A 5G energy consumption optimization method for mobile applications

Abstract

A 5G energy consumption optimization method for mobile applications, including: (1) converting RRC states; (2) defining symbols; (3) extracting features; (4) identifying EOC; (5) predicting the arrival time of the next data packet. The present invention aims at the energy consumption optimization problem of mobile applications on 5G smartphones and proposes a 5G radio energy saving method 5GSaver with two-stage learning: 1) In the first stage, 5GSaver uses random forest (RF) to identify EOC events (IEOC); 2) In the second stage, if an EOC event is identified, 5GSaver further uses deep forest (DF) to predict the next packet arrival time (PPAT) by utilizing the program execution mode and the statistical information of each application. The present invention utilizes the newly introduced inactive state in 5G NR to optimize UE energy consumption. Compared with the existing tail optimization method, the present invention has better effects in energy saving and communication delay.

Description

Technical Field

[0001] The present invention proposes a 5G energy consumption optimization method for mobile applications. This invention is specifically aimed at the energy consumption optimization problem of mobile applications on 5G smartphones, and proposes a two-stage energy-saving method 5GSaver combined with the inactive state of 5G NR to achieve a high improvement in energy saving and communication delay performance. Background Art

[0002] 5G technology is widely used in smartphones, which can significantly improve the mobile user experience with better performance and higher efficiency. However, due to 5G's higher data rate, more antennas and more powerful radio frequency modules, its energy consumption per unit time is three times higher than that of traditional radio. Studies have shown that a large amount of energy waste in 5G NR is attributed to the radio tail, that is, the period of time when the radio interface remains in a high-power state after data transmission is completed. Existing tail-cutting methods will force the radio into an idle state, resulting in non-negligible energy costs and delays. Summary of the invention

[0003] The present invention aims to optimize the high energy consumption problem of 5G NR and provides a 5G energy consumption optimization method 5GSaver for mobile applications with two-stage learning.

[0004] 5GSaver reduces 5G energy consumption by alleviating the radio tail phenomenon: When an EOC event is detected, 5GSaver will further predict the arrival time of new network communications. If new communications are expected soon, 5GSaver will help the UE enter the RRC_INACTIVE state. Otherwise, 5GSaver will choose to enter the RRC_IDLE state for long-term energy saving.

[0005] The technical solution adopted by a 5G energy consumption optimization method for mobile applications of the present invention includes the following steps:

[0006] (1) Transition RRC state. 5G NR has three radio resource control (RRC) states, namely RRC_IDLE, RRC_INACTIVE, and RRC_CONNECTED. The RRC state transition process specifically includes:

[0007] (1.1) The 5G radio on the UE is initialized from the RRC_IDLE state. Once the UE has data packets to send, the UE will send an RRC connection request to enter the RRC_CONNECTED state. The base station can initiate a state transition to suspend the RRC connection and put the UE into the RRC_INACTIVE state.

[0008] (1.2) When entering the RRC_INACTIVE state, the UE will store the UE inactive entry layer (AS) context and RRC configuration. In the RRC_INACTIVE state, when the UE needs to send a message, it will resume the suspended RRC connection, restore the above information, and quickly transition to the RRC_CONNECTED state.

[0009] (1.3) When the base station considers that the context is no longer valid or there is no more data transmission during the inactivity timer, the base station will release the RRC connection and allow the UE to enter the RRC_IDLE state.

[0010] (2) Definition of symbols. A network session is defined as the period of time during which an application sends or receives packets, so that any two consecutive packets have a small inter-packet time. Use period t s To indicate when the UE needs to release the RRC connection, an additional threshold t′ is set in 5GSaver s To distinguish whether the UE should enter the inactive state or the idle state. The period between any two consecutive messages is defined as the data segment, using t seg To represent the duration of any segment, extract the time period t from these segments d The data item of seconds is used as the basic data unit. The data item extraction includes:

[0011] (2.1)t seg <t d ,Due to the frequent packet transmission during data communication, ,the duration of a large number of segments is very short.,In 5GSaver, a trade-off between overhead and accuracy is made using data items, ,connecting as many such short consecutive segments as needed to extract an ,active data item.

[0012] (2.2)t d <t seg <t s , an active data item will be extracted from the observed data packet, and the length of each data item will be truncated to t d seconds, because it is necessary to restrict learning to the shorter features of the packet.

[0013] (2.3)t seg >t s , if the duration of the segment is greater than t s , extract a time period at t from the observed message d For EOC data items within seconds, the machine learning model can use features to learn the rules of EOC events for a long period of time after the data packet is observed. For all EOC data items, they are further divided into two cases according to their duration.

[0014] (2.4)ts <t seg <t′ s ,The UE should be inactive at this time, and the EOC data items of these transient EOC events will be marked as EOC-FR.

[0015] (2.5)t seg >t′ s , when the duration is greater than t′ s When the application is considered to have really finished data transmission, it should transition to the idle state. The EOC data items of these instantaneous EOC events will be marked as EOC-IDLE.

[0016] (3) Feature extraction. Application-related data is collected at a sampling rate of 500 Hz in the background, and the following features are extracted from each data item: system function call, data packet information, application resource usage, log information, and previous data item status.

[0017] (4) Identify EOC. After studying the classification model of smartphones, 5GSaver uses RF classifier as its EOC estimation tool. Since some features may not have a strong correlation with the running status of a specific application, these features need to be filtered out during training. The specific process includes:

[0018] (4.1) The input of the RF classifier is data items labeled ACTIVE and EOC. The number of ACTIVE data items is much less than the number of EOC data items, which makes the dataset obviously unbalanced. The ACTIVE data items are undersampled to eliminate the imbalance in sample size.

[0019] (4.2) Then all features are normalized to accelerate the convergence of the RF model. Some data items and their corresponding labels are used as training sets, and the rest are used as test sets. A grid-based search is performed on the training set to automatically find the optimal hyperparameters, and a five-fold cross-validation strategy is used to maximize the classification accuracy.

[0020] (4.3) Finally, the optimized parameters are used to construct the final RF model, which can be applied to the application of interest.

[0021] (5) Predict the arrival time of the next data packet. After determining that the current data item is marked as EOC, 5GSaver calculates the next data packet arrival time based on t′ s To determine whether to enter the idle state or the inactive state. It only needs to predict whether the arrival time of the next message is greater than t′ s Therefore, the prediction task is transformed into a binary classification problem. Because prediction is challenging, we try to use a more powerful deep learning model to solve this problem. The specific process includes:

[0022] (5.1)5GSaver uses DF as the prediction model because DF is lightweight and can achieve performance that is highly competitive with deep neural networks, while the performance is quite robust to hyperparameter settings. Therefore, DF performs well in different mobile applications.

[0023] (5.2) Similar to RF, DF also integrates a feature selection module to remove irrelevant features and reduce runtime overhead. The input of the DF model consists of EOC data items labeled EOC-FR and EOC-IDLE. The ratio between these two labels depends on the application. For applications with frequent real-time network communication such as browsers, the number of EOC-FR data items is much larger than the number of EOC-IDLE data items. For applications with less network communication, the ratio of EOC-FR data items will be greatly reduced.

[0024] (5.3) Use similar undersampling and normalization methods to process the input data, and use part of the data items to train the DF model.

[0025] Preferably, step (4.2) uses 70% of the data items and their corresponding labels as a training set, and the rest as a test set.

[0026] Preferably, 70% of the data items are used to train the DF model in step (5.3).

[0027] Preferably, the application with less network communication volume in step (5.2) may be a music application.

[0028] The present invention utilizes the 5G NR inactive state RRC_INACTIVE to achieve low power consumption and fast wake-up.

[0029] The key to 5G energy consumption optimization methods is to clearly distinguish between instantaneous EOC events and other EOC events so that the radio can enter an inactive state for fast wake-up. Based on this, a 5G radio energy saving method 5GSaver with two-stage learning is proposed: 1) In the first stage, 5GSaver uses random forest (RF) to identify EOC events (IEOC); 2) In the second stage, if an EOC event is identified, 5GSaver further uses deep forest (DF) to predict the next packet arrival time (PPAT) by utilizing the program execution pattern and the statistical information of each application. Compared with other methods, the two-stage method has higher recognition accuracy and lower computational overhead.

[0030] Experimental results show that compared with the state-of-the-art RadioJockey method, the proposed two-stage energy-saving method 5GSaver can reduce the wireless energy consumption cost by 9.5% while reducing the communication delay by 12.4% on average.

[0031] The advantages of the present invention are: a two-stage energy-saving method 5GSaver combined with the inactive state of 5G NR is proposed. Its core concept is to optimize UE energy consumption by utilizing the newly introduced inactive state in 5G NR. Compared with the existing tail optimization method, 5GSaver has better effects in energy saving and communication delay. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a state transition diagram of UE RRC state and NR in the present invention.

[0033] Figure 2 It is a diagram of extracting data items from a data packet according to the present invention.

[0034] Figure 3 It is a two-stage workflow diagram of 5GSaver of the present invention.

[0035] Figure 4 It is a comparison chart of the 5GSaver of the present invention and the existing energy-saving methods.

[0036] Figure 5 It is a comparison chart of the communication delay between the 5GSaver of the present invention and the existing method. DETAILED DESCRIPTION

[0037] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0038] The present invention provides a 5G energy consumption optimization method for mobile applications, comprising: (1) converting RRC state; (2) defining symbols; (3) extracting features; (4) identifying EOC; and (5) predicting the arrival time of the next data packet.

[0039] like Figure 3 , the present invention proposes a two-stage energy saving method 5GSaver combined with the inactive state of 5G NR. 5GSaver uses application-related functions to train the RF model for identifying EOC events and trains the DF model for predicting the arrival time of the next packet. The specific implementation is as follows:

[0040] (1) Configuring the experimental setup. All experiments were conducted using a commercial Android smartphone, ZTE Axon10 Pro. ZTE Axon10 Pro represents a mainstream 5G smartphone equipped with a powerful 5G NR module (SDX M50 5G modem). A battery status monitoring application based on Android API was developed to obtain the instantaneous current and voltage and then calculate the instantaneous energy consumption. When the target application is running, the monitoring application will run in the background at a sampling rate of 500Hz.

[0041] (2) Obtaining the energy consumption of 5G NR. The energy consumption of smartphones running two typical applications: http download and online video player was measured, and the overall energy consumption of 5G / 4G was decomposed into the following four parts:

[0042] (2.1) Get system consumption. Turn off the screen, turn on airplane mode, and close all background applications.

[0043] (2.2) Get screen consumption. Measure screen elements at maximum brightness while keeping other settings unchanged.

[0044] (2.3) Obtain the power consumption of the application. Load the application content in advance and then run the application offline.

[0045] (2.4) Obtain the energy consumption of the wireless interface. Turn on and off the 5G mode and measure the energy consumption of 5G NR and 4G LTE respectively.

[0046] (3) Feature extraction. A feature extraction module was developed to collect application-related data at a sampling rate of 500 Hz in the background. For each data item, the collected features include a variable number of system function calls, 2 packet features, 11 battery features, 8 CPU features, 7 memory features, 2 log features, and 1 previous state. These features are broken down into each individual application and can be well captured using a per-application model. Therefore, 5GSaver trains models at the application granularity.

[0047] (3) Training model. 5GSaver uses RF classifier as its EOC estimation tool, and the input is data items labeled ACTIVE and EOC. The ACTIVE data items are undersampled to eliminate the imbalance of sample size. 70% of the data items and their corresponding labels are used as the training set, and the rest are used as the test set. A grid-based search is performed on the training set to automatically find the optimal hyperparameters, and a five-fold cross-validation strategy is adopted to maximize the classification accuracy. Finally, the optimized parameters are used to build the final RF model, which can be applied to the application of interest. 5GSaver uses DF as the prediction model, and the input consists of EOC data items labeled EOC-FR and EOC-IDLE, and the ratio between these two labels depends on the application. A similar undersampling and normalization method is used to process the input data, and 70% of the data items are used to train the DF model. The RF classifier can usually be trained in 0.21 seconds using 6514 data items involving 78 features on a common smartphone (Qualcomm Snapdragon 855, a smartphone equipped with a Snapdragon 855 chip and 8GB RAM), which is sufficient to achieve good EOC recognition accuracy. Since the DF prediction model is relatively fast to train compared to other deep learning algorithms, it can also be trained in 1.64 seconds.

[0048] (4) Implement an online learning engine. Combine the pre-trained offline training model with the online learning algorithm. During online learning, 5GSaver will continuously monitor the misclassification and misprediction rates of the trained RF and DF models in a sliding time window. If the misclassification (or misprediction) rate is high, which means that the traffic pattern of the application has changed significantly, the learning engine will use the data items of the current application instance to retrain the RF (or DF) model. The lightweight model allows the online learning engine to execute directly on the mobile phone. At runtime, whenever the application sends or receives a packet (packet 1) after being idle for a long time, the runtime engine will reset the running state, form a new data item, and start collecting relevant features in the data item. If the application is idle for a long time in data item t d If the next packet (packet 2) is further sent or received within the time window t, the current cycle (from packet 1 to packet 2) is recognized as ACTIVE, and the trace collection is restarted as part of the data item started by packet 2. On the other hand, if the current cycle (from packet 1 to packet 2) is ACTIVE, the trace collection is restarted as part of the data item started by packet 2. dIf there is no network transmission, the runtime engine needs to use the trained RF classifier and the currently extracted data items to identify whether the application has reached an EOC event. If the data item is identified as ACTIVE, the UE will remain in the RRC_CONNECTED state. If the data item indicates that an EOC event has occurred, 5GSaver needs to further use the DF model to predict whether the application will restart new communications soon. If so, the UE should suspend the RRC connection and transition to the RRC_INACTIVE state. Otherwise, the engine will use fast sleep to force the radio into the RRC_IDLE state.

[0049] (5) Analyze the experimental results. Figure 4 and Figure 5 It can be seen that compared with commercial strategies, 5GSaver can significantly save energy (38.4% on average) while introducing acceptable latency. Compared with the state-of-the-art RadioJockey method, 5GSaver can reduce the wireless energy cost by 9.5% while reducing the communication latency by 12.4% on average.

[0050] This paper specifically targets the energy consumption optimization problem of mobile applications on 5G smartphones and proposes a 5G radio energy saving method 5GSaver with two-stage learning: 1) In the first stage, 5GSaver uses random forest (RF) to identify EOC events (IEOC); 2) In the second stage, if an EOC event is identified, 5GSaver further uses deep forest (DF) to predict the next packet arrival time (PPAT) by utilizing the program execution mode and the statistical information of each application. Its core concept is to optimize UE energy consumption by utilizing the newly introduced inactive state in 5G NR. Compared with the existing tail optimization methods, 5GSaver has better effects in energy saving and communication delay.

[0051] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A 5G energy consumption optimization method for mobile applications, comprising the following steps: (1) Transitioning the RRC state: 5G NR has three radio resource control (RRC) states: RRC_IDLE, RRC_INACTIVE, and RRC_CONNECTED. The RRC state transition process specifically includes: (1.1) The 5G radio on the UE is initialized from the RRC_IDLE state; once the UE has a data packet to send, the UE will send an RRC connection request to enter the RRC_CONNECTED state; the base station can initiate a state transition to suspend the RRC connection and put the UE into the RRC_INACTIVE state; (1.2) When entering the RRC_INACTIVE state, the UE will store the UE inactive entry layer (AS) context and RRC configuration; in the RRC_INACTIVE state, when the UE needs to send a message, it will resume the suspended RRC connection, restore the UE inactive entry layer (AS) context and RRC configuration, and quickly transition to the RRC_CONNECTED state; (1.3) When the base station considers that the context is no longer valid or there is no more data transmission during the inactivity timer, the base station will release the RRC connection and allow the UE to enter the RRC_IDLE state; (2) Definition of symbols; a network session is defined as the time period during which an application sends or receives packets; using period t s To indicate when the UE needs to release the RRC connection, an additional threshold t′ is set in 5GSaver s To distinguish whether the UE should enter the inactive state or the idle state; the period between any two consecutive messages is defined as the data segment, using t seg To represent the duration of any segment, extract the time period t from these segments d The data item of seconds is used as the basic data unit; the data item extraction specifically includes: (2.1)t seg <t d ,Due to the frequent packet transmission during data communication, ,a large number of segments have very short durations;,In 5GSaver, the data items are used to trade off between overhead and ,accuracy, by concatenating as many such short consecutive segments as needed to ,extract an active data item; (2.2)t d <t seg <t s , an active data item will be extracted from the observed data packet, and the length of each data item will be truncated to t d Second; (2.3)t seg >t s , if the duration of the segment is greater than t s , extract a time period at t from the observed message d For EOC data items within seconds, the machine learning model can use features to learn the rules of EOC events in the long period of time after the data packet is observed; for all EOC data items, they are further divided into two cases according to their duration; (2.4)t s <t seg <t′ s ,The UE should be inactive at this time, and the EOC data items of these transient EOC events will be marked as EOC-FR; (2.5)t seg >t′ s , when the duration is greater than t′ s When the application is considered to have really finished data transmission, it should switch to the idle state. The EOC data items of these instantaneous EOC events will be marked as EOC-IDLE. (3) Feature extraction: collect application-related data at a sampling rate of 500 Hz in the background, and extract the following features from each data item: system function call, data packet information, application resource usage, log information, and previous data item status; (4) Identify EOC. After studying the classification model of smartphones, 5GSaver uses RF classifier as its EOC estimation tool. Since some features are not strongly correlated with the running status of a specific application, these features need to be filtered out during training. The specific process includes: (4.1) The input of the RF classifier is data items labeled ACTIVE and EOC. The number of ACTIVE data items is less than the number of EOC data items, so the ACTIVE data items are undersampled. (4.2) Then all features are normalized to accelerate the convergence of the RF model; some data items and their corresponding labels are used as training sets, and the rest are used as test sets; a grid-based search is performed on the training set to automatically find the optimal hyperparameters, and a five-fold cross-validation strategy is used to maximize the classification accuracy; (4.3) Finally, the optimized parameters are used to construct the final RF model, which can be applied to the application of interest; (5) Predict the arrival time of the next data packet; after determining that the current data item is marked as EOC, 5GSaver calculates the next data packet arrival time based on t′ s To determine whether to enter the idle state or the inactive state; it only needs to predict whether the arrival time of the next message is greater than t′ s Therefore, the prediction task is transformed into a binary classification problem. Because prediction is challenging, a deep learning model is used to solve this problem. The specific process includes: (5.1)5GSaver uses deep forest DF as the prediction model; (5.2)DF also integrates a function selection module to remove irrelevant functions and reduce runtime overhead; the input of the DF model consists of EOC data items labeled EOC-FR and EOC-IDLE; the ratio between these two labels depends on the application; for applications with frequent instant network communication, including browsers, the number of EOC-FR data items is more than the number of EOC-IDLE data items; for applications with less network communication, the ratio of EOC-FR data items will be reduced; (5.3) Use the undersampling method of step (4.1) and the normalization method of step (4.2) to process the input data, and use part of the data items to train the DF model.

2. A 5G energy consumption optimization method for mobile applications as claimed in claim 1, characterized in that: Step (4.2) uses 70% of the data items and their corresponding labels as the training set and the rest as the test set.

3. A 5G energy consumption optimization method for mobile applications as claimed in claim 1, characterized in that: In step (5.3), 70% of the data items are used to train the DF model.

4. A 5G energy consumption optimization method for mobile applications as claimed in claim 1, characterized in that: The application with less network traffic described in step (5.2) is a music application.

Images