Download presentation
Presentation is loading. Please wait.
1
DPM Dynamic power management
2
DPM Tree DPM Timeout Adaptive Device dependent Predictive L-shape Exponential average Predictive wakeup Adaptive Disk shutdown Predictive shutdown OS-directed Task-based Task scheduling Stochastic Sliding window Competitive Learning tree
3
Dynamic power management (DPM) reduce power consumption of electronic systems Most common to shut down idle components. Timeout Predectiv Stochastic OS-directed power management. Advanced configuration and power interface, ACPI. Introduction
4
Adaptive timeout. 1.The ratio between τ and the latest idle period: short -> increase τ, long -> decrease τ. 2.τ is updated asymmetrically, increase with 1 s or decrease with ½ s. 3.Change according to the latest busy period : short -> decrease τ, long -> increase τ. Device dependent timeout. τ based on the break-even time of the device under control. C-competitive If τ is equal to the break-even time this algorithm is proven to be 2-competitive Timeout
5
Short busy periods are followed by long idle periods Long busy periods are followed by short idle periods Problem in the left corner, only short periods L-shape
6
Uses both the predicted and the actual lengths of a previous idle period to predict the next idle period P(n+1) = a*I(n)+(1-a)*p(n) The constant a has a value between 0 and 1 Exponential average
7
Predictive wakeup and shutdown The power manager performs a predictive wakeup, even if there is no incoming request. Hard to comput the right lenght of the idle period Take the decision to shutdown based on observation of the previous idle- and busy periods Take the decision to shutdown based on observation of the recent busy period
8
The requests clusters together to sessioner. Shut down the hard disk between sessions. It is hard to decide how long the treshold should be. An adjustment parameter decide when the disk should shut down General adaptive algorithm. Adaptive disk shutdown
9
Stochastic model 0,95 0,05 0,12 0,88 0 1 0 1 0 1 P = 0 1 0,88 0,12 0,05 0,95 Service requestor Service requestor with one transition matrix Service provider with two transition matrices Queue with four transition matrices Power manager Cost metrics
10
Sliding window 11110000 W(0) W(1) W(2) W(3) W(4) W(5) …………… W(WS-2) W(WS-1) ……………… Sliding window is based on the stochastic model It is used for non-stationary service requests The basic window operation is to shift one slot constantly every time slice The shutdown decision is evaluated each period, thus causing overhead Single- or a multi window approach
11
Learning tree 1 312 32e bcd a Adaptive learning tree can control multiple sleeping states A sequence of idle periods is transformed in to a sequence of discrete events All leaf nodes are predictions for the next idle period and store the Prediction Confidence Level (PCL)
12
ACPI Motherboard deviceChipsetCPU Platform hardware BIOS Table interface BIOS interface Register interface ACPI tablesACPI BIOSACPI Registers Device drivers Kernel ACPI drivers AML interpreter PM Application OS ACPI ACPI is a uniform HW/SW interface for power management It specifies an abstract and flexible interface between hardware components
13
Task-based power management TBPM is a software-centric approach TBPM uses a two-dimensional data structure, U, and a vector, P The matrix U stores the relation between devices and requests To update U the same approach as in exponential average is used P contains the percentage of CPU time executing task r P is updated based on sliding window
14
Task scheduling 1 1 1 2 2 2 3 3 T1T2T3T1T2T3 idle T time 1 1 1 2 2 2 3 3 T1T2T3T1T2T3 idle time This algorithm uses task scheduling and tries to make as long idle periods as possibly Every task has a required device set, RDS This algorithm can also schedule multiple devices
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.