Adv Distributed Syst & Applic
Adv Distributed Syst & Applic CS 795
Popular in Course
Popular in ComputerScienence
This 47 page Class Notes was uploaded by Leland Swift on Monday September 28, 2015. The Class Notes belongs to CS 795 at George Mason University taught by Hakan Aydin in Fall. Since its upload, it has received 49 views. For similar materials see /class/215115/cs-795-george-mason-university in ComputerScienence at George Mason University.
Reviews for Adv Distributed Syst & Applic
Report this Material
What is Karma?
Karma is the currency of StudySoup.
You can buy or earn more Karma at anytime and redeem it for class notes, study guides, flashcards, and more!
Date Created: 09/28/15
CS 795 004 LowPower Computing George Mason University Spring 2009 GMU CS 795 Operational Info Instructor Hakan Aydin Email aydincsgmuedu Office Science and Tech II Rm 401 Office Hours 0 Tuesday and Wednesday 725 PM 825 PM 0 By appointment Course homepage httpcsgmueduaydincs795 About the course Seminartype course Will discussevaluate hot research issues in the general Systems area around the general theme of LowPower Computing Prerequisites Fundamental knowledge of 0 Operating Systems eg CS 571 0 Computer Networks eg CS 555 No textbook papers from recent conference proceedings and journals will be provided GMU CS 795 13 Course Format and Grading 39 First half 0 Fundamental principles of PowerAware Computing 0 Paper presentations by the instructor 39 Second half 0 Paper and project presentations by students 39 Grading Components 0 Term Project 50 0 Paper presentation 30 0 Class Participation and Paper SummariesEvaluations 20 GMU CS 795 14 Paper Presentations 39 Two paper presentations by each student The presentation will include 0 A clear description of the problem 0 Comparison to related work 0 Solution of the paper 0 Critical evaluation Brief summary evaluation of the papers will be submitted by all the students prior to the presentation List of potential papers will be provided Suggestions welcome Get help for wise selections GMU CS 795 15 Term project 39 Several options 0 Theoretical investigation 0 Simulationbased performance analysis 0 Critical survey 0 Implementation 39 entative time table 0 March 4th Project assignments 0 April lst Progress reports due 0 April 29th Inclass project presentations 0 May 6th Project reports due GMU CS 795 16 Lowpower Computing 39 Moved to the forefront of research and development in recent years 39 For devices that have to rely on battery power 0 Mobilepervasive computing sensor systems adhoc networks realtimeembedded systems 0 The energy demand is getting higher and higher Multimedia applications Realtime wireless communication Sensors adhoc networks 39 Energy is an essential and occasionally the main constraint for many current and emerging computer systemsapplications GMU CS 795 Lowpower Computing Research 39 A partial search returns 19726 lowpowerrelated scientific publications in IEEE digital library 20002008 A dominant theme in RealTimeEmbedded Systems Design Mobile Computing and Networking conferences o RTSS RTAS EMSOFI39 ECRTS o DAC ISPLED ICCAD 0 MOBIHOC MOBICOM INFOCOM SENSYS MOBISYS 0 Various journals numerous master and PhD theses GMU CS 795 Industry Interest 39 Many research labs are actively pursuing powerenergy management projects 39 ACPI Advanced Configuration and Power Interface is an open industry specification codeveloped by Hewlett Packard Intel Microsoft Phoenix and Toshiba httpwwwacpiinfo 39 In 2005 EEMBC Embedded Microprocessor Benchmark Consortium added the energy consumption metric to the performance scores it provides for embedded processors tested against its applicationfocused benchmarks 39 International Technology Roadmap for Semiconductors lists Power Management as one of the nearterm grand challenges GMU es 795 19 LowPower Computing Implications for High Performance Servers 39 Servers connected to the electrical power grid no battery problem But powerenergy management becomes increasingly important for 0 Standalone servers 0 Server clusters Energy consumption Peak power determines the heat dissipation which in turn requires expensive cooling mechanisms GMU CS 795 110 LowPower Computing Implications for High Performance Servers 39 High powerenergy consumption 0 An average server power consumption figure is around 200 Watts 0 Internet and hosting centers often have thousands of servers Overall electricity used by servers in US doubled from 2000 to 2005 The total energy consumed by servers was more than 21 states in US in 2005 Also peak power directly translates to heat Approximately 2530 of expenses are related to energy cooling By 2010 the energy costs will emerge as the second highest operatingco1st11 GMU CS 795 LowPower Computing Dimensions 39 Target system components 0 CPU 0 Memory system 0 Disk and IO Subsystems also network interfaces and wireless radio 39 Design levels 0 Circuit and microarchitecture 0 Compiler 0 Operating system 0 Network protocols 0 Applications GMU CS 795 112 Tentative Discussion Topics 39 OSlevel power management techniques 0 Poweraware scheduling dynamic voltage scaling 0 Adaptations for multimedia applications 0 Poweraware memory and IO device management 39 Power management for multiprocessor systems 39 Powertemperature management in highperformance computing and server farms GMU CS 795 113 Tentative Discussion To ics Cont 39 Power management in wireless adhoc and sensor networks 0 Poweraware Routing 0 Topology Control 0 Sleep Schedules 0 Clustering Algorithms 0 Poweraware Multicast 39 Interplay of Fault Tolerance and Power Management 39 Interplay of Security and Power Management 39 Controltheory and gametheoryinspired techniques for Power Management 39 Power Management for Wearable and Pervasive Computing 39 Energy Harvesting techniques GMU CS 795 114 Power and Performance 39 Moore s Law the number of transistors on a microprocessor doubles about every 18 months with a similar increase in performancespeed 39 Has remained valid in the past 40 years 39 Are we approaching or did we reach the end GMU CS 795 115 Exponential Increase in Power Density 1000 Nuclear Reactor p 9r u 100 i W V E quot Wig 2 y in v E Hotplate s Pentium l g 10 Pentlum l r v 9Pentium Pro vPentium 9 O 1 I 15 1 07 us 035 015 on 013 uh Wquot From Fred Pollack s keynote address at Micro32 conference GMU CS 795 116 Power and Energy 39 Various terms 0 Poweraware energyaware 0 Powerefficient energyefficient 0 Power management energy management 0 Lowpower lowenergy Energy is the aggregate power consumption over a time interval 39 Should be aware of distinctions 0 Minimizing the power consumption is not equivalent to minimizing the energy consumption 0 Minimizing the peak power is not equivalent to minimizing the average power GMU CS 795 117 Energyefficient Operation 39 Aim Maximize energy savings while satisfying some performance constraint 9 Powerawareness is the key 39 How to save power 0 Switch to standbysleep modes when possible Also known as Dynamic Power Management 0PM Should consider transition overheads 0 Switch to lowpowerlowperformance modes when possible 0 Use applicationlevel adaptations GMU CS 795 Componentlevel Power Management 39 Many power management techniques are based on adjusting control knobs 0 Different system componentsdevices can offer multiple operation modes with different powerperformance characteristics 0 Adjust speed memory access latency sleep mode GMU CS 795 Componentlevel Power Management 39 Adjusting control knobs involves determining the transition strategy 0 When to switch to a different mode 0 Which mode to switch to Standby idle deep sleep Medium speed low speed 0 Can we amortize the overhead of mode transitionquot GMU CS 795 Exploiting performance tradeoffs 39 OS and networklevel power management techniques usually exploit performance tradeoffs 0 Trade response time for power but still meet the deadlines 0 Trade transmission power range for power use multiple hops but still deliver the packet to the destination or keep the network connected 0 Trade the frame rate for power but still deliver the entire video stream to the destination GMU CS 795 121 EnergyConstrained Operation Missionoriented energy is a m constraint The system has a fixed energy budget Must remain functional during a mission operation flight meeting commute Make the best use of available energy to maximize some performance metric 0 Give preference to more important taskspacketsfra mes 0 Minimize the quality degradation GMU CS 795 122 CPU Power Management 39 Power consumption in CMOS circuits has two components 0 Dynamic Power Ce Vddzf Ce Effective switching capacitance Vdd The supply voltage f The clock frequency 0 StaticLeakage Power 39 Most of the early research assumed the staticlea kage power negligible GMU CS 795 123 D namic Volta e Scalin DVS Idea To reduce dynamic power dissipation adjust the supply voltage and the clock frequency onthefly Convex relationshi between the speed and CPU power consumption usua ly quadratic or cubic We will assume that the speed values are normalized SW 10 GMU CS 795 124 D namic Volta e Scalin DVS DVSenabled processors are now commercially available I Transmeta CrusoeEfficeon I Intel XScaIe architecture I AND K6K7 PowerNow Specs for Transmeta TM 5400 Processor Crusoel Frequency MHZ Voltage V Relative Power 700 165 100 600 160 81 500 150 59 400 140 41 300 125 24 200 110 12 GMU cs 795 125 DVS Res onse time Ener tradeoff execution energy time power consumed 1 64 64 2 16 32 3 7 21 4 4 16 5 25 125 39 RT DVS problem Determine task speed assignments so as to minimize total energy consumption WHILE still meeting the deadlines preserving the feasibility 0 Requires the control of the operating system 0 Subject to extensive research in recent years GMU CS 795 126 RT DVS Solutions depend on the task and system model as well as scheduling policy Many realtime tasks are peroa l39cin nature 0 Multimedia processing networking 0 Controlsensing tasks GMU CS 795 127 Periodic RT Task Model 39 Each task 7 has a period P equal to relative deadline D l l l l l l l l l l l l Under what condition can all the deadlines be met The worstcase execution time of task Ti Ci The utilization of task Ti ui Ci Pi The total utilization Uh0t 2 Ci Pi A wellknown result of realtime scheduling theory states that the feasibility will be preserved with EarliestDeadlineFirst EDF policy if and only if Utot 10 GMU CS 795 128 RTDVS for Periodic Task Model In variablespeed settings the worstcase workload of a task Ti depends on the current task speed Si 0 The worstcase execution time under max speed Ci 0 The worstcase execution time under speed Si Ci Si The nominal utilization ui Ci Pi The total nominal utilization Uh0t 2 Ci Pi What are the optimal speed assignments to individual tasks to minimize the energy consumption while meeting all the deadlines GMU CS 795 129 Solving Periodic RTDVS for Single Processor 39 For a periodic task set the speed that minimizes the total energy consumption while meeting all the deadlines is the same for all the task instances Sopt Z 39 EarliestDeadlineFirst EDF policy can be used to obtain a feasible schedule with SDIDt 39 Observe 0 Sopt Uh0t Utilization under maximum speed 9 Fully utilized schedules GMU CS 795 130 Static Optimal Solution Example P1 P2 10 C1 C2 2 P3 30 C3 3 SmDt 05 T o 2 1o 12 20 22 30 No Power I i E i D i Control T2 0 4 1o 14 20 24 39 T 4 3 0 7 30 T1 Static Optimal 1 1 3 2 3 ScheduleSopt05 T2 i o s 10 1s 20 28 30 T 3 0 1o 20 30 GMU cs 795 131 Three dimensions of RTDVS 1 Static Optimal Solution Optimal speed assignments assuming the worstcase workload 2 Dynamic Reclaiming Detect early completions and adapt to the actual workload by dynamically reducing the CPU speed 3 Aggressive Speed Reduction Using the expected workload information speculativeh reduce the CPU speed The components should be designed and combined so as to preserve the feaSIbIIW of the schedule GMU CS 795 132 Conseguences of Na39ive Reclamation T1 Static Optimal 0 I4 14 Schedule sopto5 T2 0 8 1390 18 20 2s 36 T3 IEED 0 10 20 30 0 4 10 18 2s 30 Actual T2 g Execution 0 8 1b 14 22 30 T Qeadlmeiviigs 3 0 10 N completes arly GMU cs 795 133 Aggressive Speed Reduction Make the common case more efficient S 05 173 i T3 Static Optimal s 03933 Til Average Case 7 WorstCase s103933 s2s31 deadline Must still guarantee the feasibiligt if needed I We may have to use high the speed if speculation fails The best aggressiveness level GMU cs 795 134 RT DVS for Multiprocessors 39 ENERGYPARTITION Problem Given m homogenous processors each equipped with DVS capability where the power consumption function is gS and a set of periodic realtime tasks Find a tasktoprocessor assignment and the speed of each CPU such that the feasibilig is guaranteed and the total energy consumption is minimized 39 The solution to this problem is energyoptimal partitioning 9 task assignment affects the energy consumption GMU CS 795 135 Example 39 Three tasks to be scheduled on 2 processors u1 05 u2 025 u3 015 gSS2 075 09 E 081 E E 0585 GMU CS 795 136 Some RT DVS research issues 39 Staticleakage power dissipation is gaining importance what are the implications for the existing DVS schemes 39 Different taskprocessorsystem models 39 If the system is energyconstrained then how to select tasks for execution to provide an acceptable performance 39 What if the system has both realtime and non realtime tasks Energy Response Time 39 Effect of DVS on other system components 9 Towards a global view of energy management GMU CS 795 137 Power Management for Memory SubSystem 39 Memory is an important power consumer 0 Small handheld and embedded devices 0 High performance servers 39 Effects of the various power management techniques on memory power consumption is an active research topic GMU CS 795 138 Power Management for Memom SubSystem Main technigues 39 Exploit the newlyavailable multiple power states of the memory chips 0 Faster access higher power consumption 0 Limited number of sleep levels 39 Use the operating system to make page allocation in powerefficient manner 0 Cluster accesses to active chip units 39 HardwareSoftware hybrid solutions GMU CS 795 139 Rambus DRAM Power States Active 300mW 5ns Standby 180mW GMU CS 795 140 PowerAware Main Memory When to switch from one state to another Software control All chips subject to the same threshold times Hardware control Virtual Memory Page Allocation can 39 make a significant Active Standby Power impact Down GMU CS 795 141 PowerAware Page Allocation 39 Random Allocation Conventional Solution 39 eguential FirstTouch Allocation Uses minimal number of chips 39 reguencybased Allocation Allows page migration between chips GMU CS 795 142 Power Management for Memory Subsystem Issues 39 Delay versus Energy tradeoffs 39 Energy Delay product 39 How to choose goodadaptive thresholds 39 How do the poweraware memory management and DVS interact GMU CS 795 Power management for Disk Subsystem 39 One of the earliest targets for Power Management still important for mediumlarge systems Disk SpinDown algorithms have been around for more than 10 years still open issues 0 Turnoff the unit when the disk has remained idle for an inactivity threshold time 0 Spinup time 16 s 0 Spinup energy 530 J 0 Spinning down as soon as the disk is idle yields poor performance GMU CS 795 Power management for Disk Subsystem 39 Choosing the inactivity threshold 39 Fixed 0 110 seconds give the best results 0 Double energy savings with respect to 35 minutes thresholds Short inactivity thresholds result in perceived increased delay 0 Frequent spin upspin down cycles may result in premature disk failure 39 Competitive analysis Any deterministic threshold can fail against a nasty series of disk access patterns GMU CS 795 145 Power management for Disk Subsystem 39 Variable threshold 0 Disk access patterns may change over time from process to process 0 Keep a list of candidate thresholds 0 After each disk access candidates weights are increaseddecreased according to how well they would have performed relative to the optimal on line strategy over the last period 39 Alternative is to avoid an explicit inactivity threshold 9 Predict the actual time of the next disk access to determine when to spin down the disk GMU CS 795 146 Power management for Disk Subsystem 39 Other techniques 0 Change the configuration or usage of the disk cache 0 Use prefetching Fill the disk cache with data that is likely to be requested in the future before spin down 0 Reduce paging activity Smaller working set sizes and better memory access locality GMU CS 795 Power management for Disk Subsystem 39 Evaluation techniques 39 Competitive analysis performance against a clairvoyantoptimaloffline policy Average case through simulations or actual system traces Performance metric 0 Energy 0 Incorporate Delay 0 Incorporate the number of spinupspindowns 0 Effect of networking GMU CS 795 148 DVS for Systemlevel Power Management 39 The effect of DVS on Systemlevel Power Management 39 DVS essentially targets the CPU or onchip power which is frequencydependent 39 Consider a task that uses multiple IO devices during its execution these devices will also consume power which is typically frequencyindependent 39 At lowfrequency levels the overall energy can increase GMU CS 795 149 Task Energy Consumption Energy Consumption P Ji I I 339 K I I I 03 04 05 06 07 08 09 1 Task Speed Energyefficient speed 39 In general there is a speed level called energyef cient speeaQ below which DVS increases overall energy consumption 39 The exact value of energyefficient speed depends on task and system characteristics 39 How do the energyoptimal speed computations change when considering the energyefficient speed thresholds GMU CS 795 151 Power Management for HighPerformance Servers 39 Servers are designed and provisioned for peak load 9 high performance 9 high power consumption 39 Busy servers cannot switch the hardware components to lowpower modes 39 Even when the load is light a component such as disk needs to remain inactive for a long time to amortize transition costs GMU CS 795 152 Strategies for StandAlone Servers 39 DVS when applicable 39 Request batching 0 Incoming requests are accumulated in memory by the network interface processor while the host processor is kept in a lowpower state 39 Multispeed disks 39 Popular Data Concentration PDC technique migrates the most popular disk data to a subset of disks GMU CS 795 153 Strategies for StandAlone Servers Throttling 39 Aim To reduce the the peak power consumption consequently the thermal dissipation 39 The energy consumption is monitored periodically 39 If the processor consumes more energy than allowed halt cycles are introduced to put the CPU to the lowpower state GMU CS 795 154 CS 795 004 LowPower Computing Lecture 2 George Mason University Spring 2009 Contents 39 Wireless communication and networks 39 Adhoc and sensor networks 39 Wireless LANs 39 IEEE 80211 standard 39 Power management techniques GMU cs 795 22 GMU CS Wireless Networks 39 Great potential for 0 Mobile Computing 0 Pervasive Computing Various types 0 Cellular systems 0 Satellite systems 0 Broadcast systems 0 Wireless LANs 0 Adhoc and sensor networks 795 23 GMU CS 795 Wireless Communication Negative characteristics 0 Low bandwidth 0 Broadcast medium increased number of collisionsinterference 0 High error rates 0 High delays large delay variations 0 Security Signal Propagation 39 Pathloss modeling 39 How is the radio signal strength affected at the sender 39 Many parameters are important 0 The transmitter and receiver antenna gains 0 The wavelength of the carrier 0 The distance between the sender and the receiver GMU cs 795 25 Si nal Pro a ation 39 In general the received signal power at the receiver Pr varies with the distance d in an exponential fashion Pr P0 da where P0 is the received power at a reference distance from the transmitter 39 The exponent a depends on the environment 0 a 2 for freespace 0 a 4 for urban radio channels GMU cs 795 26 Signal Propagation 39 Receiving power influenced by 0 shadowing 0 reflection at large obstacles 0 refraction depending on the density of a medium 0 scattering at small obstacles 0 diffraction at edges fKn shadowing re ection refraction scattering diffraction GMU cs 795 27 Multipath propagation gt signal at sender 39 Signal can take many different paths between sender and receiver due to reflection scattering diffraction 39 The original signal is spread due to different delays GMU cs 795 28 CSMA in wireless environment 39 CSMA Carrier Sense Multiple Access Listen for other transmissions and only transmit if no other node is transmitting 39 Basis of many successful protocols for wired networks 39 Problem CSMA merely tells whether there is activity around the station sensing the carrier 39 In radio communication what matters is the activity around the receiver GMU ics 795 1 9 Hidden and ex osed terminal roblems l l WA E E are E if m le al 1 a Hidden terminal problem C starts transmitting causing collision at B b Exgosed terminal problem C defers sending a packet to D thinking that the collision may occur omuicsws 1 l MACA Multiple Access with Collision Avoidance 39 Proposed by Karn in 1990 39 Idea To reduce the number of collisions stimulate the receiver into sending a short frame so that the stations nearby can detect this transmission and avoid transmitting for the duration of the upcoming large data frame I The sender starts by sending an RTS Request to Send frame to B I B replies with a CTS Clear to Send frame I Any station hearing CTS will be silent during the upcoming data transmission GMU705795 1 ll The MACA protocol Ranger of A39s quotmamquoter Range cl E s hansmiller E IE 8 a A sending an RTS to B b B responding with a CTS to A CTS GMU705795 1 l1 MACAW MACA for Wireless 39 Bhargavan et al 1994 39 To speed up the retransmissions an ACK frame is added 39 Carrier sensing is added to further reduce the collisions 39 Exponential backoff algorithm run separately for each data stream sourcedestination pair rather than for each station GMU CS 795 213 Infrastructure versus AdHoc networks y Infrastructure M network A 539 x lquot AP Access Point wired network l i a A I y ad hgg39 iWg gw 2 4 5 J xquot 3 GMU CS 795 214 AdHoc Networks 39 A collection of wireless mobile nodes that selfconfigure to form a network without the aid of any established infrastructure 39 Applications 0 Data networks 0 Home networks 0 Device networks 0 Sensor networks 0 Distributed Control Systems GMU CS 795 215 AdHoc Networks I ISSUES 0 Neighbor discovery 0 Network connectivity 0 Routing 0 QoS 0 PowerZEnergy GMU CS 795 216 Sensor Networks 39 Network of sensor nodes that are densely deployed either inside a phenomenon or very close to it 39 A Berkeley MICA mote 39 Small in size Approximately 2 x 15 x 05 inches 39 Uses 2 AA batteries GMU CS 795 217 Sensor Networks 39 Each sensor is comprised of 0 A computing subsystem a microprocessor or microcontroller 0 A communication subsystem short range radio for wireless 0 A sensing subsystem group of sensors andor actuators 0 A power supply subsystem battery and DCDC converter GMU CS 795 218 A general picture worldwide user local area omuicsws 1 l9 Sensor Networks 39 The number of sensor nodes in a sensor networks can be several orders of magnitude higher than the nodes in other types of ad hoc networks 39 Densely populated 39 Prone to failures 39 Sensor nodes are verylimited in power computational capacities and memory GMU 5795 1 1 Sensor Networks 39 Surveillance tracking control 39 Habitat monitoring disasteraccident precautions 39 Managing inventory monitoring product quality 39 Need for aggregation 39 Existence of base station 39 Sensing and transmission ranges GMU cs 795 221 IEEE 80211 standard 39 Developed by IEEE as the wireless LAN standard 39 Also known as WiFi 39 Distinguishes two modes infrastructure and adhoc 39 Multiple variants with different modulation techniques at the physical layer still open issues GMU cs 795 222 IEEE 80211 variants 39 80211a operates in 5 GHz band delivers up to 54 Mbps 39 80211b operates in 24 GHz band delivers up to 11 Mbps 39 802119 operates in 24 GHz band delivers up to 54 Mbps GMU cs 795 223 Connectin wired and wireless networks xed 1 terminal Infrastructure mobile terminal network access point application application TCP TCP IP IP LLC LLC LLC 80211 MAC 80211 MAC 8023 MAC 8023 MAC 80211 PHY 80211 PHY 8023 PHY 8023 PHY GMU cs 795 224 IEEE 80211 CSMAZCA contention window randomized backoff mechanism IFS F 0 Starts to send only if the medium is free for the duration of an InterFrame Space IFS 0 If the medium is busy the station waits for an IFS then a random backoff time 0 If another station starts transmitting during the back off time of the station the backoff timer stops medium busy next frame direct access if medium is free 2 IFS slot time GMU cs 795 225 IEEE 80211 CSMAZCA 39 Priorities 0 de ned through different inter frame spaces 0 SIFS Short Inter Frame Spacing highest priority for ACK CTS polling response PIFS PCF IFS medium priority for timebounded service using PCF 0 DIFS DCF Distributed Coordination Function IFS lowest priority for asynchronous data service DIFS Er 39 ect access if medium is free 2 DIFS GMU cs 795 226 IEEE 80211 CSMAZCA 39 Sending unicast packets 0 The receiver acknowledges after waiting for SIFS if the packet was received correctly 0 Automatic retransmission in case of transmission errors sender receiver other stations waiting time contention GMU cs 795 227 IEEE 80211 CSMAZCA 39 Sending unicast packets 0 The sender transmits RTS with reservation parameter 0 The receiver replies with CTS 0 Other stations store medium reservations distributed via RTS and CT 5 sender receiver other stations defer access contention GMU cs 795 228 Fragmentation harshPr other stations contention GMU cs 795 229 Power Management in IEEE 80211 39 Idle listening mode of wireless cards consumes signi cant power 39 Cabletron 80211 card 0 Transmit 1400 mW 0 Receive 1000 mW 0 Idle 830 mW 0 Sleeping 130 mW GMU cs 795 230 Power Management in IEEE 80211 39 Switch off the transceiver when it is not needed 39 A station may be in either sleep or awake state Data is buffered at the sender s side 39 The sleeping station has to wake up periodically and stay awake for a certain time During this time all senders can announce the destinations of their buffered data frames 39 A station hearing its id for buffered data has to stay awake GMU cs 795 231 Power Management in IEEE 80211 39 Infrastructure networks 0 The access point buffers all frames to be sent to sleeping nodes 0 The access point along with every beacon signal transmits a traffic indication map TIM 39 Adhoc networks 0 All stations compete to announce a list of their buffered frames during ATIM window Adhoc traffic indication map 0 Relies on distributed synchronization function through beacon signals GMU cs 795 232 Power Management Technigues 39 Sleep Schedules 39 Topology Control 39 PowerAware Routing 39 MulticastBroadcast 39 Others GMU cs 795 233 Sleep Schedules 39 IEEE 80211 base Power Management algorithm is only one of the options 39 Nodes can determine their sleep schedules 0 In a coordinated or uncoordinated way 0 Assuming or not assuming clock synchronization 0 Deterministically or probabilistically 0 Statically or dynamically GMU cs 795 234 Sleep Schedules 39 Excellent prospects for energy saving 39 Also reduces the collisions 39 Potential Problems 0 Latency 0 Network partitioning 0 Buffer overflow 0 Insufficient coverage for sensors GMU cs 795 235 Sleep Schedules 39 Possible performance metrics 0 Keep the network connected at all times 0 Cover all the regions at all times with awake nodes for sensors 0 Have an awake coordinatorcluster head in all interesting regions at all times 39 Different interpretations of network lifetime GMU cs 795 236 To olo Control 39 Nodes can adjust their transmission power to determine the network topology 39 Power consumption grows exponentially with the transmission power 39 The topology will also determine the network capacity 39 Effect on network lifetime GMU CS 795 To olo Control 39 Example problem if we have to choose a unique transmission power for all nodes what should it be to maximize network lifetime while keeping it connected 39 Interplay with sleep schedules GMU cs 795 238 PowerAware Routing 39 If nodes can dynamically adjust their transmission power how to route packets in the adhoc network to save power 39 Use multiple hopes overall energy consumption will be reduced 39 Localized routing algorithms depend only on information about the location of the source destination and immediate neighbors GMU CS 795 PowerAware Routing 39 Possible performance objectives 0 Energy consumed per packet 0 Time to network partition 0 Variance in energy levels across nodes 0 Cost per packet where cost function may consider the energy levels of individual nodes GMU cs 795 240
Are you sure you want to buy this material for
You're already Subscribed!
Looks like you've already subscribed to StudySoup, you won't need to purchase another subscription to get this material. To access this material simply click 'View Full Document'