Federal learning weak network communication optimization method based on QUIC protocol

By using QUIC protocol and FEC encoding technology to optimize tail packet protection in federated learning, the network switching and head-on blocking of TCP protocol in weak network environments is solved, efficient federated learning communication is achieved, and the transmission efficiency of mobile devices is improved and performance.

CN119030935APending Publication Date: 2024-11-26FUDAN UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411133570.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing TCP protocol has network switching problems, head of queue blocking and difficult to modify kernel protocol stacks in a weak network environment of mobile devices, resulting in low efficiency of federated learning communications and ineffective optimization of data transmission.

Method used

The federated learning communication solution PS-QUIC is rebuilt by using the QUIC protocol, combining FEC encoding technology to protect only the tail packets, optimize the communication transmission strategy, adapt to the characteristics of high-frequency communication, and reduce transmission delay and redundant data overhead.

Benefits of technology

In a weak network environment, the communication time is significantly shortened, the transmission efficiency is improved, the time loss of tail packet loss and retransmission is reduced, and the overall performance of federated learning communication is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119030935A_ABST
    Figure CN119030935A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of federated learning communication transmission, and particularly relates to a federated learning weak network communication optimization method based on a QUIC protocol. The method comprises the following steps: constructing a communication framework based on a QUIC protocol, and recording the communication framework as PS-QUIC; parameter transmission is carried out under a PS-QUIC framework; optimizing a tail packet in a federated learning communication transmission scene; omnibearing design and adaptation are carried out for node management, and on the basis of ensuring that the throughput is unchanged, the parameter transmission delay is reduced, the communication duration of federated learning in a weak network environment is reduced, and the communication efficiency is improved; and meanwhile, aiming at a high packet loss scene, a tail packet optimization strategy is designed, the last tail packet in the transmission process is protected by utilizing an FEC coding technology, and unnecessary extra retransmission timeout duration is reduced. Simulation experiment results show that in a weak network scene with high packet loss, the method realizes shorter total time duration.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of federated learning communication transmission technology, and specifically relates to a federated learning weak network communication optimization method based on the QUIC protocol. Background Art

[0002] Federated learning is a machine learning framework that allows multiple participants to collaboratively train a shared model, breaking down data silos and making comprehensive use of scattered data. While improving the generalization ability of the model, data providers do not need to store or exchange their own data in a centralized manner. Instead, they only share model parameters or gradient information, effectively reducing the risk of data leakage, thereby achieving effective use of data while protecting data privacy. Currently, with the upgrade of chips, mobile devices are able to participate in complex data processing and machine learning tasks. At the same time, mobile devices can generate and instantly access a large amount of real-time data, such as location data, health detection information, etc. Such information can be used by federated learning to update data in real time and optimize models, thereby providing more personalized and timely services. Therefore, federated learning is increasingly being deployed on mobile devices.

[0003] However, due to various limitations, the network environment of mobile devices is far less stable than the network transmission between data centers. In addition, most of the current federated learning communication transmission protocols use the TCP transmission protocol, which has the following pain points:

[0004] (1) The TCP protocol is embedded in the kernel and is difficult to modify. The network environment and hardware performance of mobile devices vary greatly, and the requirements for protocols are also different. The federated learning framework that supports computing and data transmission cannot involve different client kernels, so it is particularly difficult to modify the protocol stack. Therefore, federated learning that uses TCP as the communication protocol has certain disadvantages in this scenario.

[0005] (2) Network switching problem. Mobile devices often face switching of network environments, such as switching between cellular networks and WiFi networks. The TCP protocol uniquely identifies a connection through a four-tuple consisting of source address, source port, destination address, and destination port. When the network of a mobile device is switched, its IP information will change accordingly. At this time, the connection between the central server and the mobile device based on the previous four-tuple information will be disconnected, which will cause the mobile device to disconnect and recover, which will inevitably incur additional time.

[0006] (3) Head-of-line blocking, which is also an unavoidable problem caused by TCP in order to achieve reliability. TCP only knows the length of the payload carried by each data packet, but does not know what data these bytes represent. When a TCP packet is lost in the middle of a connection, the TCP protocol is not sure whether it can take out the data in the later data packets and deliver it to the upper layer. Therefore, it can only be stored in the buffer and wait for the retransmission of the lost data packet. This is a blocking problem caused by the TCP protocol itself and cannot be avoided. In a network environment with a high packet loss rate, the extra time consumption caused by this type of blocking will increase to a significant level.

[0007] At present, most optimizations in the field of federated learning communication transmission focus on reducing the communication frequency and the total amount of data, but there are relatively few optimizations for the federated learning communication network itself. Starting with network optimization, it is possible to further shorten the communication time and improve the efficiency of federated learning communication transmission while optimizing the data volume. Summary of the invention

[0008] The purpose of the present invention is to provide a QUIC protocol-based federated learning weak network communication optimization method with short communication time and high transmission efficiency.

[0009] The present invention combines the QUIC transmission protocol to reconstruct a communication scheme based on the QUIC protocol for the parameter server architecture PS-Lite of the existing federated learning system MXNet, which is recorded as PS-QUIC. On the basis of having characteristics such as reliable transmission, flow control, and congestion control of the TCP protocol, it is possible to easily modify the content of the protocol layer, so that the overall transmission strategy is more adapted to the high-frequency communication characteristics of federated learning in a weak network environment. At the same time, since some training tasks have the characteristics of high communication frequency and small amount of transmitted data, using FEC to protect all data will generate a great burden. For ordinary data packets, using FEC for redundant protection will not have obvious advantages compared with loss retransmission, but for a single tail packet, loss will not trigger the fast retransmission mechanism, and can only wait for timeout retransmission. In addition, with the characteristics of high communication frequency, a training task will generate a large number of tail packets, and the loss and retransmission of these tail packets will bring great time loss. Therefore, the present invention uses the FEC mechanism to protect only the tail packet on the basis of the PS-QUIC framework, and strikes a balance between data overhead and time loss.

[0010] The method for optimizing weak network communication of federated learning based on the QUIC protocol provided by the present invention comprises the following specific steps:

[0011] (1) Construct a communication framework based on the QUIC protocol, denoted as PS-QUIC;

[0012] In the PS-QUIC framework, the system of a single node is mainly composed of three core modules: global management, communication transmission, and data management; among them:

[0013] The global management module is mainly responsible for the basic management operations of the nodes, including three key sub-modules: node information maintenance, node routing maintenance, and node synchronization maintenance: (1) The node information maintenance sub-module, also known as the initialization module, is mainly responsible for completing the initialization of its own node information, including address information and ID information, etc., which are used for communication addressing and gradient data identification. At the same time, it stores the preset Scheduler node information for the first communication in the startup phase, so as to ensure that it can be correctly identified and scheduled in the system; (2) The node routing maintenance sub-module is mainly used in the multi-server structure to realize the routing function between the gradient data after the Worker is segmented to the corresponding Server; (3) The node synchronization maintenance sub-module uses the thread blocking and wake-up mechanism to realize the synchronization management of the global system through the information transmission module to ensure that all nodes run on the same stage.

[0014] The communication transmission module is mainly responsible for information transmission operations between nodes, which is also one of the main contents of the QUIC network protocol. It includes five sub-modules: node registration, connection establishment, heartbeat monitoring, information confirmation response, and node termination: (1) The node registration sub-module uses the Scheduler node information provided by the information maintenance module to complete the information exchange between nodes in the startup phase, providing a basis for subsequent data transmission. At the same time, it provides support for the rejoining of disconnected nodes; (2) The transport layer information receiving and sending submodule uses the QUIC protocol to complete the connection establishment and information receiving and sending between nodes, so that the communication transmission is reliable while adding features such as 0-RTT, protocol customization, and seamless connection migration; (3) The heartbeat monitoring submodule is mainly responsible for monitoring the survival status of non-Scheduler nodes through periodic heartbeat signals during system operation, and cooperates with the node registration module to complete the recovery of disconnected nodes, thereby ensuring the reliability and continuity of the system under abnormal conditions; (4) The information confirmation response submodule is an optional functional module. Its main function is to implement an information confirmation mechanism similar to the transport layer ACK at the application layer, providing further guarantee for the transmission and delivery of information; (5) The node termination submodule is responsible for the safe termination of the system and the correct release of memory in each thread.

[0015] The data management module is the core part of the training task and is mainly deployed between the Worker node and the Server node. It is used to process the training data, including four key sub-modules: data segmentation, sending and receiving, processing and response: (1) The data segmentation sub-module is for the gradient data after the Worker node completes local training in the multi-Server architecture. In PS-QUIC, the data transmitted between the Worker and Server nodes is in the form of $(Key, Value)$. When there are multiple Servers, the Worker node will divide the data into multiple $(Key, Value)$ groups according to the number of Servers and push them to different Servers. This improves the system's fault tolerance and communication efficiency. (2) The data transceiver submodule is aimed at the information transmission and reception operations at the application level. The bottom layer still needs to use the communication transmission module to realize the information transmission and reception at the transmission layer. (3) The data processing submodule is used to process the information from the transceiver module. For the Server, it is mainly combined with the gradient information pushed by the Worker to update the global model. For the Worker, it is combined with the global model pulled from the Server for the next step of calculation. (4) The data response submodule is deployed on the Server node. For the push or pull request initiated by the Worker, it calls the pre-set processing function for processing.

[0016] (II) Parameter transmission of the communication framework based on the QUIC protocol, specific process:

[0017] (1) Each node is initialized based on the preset environment information. This process includes initializing address information, establishing processing and response functions, mapping physical IDs and role IDs, and establishing routing in the case of multiple servers. After that, the non-Scheduler node establishes a connection to the Scheduler node through the main control thread and sends a node registration request;

[0018] (2) The Scheduler node identifies the data packets of different nodes through the receiving thread, completes the connection establishment, and starts the node counting. When the preset total number of non-Scheduler nodes is reached, the registration reply packet containing its own node information and all other node information is organized and sent to each node in turn;

[0019] (3) After receiving the registration reply packet, the non-Scheduler node obtains all node information and establishes connections with nodes with different roles in turn. It then enters the synchronous blocking state and waits for other nodes to complete their tasks.

[0020] (4) The non-Scheduler starts the heartbeat thread and periodically reports its online information to the Scheduler node. When a node is offline for various reasons, it needs to be reinitialized and initiate a new node registration request to the Scheduler. At this time, the Scheduler will assign an ID to the newly registered node according to the offline node queue;

[0021] (5) The Worker node and the Server node start the formal training task, and both parties complete the push, pull, and update of their respective models. When the training is completed, each node establishes a connection with itself and sends a Terminate type message, thereby safely terminating each thread opened in the node and correctly releasing the memory.

[0022] (III) Tail Packet Optimization in Federated Learning Communication Transmission Scenario

[0023] Using FEC [1] (Forward Erasure Correction) encoding technology only encodes and protects the tail packets generated by the two types of communications, namely, the upload of gradient data in each round of federated learning and the download of the global model. The amount of additional data generated is only the size of a single data packet, which is negligible. At the same time, the tail packet can be restored at the receiving end to avoid retransmission. After a higher round of communication, the optimization of tail packet loss can generate certain benefits, as follows:

[0024] The gradient data from MXNet is serialized into a byte stream in each round of Push or Pull communication for easy network transmission. At this time, in QUIC, the corresponding data packets are organized for transmission according to the set MTU size. When the remaining data is less than the MTU, it indicates that the current data is the tail packet of the stream that needs to be transmitted this time. At this time, the corresponding data packet is sent to the selected FEC algorithm for encoding, and the obtained original data packet and redundant data packet are sent to the receiving end. If the tail packet is lost during the transmission process, the receiving end will be able to calculate the lost data packet based on the redundant packet, and integrate all the data and deliver it to the upper-layer MXNet for subsequent process advancement without waiting for the retransmission timer to time out.

[0025] The specific process is:

[0026] (1) The federated learning framework MXNet pushes the gradient data that needs to be transmitted in the current round;

[0027] (2) The PS-QUIC framework serializes the gradient data into a byte stream, packages the byte stream according to the set MTU size, and filters out the tail packet of the current round;

[0028] (3) PS-QUIC calls the FEC module algorithm RLC [2], and XOR [3] The algorithm encodes and protects the current tail packet, and the original data packet and the generated redundant data packet are added to the message transmission queue;

[0029] (4) The receiving end receives all message data. If no tail packet is lost, the data is delivered to the upper layer normally.

[0030] (5) If the receiver detects that the tail packet is lost, it uses the redundant data packet to perform FEC decoding to recover the original lost data and adds a flag bit to the transmitted ACK to indicate that the tail packet is lost but has been recovered. This prevents the retransmission of the tail packet while guiding the sender's congestion control algorithm to adjust the rate.

[0031] The parameter transmission framework PS-QUIC based on the QUIC transmission protocol proposed in this invention can realize the transmission of gradient data between nodes using the QUIC communication protocol. At the same time, a new node management logic is designed to make the communication transmission between nodes more adaptable to the QUIC protocol, and to reduce the transmission delay without affecting the throughput. At the same time, the FEC tail packet optimization strategy can be used to reduce the transmission communication time and improve the transmission efficiency in the weak network scenario with high packet loss. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is the overall structure diagram of the PS-QUIC parameter transmission system.

[0033] Figure 2 A schematic diagram of the workflow.

[0034] Figure 3 Optimize strategy graph for tail packet loss.

[0035] Figure 4 , 5 , 6, and 7 are schematic diagrams of simulation results of the method of the present invention and the framework PS-Lite based on the TCP protocol under network environments with different links. The experimental related data are all the results after 20 rounds of communication.

[0036] Figure 8 This is a comparison chart of simulation results of the parameter transmission framework PS-QUIC based on the QUIC protocol and the framework PS-Lite based on the TCP protocol under different packet loss network environments according to the method of the present invention.

[0037] Fig. 9 This is a comparison chart of simulation results of the FEC tail packet optimization strategy of the method of the present invention compared with the case where no encoding mechanism is added. DETAILED DESCRIPTION

[0038] The weak network communication optimization method for federated learning based on the QUIC protocol has the following specific steps:

[0039] Step 1: The federated learning framework calls the PS-QUIC method of the present invention to start building a communication environment between different nodes, including node registration and management, QUIC connection establishment and management, information queue establishment and management, etc. This step relies on the mutual coordination of the three modules of PS-QUIC global management, communication transmission and data management;

[0040] Step 2: After the communication environment is established, the MXNet framework uses normal training of federated learning tasks. When gradient data that needs to be globally synchronized and updated is generated, PS-QUIC is called to transmit data between nodes. PS-QUIC first searches whether there is a connection context with the target node, then serializes the gradient data, sets the format of the data packet according to the MTU size, and performs FEC encoding protection on the tail packet. Finally, the data packet is pushed to the message queue and waits for sending.

[0041] Step 3: The present invention also makes certain modifications to the underlying QUIC protocol, changing the transmission data of different rounds of the upper layer into different streams for transmission, and the termination of the stream depends on the ACK information of the tail packet;

[0042] Step 4: The target node continues to listen to the port specified during initialization. When receiving data, it first decodes the packet header data, obtains the connection ID information carried in the current data packet header, and uses the ID information to index the correct inter-node QUIC communication context, thereby correctly delivering the data packet data;

[0043] Step 5: The target node simultaneously detects all message data received by the receiving end. If there is no tail packet loss, the data is delivered to the upper layer normally. If the tail packet loss is detected, the redundant data packet is used for FEC decoding to recover the original lost data, and a flag is added to the transmitted ACK to indicate that the tail packet is lost but recovered, thereby preventing the retransmission of the tail packet while guiding the congestion control algorithm of the sender to adjust the rate.

[0044] Step 6: After the training task is completed, the PS-QUIC system sends a termination message to all currently connected nodes, releases the thread locks of all nodes, correctly releases the memory and shuts down the system;

[0045] Simulation results:

[0046] In a high-quality link with a bandwidth of 200 Mbps, a latency of 5 ms, and a packet loss setting of 0.001%, Figure 4 As shown in the figure, for the same transmission system, the overall network performance is basically the same when training with VGG16, VGG19 and AlexNet models. Figure 4-a shows that the average throughput of PS-QUIC is lower than that of PS-Lite system, close to 98.30%. This is due to the combined influence of multiple factors, one of which is that the QUIC protocol adds more detailed header information to the data packet during design, which increases some data overhead. At the same time, there are some changes in congestion control, which reduces some aggressiveness compared to TCP. In terms of average latency performance, PS-QUIC shows certain advantages, such as Figure 4 -b shows that the latency is reduced by 7.31% compared with the PS-Lite system. Therefore, in this type of scenario, the performance of the PS-QUIC system meets the design expectations.

[0047] Figure 5 The average throughput and average latency of daily links with a bandwidth limit of 50Mbps, a latency of 25ms, and a packet loss of 0.001% are shown. In this type of network, the overall performance of the PS-QUIC transmission system and the PS-Lite transmission system is similar to that of high-quality links. In terms of average throughput, PS-QUIC has a certain degree of reduction, but has achieved a small improvement in average latency performance.

[0048] Figure 6 The average throughput and average latency of a long-distance link with a bandwidth limit of 20Mbps, a latency setting of 120ms, and a packet loss rate of 0.01% are displayed. Compared with the first two scenarios, the long-distance link not only has an increase in latency, but also an increase in packet loss rate from 0.001% to 0.01%. Since the congestion control protocol set by the communication protocol is unified as CUBIC, when the link packet loss rate increases, the CUBIC algorithm will frequently reduce the congestion window, thereby reducing the sending rate, thereby alleviating the congestion indicated by the packet loss signal. Figure 6 -a indicates that the bandwidth utilization of PS-QUIC and PS-Lite transmission systems has decreased. On this basis, PS-QUIC still has some throughput deficiencies compared to PS-Lite, but it is also because of this that the average latency performance has been improved by nearly 8.2%.

[0049] When the network environment changes to a weak link with a bandwidth of 5Mbps, a delay of 80ms, and a packet loss rate of 1%, Figure 7As shown in the figure, the PS-QUIC transmission system has shown a positive improvement in average throughput compared to the PS-Lite system, with an increase of about 1.74%, and the average delay still maintains a positive improvement. In a weak network environment, when the TCP protocol has a timeout retransmission, due to the design of the protocol, the sequence number of the original message sent by the sender and the retransmitted message retransmitted due to the timeout are the same, so the confirmation ACK returned by the receiver for the data packet is also the same. At this time, the sender will not be able to distinguish whether the received ACK belongs to the original message or the retransmitted message, which will affect the RTT estimation, and then affect the packet sending rate of the sender, and ultimately make the bandwidth unable to be effectively utilized. The QUIC protocol adopts a strictly increasing packet sequence number, which avoids the retransmission ambiguity problem similar to TCPACK, allowing it to show certain advantages in such scenarios.

[0050] In terms of average throughput, Figure 8 -a shows that with the gradual increase of packet loss rate, both PS-QUIC and PS-Lite transmission systems show a downward trend. When the packet loss rate is 0.01%, the bandwidth utilization rate drops to about 30%. When the packet loss rate increases to 5%, the average throughput is only about 3% of the original. Among them, when the packet loss rate is 0.1%, the average throughput of PS-QUIC begins to improve relative to PS-Lite, and the improvement gradually increases with the increase of packet loss rate. Figure 8 -b shows the average latency performance of the two transmission systems in weak network environments with different packet loss rates. In this scenario, due to the relatively large packet loss rate, the bandwidth utilization controlled by the CUBIC algorithm cannot be effectively improved, so the backlog of packets in the link is relatively small, and the overall latency performance is consistent, all concentrated near the fixed latency of the link. At the same time, as the packet loss rate gradually increases, the latency gap between PS-QUIC and PS-Lite is gradually narrowed.

[0051] Fig. 9 The figure shows the completion time statistics of three rounds of communication under five protection measures: no coding protection, full protection using XOR coding algorithm (XOR-ALL), tail packet protection using XOR coding algorithm (XOR-Tail), full protection using RLC coding algorithm (RLC-ALL), and tail packet protection using RLC coding algorithm (RLC-Tail) under different packet loss rates.

[0052] As can be seen from the figure, as the packet loss rate increases, the completion time gradually increases due to the decrease in bandwidth utilization. At the same time, compared with the baseline uncoded protection, for the full-process protection measure, the newly added additional data packets do recover some of the lost data packets, but a large number of redundant data packets also greatly increase the network load and increase the completion time. For the tail packet protection measure, since only the last data packet needs to be encoded, the added redundant data has almost no additional impact on the link traffic. As the packet loss rate increases, the probability of tail packet loss gradually increases, and the advantage of tail packet protection also gradually increases. Compared with the normal transmission scenario without any protection measures, the transmission time of tail packet protection is slightly reduced.

[0053] References:

[0054] [1]SHANNON C EA mathematical theory of communication[J].Bell SystemTech-nicalJournal,1948,27(3):379-423.

[0055] [2]FRAGOULI C,LE BOUDEC JY,WIDMER J.Network coding:an instant primer[J].ACMSIGCOMM Computer Communication Review,2006,36(1):63-68.

[0056] [3]LI A.Rfc 5109:Rtp payload format for generic forward errorcorrection[M].RFCEditor,2007.

Claims

1. A method for optimizing weak network communication in federated learning based on the QUIC protocol, characterized in that: For the parameter server architecture PS-Lite of the existing federated learning system MXNet, a communication framework based on the QUIC protocol is rebuilt, denoted as PS-QUIC; on the basis of the reliable transmission, flow control, and congestion control characteristics of the TCP protocol, the protocol layer content is easily modified to make the overall transmission strategy more suitable for the high-frequency communication characteristics of federated learning in a weak network environment; and on the basis of the PS-QUIC framework, the FEC mechanism is used to protect only the tail packet, achieving a balance between data overhead and time loss; the specific steps are as follows: (1) Construct a communication framework based on the QUIC protocol, denoted as PS-QUIC; In the PS-QUIC framework, the system of a single node includes three modules: global management, communication transmission, and data management; among them: The global management module is responsible for the basic management operations of the nodes, including three sub-modules: node information maintenance, node routing maintenance, and node synchronization maintenance: (1) The node information maintenance sub-module, also known as the initialization module, is used to initialize its own node information, including the initialization of address information and ID information, for communication addressing and gradient data identification, and stores the preset Scheduler node information for the first communication in the startup phase, so as to ensure that it can be correctly identified and scheduled in the system; (2) The node routing maintenance sub-module is used in a multi-server structure to realize the routing function between the gradient data after the Worker is segmented and the corresponding Server; (3) The node synchronization maintenance sub-module uses the thread blocking and wake-up mechanism to realize the synchronization management of the global system through the communication transmission module to ensure that all nodes run on the same stage; The communication transmission module is responsible for the information transmission operation between nodes, and includes five submodules: node registration, transport layer information transmission and reception, heartbeat monitoring, information confirmation response, and node termination: (1) The node registration submodule uses the Scheduler node information provided by the information maintenance module to complete the information exchange between nodes in the startup phase, providing a basis for subsequent data transmission; at the same time, it provides support for the rejoining of disconnected nodes; (2) The transport layer information transmission and reception submodule uses the QUIC protocol to complete the connection establishment and information transmission between nodes, so that the communication transmission has reliability while increasing 0-RTT, and the protocol is self- Define and connect seamless migration features; (3) The heartbeat monitoring submodule is responsible for monitoring the survival status of non-Scheduler nodes through periodic heartbeat signals during system operation, and cooperates with the node registration module to complete the recovery of offline nodes, providing guarantee for the reliability and continuity of the system under abnormal conditions; (4) The information confirmation response submodule is an optional functional module, whose function is to implement an information confirmation mechanism similar to the transport layer ACK at the application layer, providing further guarantee for the transmission and delivery of information; (5) The node termination submodule is responsible for the safe termination of the system and the correct release of memory in each thread; The data management module is the core part of the training task and is mainly deployed between the Worker node and the Server node. It is used to process the training data, including four sub-modules: data segmentation, sending and receiving, processing and response: (1) The data segmentation sub-module is for the gradient data after the Worker node completes local training in a multi-Server architecture. In PS-QUIC, the data transmitted between the Worker and Server nodes exists in the form of $(Key, Value)$. When there are multiple Servers, the Worker node divides the data into multiple $(Key, Value)$ groups according to the number of Servers and pushes them. The submodule is used to send the information to different servers to improve the system's fault tolerance and communication efficiency; (2) The data transceiver submodule is used for the information sending and receiving operations at the application level. The bottom layer still needs to rely on the communication transmission module to realize the information sending and receiving at the transmission layer; (3) The data processing submodule is used to process the information from the transceiver module. For the server, it is to update the global model in combination with the gradient information pushed by the worker. For the worker, it is to perform the next step of calculation in combination with the global model pulled from the server; (4) The data response submodule is deployed on the server node. In response to the push or pull request initiated by the worker, the pre-set processing function is called for processing; (II) Parameter transmission of the communication framework based on the QUIC protocol, specific process: (1) Each node is initialized based on the preset environment information, including initialization of address information, establishment of processing and response functions, mapping of physical ID and role ID, and establishment of routing in the case of multiple servers. After that, the non-Scheduler node establishes a connection with the Scheduler node through the main control thread and sends a node registration request. (2) The Scheduler node identifies the data packets of different nodes through the receiving thread, completes the connection establishment, and starts the node counting. When the preset total number of non-Scheduler nodes is reached, it organizes the registration reply packet information containing its own node information and all other node information, and sends it to each node in turn; (3) After receiving the registration reply packet, the non-Scheduler node obtains all node information in it and establishes connections with nodes with different roles in turn; and enters the synchronous blocking state, waiting for other nodes to complete their tasks; (4) The non-Scheduler starts the heartbeat thread and periodically reports its online information to the Scheduler node. When a node goes offline for various reasons, it reinitializes and initiates a new node registration request to the Scheduler. At this time, the Scheduler will assign an ID to the newly registered node according to the offline node queue. (5) The Worker node and the Server node begin the formal training task, and both parties complete the push, pull, and update of their respective models. When the training is completed, each node establishes a connection with itself and sends a Terminate type message, thereby safely terminating each thread opened in the node and correctly releasing the memory. (III) Tail Packet Optimization in Federated Learning Communication Transmission Scenario By using FEC coding technology, only the tail packets generated by the two types of communications, namely, the upload of gradient data in each round of federated learning and the download of the global model, are coded and protected. After a higher round of communication, the optimization of tail packet loss generates certain benefits; the details are as follows: The gradient data from MXNet is serialized into a byte stream in each round of Push or Pull communication for easy network transmission. At this time, in QUIC, the corresponding data packets are organized for transmission according to the set MTU size. When the remaining data is less than the MTU, it indicates that the current data is the tail packet of the stream that needs to be transmitted this time. At this time, the corresponding data packet is sent to the selected FEC algorithm for encoding, and the obtained original data packet and redundant data packet are sent to the receiving end. If the tail packet is lost during the transmission process, the receiving end will be able to calculate the lost data packet based on the redundant packet, and integrate all the data and deliver it to the upper-layer MXNet for subsequent process advancement without waiting for the retransmission timer to time out.

2. According to the QUIC protocol-based federated learning weak network communication optimization method of claim 1, it is characterized in that: The specific process of tail packet optimization in the federated learning communication transmission scenario is as follows: (1) The federated learning framework MXNet pushes the gradient data that needs to be transmitted in the current round; (2) The PS-QUIC framework serializes the gradient data into a byte stream, packages the byte stream according to the set MTU size, and filters out the tail packet of the current round; (3) PS-QUIC calls the FEC module algorithm RLC [2] , and XOR [3] The algorithm encodes and protects the current tail packet, and the original data packet and the generated redundant data packet are added to the message transmission queue; (4) The receiving end receives all message data. If no tail packet is lost, the data is delivered to the upper layer normally. (5) If the receiver detects that the tail packet is lost, it uses the redundant data packet to perform FEC decoding to recover the original lost data and adds a flag bit to the transmitted ACK to indicate that the tail packet is lost but has been recovered. This prevents the retransmission of the tail packet while guiding the sender's congestion control algorithm to adjust the rate.