Dynamic Voltage and Frequency Management Based on Variable Update Intervals for Frequency Setting Mehrdad Najibi Mostafa Salehi Ali Afzali Kusha Massoud Pedram S. Mehdi Fakhraie Hossein Pedram Presentation given at ICCAD, Nov. 2006
Outline Prior Work Dynamic voltage and frequency management Adaptive frequency scheduling based on an effective deadline Effective deadline prediction Frequency scheduling Underload and overload conditions Experimental results
Prior work Commercial processors with DVS capability Burd et al, 2000, Arm8 processor core Nowka et al, 2002, PowerPC processor Flautner and Flynn, 2002, arm9 processor core Kihwan et a, al, 2003-05, 05, StrongArm Sto and Xscale scae... Nakai et al, embedded microprocessor
Dynamic Voltage and Frequency Management DVFM architecture [Nakai et al, JSSC-05]
Dynamic Voltage Management
Dynamic Frequency Management Frequency scheduling based on fixed interval [Nakai et al, JSSC-05]
Adaptive Frequency Scheduling Based on an Effective Deadline Fixed Interval Voltage (m mv) Freq. (MH Hz) If it is long : Slow workload tracking (fixed freq step) Max workload Slope (Upper bound) If it is short : Higher Avg. Freq Higher # of unnecessary updates 1600 1500 1400 135 120 105 0.00 0.50 1.00 1.50 2.00 2.50 3.00 Adaptive Interval Lower updates Time (ms) Power Network Fluctuations Near optimal frequency & voltage 4 up-1 down avg. freq Steady: 4 up-4 down useless changes
Effective Deadline Soft Real Time Application Periodic Workloads Arrival of the next workload Ati Active cycles Idle cycles Input rate (1) (2) Hypothetic FIFO output rate f peak = 4f eff E 1 = 16E 2 FIFONotEmpty
Prediction of the Effective Deadline Arrival time of the next workload Reported in terms of system cycles Effective Deadline Update the interval Compute the frequency Adaptive Adjustment 2 consecutive increases Interval_step * 2 2 consecutive idecreases Interval_step / 2 Up Interval Next Interval Current Interval Step Next Interval Current Interval Down Interval Step
Frequency Adjustment Frequency Up thl thh update Frequency Down thl thh update thl thh Frequency Steady update
Overload and Underload Conditions Interval without overload detection Overload (no idle cycle) 600 500 Slower freq increase (worsen the situation) 400 300 200 100 0 Interval Interval with overload detection Underload (no active cycle) 200 150 100 50 0 Slower freq decrease Interval
Results (Periodic Workload) 1600 Fixed Interval 1400 Voltage (mv) Voltag ge (mv) 1200 1600 1000 1400 800 1200 FI 80% AI 80% 600 FI 20% 1000 0.00 5.00 10.00 15.00 20.00 25.00 30.00 AI 20% 800 Time (ms) Frequency (MHz) 140 120 100 80 60 40 20 FI 80% AI 80% FI 20% AI 20% 600 120 0 2 4 6 8 10 140 Adaptive Interval Time (ms) 0 0 2 4 6 8 10 Time (ms) Freq. (MHz) 100 80 60 Voltage Frequency 40 20 0 0.00 5.00 10.00 15.00 20.00 25.00 30.00 Time (ms)
Results power (mw) 35 30 )Workload 25 20 15 10 Power Consumption Fixed Interval Adaptive Interval Frequency V DD Power Frequency V DD Power Power (MHz) (mv) (mw) (MHz) (mv) (mw) Reduction 0% 8 600 Fixed Interval 0.95 Adaptive Interval 8 600 0.5 47.37% 10% 21 875 2.16 12 810 0.92 57.41% 5 20% 41 990 4.57 25 895 2.17 52.52% 0 30% 61 1135 8.51 36 945 3.73 56.17% 0 10 40% 20 30 40 50 82 60 70 80 90 100 1285 14.01 70% 49 1050 5.89 57.96% 50% workload % 102 1430 21.07 60% 61 1121 8.38 60.23% 60% 123 1580 30.02 50% 74 1230 11.51 61.66% 70% 123 1585 30.59 40% 86 1315 15.32 49.92% 80% 123 1585 30.77 98 1400 19.52 36.56% 90% 123 1585 30.89 30% 110 1485 24.68 20.10% 100% 123 1580 30.92 20% 123 1585 30.88 0.13% 10% 0% 0 10 20 30 40 50 60 70 80 90 100
Results (Correlated Workloads) 70.00% Power/Op - Pe erfomance Reduct tion 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% Power/Operation Performance 0.00% 0% 10% 20% 30% 1.6 40% 50% 60% 70% Normmalized AI/FI Metrics Correlation Percentage 1.4 1.2 1 0.8 0.6 80% 90% Delay Power PDP EDP 0.4 0% 20% 40% 60% 80% 100% Correlation Percent
Summary Presented an efficient adaptive method to perform dynamic voltage and frequency management (DVFM) for minimizing the energy consumption of microprocessor chips Uses adaptive update intervals for optimal frequency and voltage scheduling Rapidly tracks the workload changes so as to meet soft real-time deadlines Utilizes the correlation between consecutive values of the workload for future workload prediction Because the frequency and voltage update rates are dynamically set based on variable update interval lengths, voltage fluctuations on the power network are minimized The technique leads to power savings of up to 60% for highly correlated workloads compared to DVFM systems based on fixed update intervals.