bandwidth condition;

[0009] After the selected computing nodes of the current round receive the latest to-be-processed data and model, they perform multi-round data processing according to the processing round information of the control node; the control node and the computing nodes communicate with each other by using an asymmetric communication mechanism;

[0010] After the computing nodes complete each round of data processing, they send the processing results to the control node, and if the computing nodes do not receive new to-be-processed data and model after completing the round specified by the control node, they will continue to wait;

[0011] After the control node obtains the processing results of each round, it monitors the processing results transmitted by each computing node, and if it detects that the processing result transmitted by a computing node has a negative impact on the overall result, the processing result is discarded; wherein the consistency protocol, business logic verification and clustering based on machine learning are used to monitor the processing results transmitted by each computing node, and if the deviation of a result is greater than or equal to a preset threshold, it means that the processing result transmitted by the computing node has a negative impact on the overall result;

[0012] After the negative result is discarded, the control node determines the aggregation weight of the processing result of each computing node that is not discarded according to the constraint factors of the computing nodes, and aggregates the results processed by different computing nodes according to the aggregation weight of each node; the constraint factors include but are not limited to computing node data quality, time length from the last communication, computing node data volume and difference between the current round result of the computing node and the round when the latest processing model is obtained;

[0013] After aggregating the results processed by different computing nodes, the control node redecides the selected computing nodes of the next round according to the current environmental conditions.

[0014] According to the above technical solutions, the computing nodes are deployed in an asymmetric network with uplink bandwidth smaller than downlink bandwidth, used for multi-round data processing, and the results are returned to the control node after each round of processing;

[0015] The control node is deployed in a non-symmetrical network environment close to the computing nodes, and is responsible for coordinating and monitoring the entire computing network; the control node schedules different computing nodes according to the current to-be-processed demand, computing node condition and network bandwidth condition and other factors, and determines the selection of the next round. Meanwhile, after obtaining the processing result of each round, the control node determines the aggregation weight of the result of each computing node according to the data volume and communication delay of the computing node, and aggregates the result. Since the computing nodes are usually deployed in a non-symmetrical network with uplink bandwidth smaller than downlink bandwidth, and in order to reduce network delay, the control node is also usually deployed in a non-symmetrical network environment close to the computing nodes.

[0016] The non-symmetrical communication mechanism is that the control node transmits a processing model to the computing nodes, and the computing nodes transmit the processing result back to the control node multiple times. Since the control node is usually deployed in a non-symmetrical network environment with uplink bandwidth smaller than downlink bandwidth, for the control node which needs to communicate with a large number of computing nodes at the same time, the use of symmetrical communication will cause the communication upper limit of the control node to be limited by the small uplink bandwidth, thereby wasting a large amount of downlink bandwidth resources and computing capacity of the computing nodes. Therefore, the non-symmetrical communication mechanism is used to enable the control node to maximize the use of downlink bandwidth to download the processing result from multiple computing nodes under the condition of limited uplink bandwidth.

[0017] The traditional distributed processing method usually adopts single-round communication processing, that is, the computing nodes send the result back to the control node after each round of processing, and the control node adjusts the processing model and parameters according to the result, and then sends them to the computing nodes for the next round of processing. In the present application, based on the non-symmetrical communication strategy, the computing nodes perform multiple rounds of processing after receiving a processing model and data, and return the processing result to the control node multiple times. In this process, the computing nodes adaptively modify the model according to their own processing result, and each round of processing is based on the model modified in the last round.

[0018] According to the above technical solution, when determining the weight of the result from different computing nodes in aggregation, the data volume of the computing node and the difference between the current round result of the computing node and the round in which the latest processing model is obtained are considered. The round of data processing performed by each computing node may be different.

[0019] According to the above technical solution, at the beginning of the processing round, the control node first calculates the Gaussian weight of each node according to the network condition and computing capacity of the computing node, and then calculates the probability of being selected in the current round of processing according to the Gaussian weight. The selection probability p of the computing node i is i :

[0020]

[0021] In the formula, denotes the processing duration of the computing node i, denotes the communication duration of the computing node i, denotes the processing duration of the computing node j, denotes the communication duration of the computing node j, and σ denotes a scale parameter of a Gaussian weight controlling attenuation.

[0022] The greater the Gaussian weight of a node, the more beneficial the node is to the overall data processing (the shorter the sum of the processing and communication durations, the greater the weight), where the normalized Gaussian weight is used as the selection probability, and sampling is performed based on the probability to obtain the computing node selected in the round. The method of calculating the probability first and then sampling can avoid relying on a specific client and improve system robustness. The Gaussian weight first can avoid the probability difference between different clients being too large.

[0023] In the result of each round of processing, there can be data generated by the computing node based on the latest model, or there can be results based on an outdated model. Therefore, when the control node detects that the result transmitted by a certain computing node has a negative impact on the overall result, the control node will discard the data. Subsequently, the control node determines whether to immediately transmit the latest processing model to the computing node or temporarily stop the task of the node according to the computing node selection algorithm, so as to optimize the overall processing efficiency.

[0024] According to the above technical solution, when the results are aggregated in the rth round, the aggregation weight q r,i :

[0025]

[0026] In the formula, denotes the node data amount weight of the computing node with the serial number i in the rth round, denotes the node round difference weight of the computing node with the serial number i in the rth round, and k r denotes the number of computing nodes participating in the result aggregation in the rth round, n denotes the data amount, denotes the node data amount weight of the computing node with the serial number j in the rth round, denotes the node round difference weight of the computing node with the serial number j in the rth round.

[0027] The control node dynamically adjusts the aggregation weight of the processing result of each node according to the data amount, communication delay and other factors of each computing node. For a computing node with higher data quality and shorter communication interval with the control node, the control node will give a higher weight to enhance the influence of its result on the final aggregated result, so as to more effectively correct the global processing result.

[0028] According to the above technical solution, the node data amount weight

[0029]

[0030] In the formula, k r denotes the number of computing nodes participating in the result aggregation of the rth round, n denotes the data volume, n i denotes the data volume of the computing node i. j denotes the data volume of the computing node j.

[0031] According to the technical solution, the node round difference weight

[0032]

[0033] In the formula, Δr denotes the round difference between the round corresponding to the result of the current round of the computing node and the round in which the latest processing model is obtained, and σ denotes a scale parameter of the Gaussian weight for controlling the attenuation.

[0034] According to the technical solution, the number of processing rounds performed by the computing node is limited by both the downlink bandwidth of the control node and the resources of the computing node.

[0035] Since the resources of the computing node are limited, data processing can only be performed within the range of resources that can be borne by the computing node. The control node not only needs to select an algorithm through the computing node, and according to the processing capacity, network delay and data value of the computing node and other factors, to determine whether to select a certain computing node for processing, but also needs to further reasonably allocate the processing round of the node, to ensure that the task load of the node does not exceed the resource bearing range, so as to ensure the efficient operation of the overall system. That is, the control node will accumulate multiple processing instructions sent to the computing node, to ensure that the result received in each round does not exceed the upper limit of the downlink bandwidth. On the basis of the limitation, the computing node will perform the processing of the maximum number of rounds within the range of resources that can be borne by itself. In this way, it is ensured that the computing power resources of the computing node are fully utilized, and it is also ensured that the result will not be wasted due to bandwidth limitation.

[0036] Another embodiment, an electronic device, comprising a memory and a processor, the memory stores a computer program, the computer program executes any one of the technical solutions of the above-mentioned distributed large data processing task scheduling method for non-peer-to-peer communication when running on the processor.

[0037] Also includes another embodiment, a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by the processor to realize any one of the technical solutions of the above-mentioned distributed large data processing task scheduling method for non-peer-to-peer communication.

[0038] Compared with the prior art, the beneficial effects achieved by the present application are:

[0039] 1. The asymmetric communication mechanism proposed by the present application is suitable for the asymmetric scenario where the network uplink bandwidth of the control node is smaller than the downlink bandwidth, which fully considers the asymmetry of the network uplink and downlink bandwidth, especially in the case of limited uplink bandwidth, a scheme is designed to fully utilize the downlink bandwidth. Through this mechanism, the present application effectively avoids the waste of bandwidth resources, ensures that the data processing task can run in an efficient and stable manner, and at the same time maintains the low-cost advantage of the overall network environment. This design does not require additional hardware costs, significantly optimizes the data transmission efficiency, and provides a strong guarantee for the smooth completion of long-term big data tasks.

[0040] 2. In the present application, the computing node will perform multiple rounds of processing after receiving the processing model and data once, and return the processing results to the control node multiple times. Even if the computing node fails to receive the latest processing model, it can still perform data processing and inference based on historical models. Not only does this ensure efficient operation of the asymmetric communication mechanism, but it also fully utilizes the computing power of the computing node, avoiding resource idling. In addition, this mechanism significantly enhances the autonomy and robustness of the computing node, enabling it to efficiently perform tasks and maximize its utility even in complex network conditions.

[0041] 3. In the present application, the control node dynamically adjusts the aggregation weight of the processing results of each node based on factors such as data volume, communication delay, etc., and performs resource management and allocation through the computing node selection algorithm. The aggregation weight of the processing results can effectively integrate the intermediate results generated by each computing node in the case of multi-node parallel processing, ensuring the accuracy and consistency of the data and maximizing the processing results of the ideal model. The computing node selection algorithm balances the load of each node to avoid node overload or idling, thereby significantly improving the overall processing efficiency. This ensures that the computing power resources of the computing node are fully utilized, while avoiding the waste of result transmission caused by bandwidth limitations, optimizing the overall resource allocation and utilization of the system.

[0042] 4. By combining the asymmetric communication strategy and the distributed big data processing method, the system dynamically adjusts task scheduling according to task requirements, control node conditions and network bandwidth conditions, avoids waste of bandwidth and computing power resources, improves overall processing efficiency, and can achieve efficient distributed big data processing in a network environment with large differences in uplink and downlink bandwidth, maximizing system processing capacity. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0044] Figure 1It is a large data processing mode schematic diagram using the distributed large data processing task scheduling method for non-peer-to-peer communication of the application;

[0045] Figure 2 It is a traditional distributed large data processing mode schematic diagram. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work are within the protection scope of the application.

[0047] The application proposes an asymmetric communication mechanism as shown in Figure 1 , that is, after the control node transmits a processing model to the computing nodes once, the computing nodes will perform multiple rounds of processing after receiving the processing model and setting, and return the processing results to the control node multiple times. Figure 1 In the first round, the control node selects computing node 1 and computing node 2 according to the current environmental conditions and sends a model to them for processing, computing node 4 and computing node 5 process according to their local models, when computing node 1, computing node, computing node 4 and computing node 5 complete multiple rounds of data processing, they send the processing results and the model to the control node, the control node selects computing node 3 and computing node 4 according to the current environmental conditions and sends a model to them for processing, computing node 1 and computing node 2 process according to their local models, and computing node 5 stops processing because it has not received a model for a long time, when computing node 1, computing node, computing node 3 and computing node 4 complete multiple rounds of data processing, they send the processing results and the model to the control node, and the control node performs the third round of processing according to the current environmental conditions, and so on.

[0048] In the prior art, because the control node is often deployed in an asymmetric network environment where the uplink bandwidth is less than the downlink bandwidth, for a control node that needs to communicate with a large number of computing nodes at the same time, a symmetric communication as shown in Figure 2 is adopted, that is, the computing nodes send the results back to the control node after each round of processing, and the control node adjusts the processing model and parameters according to the results and sends them to the computing nodes for the next round of processing. Figure 2In the processing round, at the beginning of the processing round, the control node selects the computing node 1 and the computing node 2 according to the current environmental condition and sends the model to them for processing, and the other computing nodes are in an idle state, when the first round of computing node 1 and computing node 2 processing is completed, the processing result and the model are sent to the control node, the control node selects the computing node 3 and the computing node 4 according to the current environmental condition and sends the model to them for the second round of processing, and the other nodes are in an idle state, when the second round of computing node 3 and computing node 4 processing is completed, the processing result and the model are sent to the control node, the control node selects the computing node 1 and the computing node 5 according to the current environmental condition and sends the model to them for the third round of processing, when the third round of computing node 1 and computing node 5 processing is completed, the processing result is sent to the control node, the control node adjusts the selection of the computing node for the next round according to the processing result of the third round, and the cycle is repeated until the data processing is completed.

[0049] From the above Figure 1 and Figure 2 It can be seen that in the prior art, due to the upper limit of symmetric communication being limited by small uplink bandwidth, the number of computing nodes that can be communicated by the control node to send instances and requirements is small in each round, which not only wastes a large amount of computing node computing resource, but also wastes a large amount of downlink bandwidth. The asymmetric communication mechanism of the present application can ensure that the computing resource of the computing node is fully utilized, and also ensures that the result will not be wasted due to bandwidth limitation.

[0050] The present application provides a technical solution, a distributed large data processing task scheduling method for non-peer communication, which comprises the following steps:

[0051] S1, at the beginning of the processing round, the control node distributes the latest data to be processed and the model to the computing nodes required for the current processing according to the current environmental condition, so as to perform local calculation and model updating; the environmental condition includes but is not limited to the current processing requirement, the computing node condition and the network bandwidth condition.

[0052] That is, the computing node selection algorithm quantifies the processing capacity of the computing node itself by using the processing time length , quantifies the network delay by using the communication time length , and the total time length of the computing node to complete the processing in each round is , the Gaussian weight of the node is calculated, and then the selection probability of the node is calculated according to the Gaussian weight of the node; the selection probability of the computing node i is p i :

[0053]

[0054] In the formula, represents the processing time length of the computing node i, denotes the communication duration of the computing node i, denotes the processing duration of the computing node j, denotes the communication duration of the computing node j, and σ denotes a scale parameter of a Gaussian weight controlling attenuation.

[0055] The processing rounds performed by the computing nodes are limited by the downlink bandwidth of the control node and the resources of the computing nodes. In other words, the control node accumulates multiple rounds of processing instructions sent to the computing nodes, and ensures that the received results of each round do not exceed the upper limit of the downlink bandwidth.

[0056] S2, after receiving the latest data to be processed and the model, the selected computing node in the round performs multiple rounds of data processing according to the processing round information of the control node; the control node and the computing node communicate by using an asymmetric communication mechanism; wherein the rounds of data processing performed by different computing nodes can also be different.

[0057] S3, after each round of data processing is completed, the computing node sends the processing result to the control node, and if the computing node has not received new data to be processed and the model after processing the specified round of the control node, the computing node will continue to wait;

[0058] S4, after obtaining the processing result in each round, the control node first uses consistency protocols, business logic verification, and clustering based on machine learning to monitor the processing result transmitted by each computing node, and if it is detected that the processing result transmitted by a computing node is greater than a set threshold P, the processing result transmitted by the computing node is discarded;

[0059] S5, after discarding the negative result, the control node determines the aggregation weight of the processing result of each computing node which is not discarded according to the restriction factors of the computing node, and aggregates the results processed by different computing nodes according to the aggregation weight of each node; the restriction factors include but are not limited to the data quality of the computing node, the duration from the last communication, the data volume of the computing node, and the difference between the result of the computing node in the current round and the round in which the latest processing model is obtained;

[0060] That is, when the result aggregation is performed in the rth round, the aggregation weight q r,i :

[0061] In the formula, denotes the node data volume weight of the computing node number i in the rth round, denotes the node round difference weight of the computing node number i in the rth round, and k r denotes the number of computing nodes participating in the result aggregation in the rth round, n denotes the data volume, denotes the node data volume weight of the computing node number j in the rth round, denotes the node round difference weight of the computing node number j in the rth round.

[0062] wherein, the node data amount weight

[0063]

[0064] wherein, k r denotes the number of computing nodes participating in the result aggregation in the rth round, n denotes the data amount, n i denotes the data amount of the computing node i. j denotes the data amount of the computing node j.

[0065] the node round difference weight

[0066]

[0067] wherein, Δr denotes the round difference between the round corresponding to the result of the computing node in the current round and the round in which the latest processing model is obtained, and σ denotes a scale parameter of the Gaussian weight controlling the attenuation.

[0068] S6, after aggregating the results processed by different computing nodes, the control node re-determines the computing nodes selected in the next round according to the current environmental condition.

[0069] It should be noted that the relational terms such as first and second and the like are used merely to distinguish one entity or action from another, without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0070] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will still be able to make modifications to the technical solutions described in the foregoing embodiments or make equivalent replacements to some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A distributed big data processing task scheduling method for non-peer communication, characterized in that, The scheduling method comprises: At the beginning of a processing round, the control node distributes the latest to-be-processed data and model to the computing nodes required for the current processing round according to the current environmental conditions, so as to perform local calculation and model updating; the environmental conditions include the current to-be-processed demand, the computing node status and the network bandwidth condition; After the selected computing nodes of the current round receive the latest to-be-processed data and model, they perform multi-round data processing according to the processing round information of the control node, and the control node and the computing nodes communicate with each other by using an asymmetric communication mechanism; After each round of data processing is completed, the computing nodes send the processing results to the control node, and if the computing nodes do not receive the new to-be-processed data and model after processing the round specified by the control node, they continue to wait; After the control node obtains the processing results of each round, it monitors the processing results transmitted by each computing node, and if it is monitored that the processing results transmitted by a computing node have a negative impact on the overall results, the processing results transmitted by the computing node are discarded; After the negative results are discarded, the control node determines the aggregation weights of the processing results of the computing nodes which are not discarded according to the constraint factors of the computing nodes, and aggregates the results processed by different computing nodes according to the aggregation weights of the nodes; the constraint factors include the data quality of the computing nodes, the time length from the last communication, the data volume of the computing nodes and the difference between the current round and the round in which the latest processing model is obtained; After the results processed by different computing nodes are aggregated, the control node redecides the computing nodes selected for the next round according to the current environmental conditions.

2. The distributed big data processing task scheduling method for non-peer communication oriented communication according to claim 1, characterized in that, The computing nodes are deployed in an asymmetric network in which the uplink bandwidth is less than the downlink bandwidth, are used for multi-round data processing, and return the results to the control node after each round of processing; The control node is deployed in the asymmetric network environment close to the computing nodes, is responsible for coordinating and monitoring the entire computing network; The asymmetric communication mechanism is that the control node transmits a processing model to the computing nodes once, and the computing nodes transmit the processing results back to the control node multiple times.

3. The distributed big data processing task scheduling method for non-peer communication oriented communication according to claim 1, characterized in that, The rounds of data processing performed by each computing node can be different.

4. The distributed big data processing task scheduling method for non-peer communication oriented communication according to claim 1, characterized in that, At the beginning of a processing round, the control node first calculates the Gaussian weight of each node according to the network condition and computing capacity of the computing node, and then calculates the probability of each computing node being selected in the current processing round according to the Gaussian weight, the selected probability of the computing node i being p i : wherein denotes the processing duration of the computing node i, denotes the communication duration of the computing node i, denotes the processing duration of the computing node j, denotes the communication duration of the computing node j, σ denotes a scale parameter of the Gaussian weight controlling the attenuation, and i, j and k are natural numbers.

5. The distributed big data processing task scheduling method for non-peer communication oriented communication according to claim 1, characterized in that, The control node determines the aggregated weight q of the processing result not discarded by each computing node according to the constraint factor of the computing node i after the negative result is discarded in the rth round r,i : In the formula, represents the node data weight of the computing node with serial number i in the rth round, represents the node round difference weight of the computing node with serial number i in the rth round, k r represents the number of computing nodes participating in the result aggregation of the rth round, n represents the data weight, represents the node data weight of the computing node with serial number j in the rth round, represents the node round difference weight of the computing node with serial number j in the rth round.

6. The distributed big data processing task scheduling method for non-peer communication oriented communication according to claim 5, characterized in that, the node data volume weight is expressed by the following equation: In the formula, k r denotes the number of computing nodes participating in the result aggregation of the rth round, n denotes the data volume, n i denotes the data volume of the computing node i j denotes the data volume of the computing node j.

7. The distributed big data processing task scheduling method for non-peer communication oriented communication according to claim 5, characterized in that, The node round difference weight In the formula, Δr represents the difference between the round corresponding to the current round result of the computing node and the round in which the latest processing model is obtained, and σ represents the scale parameter of the Gaussian weight control attenuation.

8. The distributed big data processing task scheduling method for non-peer communication oriented communication according to claim 1, characterized in that, The number of processing rounds performed by the computing nodes is limited by the downlink bandwidth of the control node and the resources of the computing nodes.

9. An electronic device, comprising: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the distributed big data processing task scheduling method for non-peer communication in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the distributed big data processing task scheduling method for non-peer communication in any one of claims 1-8.

Images