110 likes | 247 Vues
This document examines the design of experiments to optimize the Completely Fair Scheduler (CFS) in Linux, initiated since version 2.6.23. By utilizing Red-Black Trees for maintaining balance in processor time allocation and exploring machine learning techniques, we propose an energy-based model to adjust core voltage based on the ready queue length. Our strategies aim to enhance task execution efficiency while conserving energy. We evaluate power consumption and benchmark performance using L4Linux, applying various predefined strategies.
E N D
Outline • Models • Design of experiments
Current Scheduler • Completely Fair Scheduler(CFS) • Since Linux 2.6.23 • /kernel/sched.c • Maintain balance (fairness) in providing processor time to tasks. • Red-Black Tree
Our Ideas • Base on ready queue length • Base on machine learning • Energy-based model
Assumptions • Core voltage cannot be adjusted to any value, but some pre-defined value. • Level • We know N, the number of tasks waiting in the ready queue. • Each task is only in core for execution during a short period before being swept out and back to ready queue.
Base on Queue Length - 1 • Assume the time slice is fixed. • Higher core V => more cycles => more instructions executed => tasks can finish earlier. • Two methods • For every period of time, if N increases => raise core voltage. • For every period of time, if N doesn’t decreases => raise core voltage.
Base on Queue Length - 2 • Assume the instruction retired is fixed. • Higher core V => more cycles => shorter time in core. • All N tasks should be executed at least once in T. • TmaxV ≦ T ≦ TminV • Adjust core voltageaccording to N. Vhigh Nt= 6+4 = 10 Vlow T
Energy-based • Assume the V×t, energy of each task being executed once is fixed. • Similar to the previous one. • However, V×t is adjusted according to Battery Power Remaining.
Design of Experiment • First we measure the power consumption and benchmark performance using L4Linux predefined strategy. • Vhighest and Vlowest. • Apply our strategies and measure: • Only one benchmark • Multiple benchmarks • Same • Different
Paper Study • An Energy Conservation DVFS Algorithm for the Android Operating System • 2012 • An Energy Conservation DVFS Algorithm for the Android Operating System • 2010 • An Energy Conservation DVFS Algorithm for the Android Operating System • 2005