UNIVERSITY OF TWENTE. Energy-optimization for Dataflow Applications using Timed Automata. Formal Methods & Tools. Jaco van de Pol, 18 Aug 2016

Transcription

1 UNIVERSITY OF TWENTE. Formal Methods & Tools. Energy-optimization for Dataflow Applications sing Timed Atomata Jaco van de Pol, 8 Ag 206 ICT Energy Science Conference, Aalborg Joint work with Waheed Ahmad and Mariëlle Stoelinga

2 Over 0 years history of mlti-core architectres: PC, mobile HD video playback, web browsing, 3D gaming, 3D interfaces By 205, video cases 2/3 of mobile data traffic (Cisco, 20) UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

3 Over 0 years history of mlti-core architectres: PC, mobile HD video playback, web browsing, 3D gaming, 3D interfaces By 205, video cases 2/3 of mobile data traffic (Cisco, 20) Mltimedia services are energy-hngry How to get qality & performance within tight energy bonds? FP7 Sensation: Self-energy-spporting atonomos compting UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

4 Table of Contents Schedling and Mapping on Mlti-core Hardware Modeling: Synchronos Dataflow and Hardware Platform Analysis: Timed Atomata and Uppaal 2 Energy Optimization by Power Management Modeling: DPM, DVFS, VFI in Hardware Platform Model Analysis: Priced Timed Atomata and Uppaal Cora 3 Extension : Battery Management Modeling: Kinetic Battery Model Analysis: Hybrid Atomata and Uppaal SMC 4 Extension 2: Handling Uncertainty by Adaptation Modeling: Exection Time is Stochastic Analysis: Stochastic Hybrid Games and Uppaal Stratego 5 Conclsion Tool Spport Discssion UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

5 Table of Contents Schedling and Mapping on Mlti-core Hardware Modeling: Synchronos Dataflow and Hardware Platform Analysis: Timed Atomata and Uppaal 2 Energy Optimization by Power Management 3 Extension : Battery Management 4 Extension 2: Handling Uncertainty by Adaptation 5 Conclsion UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

6 Synchronos Data Flow graphs (Lee, 986) An SDF Graph is a tple G = (A, D, Tok 0, τ) where: A is a finite set of actors , 2 v, 2 w, UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

7 Synchronos Data Flow graphs (Lee, 986) An SDF Graph is a tple G = (A, D, Tok 0, τ) where: A is a finite set of actors D is a finite set of channels D A 2 N , 2 v, 2 w, UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

8 Synchronos Data Flow graphs (Lee, 986) An SDF Graph is a tple G = (A, D, Tok 0, τ) where: A is a finite set of actors D is a finite set of channels D A 2 N 2 Tok 0 : D N denotes initial tokens in each bffered channel , 2 v, 2 w, UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

9 Synchronos Data Flow graphs (Lee, 986) An SDF Graph is a tple G = (A, D, Tok 0, τ) where: A is a finite set of actors D is a finite set of channels D A 2 N 2 Tok 0 : D N denotes initial tokens in each bffered channel τ : A N assigns an exection time to each actor , 2 v, 2 w, UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

10 Synchronos Data Flow graphs (Lee, 986) An SDF Graph is a tple G = (A, D, Tok 0, τ) where: A is a finite set of actors D is a finite set of channels D A 2 N 2 Tok 0 : D N denotes initial tokens in each bffered channel τ : A N assigns an exection time to each actor , 2 v, 2 w, Actor firing: consme and prodce tokens throgh channels Actors fire concrrently; even ato-concrrency is possible UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

11 Some SDF Examples (Wiggers 09, Theelen 2) MP3, SRC, DAC, MP3 Playback VLD, IDCT, FD,2 MC, 5 RC, MPEG-4 Decoder

12 Some SDF Examples (Wiggers 09, Theelen 2) MP3, SRC, DAC, MP3 Playback VLD, IDCT, FD,2 MC, 5 RC, MPEG-4 Decoder Face Recognition (Recore) UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

13 Self-timed Exection [SDF3] (Ghamarian, Theelen, 06) 2 3 2, 2 v, 2 w, To get to a stable periodic exection: Repetition vector γ : A N, sch that p.γ(a) = q.γ(b) For this example: γ, v, w = 4, 2, 3 UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

14 Self-timed Exection [SDF3] (Ghamarian, Theelen, 06) 2 3 2, 2 v, 2 w, To get to a stable periodic exection: Repetition vector γ : A N, sch that p.γ(a) = q.γ(b) For this example: γ, v, w = 4, 2, 3 The self-timed exection garantees maximm throghpt: processors p0 p p2 p3 graph iteration v v v v w w w w w graph iteration time w UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

15 SDF for parallel, self-timed exection Advantages of SDF with self-timed exection SDF fits streaming applications (no data dependencies, SADF) Garantees maximal throghpt, transient + periodic phase Simple, polynomial algorithms (implemented in SDF3) UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

16 SDF for parallel, self-timed exection Advantages of SDF with self-timed exection SDF fits streaming applications (no data dependencies, SADF) Garantees maximal throghpt, transient + periodic phase Simple, polynomial algorithms (implemented in SDF3) Limitations of SDF with self-timed exection Leads to maximal parallelism, which is expensive So far, a homogeneos processor model is assmed Worst-case assmptions lead to over-dimensioning UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

17 Modelling Limited, Heterogeneos Resorces Some actors can be mapped to particlar processors only floating point, analog/digital,... Definition A hardware platform model is a tple (P, ζ) consisting of a finite set P of processors; and a fnction ζ : P A {0, }: actor A can rn on processor P Firing Starts Firing Ends Exection Time Claim Idle Processor Release Processor UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

18 Table of Contents Schedling and Mapping on Mlti-core Hardware Modeling: Synchronos Dataflow and Hardware Platform Analysis: Timed Atomata and Uppaal 2 Energy Optimization by Power Management 3 Extension : Battery Management 4 Extension 2: Handling Uncertainty by Adaptation 5 Conclsion UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

19 Approach sing Timed Atomata Ahmad, De Groote, Hölzenspies, Stoelinga, van de Pol ACSD 4 Application SDF Graph Architectre Translation to TA Translation to TA Mapping & Schedling by Uppaal Model- Checking Optimal Schedle UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag 206 / 44

20 Timed Atomata and Uppaal [Alr,Dill 94] [Larsen etal. 95] Ingredients of Timed Atomata States and Transitions as in finite atomata Real valed clock variables (here y) Clock constraints, invariants and resets (y 5, y := 0) Synchronisation between atomata via action labels UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

21 Timed Atomata and Uppaal [Alr,Dill 94] [Larsen etal. 95] Ingredients of Timed Atomata States and Transitions as in finite atomata Real valed clock variables (here y) Clock constraints, invariants and resets (y 5, y := 0) Synchronisation between atomata via action labels The Uppaal model-checker can then check: reachability of states (also safety, liveness properties) It can synthesize shortest or fastest traces UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

22 Translation of Processor Model and SDF graph A processor can be occpied by at most one task at the time Clock x specifies the dration of the task Model for each Processor p id : UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

23 Translation of Processor Model and SDF graph A processor can be occpied by at most one task at the time Clock x specifies the dration of the task On firing: check for reqired tokens, and consme them On ending: prodce the reqired tokens Model for each Processor p id : Model for actors a, b, c: UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

24 Synthesizing a schedle Actors are mapped on Processors non-deterministically We qery Uppaal for the fastest trace to a fll iteration. In addition, the model checker Uppaal cold check for absence of deadlocks, bffer bonds, safety, liveness,... UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

25 Synthesizing a schedle Actors are mapped on Processors non-deterministically We qery Uppaal for the fastest trace to a fll iteration. In addition, the model checker Uppaal cold check for absence of deadlocks, bffer bonds, safety, liveness,... Design Exploration Let s analyse the throghpt, and synthesize schedles, if we vary: The nmber of processors The capabilities of the processors UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

26 Reslts on Rnning Example: throghpt verss processors Self-timed exection ses 4 processors: throghpt /9 processors graph iteration graph iteration p0 p v v v v v v p2 w w w w w w p3 w w w time UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

27 Reslts on Rnning Example: throghpt verss processors Self-timed exection ses 4 processors: throghpt /9 graph iteration graph iteration p0 p v v v v v v processors p2 w w w w w w p3 w w w time Restrict to 3 processors: still the same throghpt /9 graph iteration graph iteration processors p0 v v v v v v p w p2 w w w w w w w w time UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

28 Reslts on Rnning Example: throghpt verss processors Self-timed exection ses 4 processors: throghpt /9 graph iteration graph iteration p0 p v v v v v v processors p2 w w w w w w p3 w w w time Restrict to 3 processors: still the same throghpt /9 graph iteration graph iteration p0 v v v v v v processors p w p2 w w w w w w w w time Restrict to 2 processors: small penalty, throghpt / graph iteration processors p0 p a a b a b a a b a b a a b a a c c c a c c c a time UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

29 Reslts on Rnning Example: Heterogeneos processors Assme: p 0, p can execte task, v; and p 2, p 3 only execte w Self-timed exection ses 4 processors: throghpt /9 processors graph iteration graph iteration p0 p v v v v v v p2 w w w w w w p3 w w w time Using 4 heterogeneos processors, still same throghpt /9 processors p0 p p2 p3 graph iteration v v v v w w w w w time UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

30 Table of Contents Schedling and Mapping on Mlti-core Hardware Modeling: Synchronos Dataflow and Hardware Platform Analysis: Timed Atomata and Uppaal 2 Energy Optimization by Power Management Modeling: DPM, DVFS, VFI in Hardware Platform Model Analysis: Priced Timed Atomata and Uppaal Cora 3 Extension : Battery Management Modeling: Kinetic Battery Model Analysis: Hybrid Atomata and Uppaal SMC 4 Extension 2: Handling Uncertainty by Adaptation Modeling: Exection Time is Stochastic Analysis: Stochastic Hybrid Games and Uppaal Stratego 5 Conclsion Tool Spport Discssion UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

31 Dynamic Power Managament (DPM) Idle processors can be switched off to a low power (sleep) state This is redces energy sage in idle time Don t ignore it static power consmption Switching consmes some energy as well 0.W 0.W ON 0.2W DIM 0.2W OFF P=4W P=2.5W P=2W LCD ON OFF OFF Backlight ON ON OFF UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

32 Voltage and Freqency Scaling (DFVS) (Zhravlev 3) Lower the voltage and freqency dynamically This redces energy dring active time At the expense of increasing the exection time Again, switching freqency levels has some cost UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

33 Voltage and Freqency Islands (VFI) (Ogras, DAC 07) Local verss Global DVFS VFI: a grop of processors clstered together Common clock freqency/voltage per island Examples: Intel i7, IBM Power 7 series, Samsng S5 octa core UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

34 Hardware Platform Model We extend the Hardware Platform Model to the following tple: (P, ζ, F, Time act, Pow idle, Pow occ, Pow tr ) P: set of processors, partitioned in voltage freqency islands ζ : P A {0, }: maps actors on heterogeneos processors F : finite set of freqency levels f < < f m Time act : A F N : exection time of A at freqency F Pow idle : P F R + : power consmption in idle state Pow occ : P F R + : power consmption in bsy state Pow tr : P F 2 R + : overhead of switching freqency level UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

35 Example: Samsng Exynos 420 (Park, TCAD 3) Level Voltage (V) Freqency (MHz) Exper. P idle (W) Exper. P occ (W) UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

36 Table of Contents Schedling and Mapping on Mlti-core Hardware 2 Energy Optimization by Power Management Modeling: DPM, DVFS, VFI in Hardware Platform Model Analysis: Priced Timed Atomata and Uppaal Cora 3 Extension : Battery Management 4 Extension 2: Handling Uncertainty by Adaptation 5 Conclsion UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

37 Approach with Priced Timed Atomata Uppaal Cora Ahmad, Hölzenspies, Stoelinga, van de Pol DSD 5 Application SDF Graph Architectre Translation to PTA Translation to PTA Uppaal Cora Throghpt Reqirement Model- Checking Energy Efficient Schedle UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

38 Priced Timed Atomata [Behrman, Larsen, et al. 2005] Specification of a lamp: Costs accmlate over transitions and while residing in states Uppaal Cora can synthesize fastest or cheapest traces UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

39 Translation of Hardware Platform Model UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

40 Using Uppaal Cora Uppaal Cora spports compting fastest and cheapest traces Obtain the completion time via fastest trace as before Incorporate the fond completion time in the model Obtain Energy Efficient Schedle via cheapest trace UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

41 Using Uppaal Cora Uppaal Cora spports compting fastest and cheapest traces Obtain the completion time via fastest trace as before Incorporate the fond completion time in the model Obtain Energy Efficient Schedle via cheapest trace Design Exploration Let s analyse the energy sage when we vary The reqired throghpt The nmber of processors and freqency islands UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

42 Reslts of Power Management Optimization for MPEG-4 Energy Consmption (mws) /3 5/3 4/3 3/2 2/2 / Frames per second Energy/frame verss Frames/second. Legend: nmber of processors/vfis. High throghpt: sing more processors is cheaper (rn at decent speed) UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

43 Reslts of Power Management Optimization for MPEG-4 Energy Consmption (mws) /3 5/3 4/3 3/2 2/2 / Frames per second Energy/frame verss Frames/second. Legend: nmber of processors/vfis. High throghpt: sing more processors is cheaper (rn at decent speed) UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

44 Reslts of Power Management Optimization for MPEG-4 Energy Consmption (mws) /3 5/3 4/3 3/2 2/2 / Frames per second Energy/frame verss Frames/second. Legend: nmber of processors/vfis. Lower throghpt: sing less processors is cheaper (can switch off others) UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

45 Reslts of Power Management Optimization for MPEG-4 Energy Consmption (mws) /3 5/3 4/3 3/2 2/2 / Frames per second Energy/frame verss Frames/second. Legend: nmber of processors/vfis. Very low throghpt: ses more energy per frame ( idle time dominates) UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

46 Table of Contents Schedling and Mapping on Mlti-core Hardware Modeling: Synchronos Dataflow and Hardware Platform Analysis: Timed Atomata and Uppaal 2 Energy Optimization by Power Management Modeling: DPM, DVFS, VFI in Hardware Platform Model Analysis: Priced Timed Atomata and Uppaal Cora 3 Extension : Battery Management Modeling: Kinetic Battery Model Analysis: Hybrid Atomata and Uppaal SMC 4 Extension 2: Handling Uncertainty by Adaptation Modeling: Exection Time is Stochastic Analysis: Stochastic Hybrid Games and Uppaal Stratego 5 Conclsion Tool Spport Discssion UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

47 Kinetic Battery Model (KiBaM) [Manwell, McGowan, 993] Inclde battery schedling considerations. Needs a model. This model explains the apparent recovery effect of batteries KiBaM distingishes two wells: bond and available charge Two differential eqations describe the evoltion over time c c ȧ(t) = i(t)+k(h b h a ) ḃ(t) = k(h b h a ) h b b(t) k a(t) h a i(t) h a = a(t) c h b = b(t) c UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

48 Approach sing Hybrid Atomata Uppaal SMC Ahmad, Jongerden, Stoelinga, van de Pol ACSD 206 T trn thermostat off on T 27 T = T + 30 off T 23 T = T T off off trn thermostate on off T 23 HA for Thermostat q = on on on Simlation of HA evoltion Hybrid Atomata [Henzinger, 996] Basis: states and transitions as in atomata Clock constraints, invariants, synchronisation via action labels Real-time variables evolve with a rate governed by ODEs UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

49 Analysis with Uppaal SMC [David, Larsen, Legay, et al. 2] The Uppaal SMC model checker Spports (Stochastic) Hybrid atomata Statistical Model Checking (cf. Monte Carlo simlations) Visalisation of simlated rns with line or bar plots We (and others) have sed this to investigate: Varios battery schedling policies [Wognsen, Haverkort, et al. 5] Variation in nmber of batteries, processors Srvivability of the system nder varios circmstances Scales better than previos approach sing discretisation and PTA UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

50 Table of Contents Schedling and Mapping on Mlti-core Hardware Modeling: Synchronos Dataflow and Hardware Platform Analysis: Timed Atomata and Uppaal 2 Energy Optimization by Power Management Modeling: DPM, DVFS, VFI in Hardware Platform Model Analysis: Priced Timed Atomata and Uppaal Cora 3 Extension : Battery Management Modeling: Kinetic Battery Model Analysis: Hybrid Atomata and Uppaal SMC 4 Extension 2: Handling Uncertainty by Adaptation Modeling: Exection Time is Stochastic Analysis: Stochastic Hybrid Games and Uppaal Stratego 5 Conclsion Tool Spport Discssion UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

51 Adaptive Schedling with Stochastic Hybrid Games Ahmad, van de Pol Isola 206 (forthcoming) Can we avoid overdimensioning de to worst-case exection time? Strategy can be adaptive: If task A happens fast, we can relax for task B Model: replace WCET by lower/pper bond Goal: Garantee reqired throghpt, minimize expected energy 0 simlations: Safe Strategy 0 simlations: Optimized Strategy UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

52 Uppaal Stratego (David, Jensen, Larsen, Mikčionis, Taankvist, 5) Stochastic Hybrid Atomata Uppaal Stratego Extends Hybrid Atomata Distingish controllable transitions (schedling decisions) from non-controllable transitions (stochastic exection times) Synthesize safe strategies, garanteeing a reqired throghpt Refine to near-optimal strategies, optimizing energy sage Means: game strategy synthesis, statistical model checking, reinforcement learning UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

53 Table of Contents Schedling and Mapping on Mlti-core Hardware Modeling: Synchronos Dataflow and Hardware Platform Analysis: Timed Atomata and Uppaal 2 Energy Optimization by Power Management Modeling: DPM, DVFS, VFI in Hardware Platform Model Analysis: Priced Timed Atomata and Uppaal Cora 3 Extension : Battery Management Modeling: Kinetic Battery Model Analysis: Hybrid Atomata and Uppaal SMC 4 Extension 2: Handling Uncertainty by Adaptation Modeling: Exection Time is Stochastic Analysis: Stochastic Hybrid Games and Uppaal Stratego 5 Conclsion Tool Spport Discssion UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

54 Model-Driven Framework UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

55 Meta Models, Transformations, Graphical Editors Meta-model for Hardware Platform Model Similar, we designed (or adapted): Graphical editor for Hardware Platform Model Meta-models for SDF, compatible with SDF3 tool site Meta-models for PTA, compatible with Uppaal tool site Model2Model Transformations: SDF + HPM PTA UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

56 Tool Spport Tool Chain is compatible with SDF3 and Uppaal -family Mappings are compositional model transformations in Epsilon Self-timed exection SDF graphs SDF3 Mapping & Schedling Timed Atomata Uppaal Power Management Priced Timed Atomata Uppaal Cora Battery Management Hybrid Atomata Uppaal SMC Handling Uncertainty Hybrid Stochastic Games Uppaal Stratego Main challenge: scalability. Better Algorithms (heristics, statistical methods) More clever modeling Leverage High-performance Model Checking UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

57 Visalizing the reslt: UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

58 Table of Contents Schedling and Mapping on Mlti-core Hardware 2 Energy Optimization by Power Management 3 Extension : Battery Management 4 Extension 2: Handling Uncertainty by Adaptation 5 Conclsion Tool Spport Discssion UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

59 Smmary Schedling and Mapping on Heterogeneos Mlti-core Schedle maximal throghpt on small nmber of processors Handling heterogeneos processor application models UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

60 Smmary Schedling and Mapping on Heterogeneos Mlti-core Schedle maximal throghpt on small nmber of processors Handling heterogeneos processor application models Power Optimization Exploit and combine modern power management strategies Use model checking PTA for optimal mapping & schedling, trading off throghpt verss energy UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

61 Smmary Schedling and Mapping on Heterogeneos Mlti-core Schedle maximal throghpt on small nmber of processors Handling heterogeneos processor application models Power Optimization Exploit and combine modern power management strategies Use model checking PTA for optimal mapping & schedling, trading off throghpt verss energy Extensible method Battery Management Adaptation to Uncertainty Hybrid Atomata Stochastic Hybrid Games UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

62 ... Schedling & Mapping Power Optimization Batteries Adaptation Conclsion... Applications and Ftre Work I FP7 SENSATION had several Cool and Hot applications: I Face Recognition (Recore, Twente) I Nano-satellites (Gomspace, Aalborg) I E-bikes (EnergyBs consortim) UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

63 ... Schedling & Mapping Power Optimization Batteries Adaptation Conclsion... Applications and Ftre Work I FP7 SENSATION had several Cool and Hot applications: I Face Recognition (Recore, Twente) I Nano-satellites (Gomspace, Aalborg) I E-bikes (EnergyBs consortim) I Get closer to engineering I More extensive experimental validation is still reqired I Generate SDF + HPM from instrmented experiments I Extensions I Inclde: commnication costs, processor pinning/affinity,... I Inclde: battery charging: flly energy-self-aware systems I Scalability (heristic algorithms + parallel implementation) UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44

64 Bibliography This presentation is based on for papers with PhD stdent Waheed Ahmad and co-spervisor Mariëlle Stoelinga: Ahmad, De Groote, Hölzenspies, Stoelinga, van de Pol, Resorce-constrained optimal schedling of synchronos dataflow graphs via timed atomata ACSD 4 Ahmad, Hölzenspies, Stoelinga, van de Pol, Green compting: Power optimisation of VFI-based real-time mltiprocessor dataflow applications DSD 5 Ahmad, Jongerden, Stoelinga, van de Pol, Model checking and evalating QoS of batteries in MPSoC dataflow applications via hybrid atomata ACSD 6 Waheed Ahmad and Jaco van de Pol, Synthesizing Energy-Optimal Controllers for Mltiprocessor Dataflow Applications with Uppaal Stratego.. ISOLA 6 UNIVERSITY OF TWENTE. Energy-optimization with Timed Atomata 8 Ag / 44