1. Fast Energy Evaluation of Embedded Applications for Many-core Systems Felipe Rosa, Luciano Ost, Thiago Raupp, Fernando Moraes, Ricardo Reis.

2. Outline 1. Introduction 2. Open Virtual Platforms 3. Proposed Energy Model 5. Conclusion and Future Works 4. Exploration in Large Scale Systems

3. 1. Introduction  High performance many-core systems are a reality up to 256 cores available today What about software development?

4.  Software development challenges comprise: inter-processor communication protocol stacks definition OS porting and analysis parallel programming model porting drivers development  Software development costs is increasing… Source: IBS 2013 Simulation is the key tool for many- core research 1. Introduction

5.  Full-system simulators are one good option virtual platforms that emulate hardware behaviour, making software believe that it is running on a real physical hardware  Support concomitant HW and SW development improve time-to-market  Examples of such simulators are:  Why OVP simulator? SimulatorAccuracySupported processor architecturesLicenseActive support SimicsFunctionally- accurate Alpha, ARM, MIPS, PowerPC, SPARC, and x86 PrivateYes PTLsimCycle-accurateX86OpenNo SimpleScalarCycle-accurateAlpha, ARM, PowerPC, and x86OpenNo GEM5Cycle-accurate Alpha, ARM, MIPS, PowerPC, SPARC, and x86 OpenYes QEMUInstruction-accurate ARM, MicroBraze,MIPS, PowerPC, SPARC, x86, and others OpenYes OVPsimInstruction-accurate Alpha, ARC, ARM, MIPS, PowerPC, MicroBraze, and others Open and PrivateYes

6. 2. Open Virtual Platform - OVP  Large number of processor architectures (ISAs) supported (e.g. MIPS, ARM, x86, PowerPC )  Simulation speeds of up to 2200 MIPS relies on just-in-time (JIT) dynamic binary translation  Complete development environment with APIs  Open source license  Extensive documentation and active forum  Powerful debug environment  Limitation: OVPsim provides instruction accuracy only, resulting in inaccurate software performance estimation  Contribution: propose and integration of an fast and accurate energy models into the OVPsim

7. 2. Open Virtual Platform - OVP Source: [Davidmann and Graham 2014] Proposed Model Location Callback Software stack separated from the simulator Target instructions are translated to host machine binary code

8. 3. Energy Model  Characteristics Instruction-driven energy model Developed on the basis of OVP APIs Run-time based approach ISA-based approach  Advantages It avoids huge amount of memory No trace files are required Model is transparent to the software engineer no pre- or post-processing application/software is required The approach can be applied to other processor architectures  Model called Watchdog

9.  The instruction energy cost information is not available Characterization phase Instructions are organized in groups according their energy cost similarity Less complexity/computation during the characterization and simulation phases  Reference CPU PLASMA Core MIPS Architecture 3-Stage pipeline 100 MHz 65nm low power library from ST Microelectronics  Using Cadence Tools static and dynamic energy 3. Energy Model

10. 3. Energy Model - Watchdog the parser module disassembles the binary code and identifies the instruction that must be executed 2 1 identified instruction is used as a hash table key to ascertain to which class such instruction belongs 3 The energy cost is computed and the instruction is executed in the CPU 4 4 4 2 3

11. - Benchmark Conception - Activity Measurement - Power Acquisition - Energy Calculation Energy Model Creation 1 2 3 4 1 2 3 4 GroupsPower (mW)Exec Time (us)Energy (nJ)# of instEnergy per Inst (nJ) Arithmetic6,456342,7552212,826347640,0636528098 Jump6,046102,600620,320102240,0606728873 Load-Store4,0941042,8004269,223485610,0879146476 Logical4,469349,7351562,966354620,0440743815 Move3,129480,7251504,189393630,0382132593 NOP2,141257,155550,569261300,0210703733 Shift3,824298,7351142,363303620,0376247494 Groups Power (mW) Exec Time (us) Energy (nJ) # of inst Energy per Inst (nJ) Arithmetic6,456342,7552212,826347640,0636528098 Jump6,046102,600620,320102240,0606728873 Load-Store4,0941042,8004269,223485610,0879146476 Logical4,469349,7351562,966354620,0440743815 Move3,129480,7251504,189393630,0382132593 NOP2,141257,155550,569261300,0210703733 Shift3,824298,7351142,363303620,0376247494 Characterization Flow 3. Energy Model 5 5 What about accuracy?

12. 3. Energy Model – Experimental Setup  Benchmarks 19 applications from different research domains WCET and other benchmarks created in house Model estimation compared with a gate-level implementation (PLASMA) #NameSuite ABFSH In-House production B BinarySear ch Mälardale n WCET C BitManipul ation Mälardale n WCET DBubble Mälardale n WCET ECounts Mälardale n WCET FCrc In-House production GEdn Mälardale n WCET HExpint Mälardale n WCET IFactorial In-House production #NameSuite JFftMälardalen WCET KFibIn-House production LHanoiIn-House production MHarmIn-House production NInsertSortMälardalen WCET OMatrixInverMälardalen WCET PMdcIn-House production QPeakspeedImperas RUdMälardalen WCET SUsqrtMälardalen WCET

13.  Benchmarks 19 applications from different research domains MiBench and other benchmarks created in house Model estimation compared with gate-level 3. Energy Model – Accuracy Evaluation  Mismatch is below 6% in 15 out of the 19 adopted benchmarks 0.01 8.56 What about speedup gain ?

14. 3. Energy Model – Achieved Speedup  Comparing each benchmark watchdog estimation execution time with gate-level execution time Achieved speedup varying from 461 to 1577 Mean relative gain 1118 the large application code the more relative gain 461.96 1577,16

15. 3. Energy Model – Scalability  Scenario exploration with up to 1000 CPUs  Each CPU has one Watchdog associated Around 1.8 MIPS Improvement Watchdog Model

16. 4. Exploration in Large Scale Systems  The proposed instruction-driven energy model was integrated into a NoC-based MPSoC model proposed in [Mandelli et al. 2013] Case study: mapping process cost evaluation Nearest Neighbor (NN), first free (FF) and LECDN 8x8 MPSoC size organized in 4x4 clusters Only the heuristic algorithm was observed 5 applications instances: 4 partial MPEG decoder containing 5 tasks 1 DTW containing 10 tasks

17. 5. Conclusion and Future Work  Inclusion of a fast and accurate energy models into OVPsim  Extensive evaluation of both models considering several benchmarks, while comparing it to a gate-level simulation  Approach is ISA/CPU-oriented thus everything is transparent to the software engineer  Programmers can use the same simulator to have fast simulation and accurate software performance evaluation  Limitation of our approach:  we are not considering processors with cache

18. 5. Conclusion and Future Work  Consider memory access power cost Calibrate our model considering NVSim/CACTI  Porting the proposed model to the OVPSim morphing phase  Improve overall model accuracy evaluate load and stores patterns enhance the division and multiplication algorithm estimation  Complex processor architectures as Out-Order or Super Pipeline

19. Questions?

20. Fast Energy Evaluation of Embedded Applications for Many-core Systems Felipe Rosa, Luciano Ost, Thiago Raupp, Fernando Moraes, Ricardo Reis.