Using Imprecise Computation Techniques for Power Management in Real-Time Embedded Systems

Transcription

1 Usig Imprecise Computatio Techiques for Power Maagemet i Real-Time Embedded Systems Geovai Ricardo Wiedehoft ad Atôio Augusto Fröhlich Abstract Embedded systems preset severe limitatios i terms of processig ad memory capabilities ad are ofte powered by batteries, makig eergy a importat resource to be maaged. This work explores eergy as a parameter for Quality of Service (QoS) of embedded systems. The goal is to guaratee the battery lifetime specified by the applicatio ad yet preserve the deadlies of essetial (hard realtime) tasks. We propose equatios to check at project-time if a give set of tasks are schedulable. At executio-time, a preemptive scheduler for imprecise tasks based o the EDF algorithm prevets the optioal subtasks executio whe ever there is the possibility of deadlie loss or battery exhaustio. A prototype was developed i EPOS usig power maagemet mechaisms provided by the system. 1 Itroductio Embedded systems are computatioal platforms dedicated to execute a usually kow set of tasks with specific objectives. Typically, these systems preset severe limitatios i terms of processig ad memory capabilities. Some of them, due to the mobile ature of their applicatios, are also powered by batteries with a limited supply of eergy. Cosiderig all these limitatios, it is importat for the mobile embedded system to be able to maage eergy cosumptio without compromisig system s performace. Embedded systems hardware ca rely o several mechaisms to maage eergy cosumptio. Amog them, are techiques of DVS (Dyamic Voltage Scalig) ad resources hiberatio. Some works i the literature explore the itegratio of these techiques with approaches that guaratee quality of service (QoS). Most of these Geovai Ricardo Wiedehoft Atôio Augusto Fröhlich Laboratory for Software ad Hardware Itegratio, Federal Uiversity of Sata Cataria, PO Box Floriaópolis, SC, Brazil grw,guto@lisha.ufsc.br 121

2 122 Geovai Ricardo Wiedehoft ad Atôio Augusto Fröhlich approaches, however, oly seek to miimize eergy cosumptio with the mai focus o traditioal QoS metrics for processig, memory ad commuicatio. I a previous work [10], we argue that it is ot eough just esure traditioal QoS metrics if, by doig so, the system rus out of battery ad is uable to complete its computatios. We cosider eergy as a QoS parameter to meet the battery lifetime specified by the system developer, thus usig QoS i terms of eergy. I this work, the goal is ot oly to reduce eergy cosumptio, but to improve the applicatio utility i a system with limited eergy charge, esurig the battery lifetime ad the deadlies of hard real-time tasks. The proposed approach expects the developer to defie the period that the embedded system must be operatioal. By moitorig battery lifetime, the scheduler is able to select the tasks that will be executed or it ca decrease QoS levels i order to reduce eergy cosumptio ad ehace system lifetime. To achieve the proposed goal, the QoS cotrol of applicatios was ispired by imprecise computatio [5]. Imprecise computatio divides tasks ito two subtasks: oe implemetig a madatory executio flow ad aother implemetig a optioal flow. The madatory flow is the hard real-time part of the task, ad it must always be executed with i its deadlie. The optioal flow is the best-effort part of the task, which is oly executed if the desired timig requiremets ca be met. The imprecise computatio scheduler does ot execute the optioal subtasks whe there is the possibility of ay madatory subtask deadlie to be lost, thus reducig the demad for system processig. Moreover, i our scheduler, we propose that the optioal subtasks be preveted from executig whe the eergy level will ot be sufficiet to meet the time specified by applicatio. This cotrol creates more idle periods i the system, ad the scheduler ca use power maagemet techiques to reduce the eergy cosumptio of compoets durig these idle periods. The proposed scheduler is based o EDF (Earliest Deadlie First) [4] scheduler, which the tasks with the lowest deadlies have the highest priorities. A prototype of this proposal was implemeted i EPOS [6], a compoet-based embedded operatig system. EPOS provides a set of mechaisms for power maagemet, such as a ifrastructure which allows applicatios to achieve appropriate power maagemet [3] ad a power maager with differet operatig modes that realize power maagemet for applicatio [9]. Moreover, EPOS provides a battery moitorig system, which iforms the remaiig eergy i the platform. 2 Backgroud This work aims at guarateeig that the batteries used i a embedded system ca last at least the time required by the applicatio ad yet preservig the deadlies of essetial tasks, i.e., the deadlies of hard real-time tasks. Our scheduler starts to decrease QoS levels i order to save eergy whe it detects that batteries will ot last log eough to satisfy a previously defied expected system lifetime. The decreased cotrol of applicatio QoS levels is based o imprecise computatio mech-

3 Usig Imprecise Computatio Techiques for Power Maagemet 123 aisms [5], which divide tasks ito two subtasks: a madatory oe ad a optioal oe. The proposed scheduler is based o the EDF schedulig algorithm. 2.1 Imprecise Computatio Imprecise computatio is a schedulig techique origially proposed to satisfy timig requiremets of real-time tasks through decreasig levels of QoS. The cotrol of applicatio QoS levels doe by imprecise computatio worses quality of results by ot executig optioal subtasks i order to guaratee that o madatory subtask deadlies will be lost. With the divisio of each task ito two parts, imprecise computatio uites realtime computig ad best effort techiques for, respectively, the madatory ad optioal subtasks. The madatory subtask of imprecise tasks geerates imprecise results which reflect the miimum of QoS to guaratee that these results are useful. These imprecise results have their quality ehaced whe the optioal subtask executes, geeratig the precise results. The imprecise computatio showed us favorable to use i our proposal i relatio to eergy. Suppose that a task cosumes X eergy uits obligatorily. Whe it is divided ito madatory subtask (Y eergy uits) ad optioal subtask (Z eergy uits) the scheduler ca save Z eergy uits if the optioal subtask is ot executed. 2.2 EDF The EDF (Earliest-Deadlie First) [4] algorithm is a real-time schedulig mechaism based o dyamic priorities ad widely used i the literature. EDF distributes the highest priorities to the tasks with the shortest deadlies. At projecttime a schedulability test evaluates the possibility of ay task lose its deadlie. At executio-time a preemptive scheduler selects to execute the highest priority task i READY state. A exact schedulability test of the EDF algorithm is preseted below. The realtime system cosidered cotais periodic ad idepedet tasks, t = {t 0,t 1,..., t 1 }. Each t i is characterized by three parameters, (P i,,c i ), where P i is the period i which the task i is scheduled, is the max relative deadlie of coclusio i relatio to istat of the task i release ad C i is the task i executio time i the worst case which icluded times waitig by the priorities reversal. I this test is supposed that 8t i, = P i. The utilizatio U i of the task i i processig terms is represeted by equatio U i = C i. The processor s capacity is set to 1, i.e., 100%. A system with w processors has w capacity. Thus, i order to tasks to be schedulable i the EDF algorithm, the utilizatio sum of all the tasks must be less tha or equal to the processors capacity, i.e.,

4 124 Geovai Ricardo Wiedehoft ad Atôio Augusto Fröhlich Ci w (1) where w = 1 o a system with sigle-processor. If U i > w, the processor will be overloaded ad the tasks will ot be schedulable. 3 The Proposed Schedulig Strategy Our scheduler, based o EDF, guaratees the executio of madatory subtasks with their deadlies respectively met, idepedetly of the system eergy level. However, the optioal subtasks executio is ot guarateed. The optioal subtasks are executed oly if the madatory subtasks deadlies ad the system s batteries lifetime desired by applicatio are met. The objective of this scheduler is ot oly save the eergy cosumed i the system otherwise, the techique would simply ever execute the optioal subtasks but to meet the battery lifetime specified by the applicatio ad to meet the madatory subtasks deadlies with the executio of the maximum possible of the optioal subtasks, thus optimizig the applicatio utility. Figure 1 presets proposed scheduler algorithm, which the subtasks are treated as tasks i terms of schedulig. p is the iterval amog battery charge measuremets that ca be specified by the applicatio programmer ad must take ito cosideratio that each measuremet cosumes eergy to be performed. This iterval depeds o the battery power state foud i the last measuremet. 1: For every task that eters i READY state: 2: Determie the ew absolute deadlie i accordace with the elapsed time 3: Determie the priority based o absolute deadlie 4: Add to the queue accordig to calculated priority 5: 6: For each p time uits: /* p specified by the programmer ad it depeds o the eergy state */ 7: Measure the battery 8: Check if there is eough eergy to meet the time required by applicatio 9: 10: For each reschedulig: 11: Select the highest priority task i READY state 12: if, task is hard real-time, the 13: Execute the selected task 14: else, /* task is best effort */ 15: if, there is eough eergy to meet the system lifetime required, the 16: Execute the selected task 17: else, /* Battery does ot have sufficiet eergy */ 18: Use power maagemet techiques 19: Fig. 1 Proposed scheduler algorithm.

5 Usig Imprecise Computatio Techiques for Power Maagemet Schedulability Tests at Project-Time The proposed scheduler is based o the EDF algorithm, thus it is possible to follow the same logic to calculate the tasks schedulability at project-time with a few adjustmets. Suppose that the real-time system cosidered has periodic ad idepedet tasks, t = {t 0,t 1,...,t 1 }, where 8t i, = P i. I the imprecise computatio model, each t i is divided ito madatory ad optioal subtasks with executio times i the worst cases of i ad q i, respectively. Therefore, the total executio time of t i, i the worst case, is C i = i +q i. I order to guaratee that o madatory subtasks deadlies will be lost, equatio (2) must be respected. i + s w (2) Where w = 1 for a system with a sigle-processor ad s represets the iterferece i the worst cases, which icludes: time spet i the operatig system, cotext switch, scheduler algorithm. Equatio (2) must be met i order for the tasks to be schedulable i relatio to madatory subtasks deadlies, otherwise, the processor is overloaded. With the iclusio of the optioal subtask executio time i equatio (2), we ca determie if the tasks as a whole will be executed, madatory ad optioal subtasks. However, it is importat to ote that equatio (3) is ot a obligatory requiremet i our algorithm ad oly will be relevat whe equatio (2) is true, otherwise, the tasks are ot schedulable. i + q i + s w (3) Madatory ad optioal subtasks are schedulable i relatio to their deadlies whe equatio (3) is respected. Otherwise, a certai fractio c of optioal subtasks is discarded. Equatio (4) presets how to fid the fractio c. c = i +q i + s w qi (4) The eergy-related objective ca be achieved by followig the same kid of logic preseted thus far, but takig ito accout the tasks eergy cosumptio rate. The t i eergy cosumptio i the wort case, E i, is give by the sum of the eergy cosumptio i the madatory ad optioal subtasks worst cases times E i e E qi, respectively, (E i = E i +E qi ). We suppose that, as with worst cases times, the worst cases eergy cosumptios are previously kow by the applicatio developer. These values ca be obtaied by eergy profilig or aother techiques. The maximum umber of possible executios h i of t i i the time required by applicatio T t is give by divisio betwee the time required ad the executio iterval of t i, i.e., h i = T t P i. T t is give by the applicatio developer based o battery capacity. I order to meet at

6 126 Geovai Ricardo Wiedehoft ad Atôio Augusto Fröhlich least the madatory parts of the tasks, we have equatio (5) which idicates if the set of tasks will be schedulable with respect to eergy. Ei h i + e 1 (5) E t Where E t is the total eergy of the system (battery specificatio), i.e., battery capacity, e represets eergy cosumptio i the worst case of differet factors such as the eergy cosumed by the operatig system, the cotext switch, the scheduler algorithm itself. The battery s capacity is set to 1, i.e., 100 %. Substitutig h i i the equatio (5) we have equatio (6). Ei T t P i E t + e 1 (6) The tasks are schedulable i relatio to eergy i our algorithm if equatio (6) is respected. Otherwise, the system will ot meet the battery lifetime required by applicatio for this set of tasks. The iclusio of the eergy cosumed by optioal subtasks i equatio (6) allows us to check if the tasks as a whole will be executed. As discussed previously, this is ot a obligatory requiremet ad equatio (7) oly should be calculated if equatio (6) is respected, i.e., madatory subtasks met.! Ei + E qi Tt + e 1 (7) P i E t All madatory ad optioal parts of the tasks are executed i relatio to system eergy if equatio (7) is respected. Otherwise, a certai fractio g of optioal subtasks will ot be executed because the system would ot meet the battery lifetime specified by the applicatio. Equatio (8) provides a fractio g of optioal subtasks discarded i relatio to eergy. g = (Ei +E qi) T t P i E t + e 1 (8) Eqi T t P i E t I this algorithm, the objective is to meet the two parameters i relatio to time ad eergy, i.e., the madatory subtasks deadlies ad battery lifetime specified by the applicatio, respectively. Thus, (9) is the full equatio of our scheduler that must be true i order to tasks will be schedulable. " i # + s w ^ " Ei T t P i E t # + e 1 The madatory subtasks have their executios guarateed i our scheduler i relatio to time ad eergy if equatio (9) is respected. The maximum fractio l possible of optioal subtasks lost i relatio to time ad eergy ca be obtaied by equatio (10). (9)

7 Usig Imprecise Computatio Techiques for Power Maagemet 127 l = max(c,g) (10) 3.2 Schedulability Test at Executio-Time I order to provide QoS i terms of eergy ad make better use the resources with the optioal subtasks executio it is ecessary periodically to check at executiotime if the battery lifetime specified by the applicatio T tk i the istat k ca be achieved. Therefore, T tk is recalculated i the istat k accordig to the elapsed time. The total eergy of the system (battery charge) E tk also must be recalculated i the istat k. The embedded systems platforms usually provide mechaisms to get the battery charge. Equatio (11) ca be recalculated with the ew values i order to check if T tk ca be met i the istat k. Ei T tk P i E tk + e 1 (11) All madatory subtasks are executed ad optioal subtasks will be scheduled if equatio (11) is respected because this equatio idicates there is sufficiet eergy to meet T tk. Otherwise, some optioal subtasks will be discarded. The scheduler calls a power maager i the time that the optioal subtasks would be i executio. Thus, it takes the idle time of the system i order to save eergy. The optioal subtasks retur to be executed whe it is observed that equatio (11) returs to be true. 4 Implemetatio A prototype was developed i order to test the proposed scheduler usig EPOS (Embedded Parallel Operatig System) [6]. EPOS is a framework of hierarchically orgaized compoets that geerates applicatio-specific rutime support systems. To do that EPOS aalyzes the set of dedicated applicatios it must support prior to system geeratio time, thus cofigurig the system accordigly. Furthermore, through the separatio of system abstractios, hardware mediators ad sceario aspects, EPOS allows the developmet of fully platform-idepedet applicatios. I EPOS, every system compoet implemets a uiform power maagemet iterface [3]. This ifrastructure allows applicatios to iteract with the system to implemet proper eergy cosumptio maagemet for embedded systems. Through the use that EPOS provides a low-overhead dyamic power maager [9]. This power maager uses re-pluggable heuristics, allowig cofiguratio ad adaptability to specific applicatios. The EPOS power maager has differet operatio modes, such as the possibility to choose if the maager will be o or off, the possibility of cofigurig oly the desired compoets by the applicatio for the power maagemet, ad if the maager will be active or passive i the power maagemet.

8 128 Geovai Ricardo Wiedehoft ad Atôio Augusto Fröhlich EPOS also provides a battery charge moitor, which cotributes to achieve the objectives of this work. The EPOS moitor is based o the battery voltage observatio i order to get the battery charge, because the battery characteristic is to have its tesio reduced as the use. However, there are some details to be observed, because the sampled voltage is ot liearly related to battery discharge rate, the system does ot have the ability to covert all provided tesio i usable resource ad also there is a miimum voltage that the system works. Thus, the moitor establishes a discreet relatioship betwee the voltage ad battery charge through the divisio of the obtaied voltages i 10 time slices, which the voltages have differet variatios. Each slice correspods to a omial capacity percetage of the used battery. The EPOS moitor does ot implemet a costat trackig of the real battery voltage, as each samplig cosumes eergy to be realized, i additio to cosiderably overhead for the applicatio. I order to reduce these effects, the moitor uses a structure with iformatio previously kow which allows trackig the eergy cosumptio i a approximate way. The iformatio are i relatio to specific characteristics of the battery ad eergy cosumptio by the system hardware compoets that will be moitored. The moitor verifies the battery charge through the voltage i the begiig of the executio, ad durig the executio updates the value with eergy cosumed by system peripherals. We exteded EPOS to support our scheduler with imprecise tasks ad coditioal executios to time ad eergy parameters. The tasks model i EPOS was based o mootoe imprecise tasks. I this model, the mootoe tasks improve the result quality at the time i executio ad the worst case do ot chage the result. Thus, the madatory subtasks geerate results with the miimum QoS ecessary to guaratee that these results are useful, ad the optioal subtasks realize successive refiemets that results. The completio of these tasks ca occur at ay executio time without cause itegrity problems i the results. Thus, the scheduler ca decide at ay istat to fialize the optioal subtask executio. The applicatio is resposible for the results itegrity by differet methods such as the use of cotrol bits or the use of last data update timestamps. The imprecise tasks implemetatio i EPOS was realized through the creatio of two threads: oe cotaiig executio flow to hadle the madatory part ad aother with the executio flow to hadle the optioal part. The system creates these threads i a trasparet maer to the programmer. This approach oly expects the programmer to specify, whe creatig a imprecise thread, two etry poits: oe for the madatory subtask ad aother for the optioal subtask with their parameters. The scheduler i executio always chooses the highest priority subtask i accordace with the deadlies as our algorithm is based i EDF. The optioal subtasks are scheduled if there are ot madatory subtasks i READY state ad if there is eergy eough to meet battery lifetime specified by the applicatio, i.e., optioal subtasks have lower priorities tha madatory subtasks. Whe a madatory subtask eters i READY state ad its optioal subtask is ot yet fiished the executio i the previous period, the scheduler immediately suspeds this optioal subtask executio. These characteristics prevet madatory subtasks deadlies losses. The optioal subtasks cotexts are always restarted i a ew task period.

9 Usig Imprecise Computatio Techiques for Power Maagemet 129 The scheduler also updates at executio-time the T tk with elapsed time ad the E tk usig the EPOS eergy moitor. Scheduler recalculates the equatio (11) i periods p with these ew values i order to check if the system is able to sustai the curret workload without ruig out of battery before the required lifetime T tk is achieved. p will deped o the last eergy aalysis. I the best case, equatio (11) is respected ad optioal subtasks ca be scheduled. Otherwise, optioal subtasks are discarded ad, takig advatage of the idle period created, the scheduler executes the EPOS power maager i passive mode. I additio to savig eergy by ot execute the optioal subtasks, the power maager reduces the system eergy cosumptio through the use of power maagemet techiques. The optioal subtasks retur to executio whe the scheduler idetify T tk+i ca be met agai i istat k + i. 5 Related Work GRACE-OS [11] is a eergy-efficiet operatig system for mobile multimedia applicatios. This system uses a cross-layer adaptatio techique to guaratee QoS o systems with adaptive software ad hardware. It combies real-time schedulig with DVS mechaisms to dyamically maage eergy cosumptio. It was implemeted over the LINUX operatig system ad it oly supports soft real-time tasks. GRUB-PA [8] is somehow similar to GRACE-OS. The mai differece is GRUB- PA supports both soft ad hard real-time tasks. Niu [7] proposed to miimize eergy cosumed by soft real-time systems while guarateeig QoS requiremets. This goal is achieved by a hybrid static/dyamic schedulig algorithm that it uses DVS mechaisms ad it partitios the set of tasks i madatory ad optioal tasks. I this work, the QoS requiremets are qualified by (m,k) costraits which it specifies that tasks must meet at least m deadlies i ay k cosecutive task releases. I a similar work, Harada [2] proposed to resolve the trade-off betwee QoS maximizatio ad eergy cosumptio miimizatio. It uses a allocatio of processor cycles ad frequecy with QoS guaratees ad it divides each task ito madatory ad optioal parts. Other projects explore trade-off betwee applicatio s QoS ad eergy cosumptio through adaptatios i the applicatios aimig to meet the time specified by applicatio. ODYSSEY [1] uses that idea. It moitors the eergy budget ad with this iformatio it ca select the correct state betwee eergy savig ad quality of applicatio. This work also demostrates how the applicatios ca dyamically chage their behavior ( fidelity of the data) with the goal of savig eergy. ECOSYSTEM [12] is aother operatig system that supports applicatio adaptatio. This system is based i a currecy that the applicatios use to allocate ( to pay ) system resources (e.g., access to memory, etwork or disks), called curretcy. The system distributes curretcies periodically to tasks accordigly to a equatio that defies the discharge rate that the system battery ca assume to force the system to last for a defied period of time. This allows applicatios to adapt their executio based o their curretcy balace. This model uifies the calculatio of eergy o

10 130 Geovai Ricardo Wiedehoft ad Atôio Augusto Fröhlich the various hardware devices ad it provides a satisfactory eergy allocatio amog the applicatios. 6 Coclusio This work proposed a approach to exploit eergy as a QoS parameter i order to guaratee that battery lifetime ca last time desired by mobile embedded system ad yet preserve the deadlies of hard real-time tasks. Our approach was ispired by imprecise tasks cocepts, accordig to tasks ca be divided ito madatory ad optioal parts. I this article, equatios at project-time were preseted with objective the of applicatio programmer to check if a set of tasks will be schedulable i our algorithm i relatio to two parameters desired, i.e., time ad eergy. At executiotime, our scheduler based o EDF algorithm esures the madatory subtasks deadlies ad recalculates the equatio of eergy i order to check if the required battery lifetime will be met. The optioal subtasks are preveted from executig, i.e, decreasig QoS levels if ay required parameter will ot be met. A prototype was developed i EPOS, which allowed the executio of a power maager i idle periods created by o-executio of the optioal subtasks, thus reducig eergy cosumptio by stoppig or slowig dow system compoets durig these idle periods. Refereces 1. Fli, J., Satyaarayaa, M.: Eergy-aware adaptatio for mobile applicatios. I: ACM SOSP 99, pp ACM Press, New York, NY, USA (1999) 2. Harada, F., Ushio, T., Nakamoto, Y.: Power-aware resource allocatio with fair qos guaratee. I: IEEE RTCSA 06, pp IEEE Computer Society, Washigto, DC, USA (2006) 3. Hoeller, A.S.J., Waer, L.F., Fröhlich, A.A.: A Hierarchical Approach For Power Maagemet o Mobile Embedded Systems. I: 5th IFIP DIPES, pp Braga, Portugal (2006) 4. Liu, C.L., Laylad, J.W.: Schedulig algorithms for multiprogrammig i a hard-real-time eviromet. J. ACM 20(1), (1973). DOI 5. Liu, J.W., Shih, W.K., Li, K.J., Bettati, R., Chug, J.Y.: Imprecise computatios. Proceedigs of the IEEE 82(1), (1994) 6. Marcodes, H., Juior, A.S.H., Waer, L.F., Fröhlich, A.A.: Operatig Systems Portability: 8 bits ad beyod. I: 11th IEEE ETFA, pp Prague, Czech Republic (2006) 7. Niu, L., Qua, G.: A hybrid static/dyamic dvs schedulig for real-time systems with (m, k)-guaratee. rtss 0, (2005) 8. Scordio, C., Lipari, G.: Usig resource reservatio techiques for power-aware schedulig. I: ACM EMSOFT 04, pp ACM Press, New York, NY, USA (2004) 9. Wiedehoft, G.R., Hoeller, A.S.J., Fröhlich, A.A.: A Power Maager for Deeply Embedded Systems. I: 12th IEEE ETFA, pp Patras, Greece (2007) 10. Wiedehoft, G.R., Hoeller, A.S.J., Fröhlich, A.A.: Quality-Of-Service: the Role of Eergy. I: 9th Workshop o Real-Time Systems, pp Belem, Brazil (2007) 11. Yua, W.: Grace-os: A eergy-efficiet mobile multimedia operatig system. Ph.D. thesis, Uiversity of Illiois at Urbaa-Champaig (2004) 12. Zeg, H., Ellis, C.S., Lebeck, A.R., Vahdat, A.: Ecosystem: maagig eergy as a first class operatig system resource. I: ACM ASPLOS-X, pp ACM, New York, NY (2002)