A cloud data center-oriented communication overhead optimization method
By constructing a federated learning architecture in a cloud data center and adopting a full-duplex communication transmission protocol and ternary vectorized model parameters, the problems of high communication overhead and privacy leakage in federated learning are solved, thereby improving communication efficiency and protecting privacy.
Patent Information
- Application Number
- CN202310283873.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-22
- Publication Date
- 2025-12-09
- Estimated Expiration
- 2043-03-22
AI Technical Summary
Existing cloud data centers, after introducing federated learning architecture, suffer from high interactive communication overhead, which affects communication bandwidth resources, and user privacy information may be obtained by untrustworthy end users or intermediate operators, leading to privacy leaks.
A federated learning architecture based on a global cloud data center, a local data center, and end users is constructed. A full-duplex communication transmission protocol is adopted. Through the communication transmission of adaptation functions and ternary vectorized model parameters, unselected end users send the corresponding communication transmission protocol to the communication transmission protocol in the form of ternary vectors. Unselected end users transmit to the communication protocol in the form of ternary vectors. The communication is carried out using the TCP/IP protocol. During the global model update process, the optimal local model and ternary vectorized model parameters are combined for synchronous iterative updates to reduce communication bandwidth consumption.
It effectively reduces the interaction and communication overhead of the federated learning architecture, protects the privacy of end users, shortens the global iteration time, and improves communication efficiency.
Smart Images
Figure CN116346821B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of big data and cloud computing, and particularly relates to a communication overhead optimization method for a cloud data center. BACKGROUND
[0002] In recent years, due to the large-scale development of cloud computing and big data technology, the amount of data obtained, processed and stored also increases exponentially, which makes the demand for data centers increase continuously, and the establishment of a new data center and the overall optimization of the data center need to consider the problems of a large amount of resources and costs consumed by data storage, processing and management services. Research shows that a large number of enterprise users choose cloud data center servers to process business requirements, so as to reduce the resource overhead of data centers in maintenance and management. Cloud data centers have become one of the key participants in emergency demand response due to their flexibility in demand response and management, but in the communication transmission process of the existing demand response mechanism, the privacy information of users may be obtained by other terminal users or intermediate operators who are not worthy of trust, causing serious privacy leakage problems. In order to solve the privacy security concerns of users participating in demand response, the present application introduces a federated learning architecture to process the problem under the cloud data center scenario, and proposes a federated learning architecture for a cloud data center. However, a large amount of communication overhead will be generated in the multi-round interactive communication of the federated learning model, causing a serious communication burden problem to the communication bandwidth resources. SUMMARY
[0003] The present application aims to provide a communication overhead optimization method for a cloud data center, which aims to protect the transmission communication privacy between terminal users and data centers, and effectively solve the problem of high interactive communication overhead generated by introducing a federated learning architecture.
[0004] To achieve the above-mentioned purpose, the present application provides a communication overhead optimization method for a cloud data center, comprising the following steps:
[0005] A federated learning architecture for a cloud data center is constructed based on three nodes of a global cloud data center, a local data center and a terminal user;
[0006] A communication transmission protocol is established in the federated learning architecture;
[0007] The last round of global model parameters are initialized in the global cloud data center, and the initialized global model parameters are distributed to each terminal user;
[0008] Each terminal user obtains a local model and an evaluation cost based on the initialized global model parameters using its own local data set, and sends the evaluation cost to the local data center through the communication transmission protocol;
[0009] The local data center quantifies the effects of the local models trained by different end user data sets through an adaptation function based on the evaluation cost, thereby selecting the optimal local model and the corresponding end user;
[0010] The unselected end users send the corresponding ternary vectorization model parameters and evolution direction to the local data center in the form of ternary vectors;
[0011] The global cloud data center updates the global model in combination with the optimal local model and the ternary vectorization model parameters and evolution direction; during the global model updating process, the synchronization parameters are modified to adjust the synchronous iterative updating of the optimal local model and other ternary vectorization model parameters, to obtain the final global model, and the final global model is used as the initialized global model for the next round of training.
[0012] The global cloud data center is configured to aggregate model parameter instances and send the initialized global model parameters.
[0013] The local data center is configured to provide data resources and management services.
[0014] The end user is configured to participate in local model training according to local data conditions.
[0015] The federated learning architecture is configured to monitor the communication transmission conditions between the global cloud data center, the local data center, and the end user.
[0016] The communication transmission protocol is built on the TCP / IP protocol, and the communication transmission protocol adopts a full-duplex communication transmission mode for message transmission between the local data center and the end user.
[0017] Each of the end users obtains a local model and an evaluation cost based on the initialized global model parameters using the local data set trained by itself, and sends the evaluation cost to the local data center through the communication transmission protocol, including:
[0018] Each of the end users completes training based on the initialized global model parameters using the local data set trained by itself, and judges using the local data set to obtain a local model and an evaluation cost;
[0019] The evaluation cost is sent to the local data center through the communication transmission protocol.
[0020] The local data center quantifies the effects of the local models trained by different end user data sets through an adaptation function based on the evaluation cost, thereby selecting the optimal local model and the corresponding end user, including:
[0021] The local data center quantifies the effect of the local model trained by different end user data sets based on the evaluation cost through an adaptation function, and the product of the weight of the local data set of any end user in the data size of the end user set participating in the training of this round and the evaluation cost is taken as the evaluation standard of the training effect of the local model, so as to select the optimal local model and the corresponding end user.
[0022] The unselected end user sends the corresponding ternary vector model parameters and evolution direction to the local data center in the form of a ternary vector, including:
[0023] The unselected end user will not send the entire local model instance to the local data center for aggregation, but will only send the evolution direction of the corresponding model parameters in the form of a ternary vector by means of the communication transmission protocol.
[0024] The global cloud data center combines the optimal local model and the ternary vector model parameters and evolution direction to update the global model; in the global model updating process, the synchronization parameters are modified to adjust the synchronous iterative updating of the optimal local model and other ternary vector model parameters, to obtain the final global model, and the final global model is taken as the initial global model of the next round of training, including:
[0025] The global cloud data center combines the optimal local model selected by the local data center and the ternary vector model parameters of other unselected end users to perform global iterative updating, adjusts the synchronization parameters of the optimal local model and the ternary vector model parameters in the global model updating process, and after the completion of the global iterative updating of this round, the global cloud data center will take the global model aggregated in this round as the global model for the start of the next round of training.
[0026] The application provides a communication overhead optimization method for a cloud data center. BRIEF DESCRIPTION OF DRAWINGS
[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, a brief introduction will be given to the drawings needed in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.
[0028] Fig. 1 is a step diagram of the communication overhead optimization method for the cloud data center provided by the application.
[0029] Fig. 2 is a framework diagram of the communication overhead optimization method for the cloud data center provided by the application.
[0030] Fig. 3 is a flowchart of the communication overhead optimization method for the cloud data center provided by the application. DETAILED DESCRIPTION
[0031] Embodiments of the present application are described below in detail with reference to the drawings, examples of which are shown in the accompanying drawings, wherein the same or similar numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.
[0032] Referring to Figs. 1 to 3 The present application provides a communication overhead optimization method for a cloud data center, comprising the following steps:
[0033] S1, a federated learning architecture for a cloud data center is constructed based on three nodes of a global cloud data center, a local data center and a terminal user;
[0034] Specifically, the global cloud data center is used to aggregate model parameter instances and send initialized global model parameters; the local data center is used to provide data resources and management services; the terminal user is used to participate in local model training according to local data conditions; and the federated learning architecture is used to monitor the communication transmission conditions of the global cloud data center, the local data center and the terminal user.
[0035] The federated learning architecture for a cloud data center mainly includes three roles of a global cloud data center, a local data center and a terminal user. The global cloud data center is generally used to aggregate model parameter instances or send initialized model parameters, the local data center is usually used to provide data resources and management services, and the terminal user is used to participate in model training according to local data conditions. The main process of the federated learning architecture for a cloud data center is as follows: when the t-1th round of model training is performed, the number of terminal users participating in the training is n∈{1,…,N}. When the tth round of training is performed, the local data center obtains the global model parameters initialized by the global cloud data center in the last round, transmits the initialized model parameters to each terminal user, and the terminal user trains the model according to the local data set. When the training model precision converges to a threshold value, the local data center aggregates the model parameters generated by the terminal user training to form a local model, and uploads the local model to the global cloud data center. Finally, the global cloud data center aggregates the local models to form a global model, and when the global model converges to an expected threshold precision θ t , the global model is used as the initialized global model for the next round of training, and a new round of training is started.
[0036] S2, a communication transmission protocol is established in the federated learning architecture;
[0037] Specifically, in order to better protect the personal privacy information of the end user in the demand response process, and at the same time reduce the communication burden on the end user in the response demand process, a communication transmission protocol is established in the system architecture, which is built on the TCP / IP protocol, and mainly realizes message transmission between the local data center and the end user.
[0038] The main process of transmission is: the end user first sends an evaluation cost to the local data center according to the evaluation of its own local data, the local data center selects the optimal local model through the adaptive function, and the remaining end users participating in the training of this round send the ternary vectorization model parameters to the local data center through the communication transmission protocol. In the transmission process of the communication transmission protocol of the system architecture, only when the evolution direction of the ternary vector of the model parameter sent by the end user changes, the corresponding data packet instruction needs to be sent to the local data center, so in the communication transmission process, the bandwidth occupation of the communication transmission protocol of the system architecture is effectively reduced. The protocol uses full-duplex communication transmission mode, in which the end user and the local data center assume the roles of message sender and receiver, and the system architecture acts as a message proxy to supervise and manage the communication process in real time.
[0039] S3 initializes the global model parameters of the last round in the global cloud data center, and distributes the initialized global model parameters to each end user;
[0040] Specifically, when the tthround of training is performed, the local data center obtains the global model parameters initialized by the global cloud data center, transmits the initialized model parameters to each end user, and the end user trains the model according to the local data set.
[0041] S4 each end user uses the local data set trained by itself to obtain a local model and an evaluation cost based on the initialized global model parameters, and sends the evaluation cost to the local data center through the communication transmission protocol;
[0042] Specifically, each end user uses the local data set trained by itself to complete training based on the initialized global model parameters, obtains a local model, and uses the local data set to judge to obtain a local model and an evaluation cost; the evaluation cost is sent to the local data center through the communication transmission protocol.
[0043] Each end user obtains a local partial model after completing training locally, and obtains an evaluation cost according to a judgment using a local data set trained, which can be a function loss value of model training or a reconstruction error caused by model training. In the interactive communication between the end user and the local data center, due to the difference of the local data of the end user, the effect of the partial model after training is also uneven, and the data size of the partial model is also different, which also causes a large communication overhead, so the end user will first send the evaluation cost to the local data center.
[0044] The local data center quantifies the effect of the partial model trained by different end user data sets based on the evaluation cost through an adaptive function, so as to select the optimal partial model and the corresponding end user.
[0045] Specifically, the local data center quantifies the effect of the partial model trained by different end user data sets based on the evaluation cost through an adaptive function, and the product of the weight of the local data set of any end user in the data size of the end user set participating in this round of training and the evaluation cost is taken as the judgment standard of the training effect of the partial model, so as to select the optimal partial model and the corresponding end user.
[0046] The adaptive function is used to quantify the effect of the partial model trained by different end user data sets, so as to determine the optimal partial model and the corresponding training end user. The adaptive function is defined as shown in formula (1).
[0047]
[0048] Wherein, L n is the data set size of any end user n, L N represents the data size of the end user set N participating in this round of training, and the evaluation cost of the end user. Considering the influence of the data set size difference between the end users on the evaluation cost, the product of the weight of the local data set of any end user in the data size of the end user set participating in this round of training and the evaluation cost is taken as the judgment standard of the training effect of the partial model, which effectively avoids the error caused by the data set difference to the selection of the optimal partial model. When in the first round of training, the partial model with lower unit evaluation cost is selected. However, from the second round of training, the training progress of the historical round needs to be considered, and since the evaluation cost information of the previous round of training is mastered, the training progress can be evaluated by calculating the evaluation cost reduction amount of adjacent two rounds of training, until the optimal partial model of this round of training is determined.
[0049] S6 The unselected end users send the corresponding ternary vectorized model parameters and evolution directions to the local data center in the form of ternary vectors;
[0050] Specifically, for privacy considerations, those unselected end users will not send the entire local model instance to the local data center for aggregation, but will only send the evolution direction of the corresponding model parameters in the form of ternary vectors through the communication transmission protocol of the system architecture. The global model is updated by aggregating the optimal local model instance and the ternary evolution direction of the model parameters. The end user calculates the ternary vector definition of the corresponding model parameters according to the local data as shown in formula (2).
[0051]
[0052] Wherein, is the original parameter set randomly initialized by the global cloud data center in the first round of training, and α n is the learning rate of the end user, which is generally set to a value in the range of (0, 1). Since the end user has no historical round training parameter evolution information in the first round of training, the ternary vector value can only be calculated by the learning rate of each end user. If the local parameter changes relative to its initial value P 0 , it is considered that the ternary vector value of the parameter is -1. If the parameter value changes little, it is considered that the ternary vector value of the parameter is 0. If the local parameter changes relative to the initial parameter, it is considered that the ternary vector value of the parameter is 1.
[0053] From the second round of training, the ternary vector value needs to be calculated according to the historical round training information. The ternary vector value of the parameter m calculated by the end user n in the tth round of training is defined as formula (3).
[0054]
[0055] Wherein, and represent the performance of parameter m in the global model instance received by the end user in the t-1th and t-2th rounds of training, respectively. If parameter m changes significantly in different directions in consecutive two round iteration updates, it is considered that the ternary vector value of the parameter is -1. If the performance of parameter m in the tth round is compared with the t-1th round, or the performance of the t-1th round is compared with the t-2th round, there is no significant change. It is considered that the ternary vector value of the parameter is 0. If parameter m changes significantly in the same direction in consecutive two round iteration updates, it is considered that the ternary vector value of the parameter is 1.
[0056] Communication optimization method based on ternary evolution model parameters
[0057] The present application considers reducing communication overhead and terminal user privacy, etc., and updates the global model instance in combination with the selected optimal local model instance and ternary vectorization model parameters. For privacy considerations, those terminal users who are not selected by the adaptation function will not send the entire local model instance to the local data center for aggregation, but only send the evolution direction of the corresponding model parameters in the form of a ternary vector. The global model is updated by aggregating the optimal local model instance and the ternary evolution direction of the model parameters. Without loss of generality, the present application assumes that the federated learning training process has run for t-1 rounds and is now at the beginning of the tth round. For any terminal user n that is not selected, the ternary vector obtained at the tth round is denoted as The definition is shown in formula (7):
[0058]
[0059] Wherein represents the change of the model parameter m produced by the terminal user n in the tth round of training from the model parameter m of the t-1th round, is the local model parameter obtained by the terminal user n in the tth iteration, and the definition of the ternary vector value is as follows:
[0060] When , it means that the evolution direction of the parameter m obtained by the terminal user n in the tth iteration is different from the evolution direction in the t-1th iteration. That is, if the parameter m increases in the t-1th iteration and decreases in the tth iteration or the parameter m decreases in the t-1th iteration and increases in the tth iteration (i.e. and ), it is considered that
[0061] When , it means that the model parameter m of the terminal user n has no significant change in the continuous two iterations of training;
[0062] When , it means that the evolution direction of the parameter m obtained by the terminal user n in the tth iteration is the same as the evolution direction in the t-1th iteration, that is, if the parameter m continuously increases or decreases in the t-1th and tth iterations (i.e. or ), it is considered that
[0063] The global cloud data center combines the optimal local model and the ternary vectorization model parameters and evolution direction to perform global model updating; in the global model updating process, the synchronization parameter is modified to adjust the synchronous iterative updating of the optimal local model and other ternary vectorization model parameters, to obtain a final global model, and the final global model is used as an initial global model for the next round of training;
[0064] Specifically, the updating of the global model does not rely only on the optimal local model of a certain terminal user. For other terminal users who fail to adapt, the evolution direction of the local model parameters is often transmitted in the form of a ternary vector instead of the entire local model. In combination with the optimal local model of the terminal user n * and the ternary vectorization model parameters of other terminal users participating in training except the terminal user n * , the global model iterative updating is as shown in formula (4).
[0065]
[0066] wherein, a is the learning rate of the global cloud data center for the ternary vector representation model parameter evolution, and the learning rate is usually set to a value in the range of (0, 1). And ω is the synchronization parameter for adjusting the optimal local model and the ternary vectorization model parameters. In the first round of training, only the optimal local model and the ternary vectorization model parameters are combined to perform global model iterative updating. From the second round of training, the ternary vectorization model parameters need to be considered at the same time, and the performance of the parameters on the global model during the historical round of iteration is also considered.
[0067] Specifically, in the process of global model iterative updating, the synchronization parameter ω of the optimal local model and the ternary vectorization model parameters is adjusted. After the completion of the t-1 round of global iterative updating, the global cloud data center will take the global model instance aggregated in this round as the global model at the beginning of the t round of training.
[0068] Experimental evaluation
[0069] Specifically, to verify the communication overhead effectiveness of the proposed method and show the superiority of the proposed method in global iteration time, the application calculates the data volume in the interactive communication process and the global iteration time under the same training accuracy as the evaluation standard, selects the federated average algorithm, the parallel gradient descent method in federated learning training and the convolution-federated learning network method as the comparison scheme, and at the same time, in order to objectively evaluate the proposed algorithm, the experimental results are based on 20 independent experiments, and the average value is selected as the final experimental result. According to table 9 and table 10, it can be seen that the communication optimization method based on ternary vector evolution model parameters proposed in the application has a significant improvement in communication transmission data volume and global iteration time compared with the comparison scheme.
[0070] Table 9 Comparison of total communication transmission data volume under different terminal user numbers
[0071]
[0072]
[0073] Table 10 Comparison of global iteration time under different terminal user numbers
[0074]
[0075] The communication overhead optimization method for cloud data center provided by the application mainly guarantees the privacy information of the terminal users of the cloud data center, introduces a federated learning architecture, establishes a federated learning architecture for the cloud data center to guarantee the privacy of the terminal users. In addition, in order to reduce the interactive communication overhead problem introduced by the federated learning architecture, a ternary evolution model parameter communication optimization algorithm is proposed in the cloud data center scenario, which guarantees the transmission communication privacy between the terminal users and the data center and also effectively reduces the interactive communication overhead. Finally, the effectiveness of the proposed method is verified through experiments.
[0076] Advantages:
[0077] 1. In order to guarantee the privacy information of the terminal users of the cloud data center, the application introduces a federated learning architecture, establishes a federated learning architecture for the cloud data center to guarantee the privacy of the terminal users.
[0078] 2. In order to reduce the interactive communication overhead problem introduced by the federated learning architecture, a ternary evolution model parameter communication optimization method is proposed in the cloud data center scenario, which guarantees the transmission communication privacy between the terminal users and the data center and also effectively reduces the interactive communication overhead.
[0079] 3. The effectiveness of the proposed method is verified by experimental simulation. The experimental results show that, under the premise of meeting the expected accuracy of training accuracy, compared with the federated average algorithm, the proposed algorithm not only effectively reduces the interactive communication overhead, but also greatly reduces the global iteration time.
[0080] The above only discloses a preferred embodiment of the cloud data center-oriented communication overhead optimization method of the present application, of course, cannot limit the scope of the present application, those skilled in the art can understand that all or part of the above-mentioned embodiments are implemented, and the equivalent changes made according to the claims of the present application still belong to the scope covered by the present application.
Claims
1. A method for optimizing communication overhead for cloud data center, characterized in that, The method comprises the following steps: a federated learning architecture for a cloud data center is constructed based on three nodes of a global cloud data center, a local data center and a terminal user; a communication transmission protocol is established in the federated learning architecture; the global cloud data center initializes a global model parameter of a previous round and sends the initialized global model parameter to each terminal user; each terminal user obtains a local model and an evaluation cost based on the initialized global model parameter using a local data set trained by itself, and sends the evaluation cost to the local data center through the communication transmission protocol; the local data center quantifies the effects of the local models trained by different terminal user data sets based on the evaluation cost through an adaptive function, so as to select an optimal local model and a corresponding terminal user; terminal users not selected send corresponding ternary vectorized model parameters and evolution directions to the local data center in a ternary vector manner; the global cloud data center updates a global model in combination with the optimal local model and the ternary vectorized model parameters and evolution directions; in the global model updating process, a synchronization parameter is modified to adjust the synchronous iterative updating of the optimal local model and other ternary vectorized model parameters, so as to obtain a final global model, and the final global model is used as an initialized global model for a next round of training.
2. The cloud data center-oriented communication overhead optimization method of claim 1, wherein the global cloud data center is configured to aggregate model parameter instances and send the initialized global model parameter; the local data center is configured to provide data resources and management services; the terminal user is configured to participate in local model training according to local data conditions; the federated learning architecture is configured to monitor communication transmission conditions among the global cloud data center, the local data center and the terminal user.
3. The cloud data center-oriented communication overhead optimization method of claim 2, wherein the communication transmission protocol is constructed on the basis of a TCP / IP protocol, and the communication transmission protocol adopts a full-duplex communication transmission mode for message transmission between the local data center and the terminal user.
4. The cloud data center-oriented communication overhead optimization method of claim 3, wherein the process in which each terminal user obtains a local model and an evaluation cost based on the initialized global model parameter using a local data set trained by itself, and sends the evaluation cost to the local data center through the communication transmission protocol comprises: each terminal user completes training based on the initialized global model parameter using a local data set trained by itself, and judges using the local data set to obtain a local model and an evaluation cost; the evaluation cost is sent to the local data center through the communication transmission protocol.
5. The cloud data center-oriented communication overhead optimization method of claim 4, wherein The local data center quantifies the training effect of the local model of different end user data sets based on the evaluation cost through an adaptation function, so as to select the optimal local model and the corresponding end user, including: The local data center quantifies the training effect of the local model of different end user data sets based on the evaluation cost through an adaptation function, and the product of the weight of the local data set of any end user in the data size of the end user set participating in the training and the evaluation cost is taken as the evaluation standard of the training effect of the local model, so as to select the optimal local model and the corresponding end user.
6. The cloud data center oriented communication overhead optimization method of claim 5, characterized in that, The unselected end user sends the corresponding ternary vector model parameters and evolution direction to the local data center in the form of a ternary vector, including: The unselected end user does not send the entire local model instance to the local data center for aggregation, but only sends the evolution direction of the corresponding model parameters in the form of a ternary vector by means of the communication transmission protocol.
7. The cloud data center oriented communication overhead optimization method of claim 6, characterized in that, The global cloud data center combines the optimal local model and the ternary vector model parameters and evolution direction to update the global model; in the global model updating process, the synchronization parameters are modified to adjust the synchronous iterative updating of the optimal local model and other ternary vector model parameters, to obtain the final global model, and the final global model is taken as the initial global model of the next round of training, including: The global cloud data center combines the optimal local model selected by the local data center and the ternary vector model parameters of other unselected end users to perform global iterative updating, adjusts the synchronization parameters of the optimal local model and the ternary vector model parameters in the global model updating process, and when the global iterative updating of the current round is completed, the global cloud data center will take the global model aggregated in the round as the global model for the start of the next round of training.