1. Runtime Software Power Estimation and Minimization Tao Li

2. Power-aware Computing

3. Power: Software Perspective & Impact  Power estimation: the first step to power management & optimization  Software contributes to & largely impacts power consumption

4.  It is crucial to model power from the perspective of software  Evaluate software energy in early design stage  Understand impact of software optimizations on energy  Support run-time power management and optimizations Power: Software Perspective & Impact (Contd.)

5.  Instruction level modeling  Computation intensive  High level macro-modeling  Difficult to apply to general code  Event counting based modeling  Impacted by the availability of performance counters  Architecture level simulation  Large slowdown Software Power Estimation: Current Techniques

6. Challenges in Run-time Power Estimation  High fidelity & fast speed  On-the-fly estimation capability, non- intrusive & low overhead  Simplicity, availability and generality

7. Experimental Methodology  SoftWatt: cycle-accurate & full-system power simulation framework  SimOS infrastructure, Wattch power model  Commercial OS & real applications  Out-of-order superscalar processor  Caches & memory hierarchy  Low-power disk

8. Experimental Methodology (Contd.)  Applications  E-mail and file management (sendmail, fileman)  Java (SPECjvm98: db, jess, javac, jack, mtrt, compress)  SPECInt95 (gcc, vortex)  Database (Postgres: select, update, join)  Miscellaneous (pmake, osboot)

9. OS Power Characterization  OS power varies from one application to another  29 Watt (gcc) ~ 66 Watt (fileman)  Variance of power consumption in OS service routines & invocations

10. OS Power Characterization (Contd.)  OS routine power correlates with its performance  Circuits used to exploit ILP burn significant portion of power  The number of in-flight instructions that flow through impacts circuit switching activity  For a given OS routine, similar IPC indicates similar circuit switching activity and therefore, similar power

11. OS Routine Power-Performance Correlation SCSI Disk Interrupt Handler Read File System Call

12. Routine Level OS Power Model  Idea: use a linear regression model P routine =k 1 *IPC routine +k 0 to track the OS routine power showing different performance  Energy(OS)= Sum [ Energy(OS routines) ] = Sum [ Power(OS routines)*Time(OS routines) ]

13. Routine Level OS Power Model (Contd.)  : Model Fitting Error

14.  Pre-characterization  Low level energy simulation  Model fitting  Run-time estimation  OS routine boundaries  Evaluation using counter values Routine Level OS Power Modeling

15. Routine based Regression Model P routine =k 1 *IPC routine +k 0 Flat Regression Model P OS =g 1 *IPC OS +g 0 Cumulative Estimation Error

16. Flat Regression Model P OS =g 1 *IPC OS +g 0 Per-routine Estimation Error

17. Routine based Regression Model P routine =k 1 *IPC routine +k 0 Per-routine Estimation Error (Contd.)

18. OS Energy Dissipation 92% 89%

19. Phases in Programs (8-issue machine) Benchmark: SPECjvm98 jess  Resources are utilized differently during different phases of program execution  Average IPC - User: 2.1, OS: 1.1

20. Power Minimization via Processor Resource Adaptations  Adapt processor resources to program needs  What can be adapted?  Bandwidth of fetch/decode/issue/retire…  Size of instruction window, re-order buffer, load store queue…  Reduce power, retain performance

21. Effects of Tuning Processor Resource for the OS 8-issue -> 4-issue OS Performance degradation: 4% OS Power savings: 50%

22. Previous Approach for Adaptations  Sampling Cycles Sampling Window IPC (Inst. Per Cycle) Adaptation ABCDEF

23. Problems with Sampling based Adaptations (Contd.)  OS executions  Short-lived

24. OS-aware Routine based Adaptations  OS-aware:  Identify OS executions via processor execution modes  Just-in-time & full coverage of OS activities  Routine-based:  Adapt processor resources at OS routine boundaries  Precise exceptions: drained pipeline  Achieve minimum adaptation overhead

25. OS-aware Routine based Adaptations (Contd.)  Apply optimal adaptation for individual OS routine  Exploit the routine level Energy-Delay Product variance OS Services

26. Routine based Adaptations: OS Power

27. OS Performance

28. OS Power & Performance Tradeoff