Power Management of MEMSBased Storage Devices for Mobile Systems

1
Power Management of MEMS-Based Storage Devices
for Mobile Systems

• M.G. Khatib Univ. of Twente
• P.H. Hartel Univ. of Twente

2
Outline

• Motivation
• Why a new storage technology
• MEMS-based storage
• Brief introduction
• Problem
• Idle energy is significant
• How optimal is a fixed-timeout policy
• By simulations
• Near optimality of the fixed-timeout policy
• Energy-efficient shutdown policy
• Conclusions

3
Motivation
Streaming on iPAQ PDA (192 Kbps)
energy hungry cheap 0.5 /GB
energy efficient expensive 6 /GB
4
Background
3D view
2D view
5
Background

• MEMS devices have potentially low cost
• Well-established MEMS fabrication techniques
• Reuse micron-scale fabs installed 10 years ago
• Few requirements on lithography to produce future
  MEMS generations
• State of the art
• IBM 26 nm x 26 nm bit pitch (realized 840
  Gb/in2)
• Nanochip 15 nm x 15 nm
• project to 2 nm x 2 nm

6
Problem

• MEMS devices waste a considerable amount of
  energy when idling
• Thus, need a power-management policy that
• maximizes energy saving
• minimizes influence on the performance

File system ext3 Block size 4 KB
7
Architectural characteristics

• Spring-suspended medium
• Z nano-positioners
• ? No startup overhead
• Independent motions along X Y
• ? tseek max (tX-seek,tY-seek)
• Many probes sharing one medium
• ? Short seek times

8
Physical characteristics

• Light-weight medium (0.1g vs. 1.0g of 0.85 disk)
• High storage density (1Tb/in2 vs. 30Gb/in2)
• Small dimensions (41mm2 vs. 366mm2)
• Devices of a small form factor
• MEMS unique characteristics lead to
• no startup overhead
• small seek and shutdown overheads
• Investigate simple power management first

9
Power management

• Power breakdown
• 1 probes 0.25 mW
• X-Actuation 60 mW
• Y-Actuation 60 mW
• 4096 probes 1 W
• Startup 0 mW

Power State Machine (PSM)
10
Power management

• If a workload exhibits locality of reference,
• shutdown incurs longer seek distances from the
  center
• increasing seek energy
• long inactive periods to compensate for
• increasing seek time
• long-enough timeout for confidence
• Tradeoff
• energy consumption
• vs.
• response time
• Parameter
• timeout

state transition
11
Experiments setup

• Trace-driven simulation using DiskSim
• Refine CMUs MEMS model
• Model parameters from IBM prototype 2007
• Real-world traces on an iPAQ PDA
• Various usage scenarios
• mp3, video, random
• Different I/O settings
• File system
• Block size

model key settings
12
Experiments methodology

• 4 traces
• ext3-4K, ext3-1K, ext2-4K, ext2-1K
• Each has 7 different usage scenarios (or
  benchmarks)
• Run simulation per
• Every trace as a whole
• Every benchmark of each trace
• Timeout 0-10, 20, 30, 40, 50 ms

13
Idle-period distribution
Whole ext3-4KB trace
14
Results
Count
Idle-period length ms

• 48 of idle periods are in 0,10 ms
• 10 decrease in response time for TTO10 relative
  to TTO0 ms
• Energy for TTO10 is within 6 of Emin (Emin
  ERW Eseek Einactive)
• Response time for TTO10 is within 12 of tmin
  (tmin tRW tseek)
• But the minimum (at TTO40) is within 8, thus
  only 4 is left

15
Shutdown policy

• PM policy
• When to shut down
• Shutdown policy
• How to shut down
• We deployed the PE policy previously
• We introduce the EE policy now

Performance-efficient (PE) policy
Energy-efficient (EE) policy
16
Shutdown policy
Copying scenario in the ext3-4KB trace
Whole ext3-4KB trace
Total energy consumption J
Total energy consumption J
Average response time ms
Average response time ms
17
Conclusions

• Idle energy is significant in MEMS devices
• Therefore, Power Management (PM) is necessary
• Fixed-timeout PM achieves optimal energy saving
• 95 energy saving for 4 performance degradation
• Thanks to MEMS unique characteristics
• Fixed-timeout PM achieves (near-)optimal saving
• Shutdown overhead is minimized using springs
• For workloads of locality, small timeout values
• vary noticeably in response time
• save the same amount of energy

18
Q A
Q A

• Thanks for your attention!