A Close Examination of Performance and Power

A Close Examination Of Performance And Power-Free PDF

  • Date:08 Jun 2020
  • Views:12
  • Downloads:0
  • Pages:14
  • Size:785.24 KB

Share Pdf : A Close Examination Of Performance And Power

Download and Preview : A Close Examination Of Performance And Power


Report CopyRight/DMCA Form For : A Close Examination Of Performance And Power


Transcription:

Figures Tables Data source Description, 4 1 Figure 5 6 Public deployment of 4GTest Compare LTE network performance with other mobile networks. 4 2 Figure 7 Local LTE Phone Infer LTE parameters with network based approach. 4 3 4 4 Figure 8 LTE Laptop and LTE Phone Understand the impact of various factors on LTE performance. 5 1 5 2 Table 3 4 Figure 9 10 11 Local LTE Phone Derive LTE power model and compare with WiFi and 3G. 5 3 Figure 12 Local LTE Phone Compare bulk data transfer energy efficiency of LTE with WiFi and 3G. 5 4 Table 5 Local LTE Phone Validate LTE power model with real applications. 6 1 6 2 Figure 14 Simulation on UMICH data set Analyze energy efficiency and the impact of different energy components. 6 3 Figure 15 19 20 16 17 18 Simulation on UMICH data set Study the impact of LTE parameter configurations. 7 1 Figure 22 Different local platforms Compare JavaScript execution speed in 2009 and in 2011. 7 2 Table 6 Figure 21 Local LTE Phone Study performance bottleneck of web based applications. Table 1 Summary and index of experimental results, by feeding WiFi traces into our LTE model simulation framework signed for energy saving We also study the impact of configuring. given that RTT of WiFi is close to LTE 4 1 and throughput is various LTE parameters on radio energy channel scheduling delay. less important than RTT for most web based smartphone applica and signaling overhead Similar to 3G UMTS network LTE tail. tions 7 2 timer Ttail remains a key parameter in determining UE energy us. Compared with 3G LTE significantly improves the network per age and performance as well as signaling overhead of cell towers. formance Meanwhile device processing capability and software iv We perform case studies of several popular applications on. design have also improved remarkably over the last two years com Android to understand the impact of improved LTE network per. paring with our previous study 24 To understand the potential formance and enhanced UE processing power on applications The. performance bottleneck shift for smartphone applications we per processing speed for contemporary smartphones have increased sig. form case studies of several popular applications on Android With nificantly compared with our analysis two years ago However we. the help of CPU network and power traces we devise a systematic still identify that the performance bottleneck for web based appli. methodology for application performance and power measurement cations lies more in the device s processing power than in the net. and analysis work indicated by the high average CPU usage of 79 3 in LTE. We summarize the results of this paper in Table 1 and highlight network as compared to 57 7 in 3G network as well as the un. our main contributions as follows derutilized network capacity due to small object size in typical web. i This paper is one of the first studies on commercial LTE net transactions. works We characterize LTE network performance with both local The remaining paper is organized as follows We cover back. experiments and global measurement by deploying a network per ground on RRC and DRX for LTE in 2 followed by experimental. formance measurement tool 4GTest with distributed server support setup and methodology in 3 We then present network characteri. globally Generally LTE is observed to have significantly higher zation results including 4GTest results in 4 and power measure. downlink and uplink throughput than 3G and even WiFi with a ments in 5 6 summarizes the simulation analysis on UMICH data. median value of 13Mbps and 6Mbps respectively which is even set and 7 includes the case study of several popular Android ap. higher than the advertised throughput range of the LTE provider plications We discuss related work in 8 before concluding in 9. ii We develop the first empirically derived comprehensive power 2 BACKGROUND. model of a commercial LTE network considering both uplink and. We first cover the necessary background on LTE state machine. downlink data rates in addition to state transitions and DRX It ac. behavior and corresponding power characteristics, curately quantifies the radio energy usage of UE with less than 6. error rate Our power measurement also reveals that LTE is less 2 1 Radio Resource Control RRC and Dis. power efficient than 3G and WiFi for small data transfers e g for continuous Reception DRX in LTE. transferring one packet the energy usage for LTE 3G and WiFi is. LTE has two RRC states RRC CONNECTED and RRC IDLE 10. 12 76J 7 38J and 0 04J respectively For bulk data transfer LTE. 17 as shown in Figure 1 At RRC CONNECTED state UE can be. is more power efficient than 3G yet still lags behind WiFi e g for. in one of the three modes Continuous Reception Short DRX and. downloading 10MB data 3G and LTE require 34 77 and 1 62 times. Long DRX While at RRC IDLE state UE is only in DRX mode. the energy of WiFi respectively Considering that LTE is more en. Table 2 enumerates a list of important LTE parameters which have. ergy efficient if fully utilized it is even more crucial to optimize. significant impact on UE s radio energy consumption user experi. applications using tools like ARO 28 in LTE than 3G to achieve. ence and signaling overhead for cell towers The terms in Table 2. the full LTE savings The energy model derived in this paper will. are used consistently throughout this paper, provide the necessary understanding to drive those optimizations. If UE is initially in RRC IDLE state and receives sends one packet. iii With the aforementioned power model we build a trace. regardless of the packet size a state promotion from RRC IDLE to. driven LTE analysis modeling framework which breaks down the. RRC CONNECTED occurs with a relatively stable delay similar to. total energy consumption into different components to identify the. the promotion from IDLE to DCH FACH in UTMS network 27. key contributor for energy usage It also quantifies the channel. We define the LTE promotion delay to be Tpro 1 consistently through. scheduling delay perceived by UE and signaling overhead of cell. out this paper During this period radio resources are allocated to. towers This analysis framework is applied to a comprehensive data. set UMICH consisting 5 months data from 20 real users By com. After being promoted to RRC CONNECTED UE enters Continu. paring the estimated energy usage of different networks LTE is as. ous Reception by default and keeps monitoring the Physical Down. much as 23 times less power efficient compared with WiFi and. even less power efficient than 3G despite the presence of DRX de Tpro is a measured system property different from the config. urable LTE parameters in Table 2,Symbol Full name Measured value Description.
Ti DRX inactivity timer 100ms UE stays in Continuous Reception for Ti before DRX starts when idling. Tis Short DRX cycle timer 20ms UE remains in Short DRX for Tis before entering Long DRX when idling. Ttail RRC inactivity timer 11 576s UE stays in RRC CONNECTED for Ttail before demoting to RRC IDLE. Ton RRC CONNECTED On Duration timer 1ms The on duration of UE during each DRX cycle in RRC CONNECTED. Toni RRC IDLE On Duration timer 43ms The on duration of UE during each DRX cycle in RRC IDLE. Tps Short DRX cycle 20ms The cycle period of Short DRX in RRC CONNECTED. Tpl Long DRX cycle 40ms The cycle period of Long DRX in RRC CONNECTED. Tpi RRC IDLE DRX cycle 1 28s The cycle period of DRX in RRC IDLE. Table 2 Important LTE RRC and DRX parameters,t2 Data transfer starts. Continuous,Reception 3000,Ti DRX 2000 t3 Tail starts. Short Long,DRX DRX t1 Promotion starts t4 Tail ends. 0 t1t2 5 10 t3 15 20 t4 25,RRC CONNECTED RRC IDLE,Time second. Timer Data,expiration transfer,Figure 3 Power states of LTE.
Figure 1 RRC state transitions in LTE network, For our measurement Tis coincidentally equals Tps so only one. On Duration Data transfer Ti expiration cycle of Short DRX is expected to take place before Tis expires. Tis expiration Every time UE enters Continuous Reception when there is any data. transfer UE starts the tail timer Ttail which is reset every time. a packet is sent received When Ttail expires UE demotes from. Continuous RRC CONNECTED to RRC IDLE and the allocated radio resource is. Long DRX Reception Short Long DRX released Notice that Ttail coexists with Ti and Tis. cycle DRX cycle,cycle 2 2 Smartphone power model for LTE. Given the description of LTE state machine we illustrate the. Figure 2 Illustration of the LTE DRX in RRC CONNECTED power traces of an Android smartphone in a commercial LTE net. work based on local experiments described in 3 2 We observe. that network activities match the corresponding state transitions in. link Control Channel PDCCH which delivers control messages dicated by different power levels. to UE UE also starts the DRX inactivity timer Ti which is reset Figure 3 shows the power trace of uploading at the speed of. every time UE receives sends a packet Upon Ti s expiration with 1Mpbs for 10 seconds With screen off the energy is mostly con. out seeing any data activity UE enters the Short DRX mode sumed by the radio interfaces as the power level is less than 20mW. Discontinuous Reception DRX 16 30 illustrated in Figure 2 before t1 At t1 the application sends a TCP SYN packet trigger. is adopted by LTE for UE to micro sleep to reduce power con ing RRC IDLE to RRC CONNECTED promotion and the application. sumption while providing high QoS and connectivity DRX in waits for Tpro until starting data transfer at t2 Between t2 and. RRC CONNECTED and RRC IDLE have similar mechanisms but dif t3 depending on the instant data rate the power level fluctuates. ferent parameter settings A DRX cycle includes an On Duration We notice the power level during fast data transfer is significantly. during which the UE monitors PDCCH UE rests for the rest of the higher than the base power in RRC CONNECTED which motivates us. cycle to save energy The tradeoff between battery saving and la to incorporate data rates into our LTE power model After the data. tency is the guideline for determining the parameterization of DRX transfer completes at t3 the device remains in RRC CONNECTED. cycle With a fixed On Duration a longer DRX cycle reduces en for a fixed tail time Ttail until t4 when the device goes back to. ergy consumption of UE while increasing user perceived delay and RRC IDLE The periodicity of DRX between t3 and t4 is not obvi. a shorter DRX cycle reduces the data response delay at the cost of ous due to limited sample rate. more energy consumption Short DRX and Long DRX modes hav In summary this section provides necessary background infor. ing the same On Duration and differing in cycle length are to meet mation and some initial motivating observations for the following. these conflicting requirements discussions on the network and power characterization of LTE net. When UE enters Short DRX Short Cycle Timer Tis is started work. Upon Tis s expiration if there is no data activity UE switches to. Long DRX otherwise UE goes back into Continuous Reception. 3 METHODOLOGY Packet and CPU trace collection In order to collect packet. In this section we present the methodology for network and traces on LTE Phone we cross compile tcpdump 9 With a small. power measurement as well as trace driven simulation analysis and snapshot length only to capture TCP UDP and IP headers the CPU. real application case studies overhead of the background tcpdump process is less than 5 To. capture CPU usage history we write a simple C program to read. 3 1 Network measurement proc stat in Android system every 25ms and the overhead is. In this section we discuss the design of our publicly deployed verified to be smaller than 5 on LTE Phone. tool 4GTest followed by experimental setup and methodology for Network based approach for LTE parameter inference is used. local network measurement to validate the power based approach 3 2 We use the follow. ing experiment to infer LTE state machine and RRC IDLE DRX. 3 1 1 The design of 4GTest parameters LTE Phone maintains a long lived TCP connection. Based on our previous experiences of developing 3GTest 6 with a test server For each experiment server first sends a packet. 24 we design a new mobile network measurement tool for An P1 to trigger UE s promotion to RRC CONNECTED and after X. droid devices called 4GTest 2 with higher accuracy and im seconds server sends another packet P2 to the phone to measure. proved user experience As a new feature 4GTest allows users RTT Assume the RTT at RRC CONNECTED is RT Tb ignoring the. to switch among different network types i e 3G WiFi and LTE minor impact of DRX in RRC CONNECTED When X Ttail. We also improve the network measurement methodology leverag RT T X RT Tb Otherwise RT T X RT Tb Tpro Tdrx. ing the M Lab 5 support 4GTest server suite is deployed to 46 where Tdrx Tn X and Tn is the start time of the next DRX. M Lab nodes across the world covering North America Europe on duration after P2 s arrival If P2 s arrival is inside any DRX on. Asia and Australia Each node has 4 core 2 00 GHz Intel Xeon duration Tdrx 0. CPU and our virtual slice has 4GB memory and 100Mbps Ether To understand the impact of packet size on one way delay OWD. net network access which ensures that the network bottleneck is uplink and downlink OWD is measured with varying packet size. unlikely on the wired network path For each packet size 100 samples are measured and we make. A Close Examination of Performance and Power Characteristics of 4G LTE Networks creasing interest to better understand the performance and power characteristics compared with 3G WiFi networks In this paper we take one of the rst steps in this direction Using a publicly deployed tool we designed for Android called 4GTestattracting more than 3000 users within 2 months and exten sive

Related Books