Online Strategies for DPM in Systems with Multiple PowerSaving States

1
Online Strategies for DPM in Systems with
Multiple Power-Saving States

• Presented by Xiaoyu Yao for 896ac
• 2006.2.1

2
Motivation for Energy Efficiency

• Energy Bills (up to 25 TCO)
• 8M per year for a 30,000-square-foot data center
  DoE03
• Increase as much as 25 annually
• Temperature
• Intels thermal wall
• Cooling cost
• Device Lifetime
• Sensor network
• CPU/Memory/Disk reliability

3
Dynamic Power Management

• Transition to lower power usage state when the
  device is idle.
• Additional time and energy are required to
  transition back to active state when a new
  request for service arrives.
• What is the best threshold for power transition?
• Too soon pay start-up costs too frequently
• Too late spend too much time in high-power state

4
Device Level Support

• Speed-scaling
• Multiple power states
• Multiple shutdown states
• Multiple operating states
• Power down
• Simple on/off state

5
Algorithm Issues

• Decision problem
• Offline version
• Assume an oracle knows the future idle sequence
• Online version
• Idle Predication based on previous idle sequence
• Stochastic-control based method
• Need assumption about job inter-arrival/service
  time
• Computationally expensive

6
How to evaluate?

• Competitive analysis
• C-competitive (Competitive Ratio) compared with
  offline optimal version
• Worst case analysis similar to approximation
  algorithm analysis
• Summary of previous algorithm based on
  competitive analysis
• Deterministic algorithm CR gt 2
• Probability algorithm CR lt e/(e-1)1.58

7
System Model

• Device Power States
• (S_0,,S_k)
• State Transition
• Threshold
• T_ij
• P_ij
• Idle Sequence
• Trace from simulator

S_0
a_0
b_i
S_i
a_i
b_k
b_j
S_j
a_j
S_k
a_k
8
Deterministic Algorithm - LEA

• LEA (Lower Envelope Algorithm)
• Plot ca_itb_i
• LE(t) mina_itb_i for all i
• Given this ordering, only need to determine
  thresholds
• When to transition from state S_i to state
  S_(i1).
• Lower Envelope Algorithm Transitions from one
  state to the next at the discontinuities of the
  lower envelope curve.

9
Breakeven-Time for Multiple Power Modes
Active mode
Energy Consumption
Spinup cost
Idle Period Length
From Qingbo Zhus HPCA04 slides
10
Deterministic Algorithm - LEA

• The Lower Envelope Defines an ordering of the
  states.
• Throw out states that do not appear on lower
  envelope
• Theorem LEA is 2-competitive.
• This ratio can be improved by considering input
  distribution
• Which can be learned on-line.

11
Probability-based LE Algorithm-PLEA

• PLEA offline probability p(t)
• 2-state case (1 threshold)
• Multiple-state case (K thresholds)

12
OPBA Online Probability-based Algorithm

• Learn the Probability Distribution Idle history
  window size w
• Histogram based learning n bins
• ith bin r_i, r_(i1)) lower end is the
  threshold
• Each bin maintains a counter c_i
• Estimate the distribution p(t) by the
  distribution that generates idle period of length
  r_i with probability c_i/w for each i in 0, ,
  n-1

13
Generate New PM Strategy (r_t)

• Solution
• Algorithm Complexity O(kn)
• Algorithm Efficiency
• How many states? k
• How many bins? nc
• How frequently to adjust? w

14
Evaluation

• IBM Mobile disk
• Auspex file server trace
• Series of thresholds
• LEA and OPBA
• Single value prediction
• OPT
• LAST
• EXP Decay
• Adaptive learning Tree

15
Energy and Latency
16
Relative Comparison
17
Conclusion

• OPBA is the best in terms of energy efficiency
• LAST TREE and EXP suffer at least an additional
  40 latency
• Those have lower latency than OPBA (LEA and
  pre-emptive version of LAST, TREE, EXP) use 25
  more power on average

18
Limitation of this paper

• Single Device level
• Collaborative multiple devices?
• Not addressing
• Choosing the states among possible operating
  states
• Preemptive to intermediate power states

19
Discussion

• Question?