Balancing Power Consumption in Multiprocessor Systems

1. Andreas Merkel Frank Bellosa System Architecture GroupUniversity of Karlsruhe Balancing Power Consumptionin Multiprocessor Systems EuroSys '06 April 18 - 21, 2006 Leuven, Belgium

2. State of the Art in Thermal Design Increasing power dissipation in computer systems Ever increasing cooling costs

3. State of the Art in Thermal Design Increasing power dissipation in computer systems Ever increasing cooling costs Alternatives for designing cooling systems: Worst case thermal design Overprovisioning Unnecessarily high cooling cost Lower thermal design power Throttling to handle “hot” tasks Performance penalties

4. Alternatives in SMP Systems Multiple processors, running different tasks Tasks differ in power consumption “hot” and “cool” tasks Hot and cold processors Unnecessary throttling T temperature limit t

5. Alternatives in SMP Systems Multiple processors, running different tasks Tasks differ in power consumption “hot” and “cool” tasks Hot and cold processors Unnecessary throttling T throttling temperature limit t

6. Our Approach: Migration instead of Throttling Leverage cool processors Migrate hot tasks away to a cool processor Combine hot tasks with cool tasks on a processor Balance processor temperatures T migrate temperature limit t

7. Our Approach: Migration instead of Throttling • Leverage cool processors • Migrate hot tasks away to a cool processor • Combine hot tasks with cool tasks on a processor • Balance processor temperatures T temperature limit t

8. Our Approach: Migration instead of Throttling • Leverage cool processors • Migrate hot tasks away to a cool processor • Combine hot tasks with cool tasks on a processor • Balance processor temperatures • Contributions: • Characterization of tasks • Task Energy Profiles • Policy for assigning tasks to CPUs • Energy-Aware Scheduling

9. Outline Task Energy Profiles Thermal Model Energy-Aware Scheduling Evaluation Conclusion

10. Characterizing Tasks Temperature High thermal capacitance of chip and heat sink Temperature changes slowly Short scheduling intervals CPU temperature: Mix of multiple tasks' characteristics No knowledge about individual task's characteristics Power consumption 37W to 61W on Pentium 4 Xeon (2.2 GHz) for different phases of compute-intensive tasks Characterize tasks by their individual power consumption idle

11. Task Energy Profiles Definition: Energy consumed during one timeslice Behavior of tasks depends on input data Online energy estimation required Tasks show periods of constant power consumption Use knowledge from the past: Energy consumed during last timeslice Requirement: Determine the amount of energy the CPU consumes during one timeslice

12. Event-Driven Energy Estimation Use performance monitoring counters for estimating energy: Choose set of suitable events Assign amount of energy to each event Count events Calculate linear combination of the counter values Error < 10% for real-world integer applications Neglecting floating point applications for lack of suitable counters

13. Thermal Model Model of processor and heat sink Relates power consumption to temperature Models temperature with exponential function

14. Thermal Model Model of processor and heat sink Relates power consumption to temperature Models temperature with exponential function Temperature Power

15. Energy Aware Scheduling Objectives: Minimize the need for throttling processors Avoid unnecessary migrations (cache affinity) Policy depends on number of tasks per runqueue • More than one task • Combine tasks with different characteristics • Energy balancing • One task • Migrate hot task before CPU overheats • Hot task migration

16. Linear Energy Balancing Goal: Balance CPU temperatures Approach: Balance CPU power consumption Metric: runqueue power Average of the energy profiles of all tasks in a runqueue Predicts future energy consumption Balance power consumption by balancing runqueue power

17. Problem: Runqueue power changes frequently Whenever tasks start or terminate Whenever tasks block/unblock Ping-pong effects Linear Energy Balancing migrate hot task P migrate hot task runqueue power CPU A runqueue power CPU B cool task blocks cool task unblocks t

18. Runqueue power does not consider tasks that are blocked or have terminated Heat produced by those tasks is still stored in the chip and the heat sink Runqueue power cannot distinguish between hot and cool CPUs Linear Energy Balancing

19. Fit averaging function to thermal model Use exponential average of CPU's power consumption as metric Thermal power Reflects past energy consumption  temperature Calibrate parameters to thermal model Exponential Energy Balancing temperature thermal power power t

20. Exponential Energy Balancing Problem: Task migrations have no immediate effect on thermal power “Over-balancing” migrate hot tasks P thermal power CPU A thermal power CPU B t

21. Energy Balancing Use both metrics for energy balancing Migrate a hot task from CPU Ato CPUBif Aoutranges Bin both metrics Avoids ping-pong effects Avoids over-balancing Introduce threshold to defer migrations Hysteresis migrate P t

22. Energy Balancing Use both metrics for energy balancing Migrate a hot task from CPU Ato CPUBifAoutranges Bin both metrics Avoids ping-pong effects Avoids over-balancing Introduce threshold to defer migrations Hysteresis P t

23. Hot Task Migration Only one task in a runqueue Balancing not possible Migrate task to cooler CPU if CPU temperature comes close to maximum Search for cool target CPU Idle CPU Migrate task CPU executing cool task Swap tasks

24. Evaluation Implementation of energy aware scheduling for the Linux kernel Test system: 8-way Pentium 4 Xeon, 2.2 GHz Workload scenarios Power consumption of tasks ranging from 37W to 61W Measure throughput by counting number of tasks finished per time unit Expect increase in throughput if energy-aware scheduling is enabled

25. Evaluation: Energy Balancing Disabled • Scenario: 8 CPUs executing different tasks • Energy balancing increases throughput by 4.7% • Enabled 60W 60W 40W 40W 20W 20W 400s 600s 400s 600s 0s 200s 800s 0s 200s 800s

26. Evaluation: Hot Task Migration Disabled • One hot task • Hot task migration avoids throttling • Enabled 60W 60W 40W 40W 20W 20W 100s 150s 100s 150s 200s 0s 50s 0s 50s 200s

27. Homogeneous vs. Heterogeneous Workloads Workloads consisting of different combinations of memrw (38W), pushpop (47W), and bitcnts (61W)

28. Conclusion • Need to throttle processors exceeding thermal design power • Approach: leverage cool processors instead of throttling hot processors • Characterize tasks by power consumption • Energy profiles • Use energy profiles for energy-aware scheduling • Energy balancing • Hot task migration • Reduce thermal imbalances • Minimize throttling  increase duty cycles