Alexey Rudenko` Peter Reiher` Gerald J. Popek`() Geoffrey H
Contents
1. s battery power is drained by perform ing tasks for the user Some of these by their nature must be performed locally For example information must be dis played on the local screen Other tasks however could be performed anywhere provided the results came back to the portable computer If the power cost of sending the task else where and receiving the results back is lower than the cost of This work was partially supported by the Advanced Research Agency Projects under contract DABT63 94 C 0080 Male Compading ond Conincs Rew Vole 2 Miber running it locally remote process execution could save battery life for the portable computer This possibility makes little sense without wireless devices If the user can plug in a wired Ethernet card he could probably also plug in a power cord But an untethered user can still communicate via wireless networks One realistic scenario is an office in which untethered users can move around through a ubiquitous wireless network migrating processes to server machines that have no power constraints In its simplest form remote process execution for power management would involve moving new tasks from the portable computer to a server machine before the task start running The server would execute the task and ship the results back to the portable computer In the meantime the portable computer would continue running other tasks going idle if nothing else is to be done If the user ru2. Other devices such as the Metri com communications card have reduced these figures by a sig nificant amount The Metricom card expends approximately 1 watt 4 watts and 1 watt for the same modes as the Wave lan card 22 but unfortunately at the cost of a much lower data rate than the WaveLAN Proxim RDA radios spend 0 375 watts in transmitting and receiving modes and 0 001 watt in standby mode with a data rate of 0 242 Mb sec 16 Future improvements can be expected One particular area for improvement is minimizing the amount of power spent by the local card in the power costly receiving mode The use of a broadcast medium and the re quirement that the portable computer receive messages that it might not be expecting which is very useful for any number 24 of purposes mean that some power must be expended exam ining all incoming packets to see whether they are destined for the local node Designs that are able to synchronize communi cation and allow the local card to spend some time in sleep mode without the loss of messages have promise for mini mizing these power costs Such designs will not only benefit power consumption for remote execution but will generally allow portable computers with wireless cards to remain truly idle at minimal power cost in a shared wireless network envi ronment Communication cards that are able to remain alert even when the portable itself is in suspend mode waking only when its own message
3. They represent tasks that are typically performed by ordinary users every day Thus assum ing a suitable environment remote execution has promise for providing better battery life for future users This technique is largely orthogonal to other power saving techniques adding to any benefits they provide The savings shown in this study are not by any means the maximum savings possible Larger tasks are likely to benefit even more More efficient ways of moving the data may also provide greater savings Improvements in wireless devices and power management will provide further benefits Preliminary experiments under more optimistic conditions have shown up to a five fold increase in battery life 25 References 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 26 AT amp T Corp WaveLAN PCMCIA User s Guide Daniel Barbar and Tomasz Imielinski Sleepers and Worka holics Caching Strategies in Mobile Environments ACM SIGMOD International Conference on Management of Data Minneapolis Minnesota May 1994 James W Davis Power Benchmark Strategy for Systems Em ploying Power Management JEEE International Symposium on Electronics and Environment 1993 Fred Douglis P Krishnan Brian March Thwarting the Power Hungry Disk USENIX Conference Proceedings Win ter 1994 Duracell Corp Intel Corp Smart Battery Data Specification Revision
4. behavior or limit their use of the portable computer to pre serve the battery s charge Any user whose portable computer has run out of power while in the middle of a long air flight understands the impact of insufficient batteries If the battery s power capacity cannot be improved the other alternative is to find ways to use less power preferably with no impact on the user Many researchers have looked at this problem 3 14 19 Solutions range from intelligent management of the disk and screen 4 7 8 13 to slow ing down the CPU clock rate 6 23 or powering down com ponents of the computer not currently in use Many of these innovations have already found their way into commercial use a strong indication of the importance of the problem Wireless communication devices are becoming increasingly common in portable computers since they help solve one of the other fundamental problems of portable computing that is when wired connections are unavailable wireless devices can maintain network connectivity allowing remote file ac cess sending and receiving of email and web browsing Gen erally wireless communication devices are regarded as con tributing to the power management problem rather than the solution as they use significant power when sending and re ceiving However this paper demonstrates that wireless de vices can sometimes be used to save a significant amount of battery power A portable computer
5. in the matrix The entire matrix had to be shipped and solution vector returned in the case of remote execution For relatively small matrices less than 500x500 the costs of moving the computation were greater than the benefits or the results were statistically indistinguishable But for larger ma trices the savings were as great as 45 Unlike compilation the Gaussian application performed very little disk I O The savings shown on a large Gaussian solution thus demonstrate that performing tasks remotely can offer benefits even if the task is largely compute bound Trans portation expenses for the shipment of the data grow slower than the workload because they are quadratic with respect to the size of problem while the workload grows at a rate of O N3 where N is the number of simultaneous equations Figure 4 shows the results of migrating a text formatting application This application used LaTeX to format 30 to 200 page documents containing multiple figures references and equations making it a moderately large text formatting process Remote execution did not improve the power con 3The regression equation is y fe 4 4 for local execution and y 172 4 for remote execution The F values are 0 99 and 0 99 respectively 4The regression equation is y 0 24x 0 for local execution and y 0 2x 0 1 for remote execution The 2 values are 0 97 and 0 84 respectively 145 12 5 Local 10 5 Remote Pe
6. so there were signif icant power costs to not turning the card off However if the card is turned off timeout or human intervention would be re quired to turn it back on when the results were to be shipped back from the server Doing so would have been neither real istic nor convenient for running the experiments Measurement of the power consumed by a task caused some problems The most accurate way to measure power consump tion would have been to insert appropriate electronic instru mentation between the battery and the computer it was driv ing Practical problems with such instrumentation suggested the use of less direct methods The readily available metrics for power currently in a battery are rather unreliable The Dell Latitudes used in these experiments use the Advanced Power Management APM tools which will report a battery s charge as a percentage of maximum However the APM measure ments are not very reliable In some cases the amount of power reported by APM will go up over time even though the machine has remained untethered In addition the amount of power expended to perform a particular task as reported by APM varies widely Mote Compading and Commuiucalons Review Vole 2 Minier We considered two options for measuring the amount of power consumed by a task First we could interrogate the APM metric before the task interrogate it again after the task completes and report the difference Alternately we cou
7. speed with limited negative im pact on performance The authors of 2 use predictive caching to reduce contention on the narrow bandwidth wireless chan nel consuming less power and allowing a mobile laptop to keep working in circumstances of long and frequent discon nections 6 proposes using idle periods to achieve a fairer Maule Compading ond Commuuuedions Rew Vole 2 Miber distribution of the workloads of busy periods This paper also includes a taxonomy of idle detection algorithms and idleness predictors The idle periods can be used to run some tasks whose results are needed in future as well as for the discon nection of unused power costly devices Wireless communication devices appear to be highly power consumptive 20 discusses a technique of transmission sus pension at the moment when interference in the channel is de tected It is presumed that interference is stationary and er godic During the interference the communication device can be suspended and power consumption by this device reduced 9 considered wireless data broadcasting as a way of dissem inating information to a massive number of clients equipped with battery powered laptops The user must periodically lis ten to the channel to obtain a consistent schedule of the data that will be transmitted in the near future At other times the user can disconnect his communication devices and save power Laptop and battery manufacturers have provided anoth
8. there is a significant difference in the client portable computer s power consumption for actual remote execution versus pure transmis sion of data and results Figure 2 shows that the amount of power expended purely on waiting is quite significant Other experiments done earlier in our work suggested that if the user manually turns the portable completely off and turns it back on again when the computation is complete the amount of power consumed by the portable is much much lower Of course this is not a practical way to operate in the real world but if the power saving mode can come closer to expending zero power remote execution will show more benefits 4 Improving Power Management The high cost of pure data transmission shown in Figures and 6 and the cost of listening to the channel suggest that certain improvements in wireless communications devices could also improve the performance of remote execution Wavelan wire less cards and wireless communications cards in general are fairly new phenomena that are not yet ubiquitous As a result they have not been heavily optimized The Wavelan card for example operates in three modes transmitting receiving and sleeping These modes take approximately 3 watts 1 48 watts and 18 watts respectively 1 The transmission and receiving modes consume as much power as a typical disk drive which is known to be one of the most power consumtive devices in a typical portable computer
9. 1 0 February 15 1995 Richard Golding Peter Bosch Carl Staelin Tim Sullivan and John Wilkes Idleness is Not Sloth HP Laboratories Techni cal Report HPL CCD 95 1 Palo Alto CA January 1995 Paul M Greenawalt Modeling Power Management for Hard Disks Mascots 94 International Workshop on Modeling Analysis and Simulation of Computer and Telecommunication Systems June 1994 David P Helmbold Darrell D E Long and Bruce Sherrod A Dynamic Disk Spin down Technique for Mobile Computing Proceedings of the Second Annual ACM International Confer ence on Mobile Computing and Networking Rye NY Novem ber 1996 Tomasz Imielinski S Viswnathan B R Badrinath Energy Efficient Indexing on Air Proceedings of ACM SIGMOD Conference on Management of Data May 1994 Intel Corp Pentium Processor 610 75 Power Consumption Revision 1 1 Technical Marketing October 1994 Intel amp Microsoft Corp Advanced Power Management APM BIOS Interface Specification Revision 1 2 February 1996 Deitmar A Kottmann Ralph Wittman and Markus Posur Del egating Remote Operation Execution in a Mobile Computing Environment Mobile Networks and Applications Vol 1 No 4 pp 387 397 January 1997 Kester Li Roger Kumpf Paul Horton and Thomas Anderson A Quantitative Analysis of Disk Drive Power Management in Portable Computers USENIX Conference Proceedings Win ter 1994 Brian Marsh and Bruce Z
10. 2 or more of the total battery power leaving more room for paying an up front penalty to reduce the overall costs The results presented so far were performed on an otherwise unused medium The only traffic in the Wavelan frequencies in the testing area was generated by the migration of tasks and results In a realistic environment the Wavelan frequencies would be used for other tasks including other machines also trying to execute processes remotely We performed further experiments to determine the power saving characteristics of large applications in the presence of noise on the wireless net 22 Mode Compading md Connmuvucrhons Renew Vole 2 Minier 145 eS Local 10 Remote Percentage g of Battery Consumed 6 4 J 2 4 l e 0 c 0 125 250 375 Number of Kilobytes Changed 500 Figure 5 Power savings for remote execution of a compilation with noise in background work However since text formatting did not show improve ments in the noiseless case we did not perform experiments for that application in the noisy environment E Experimental data in nolzy environment Network noise has many different characteristics Assuming multiple senders and receivers and multiple patterns of traffic presented to the network a vast number of different exper iments are possible Our goal was to examine a reasonable case not to exhaustively examine the
11. Saving Portable Computer Batiery Power through Remote Process Evecution Alexey Rudenko arudenko fmg cs ucla edu Peter Reiher reiher fmg cs ucla edu Gerald J Popek popek platinum com Geoffrey H Kuenning geoff fmg cs ucla edu Computer Science Department University of California Los Angeles CA USA Platinum technology inc Inglewood CA USA We describe a new approach to power saving and battery life extension on an untethered laptop through wireless remote processing of power costly tasks We ran a series of experiments comparing the power consumption of processes run locally with that of the same processes run remotely We examined the trade off between communication power expenditures and the power cost of local processing This paper describes our methodology and results of our experiments We suggest ways to further improve this approach and outline a software design to support remote process execution IL Intreduction Power management is one of the most challenging problems in making portable computers more useful Portable computers have their greatest utility when they can truly be used any where at any time and one of the greatest limitations to that goal is battery power Often AC power connections are not available and the portable computer must run off its battery However the battery life of existing and expected batteries is not sufficient for many situations Users must either alter their
12. arrives would also be helpful E A Power Conserving Infrartructure Although power savings are possible through remote exe cution considerable work remains before these savings be come available to average users Our experiments used spe cial scripts that performed each migration and sent the results back Other scripts ensured that the server stored the appro priate data and that only the required data was sent from the client portable to the server However ordinary users cannot be expected to be able to create such tricky scripts themselves Simply using normal remote execution has disadvantages as well If the data files are stored on the portable it will spend power moving them to the remote machine If not the portable cannot operate disconnected Generally a transparent facility would be preferable Common use of remote execution for power management will require a user friendly infrastructure This infrastructure will require several important compo nents First it will require the ability to remotely execute a task and deliver the results back in an efficient way Prelimi nary tests have shown that the mechanism used to transport the data from client to server and back can have a dramatic impact on the power expended Second the infrastructure will require simple replication mechanisms that allow the client and server to synchronize replicas of the required data Such replication mechanisms will allow users to ignore the d
13. ause an undesirable noisy impact on power cost measurements This decision allowed us to isolate the actual costs and benefits of executing the tasks locally and re motely In normal conditions these daemons would slightly shorten the battery s life Existing power management techniques can have a dra matic impact on the amount of power consumed by a portable computer running off its battery so we controlled these tech niques carefully Our research showed that screen timeout and disk spindown were the most important power management tools We set the screen timeout and disk spindown intervals to one minute on the client machine We also allowed the client to go into idle mode during remote execution We did not per mit the client machine to suspend since the server would be unable to ship the results back to a suspended machine Since the server was connected to AC power it performed no power management Usually the client screen timeout would occur during local execution but disk spindown would not Typi cally both screen timeout and disk spindown occurred on the client during remote execution The client generally went into idle mode during remote execution These power management settings are similar to the ones our portable computer users typically set for normal situations We did not attempt to turn off the Wavelan card at any point in the experiment The Wavelan consumes 1 48 watts even when it is neither sending nor receiving
14. chitecture MIT Press 1985 and has written more than 70 professional articles concerned with computer secu rity system software and computer architecture Dr Popek was the founder of Locus Computing Corporation and is currently also the Chief of Technology Officer for PLATINUM technology inc Geoffrey H Kuenning received B S and M S degrees in computer science from Michigan State University and a Ph D in computer sci ence from UCLA He is currently a researcher in the File Mobility Group at UCLA Prior to joining UCLA he worked as a industry consultant in the fields of operating systems embedded systems and graphics Dr Kuenning has published over 10 papers on various sub jects He is currently performing research in mobile computing dis tributed file systems prediction and clustering methods Dr Kuen ning is a member of IEEE CPSR ACM and a number of ACM SIGs Mode Compading ond Commuiucalons Renew Vole 2 Minier
15. d in our laboratory The LaTeX texts were real papers and dissertations written in our laboratory As a result some file sizes are not round numbers A Nokel sz environment The first part of the experiments was run in a noiseless environ ment All major sources of noise like other laptops equipped with wireless cards were isolated from the room where the laptop experiments were performed Figure 1 shows the power consumption for local and remote execution of the compilation process The left bar of each pair shows the amount of power used for local execution of the task and the right bar for full remote execution The X axis shows the number of kilobytes of C source code that were altered The amount of work re quired to perform the compilation thus varies from point to point Also the amount of data shipped over the wireless link for remote execution varies with the amount of altered code since only altered modules were shipped In the remote execution case the server shipped back only stripped executa bles object files were not shipped In the local case stripped executables 6 9MB were saved directly to disk We per formed the same experiment with unstripped executables The results were qualitatively the same though the transmission costs for the larger executables made the percentage improve ments smaller 1 The regression equation is y 17 A for local execution and Y 1 22 4 for remote execution The 2
16. e 3 values are 0 99 0 97 and 0 88 respectively 21 145 12 Local Remote 10 E Transmission Percentage of 8 Battery Consumed N 1 HH b lini 250 375 125 500 Kilobytes of Changed Source Code Figure 2 Total power cost of local and remote execution com pared to transmission costs done on the server to verify through a make utility that no re compilation is required In the other cases however note that the transmission costs are half or less of the costs of the re mote execution The remainder of the costs occurs because of various inefficiencies in how the portable computer behaves while waiting for the compilation to complete For example its Wavelan card enters idle mode for this time but the idle mode still consumes about 1 48 watts of power The disk con tinues to spin for at least one minute after it was last accessed the screen displays an image for at least one more minute af ter the last key was pressed and some other devices must be considered in that total as well Thus if the computer had a more efficient power saving mode remote execution would have saved significantly more power bounded by the transmis sion costs shown in this figure Figure 3 shows the power saved by remote execution for the Gaussian solution of the system of algebraic equations In this case the size of the task is controlled by the number of rows
17. ecture that the power consumed by increased transmissions is similar in the normal remote execution case These costs are significant and they almost fully explain The regression equation is y 3 8x 2 0 for the execution in a noisy envi ronment and y 1 6x 0 48 for the execution in a noiseless environment The R3 values are 0 90 and 0 88 respectively 23 the increased costs for remote execution in the face of noise as shown in Figure 5 The remaining component of cost is extra battery power consumed in the noisy case while waiting for the server to complete This extra cost appeared to be almost negligible so the portable s cost of waiting for a unit of time is approximately the same in the noisy and noiseless cases Due to space limitations we do not show the results from the Gaussian elimination experiments with noise They are similar to the results from the compilation experiments IV Discussion and Future Work Our results show that remote execution of large tasks can re duce their power consumption by up to 50 Better power management features added to portable computers in the fu ture and wireless communications devices that consume less power will make the improvement even greater The power expended by the portable computer while waiting for results to come back could be minimized by having idle modes that are entered either on command or very aggressively and that consume very small amounts of power The fact that
18. emote execution and its savings were only around 20 as opposed to a 51 improvement without noise Clearly a large amount of noise in the environment has a major effect on the power savings achievable by remote execution 5The regression equation is Y 72 A for local execution and Y 1 32 A for remote execution in noisy environment The R values are 0 99 and 0 99 respectively Maule Compuding ond Cominus Rea Vole 2 Miber 6 5 No Noise Noise 4 al Percentage of Battery 3 4 Consumed 2 al 14 4 a T T 0 0 125 250 375 500 Number of Kilobytes Changed Figure 6 Power cost of transmission with and without noise for compilation Again the power consumed in this experiment is partly due to transmission costs and partly due to the power expended by the portable while waiting for its results Given that the portable and the server now have to contend for the Wavelan radio spectrum with a very heavy consumer more collisions and retransmissions occur while transmitting the remote exe cution data and results over the network thus increasing the amount of power consumed The remainder of power con sumed is spent by the portable while waiting for the server to complete its task This distinction is relevant because dif ferent techniques would be required to reduce the costs of each component Reducing the power costs of transmitting in a noisy environment wo
19. enel Power Measurements of Typi cal Notebook Computers Matsushita Information Technology Laboratory Technical Report MITL TR 110 94 May 1994 Shankar Narayanaswamy et al Application and Network Sup port for InfoPad IEEE Personal Communications Vol 3 No 2 pp 4 17 April 1996 Proxim Corp Spread Spectrum RF Modem User s Manual M Ranganathan Anurag Acharua Shamik D Sharma and Joel Saltz Network Aware Mobile Programs USENIX 1997 An nual Technical Conference January 1997 18 Michael Richmond and Michael Hitchens A New Process Mi gration Algorithm Operating Systems Review Vol 31 No 1 November 1997 19 Chris Ruemmler and John Wilkes UNIX Disk Access Pat terns Technical Report HPL 92 152 Computer Systems Lab oratory HP Laboratories Palo Alto 20 John M Rulnick and Nicholas Bambos Mobile Power Man agement for Maximum Battery Life in Wireless Communica tions Networks Proceedings of IEEE Infocom 1996 21 Yoichi Shinoda http spider fortune co jp wildboar Wildboar project 22 Mark Stemm Randy H Katz Measuring and Reducing En ergy Consumption of Network Interfaces in Hand Held De vices Proceedings of 3rd International Workshop on Mo bile Multimedia Communications MoMuC 3 Princeton NJ September 1996 23 Mark Weiser Brent Welch Alan Demers Scott Shenker Scheduling for Reduced CPU Energy USENIX Association Fi
20. entire realm of possi bilities However we did want a challenging case not a triv ial one since the noise free experiments had already given us best case results We chose to introduce two new machines in the environment One of these machines opened a socket to the other new machine and sent data down that socket as fast as it could We refer to this as saturated socket noise Since the Wavelan cards use an Ethernet style protocol where colli sions cause backoff and retransmission we expected that the interfering with communications would significantly impede attempts to move data and results but that it would not en tirely block those activities Of course it would be possible to introduce multiple pairs of communicating machines or to direct some of the traffic to either the portable computer being tested or its server Ei ther of these options would be expected to cause even more problems and either is a defensible realistic situation so in the future we may expand our experiments to include them Other variations in the noise experiments are also possible in cluding more realistic forms of communication between the noise making machines such as file transfer or attempts by these machines to perform their own remote executions Figure 5 presents the effects of saturated socket noise on the compilation task With this background noise only the largest compilation saved a statistically significant amount of power by using r
21. er ef fort towards reducing power consumption Several specifi cations 10 5 11 address issues of power consumption They are focused mainly on two points providing system functions allowing connection disconnection of any particu lar device from the power source and getting statistics about current status of power consumption in the system including remaining battery capacity These functions serve as a hard ware software basis for multiple packages such as Wildboar 21 which supplies a user with power management utilities that can be used in his scripts and applications 15 discussed a system designed around the InfoPad portable terminal a network I O device with no computation power relying on network servers to run major processes Lo cal computing is not possible here even when it is less power costly For many years process migration and delegation have been discussed in the computer community for various purposes dynamic load balancing improved reliability reduced network traffic 18 17 and 12 discuss these issues However power consumption has not previously been identified as a benefit of process migration VI Conclusions The experimental results presented in this paper demonstrate that portable computers that execute their large tasks remotely can save significant amounts of battery power Savings of up to 51 were observed While the tasks in our experiments were large they were not unrealistic
22. eth ered client machine and either running it locally or migrating 19 it to a tethered server machine In the latter case the server ran the application and shipped the results back to the client when completed This brief description ignores many important issues that might influence the outcome of the experiment Here we de scribe all the conditions under which we ran the experiments The results of the experiments are presented in Section 3 The experiments were conducted on Dell Latitude XP portable computers running the Linux operating system Dell s rechargeable Li ion battery rating is 14 4V 2200mAh 2A Both the client and the server were Dell Latitude XPs in this experiment In realistic situations the server would likely be a more powerful machine We chose to use the same machine for both client and server for two reasons First it was con venient for experimental purposes since the wireless devices available to us did not fit into any of our other machines and it allowed us to avoid issues of possible incompatibility between the client and server machines Second it was a conservative choice since in realistic circumstances the server would run faster thus causing the client to lose less power while waiting for its results The wireless device used for the experiment was the 915 MHz AT amp T Wavelan card 1 This network adapter con sumes 250 mW to achieve a data rate of 2 Mb sec at ranges up to 800 feet In the bas
23. ic experiments we took great care to insure that the Wavelan cards in the client and server were the only devices using that portion of the electromagnetic spec trum within range of each other The experiments were per formed in a location containing no other Wavelan cards or other wireless communications devices For the baseline re sults we needed to know whether remote execution could ever win even in rather favorable circumstances We added inter ference to the channel in later experiments as described in Section 3 2 We chose three applications for test cases These applica tions were chosen because they were known to be fairly large and time consuming making them good candidates for power savings by remote execution The three tasks were a compi lation of a large program text formatting of a 200 page docu ment using LaTeX and Gaussian solution of a system of linear algebraic equations The compilation task performed signifi cant CPU processing along with a good deal of disk activity to read the source write and read temporary files and write the resulting object and executable files The text formatting task performed a moderate amount of CPU processing but rel atively little disk activity The Gaussian elimination problem performed trivial amounts of file access but made very heavy demands on the CPU and on memory The size of the ma trix varied but was never large enough to cause significant amounts of virtual memory ac
24. ifficult issues of exactly which pieces of data need to be sent to the server Moreover by running part of the replication algorithms when the portable computer is tethered at least to power perhaps to a wired network the infrastructure can minimize the amount of data that would have to be sent to perform the task remotely Intu itively if an earlier replication operation had already moved all the data associated with one of the compilations in Figure 1 for example the transportation costs shown in Figure 2 would have been lowered Third proper remote execution of a job requires ensuring that all conditions at the server machine are the same as at the client For example if the server has an older version of a library than the client does the resulting program will be have differently if compiled remotely than if compiled locally Many other issues related both to the user s personal environ ment and preferences and to the general system environment on the two machines can make providing a consistent execu tion environment challenging We hope to obtain insight on this problem from projects that execute processes remotely for different purposes 12 17 18 These and similar projects have dealt successfully with these challenges Mode Compading and Commuiuedlons Renew Vole 2 Minier Fourth the infrastructure will need to assist the user in deter mining when to execute a task remotely As the results in Sec tion 3 sho
25. lation faster To isolate this effect we measured the costs of pure transmission of the required data and results We pre compiled the source code for each case and stored the results on the server When the client portable requested remote exe cution of the task instead of compiling we shipped the saved results back immediately This experiment shows the effect of an extremely powerful server and gives some insight into the amount of power the portable computer uses for data transport versus the amount of power spent waiting idly for the server to complete the task Figure 2 shows the difference between the power consumed by local execution by remote execution and by simply ship ping data and results back and forth using the compilation ap plication shown in Figure 1 The left bar of each group shows the amount of power used for local execution of the task the middle bar for full remote execution and the right bar for sim ply shipping the data and results over the wireless network For zero bytes changed the cost of remote execution is sta tistically indistinguishable from the cost of transmitting the data and the results Since the server stores the results of the compilation with no changes made this case is identical to the pure transmission case though minor amounts of work are The regression equation is 272 0 for local execution Y 1 42 G amp A for remote execution and y AL 4 for transportation cost Th
26. ld completely charge a battery repeatedly run the task until the battery dies and divide 100 battery life by the number of ex ecutions required to drain the battery Our experiments showed that the two methods produced substantially similar results but the use of the APM metric gave more stable variances then the alternative method for the same number of runs The re sults presented here rely on the APM metric Our experimental methodology was to fully charge the bat tery of the client portable then repetitively perform tasks ei ther locally or remotely until no battery power remained We measured and recorded the power consumed for each task Be cause of the noisiness of the APM metric we performed nu merous runs to achieve sufficiently low variance Since the number of runs required caused the battery to discharge and recharge hundreds of times we were concerned that the bat tery s power storage and consumption characteristics might change over time However measurements done at the be ginning and the end of the experimental period showed no sta tistically significant difference in the battery s capacity OI Experimental Resulte We ran 220 experiments consuming approximately 900 hours to obtain the data presented here Typically each point plot ted on the curves represents four to eight hours of experimen tation All results are shown with 95 confidence intervals The compilations used real software packages designe
27. ns many tasks that drain a lot of power and the costs of moving the tasks to and from a remote server are low enough remote execution could save a large amount of power and allow portable computers to run untethered for much longer Since wireless cards themselves consume significant amounts of power there is no guarantee that migrating typical realistic tasks would actually save power This paper describes experiments that have proved that such power savings are pos sible and that the size of the savings can be very significant We compared the amount of power consumed by migrating various large tasks off portable computers over a wireless de vice to the amount of power consumed by running the tasks locally on the same portables We discovered that significant power savings are possible for certain common tasks of realis tic size This paper describes our experimental methodology presents our results analyzes those results and suggests what would be required to make remote process execution a feasible power management tool for realistic environments Tl Experimental Methodology We ran an extensive series of experiments to determine if re mote process execution could save significant amounts of bat tery power We identified a set of applications likely to profit from remote execution and ran them in both local and remote modes measuring the battery power consumed in each case The experiments consisted of requesting a task on an unt
28. on clearly show that a suffi ciently large CPU load alone may be enough to make a task power expensive A better understanding of which activities in which quantities consume a great deal of power must guide any approach to choosing jobs suitable for remote execution Design of a suitable infrastructure for remote execution is the next phase of our research Substantial questions also exist in the realms of failure de tection and recovery security and server design for remote execution support Y Eelated Work Much research has been performed on power management in cluding measurement techniques approaches methods tech nical tools etc Power measurement techniques for laptop devices and applications and benchmark strategies were dis cussed in 3 14 and 22 Many techniques to save laptop power are based on switching off or slowing down the most power costly devices such as the hard drive CPU and wire less network devices when they are not being used 24 19 13 4 7 and 8 discuss different strategies to reduce hard drive power cost Measurement results show significant power savings where real hard drive access patterns were success fully predicted The prediction of the moment when a hard drive will be in use again is essential for all techniques based on this idea and it is relevant to all other devices having inertia floppy CD etc 23 suggests that power also can be saved through slowing down clock
29. rcentage g of Battery aH Consumed h I 500x500 HH 700x700 800x800 900x900 1000x1000 Size of System Solved Figure 3 Power savings for remote execution of Gaussian so lution of a system of linear algebraic equations 1 2 5 1 0 5 0 8 J Percentage of Battery 0 6 Consumed 0 4 Local Remote 0 2 0 0 35 112 439 Kilobytes of Altered Text 843 Figure 4 Power savings for remote execution of text format ting sumed by this application In most cases there was no sta tistically significant difference between local and remote exe cution Only for the case of 439 kilobytes of altered text was the difference significant at the 95 level and in this case re mote execution performed worse than local execution There are several possible explanations for the text formatting appli cation s failure to benefit from remote execution The most obvious and most likely is that the application consumed less than 1 of the total battery power even at its heaviest work load With such minor power consumption adding anything that itself consumes significant power such as moving the re sult files back to the portable computer over the wireless link is likely to have a major impact on the total power consump tion Note that the compilation and Gaussian elimination ap plications tended to consume
30. rst Symposium on Operating Systems Design and Implemen tation Monterey CA November 1994 24 John Wilkes Predictive power conservation Tecnical report HPL CSP 92 5 Concurrent Systems Project Hewlett Packard Laboratories 14 February 1992 Biographies Alexey Rudenko is a research assistant in the File Mobility Group at the University of California Los Angeles He received his M S de gree in mathematics and mechanics from Kiev University Ukraine in 1984 and the degree of Candidate of Science from Kiev Civil Aviation Institute in 1990 He published a number of papers on mathematical modeling Before joining UCLA he worked as a software designer in fields of mathematical modeling distributed DBMS and networking His research interests focus on mobile computing and distributed file systems Peter Reiher is an Adjunct Associate Professor of Computer Sci ence at UCLA Dr Reiher received his Ph D from UCLA in 1987 He has worked on several distributed operating system projects at JPL and UCLA Dr Reiher has published 30 professional articles on his research interests which include mobile computing distributed oper ating systems optimistic and predictive computing and security for distributed systems Gerald J Popek has been a Professor of Computer Science at UCLA since 1973 His academic background includes a doctorate in com puter science from Harvard University He co authored The LOCUS Distributed System Ar
31. tivity In all cases we were able to vary the amount of information that had to be sent from the client to the server machine In the cases of the compilation and text formatting tasks we ar ranged it so that the client and server each had copies of the source but that the client had altered some varying fraction of the source Thus to run the process at the server the client had to move only the altered fraction of the source The client and server stored previously computed object files correspond ing to unchanged sources so the amount of work required to perform these tasks varied depending on the fraction of data 20 changed The amount of data shipped back by the compilation and text formatting jobs was constant since the executable or Postscript document was always the same size In the case of the Gaussian solution process the entire source matrix had to be moved to the server but we varied the size of the matrix thus varying the amount of data shipped and the amount of work done A larger source matrix required shipping a larger result vector as well Neither the client nor the server performed any other user level activity during the course of the experiment We made no attempt to prevent the operating system from performing its normal housekeeping activities but we turned off many of the Linux daemons that would typically run periodically in the background The periodic intrusion of these daemons into the experiment could c
32. uld involve changing the wireless protocol to cause fewer collisions or to make collisions less costly Reducing the power costs of waiting would involve ei ther improvements in the portable s power management soft ware improvements in the wireless communications device s use of power or reducing waiting times by using faster servers We believe that there might be also other ways to improve the power cost characteristics of the remote processing The power management tools available to us did not give any indication of which effect causes the use of power just the amount of power used Thus we could not directly mea sure the fraction of power used for retransmissions versus the fraction wasted waiting for results Instead we indirectly mea sured the contributions of these effects The cost of retransmissions was approximated by measur ing the power consumption of simple transmission Figure 6 shows these costs As in Figure 2 we measured the costs of transmitting the data fetching the results off the server with out recompiling and transmitting the results back Figure 6 shows the power consumed by this process without noise com pared to the power consumed with noise These results are a reasonable approximation for determining the power cost of collisions and retransmissions due to noise The figure shows that the increased cost of transmissions in the noisy medium consumed an additional 1 5 to 2 5 of the battery We con j
33. values are 0 99 and 0 97 respectively Male Compading ond Cominus Rea Vole 2 Miber 14 5 12 Local Remote 10 5 Percentage g of Battery Consumed 6 4 1 j j 44 rt 2 4 0 r 0 125 250 375 Kilobytes of Changed Source Code 500 Figure 1 Power savings for remote execution of a compilation For small amounts of changed code which corresponds to small amounts of work to be done by the compilation mov ing the task to the remote server consumed more power than local execution The power cost of receiving the executables from the remote machine combined with power wasted by the portable while waiting for the result dominated any benefits However as the amount of work increased the value of mov ing the work off the portable computer became clear For 500 kilobytes of altered source shipping the task to the remote server consumed less than half the battery power needed to compile locally Some of the power cost of remote execution is due to using the Wavelan card to move the data to the server and the results back to the portable computer while other components of the cost reflect power wasted by the portable computer while wait ing for the results to come back If one assumed that the server had significantly more compute power than the portable com puter the waiting period would have been much shorter since the more powerful server would have completed the compi
34. w only jobs above a certain size benefit from remote execution and small jobs can actually waste power by execut ing them remotely Therefore the system must not execute all jobs remotely and not even all jobs of particular types The infrastructure can supply varying levels of support for remote execution A very simple form of support would be to provide users with a command to execute a job remotely While sim ple this puts a heavy burden on the user to decide whether a particular job is suitable for remote execution A more com plicated alternative would be for the user to provide the sys tem with hints about when a job is and is not likely to profit from remote execution For example makefiles could be aug mented with hints about whether or not particular targets are likely to profit In its most complex form the infrastructure could attempt to deduce automatically whether particular jobs were likely candidates for remote execution A major component of this last problem is identifying ex actly which characteristics of a task are likely to cause it to consume major amounts of battery power Clearly disk ac cesses are important but it is less clear how many disk ac cesses are required for a job to be a good candidate Very large processes that will require substantial amounts of vir tual memory activity may perform few file system accesses but may actually exercise the disk heavily The results we obtained for Gaussian eliminati
Download Pdf Manuals
Related Search