Real- time Dynamic Voltage Scaling for Low- Power Embedded Operating Systems Written by P. Pillai and K.G. Shin Presented by Gaurav Saxena CSE 666 – Real Time Systems

Introduction  Energy consumption is an important issue in embedded systems Mobile and portable device Laptops, PDA’s Cellular phones, camcorders  Real Time Dynamic Voltage Scaling (RT-DVS) Reduces Energy consumption by lowering the supply voltage and operating frequency. Modify OS’s real-time scheduler and task management service to maintain real-time deadline

Facts  High performance is needed for a small fraction of time, while for the rest of time, a low performance, a low power processor would suffice  Low performance is achieved by lowering the operating frequency of the processor when full speed is not needed.  Processors are based on CMOS Logic - has voltage dependent maximum operating frequency– V α f ( so at lower frequency, the processor can operate at lower voltage)  Energy dissipated per cycle E α V 2 (Higher the frequency level, higher the energy consumption)

RT-DVS Algorithms  Static Voltage Scaling EDF  Cycle conserving RT- DVS  Look Ahead RT - DVS

Static Voltage Scaling EDF: Motivation wc1wc2 wc3 wc4 Holes in the pre-run schedule imply: EDF Test: ∑(wc i /p i ) < 1 at frequency = F max In other words, whenever ∑(wc i /p i ) < 1 there are holes in the EDF schedule Next arrival of T1 WCi = worst case computation Fmax

Static Voltage Scaling EDF: exploiting holes wc1 wc2wc3wc4 Next arrival of T1 Pre-run schedule with holes WC i = worst case computation F max Processor typically idles during holes. Instead, the holes can be exploited to slowdown the processor to save energy

Static Voltage Scaling EDF wc1 wc2wc3wc4 K*wc1K *wc2K * wc3K * wc4 EDF Test: ∑(wc i /p i ) < 1 at maximum frequency = F max Static-VS EDF Test: K* [∑(wc i /p i )] = 1 at frequency = F max /K Next arrival of T1

Static EDF: example Available frequencies: 1.00, 0.75 D T1 : 16 D T2 : 10 D T3 : 14 D T1 : 24 D T2 : 20 D T3 : 28 D T1 : 16 D T2 : 20 D T3 : 28 D T1 : 16 D T2 : 20 D T3 :

What if (C i < WC i )?? K*c1K *c2K * c3K * c4 Next arrival of T1 More holes left unexploited Actual computatio n time

What if (C i < WC i )?? K*c1K *wc2K * wc3K * wc4 Next arrival of T1 Actual computatio n time Task T1 completes Slow down all these tasks proportionally Hole of size = (wc1 – c1)

What if (C i < WC i )?? (contd..) K*c1 K’ *wc2K’ * wc3K’ * wc4 Next arrival of T1 CPU Cycles are conserved by slowing down the remaining tasks

Cycle Conserving RT-DVS  When a task set completes its first release, compare real execution time with worst case specified initially.  Any idle time in that period can be used to conserve energy  Rescale frequency that avoid idle cycles, surplus time is used to run other remaining tasks at lower frequency

Cycle conserving EDF: Example Task (Fmax): T1 = (3,6) and T2 = (6,12) U = 3/6 + 6/12 = (Fmax) T1 013 Frequency FmFm Time T2 10 New utilization = 1/6 + 6/12 = 0.67 Finding the right “k” (1*k)/6 + (6*k)/12 = 1 K = (1/0.67) New freq = (0.67) Fmax

DVS-RT Problem Definition  N Tasks each caracterized by: A fixed period P i A deadline D i Worst-case Execution Time (WCET) C i Actual Execution Time AC i Objective: Find the best task scheduling and clock frequency in order to: Minimizing energy consumption. Meeting all the deadlines.

Task set: Task Worst case computation time Period/ Deadline Invocation 1 (ac1) Invocation 2 (ac2) 13ms 8ms2ms1ms 23ms10ms1ms 3 14ms1ms At freq. : f m

Look-Ahead EDF  Defer as much works as possible and set initially to the minimum possible frequency.  Hence at later stage if a task uses much less than it worst case, deferred work may never be needed.  It ensures that there are sufficient cycles available for each task to meet its deadline after reserving cycles for higher priority jobs.  Best saving of energy.

(a)Plan to defer T3’s execution time until after D1 but, by D3 Look-ahead EDF step 1 of ms D freq. T3 D2D3 Reservation for T1 Reservation for T2

(b) Find u so as to finish T1 by D1 and T2 by D2 Look-ahead EDF step 2 of ms T2 D freq. T3 D2D3 Reservation for T1 Reservation for T2T1 T2

(c) u = 0.75, T1 finishes earlier, find new u for T2 to finish by D2 Look-ahead EDF step 3 of ms T2 D freq. T3 D2D3 Reservation for T1 Reservation for T2 T1 T2 t = 2.67

Look-ahead EDF step 4 of ms D freq. T3 D2D3 Reservation for T1 Reservation for T2 T1 t = 4.67 T2 (d) α = 0.5, T2 finishes earlier, enough time until D1, but EDF is work conserving, launch T3 at u = 0.5

Look-ahead EDF step 5 of ms freq.D2D3 T1 t = 8.00 Reservation for T2 T1 T2T3 (e) Guess for T1 again,

Look-ahead EDF step 6 of ms freq.D2D3 T1 t = 8.00 T1T2T3T2T3 (f) u= 0. 5, every task is dynamically scheduled successfully

Relative performance schemes Energy savings Look Ahead EDF Cycle Conserving EDF Static Voltage Scaling EDF

Implementation  Scheduler hooking in the kernel

References and Credits  Real time dynamic voltage scaling for low power embedded operating systems, P. Pillai and K.G. Shin, in ACM SOSP, pages ,  Dynamic Voltage Scaling Algorithms for Hard Real Time Systems, Andrea Romei, Università di Pisa  RT- DVS Algorithms by Sungwuk Jung  Energy aware real time systems by G. Sudha Anil Kumar, Iowa State University