DYNAMIC TRANSMISSION CONTROL FOR IMPROVING MOBILE DATA USAGE

This paper describes techniques for dynamically calculating values of one or more Transmission Control Protocol (TCP) memory buffer size variables (e.g., “tcp_mem,” “tcp_rmem,” and/or “tcp_wmem” variables), based on consideration of real-time network conditions, to achieve improved data usage of a mobile computing device. By dynamically determining and/or adjusting the values of TCP memory buffer size variables, the described techniques enable a mobile computing device (e.g., a mobile phone, tablet computer, wearable and/or headset device) to avoid sending too many in-flight packets that exceed network capacity, thereby reducing packet loss and the need for data retransmission from the mobile computing device. In some cases, the described techniques introduce and utilize a machine-learning model to predict suitable values of the dynamically determined TCP memory buffer size variables. The machine-learning model accepts a number of different features as inputs in order to produce a predicted output value of a memory buffer size variable. These features may include, for example, a specified time frame, real-time network allocated bandwidth, a geographic region (e.g., cell tower identifier or Global Positioning Satellite (GPS) location), and/or a packet loss rate, to name only a few examples. DESCRIPTION With the modern Internet, TCP is one of the main protocols that provides a reliable data exchange between communication devices. Various congestion control algorithms that have been widely deployed interpret packet loss as a signal indicating congestion. Network 2 Lin and Lee: DYNAMIC TRANSMISSION CONTROL FOR IMPROVING MOBILE DATA USAGE Published by Technical Disclosure Commons, 2019 congestion may occur when packets sent from a source device exceed the number of packets that a receiving device can handle, based on current network conditions. Packets may be stored in memory buffers on both the source and receiving devices, but even a temporary overflow of these buffers can lead to congestion, packet loss, data retransmissions, reduced data throughput, and/or performance degradation. The Congestion Window (CWND) variable is one of the important TCP variables for rate control that determines the number of bytes that can be outstanding at any time for a TCP connection. More specifically, the CWND variable is a TCP state variable that limits the amount of data that a source device can send into the network via a TCP connection before receiving an acknowledgment from the receiving device. The use of the CWND variable may regulate data flow in TCP connections, minimize congestion, and improve network performance. For a given connection or communication session, TCP typically increases the value of the CWMD variable over time to take advantage of available bandwidth between the source and receiving devices. However, once the size of the CWMD variable exceeds a certain value, the source device may send a volume of packets that exceeds current network capacity, which may result in dropped packets due to network congestion, and which may further then trigger data retransmission between the source and receiving devices in order to recover from packet loss. Once packet loss has occurred, TCP may drastically reduce the size value of the CWMD variable to avoid further loss. In practice, one or more TCP memory buffer size variables (e.g., “tcp_mem,” “tcp_rmem,” and/or “tcp_wmem” variables) are variables that may be used to control and limit the growing size of the CWND variable. A properly determined size of TCP memory buffer size variables helps prevent packet loss and overshooting of in-flight packets, and thus mitigates the 3 Defensive Publications Series, Art. 2368 [2019] https://www.tdcommons.org/dpubs_series/2368 problem of data retransmission. In current designs, the TCP memory buffer size variables typically have static or fixed values that are based on maximum theoretical throughput conditions for different service providers with various radio access technologies (e.g., LongTerm Evolution (LTE), Enhanced Data for Global Evolution (EDGE)). For example, Figure 1 below shows an assignment of various static/fixed TCP memory buffer sizes for various Radio Access Technology (RAT) for a certain carrier, and other carriers can provide their own configurations. Figure 1: RAT-based plus Carrier-based Configuration for TCP Buffer Sizes In the example of Figure 1, two different memory buffer size variables are shown: “tcp_rmem” and “tcp_wmem”. Each variable may have three different values: a minimum value, a default value, and a maximum value. The “tcp_rmem” variable indicates the size of receive memory buffers used by a receiving device. The minimum receive memory buffer size is indicated by the value “rmem_min,” the default receive memory buffer size is indicated by the value “rmem_def,” and the maximum receive memory buffer size is indicated by the value “rmem_max.” 4 Lin and Lee: DYNAMIC TRANSMISSION CONTROL FOR IMPROVING MOBILE DATA USAGE Published by Technical Disclosure Commons, 2019 Similarly, the “tcp_wmem” variable indicates the size of write memory buffers used by a source device. The minimum write memory buffer size is indicated by the value “wmem_min,” the default write memory buffer size is indicated by the value “wmem_def,” and the maximum write memory buffer size is indicated by the value “wmem_max.” As shown in Figure 1, with various existing designs, all of the values of the “tcp_rmem” and “tcp_wmem” variables (e.g., the minimum, default, and maximum values) are hard-coded, fixed values based on maximum theoretical throughput conditions for different service providers with various radio access technologies (e.g., UMTS, HSPA, HSUPA, HSDPA, HSPAP, EDGE, GPRS, LXRTT, EVDO, LTS). However, these fixed and hard-coded values are less helpful in addressing the overshooting problem in live networks, because they are static, non-changing values that are based on maximum theoretical conditions. In practice, network bandwidth or capabilities may change with time, location and/or environment conditions. Static values of TCP memory buffer variables such as “tcp_rmem” and “tcp_wmem” cannot typically address all live network situations, and the use of static values may introduce packet retransmission due to the window overshooting problem identified above. For mobile users with limited mobile data plans, packet retransmission may unnecessarily consume traffic and may also reduce utilization efficiency. For example, Figure 2 below shows a graph of time (seconds) versus window size (bytes) for live network activity, where each plot indicates the number of in-flight packets. 5 Defensive Publications Series, Art. 2368 [2019] https://www.tdcommons.org/dpubs_series/2368 Figure 2: Time Versus Window Size For In-flight Packets Over time, as shown in Figure 2, the window size (i.e., CWND) increases in size during packet transmission to take advantage of available bandwidth between the source and receiving devices. At point 1 (“1. Congestion Around Here”), the source device may send a volume of packets that exceeds current capacity, which may result in dropped packets due to network congestion. Once packet loss has occurred, the source device may drastically reduce the size value of the CWMD variable to avoid further loss (“2. Detection on Congestion”) and perform data retransmission between the source and receiving devices in order to recover from packet loss (“3. Retransmission and TCP Slow Start from scratch”). As a result, packet retransmission due to window overshooting will impose additional data packet consumption. In the particular example of Figure 2, a 2.08MB upload task may actually, and unnecessarily, consume 2.53MB from a user’s data plan and reduce utilization efficiency. To address these issues, the techniques described in the present paper enable the use of dynamically determined values of TCP memory buffer size variables that account for real-time network conditions. By dynamically determining and/or adjusting TCP memory buffer size 6 Lin and Lee: DYNAMIC TRANSMISSION CONTROL FOR IMPROVING MOBILE DATA USAGE Published by Technical Disclosure Commons, 2019 variables, the described techniques enable a mobile computing device to avoid sending too many in-flight packets over and above current network capacity, even during times of network congestion, thereby reducing data retransmission from the mobile computing device. As a result, the disclosed techniques can help reduce packet loss and data retransmission issues. In certain cases, such as that shown below in Figure 4, the disclosed techniques may also increase data throughput and performance. In certain other cases, data throughout may be maintained or slightly lowered in order to prevent packet loss. In some cases, the described techniques introduce and utilize a machine-learning model to predict suitable values of the TCP memory buffer size variables. The machine-learning model can be trained with labeled datasets based on historical throughput data, which may be converted or otherwise used to determine associated TCP memory buffer size information for inclusion in the model. After the machine-learning model is trained, it is configured to accept a number of different features as inputs in order to produce predicted output values of one or more memory buffer size variables (e.g., “tcp_mem,” “tcp_rmem,” and/or “tcp_wmem” variables). These features may include, for example, a specified time frame, such as particular times, hours, days, weeks, months, or other time periods or ranges. These input features may also include real-time network allocated bandwidth information for a given network, and/or or geographic region information (e.g., cell tower identifier information or Global Positioning Satellite (GPS) location information) that is associated with the source device, receiving device, or network area in general. Additional example features may include real-time packet l