1. NC STATE UNIVERSITY 1 Feedback EDF Scheduling w/ Async. DVS Switching on the IBM Embedded PowerPC 405 LP Frank Mueller North Carolina State University, Center for Embedded Systems Research (CESR) Departments of Computer Science

2. NC STATE UNIVERSITY 2 Embedded Systems Energy management –battery-constrained embedded systems –vital design constraint Dynamic Voltage Scaling (DVS)

3. NC STATE UNIVERSITY 3 Dynamic Voltage Scaling (DVS) Processor core voltage ↑ or ↓ depending on computation demands –V ↓ f ↓ – DVS in real time systems –calculate safe operation frequency –guarantee all tasks meet deadlines

4. NC STATE UNIVERSITY 4 Why PowerPC 405LP?  Hardware support for DVS  Software support to change voltage –via user-defined operating points for e.g., 266 MHz @ 1.8 V to 33 MHz @ 1 V  Ability to execute instructions while frequency/voltage is changed –doesn’t enter sleep mode during frequency/voltage transitions

5. NC STATE UNIVERSITY 5 PPC 405LP Development Board  Embedded Linux variant  Fast DVS switching (≤ 200 μs)  Two DVS modes –synchronous (blocking) conventional DVS-capable processors –asynchronous (non-blocking) execution may proceed during switch unique feature in 405LP f low f high exec sleep exec f high f low

6. NC STATE UNIVERSITY 6 Asynchronous Switching voltage current System call for asynchronous switch Voltage ramped up to max –time to reach new voltage level is estimated High-resolution timer interrupt triggered –power management unit reprogrammed –new processor frequency activated –voltage settles in a controlled manner to new operating point frequency 1 2 3 1 2 3

7. NC STATE UNIVERSITY 7 Dynamic Power Management (DPM)  Software and hardware components  Define stable operating points (freq./voltage pairs)  Support both synchronous and asynchronous DVS

8. NC STATE UNIVERSITY 8 EDF-DVS on PPC 405LP  User-level threads package under Linux  Real-time earliest deadline first (EDF) scheduler  Several hard real-time software DVS techniques –static –cycle-conserving –look-ahead 1+2 –Simple + PID feedback –AGR-2  Voltage and current measurements –analog data acquisition board –oscilloscope Pillai and Shin, 2001 Zhu and Mueller, 2002 Aydin et al., 2002

9. NC STATE UNIVERSITY 9 Threads Library  Kernel-level threads –complex kernel modifications – error-prone –more control over execution  User-level threads –simplicity of design –develop generic and portable threads library –facilitate easier debugging –less control over execution and resources –however, OS background activity minimal

10. NC STATE UNIVERSITY 10 Threads Library Implementation main () Spawn T 1 Spawn T 2

11. NC STATE UNIVERSITY 11 time Task A Task B EDF B1B2B3B4 A1A2 period A period B release A1 deadline A1 release A2 release B1 deadline B1 release B2 B1A1B2B3B4A2 EDF schedulability test EDF Scheduling

12. NC STATE UNIVERSITY 12 Example Task Set TaskPeriodWCETActual Exec Time 11021 2 21 3 42 U = α = 2/10+2/10+4/10 =0.8 f = α x f max = 0.8 x f max

13. NC STATE UNIVERSITY 13 Static RT-DVS EDF f max αf max α = WC utilization 0 ≤ α ≤ 1 static slack dynamic slack

14. NC STATE UNIVERSITY 14 Cycle-Conserving RT-DVS EDF f max αf max α = WC utilization 0 ≤ α ≤ 1 static slack dynamic slack

15. NC STATE UNIVERSITY 15 Look-Ahead RT-DVS EDF f max αf max α = WC utilization 0 ≤ α ≤ 1 static slack dynamic slack

16. NC STATE UNIVERSITY 16 Feedback RT-DVS EDF f max αf max α = WC utilization 0 ≤ α ≤ 1 static slack No remaining dynamic slack

17. NC STATE UNIVERSITY 17 Experimentation Methodology  Need to measure voltage/current at a high rate  Multimeters – coarse-grained and low precision High frequency analog data acquisition board –current measured as voltage level over a resistor with 1V drop per 360 mA  Oscilloscope –visualizing voltages and currents –high precision

18. NC STATE UNIVERSITY 18 Valid Frequency/Voltage Pairs Setting01234 CPU Frequency (MHz)334466133266 Bus Frequency (MHz)334466133 Voltage (volts)1.0 1.11.31.7

19. NC STATE UNIVERSITY 19 DVS Overhead ActivitySynchr. DVS Async. DVS Signal handler Overhead117-162 μs8-20 μs0.07-0.6 μs

20. NC STATE UNIVERSITY 20 Scheduler Overhead DVS Sched. CPU Freq. StaticCcLook- ahead PID- feedback 33 MHz217 μs487 μs2296 μs3612 μs 44 MHz170 μs366 μs1714 μs2943 μs 66 MHz100 μs232 μs1112 μs1728 μs 133 MHz52 μs120 μs546 μs801 μs 266 MHz36 μs76 μs229 μs472 μs

21. NC STATE UNIVERSITY 21 Task Sets Task Task Set 1 (msec) Task Set 2 (msec) Task Set 3 (msec) Period (P i ) WCET (C i ) Period (P i ) WCET (C i ) Period (P i ) WCET (C i ) 12400400600809012 224006003201204818 3120020040040606

22. NC STATE UNIVERSITY 22 Energy Savings of DVS Techniques

23. NC STATE UNIVERSITY 23 Energy Savings of fast DVS modulation

24. NC STATE UNIVERSITY 24 Energy Consumption for 3 Tasks

25. NC STATE UNIVERSITY 25 Energy Consuption for 30 Tasks

26. NC STATE UNIVERSITY 26 Oscilloscope 360mA 1V 0mA 2V 360mA 1V 0mA 2V 360mA 1V 0mA 2V 360mA 1V 0mA 2V time

27. NC STATE UNIVERSITY 27 Conclusions  State-of-the-art EDF-DVS on PowerPC 405LP  Measured energy oDiffers from simulations oBut match trends from simulations  Feedback EDF-DVS obetter than any other scheme oup to 64% reduced energy, 24% better than others  Async. Switching  only 1-5% energy savings  First comparative study on actual processor

28. NC STATE UNIVERSITY 28 Acknowledgements  Bishop Brock – IBM Research (Austin) –DPM S/W extension –Linux kernel support –H/W modifications –DVS overhead data  IBM Research (Austin) –Power PC 405LP development board  Students –Yifan Zhu –Ajay Dudani –A. Anatharam, A. Mahmoud, R. Venkatesan

29. NC STATE UNIVERSITY 29 Questions?

30. NC STATE UNIVERSITY 30 Backup Slides

31. NC STATE UNIVERSITY 31 Real-Time Scheduler Integration

32. NC STATE UNIVERSITY 32 Worst-Case Timing Analysis  No tool support to derive WCETs of tasks on PPC405LP  Timing loop computes average execution of N instances of the task taking into account loop overhead and cold cache misses t1 = gettimeofday(); for ( i = 0; I < N; i++); t2 = gettimeofday(); work(); t3 = gettimeofday(); for ( i = 0; I < N; i++) work(); t4 = gettimeofday();

33. NC STATE UNIVERSITY 33 DVS Overhead Activity Sync. DVS Async. DVS Signal handling Overhead117-162 μs8-20 μs0.07-0.6 μs

34. NC STATE UNIVERSITY 34 Oscilloscope – Tight Task Set 360mA 1V 0mA 2V 360mA 1V 0mA 2V 360mA 1V 0mA 2V 360mA 1V 0mA 2V time