Presentation is loading. Please wait.

Presentation is loading. Please wait.

Power Capping Via Forced Idleness ANSHUL GANDHI Carnegie Mellon Univ. 1.

Similar presentations


Presentation on theme: "Power Capping Via Forced Idleness ANSHUL GANDHI Carnegie Mellon Univ. 1."— Presentation transcript:

1 Power Capping Via Forced Idleness ANSHUL GANDHI Carnegie Mellon Univ. 1

2 Server Rack Power Consumption 2 Problems: (1) Power delivery constraints (2) Cooling requirements infeasible kW/racks 30 kW 2 kW 10 kW Source: APC White Paper #46, 2005

3 3 Limits average power consumption of servers over a time interval to stay below specified threshold. Power Capping Use clock-throttling or DVFS to control power consumption. Frequency (GHz) Power (watts) Temperature (⁰F) Lefurgy, Wang and Ware; ICAC, 2007. Wang and Chen; HPCA, 2008. Clock-throttling Use clock-throttling or DVFS to control temperature.

4 Effect on performance 4 Significant performance loss across workloads Mean Response Time (secs) Power cap (watts) Mean Response Time (secs) Power cap (watts) CPU bound “DAXPY”Memory bound “STREAM” 7X worse 3X worse Source: Our experimental results on an IBM Blade Clock-throttling

5 Goal 5 Dual Goal Power CappingReduce Mean Response Time IdleCap IdleCap can reduce mean response time 2X-4X across workloads over existing power capping. (Results based on clock throttling)

6 6 How IdleCap works Frequency (GHz) Power (watts) C1E Clock-throttling IdleCap Existing power capping: Dithers between adjacent states. IdleCap: Dithers between extreme states: C1E, 3 GHz. IdleCap achieves higher frequency for any power cap.

7 7 Example: 170 Watts Frequency (GHz) Power (watts) Clock-throttling IdleCap Clock throttling: 170 Watts  0.7 GHz IdleCap: 170 Watts  1.5 GHz Same power cap, twice the frequency. = 1.5 GHz

8 8 Analysis Power Cap IdleCapr 170 W1.5 GHz1/2 195 W2.25 GHz3/4 r: Fraction of time spent in 3 GHz state Frequency (GHz) Power (watts) Clock-throttling IdleCap

9 9 Analysis Power Cap IdleCapr 170 W1.5 GHz1/2 195 W2.25 GHz3/4 r: Fraction of time spent in 3 GHz state Mean Response Time when running at 3 GHz Mean Response Time with IdleCap Entirely Predictable

10 Experimental Setup 10 IBM BladeCenter HS21 3GHz, quad core, 4GB RAM C1E idle state Workload CPU bound (LINPACK, DAXPY) Memory bound (STREAM) Alternation period: Time between successive entries to C1E state 1millisec, 10millisec, 100millisec, 1sec, 10sec

11 Results: “DAXPY” 11 Mean Response Time (secs) Power cap (watts) Alternation period (secs) Predicted E[T] = Observed E[T] Up to 4X reduction in Mean response time. Overheads due to alternations: 6% IdleCap Clock-throttling IdleCap IdleCap theory Alternation period: 1 second

12 Results: “LINPACK” 12 Mean Response Time (secs) Power cap (watts) Alternation period (secs) Predicted E[T] = Observed E[T] Up to 3X reduction in Mean response time. Overheads due to alternations: 15% IdleCap Clock-throttling IdleCap IdleCap theory

13 Results: “STREAM” 13 Mean Response Time (secs) Power cap (watts) Alternation period (secs) Predicted E[T] = Observed E[T] Up to 2X reduction in Mean response time. Overheads due to alternations: 10% IdleCap Clock-throttling IdleCap IdleCap theory

14 IdleCap and DVFS 14 Power (watts) Frequency (GHz) IdleCap Use advanced idle states (C6 on Intel). DVFS C1E Problem: Concave upwards Future: Concave downwards C6

15 IdleCap and DVFS 15 Power (watts) Frequency (GHz) DVFS IdleCap Non-linear IdleCap applies to lower frequency range. C6

16 Conclusions 16 Response Time (secs) Mean Response Time (secs) Power cap (watts) Alternation period (secs) IdleCap is superior to existing power capping techniques based on clock throttling. Applies to various workloads under alternation periods as small as 1 millisecond. Response Time (secs) DAXPY LINPACKSTREAM LINPACK


Download ppt "Power Capping Via Forced Idleness ANSHUL GANDHI Carnegie Mellon Univ. 1."

Similar presentations


Ads by Google