Slide 1Slide 2Power management techniquesFrequency/voltage scalingDifferent goals of power managementSlide 6Implementation of Power Management PointsExample of compiler/OS collaborationCompiler/OS collaborationDVS for multiple coresSlide 11Mapping a linear task graph onto a linear pipelineSlide 13Slide 14Turn OFF some PEsDVS using Machine LearningSlide 17Energy-Reliability tradeoffOptimal number of checkpointsFaults are rare eventsNon-uniform check-pointingTriple Modular Redundancy vs. DuplexAdd memory power to the mixOS assisted Memory Power Management?Example of compiler assisted Memory Power Management?Example of compiler assisted Memory Power Management?Phase Change Memory (PCM) A power saving memory technologyProperties of PCMSo, where is the catch?Slide 30Dealing with asymmetric read/writeDealing with low write endurance (write minimization)Wear levelingSlide 34Conclusion10 years of research on Power Management(now called green computing) Rami Melhem Daniel MosseBruce Childers•Introduction•Power management in real-time systems•Power management in multi-core processors•Performance-Resilience-Power Tradeoff •Management of memory power•Phase Change MemoryPower management techniquesTwo common techniques:1) ThrottlingTurn off (or change mode of) unused components (Need to predict usage patterns to avoid time and energy overhead of on/off or mode switching)2) Frequency and voltage scalingScale down core’s speed (frequency and voltage)Designing power efficient components is orthogonal to power managementFrequency/voltage scaling•Gracefully reduce performance•Dynamic power Pd = C f 3 + Pind•Static power: independent of f.powerStatic powertimeC f 3 PindtimeWhen frequency is halved:•Time is doubled•C f 3 is divided by 8•Energy caused by C f 3 is divided by 4•Energy caused by Pind is doubledIdle time•Minimize total energy consumption- static energy decreases with speed- dynamic energy increases with speed•Minimize the energy-delay product –Takes performance into consideration•Minimize the maximum temperature•Maximize performance given a power budget•Minimize energy given a deadline•Minimize energy given reliability constraintsDifferent goals of power managementEnergy*delayfPind / f 2C f energySpeed (f)C f 2totalPind / fDVS in real-time systemsCPU speedtimedeadlineSmaxSminWorst case execution P M PRemaining time•Utilize slack to slow down future tasks (Proportional, Greedy, aggressive,…)timeStatic scalingP M PP M P(power management points)Dynamic scalingRemaining timeP M PImplementation of Power Management Points•Can be implemented as periodic OS interrupst•Difficulty: OS does not know how much execution is remaining •Compiler can insert code to provide hints to the OSmin average maxbranchloopExample of compiler/OS collaborationP M H•Compiler records WCET based on the longest remaining pathAt a power management hint min average maxAt a power management point•OS uses knowledge about current load to set up the speed P M HP M HRun-timeinformationOS/HW(knows the system)Compiler/OS collaborationCompiler(knows thetask)Static analysisApplication Source CodeApplication Source CodePMHs: Power management hintsPMPs: Power management pointsInterrupts for executing PMPsPMHs timeDVS for multiple coresManage energy by determining:•The speed for the serial section•The number of cores used in the parallel section•The speed in the parallel sectionOne coreTwo coresSlowing down the coresSlowing down the parallel sectionTo derive a simple analytical model, assume Amdahl’s law: - p % of computation can be perfectly parallelized.pUsing more coress•Streaming applications are prevalent–Audio, video, real-time tasks, cognitive applications•Constrains:–Inter-arrival time (T)–End-to-end delay (D)•Power aware mapping to CMPs–Determine speeds–Account for communication–Exclude faulty coresTDMapping streaming applications to CMPsMapping a linear task graph onto a linear pipelineIf the # of stages = # of coresCoreCore CoreCoreDtstagenstagestage)(0Ttstagestagestage )(max)(0stagenstagestageeminimizeSubject toei : energy for executing stage iei : energy for moving data from stage i-1 to stage iti : time for executing stage i ti : time for moving data from stage i-1 to stage iFind tstage1) Group the stages so that the number of stages equals the number of cores2) Use a dynamic programming approach to explore possible groupings3) A faster solution may guarantee optimality within a specified error bound.Core Core Core CoreIf the # of stages > # of coresMapping a linear task graph onto a linear pipeline•Timing constraints are conventionally satisfied through load balanced mapping•Additional constraint–Minimize energy consumption–Maximize performance for a given energy budget–Avoid faulty coresinstanceinstanceABCEDFGHIJKMapping a non-linear task graph onto CMPABCDFEGHIJKMaximum speedMedium speedMinimum speedTurn OFF some PEsMaximum speed/voltage (fmax)instanceinstanceABCEDFGHIJKABCDFEGHIJKMedium speed/voltage Minimum speed/voltage (fmin) PE OFFDVS using Machine Learning Characterize the execution state of a core by•Rate of instruction execution (IPC) •# of memory accesses per instruction•Average memory access time (depends on other threads)During training, record for each state•The core frequency •The energy consumptionDetermine the optimal frequency for each stateDuring execution, periodically,Estimate the current state (through run-time measurements)Assume that the future is a continuation of the presentSet the frequency to the best recorded during trainingMMCcoreL1 $$core core coreL1 $$ L1 $$ L1 $$L2 $$ L2 $$ L2 $$ L2 $$17Training phaseRuntimeLearning enginedetermine freq. & voltagesIntegrated DVS policyAuto. policy generatorStatistical learning applied to DVS in CMPs.deadlineIf you have a time slack:1) add checkpoints2) reserve recovery time3) reduce processing speedFor a given number of checkpoints, we can find the speed that minimizes energy consumption, While guaranteeing recovery and timeliness.SmaxUsing time redundancy (checkpointing and rollbacks)Energy-Reliability tradeoffMore checkpoints = more overhead + less recovery slackDCrOptimal number of checkpointsFor a given slack (C/D) and checkpoint overhead (r/C),we can find the number of checkpoints that minimizes energy consumption, and guarantee recovery and timeliness.# of checkpointsEnergyFaults are rare