Execution Context Optimization for Disk Energy

1
Execution Context Optimization for Disk Energy

• Jerry Hom Rutgers University
• Ulrich Kremer Rutgers University

2
Motivation

• Energy, Power, Thermal concerns
• energy gt tangible cost (, battery life)
• power gt design constraints
• thermal gt density, cooling
• Trend high density, multi-core

3
Motivation

• Disk among top energy consumers
• up to 70 in servers Zhu05
• up to 10 in desktops Karabuto05
• 5-15 in laptops Lorch98
• 5-25 in handhelds specs

4
Disk Hibernation

• no power management
• threshold-based
• oracle

5
Hibernation Techniques

• Enabling
• Caching
• Remote Accesses
• Clustering (OS or application)
• Exploiting
• Fixed or Adaptive Threshold
• Compiler Directed

6
Clustering ExampleUniprogramming
54 disk energy saved
PACT02, IEEETC04
7
Hibernation Techniques

• Enabling
• Caching
• Remote Accesses
• Clustering (OS or application)
• Exploiting
• Fixed or Adaptive Threshold
• Compiler Directed

What happens in multiprogramming?
8
Clustering ExampleMultiprogramming
5 disk energy saved
Synchronization
9
Clustering ExampleMultiprogramming
25 disk energy saved
Synchronized Proportional Buffers
10
Idea

• Definition Execution Context - a set of
  concurrently running programs
• Idea Execution Context optimizations can
  significantly reduce disk energy consumption.
• New Optimization Framework
• Synchronize disk requests
• Optimize buffer sizes
• Adapt to changing execution contexts

11
Outline

• Introduction
• Optimization Framework
• User Study
• Experiments
• Closing Remarks

12
Optimization FrameworkSynchronization
PACS03
13
Optimization FrameworkFile Descriptor Attribute
Extensions
14
Optimization FrameworkFile Descriptor Attribute
Extensions
15
Optimization FrameworkFile Descriptor Attribute
Extensions
16
Optimization FrameworkFile Descriptor Attribute
Extensions
SYNC_SEND, SYNC_RECV, BUFFERED, STREAMED FILE
int
17
Optimization FrameworkExecution Context model

• Model execution contexts as states in finite
  state machine.
• Transitions represent a program starting or
  exiting.

18
Outline

• Introduction
• Optimization Framework
• User Study
• Experiments
• Closing Remarks

19
User Study

• Survey what execution contexts users run
• Conjecture
• Many contexts rarely occur
• Most users run small number of programs
  concurrently

20
User Studyby the numbers

• Linux Trace Toolkit next generation (LTTng)
• Forty Pentium 4 desktops (2.8-3.2 GHz, 512-1024
  MB) running Fedora Core 3 (TA offices, labs)
• Four weeks (April 1-28, 2007) of daily traces
  during peak activity (1000-2000)
• Seventy-three CS grad students, 860 login hours,
  12 idle, 760 active hours, 50 programs

21
User Study50 programs

• abiword, acroread, bc, cvs, eclipse, emacs, eog,
  epiphany, evolution, festival, firefox, gaim,
  gdb, gedit, gv, gnotski, gnumeric, gthumb,
  gucharmap, hxplay, java, kaboodle, kate, kmail,
  konqueror, konquest, korganizer, ksame, kword,
  lskat, man, matlab, mozilla, mplayer, mysql,
  nedit, noatun, ooffice, pine, play, python,
  realplay, rhythmbox, screen, sftp, skype, svn,
  terminal, theseas, thue, thunderbird, vnc,
  xboard, xmms, vi, xdvi, xpdf

22
User StudySize of Execution Contexts

• 1-2 programs cover 69
• 1-3 programs cover 85
• 1-4 programs cover 94

23
User StudyPopular Apps and Contexts
14 states gt2 time cover 68
24
Outline

• Introduction
• Optimization Framework
• User Study
• Experiments
• Closing Remarks

25
ExperimentsTest Programs
26
ExperimentsSynthetic Traces
27
ExperimentsMeasurement Infrastructure

• Oscilloscope with current probe
• Xnee records and replays keyboard and mouse
  events
• Control program logs oscilloscope data while
  running experiment scripts

28
ExperimentsDisk Specs
29
ExperimentsSample Trace (OPG)
30
Experiments
SS saves 3 to 63 21 avg
Hitachi E7K60
31
Experiments
SS saves -33 to 61 8 avg
Hitachi E7K60
32
Experiments
UNI 41 to 52 MULTI 51 to 65
UNI 54 to 65 MULTI 65 to 77
33
Outline

• Introduction
• Optimization Framework
• User Study
• Experiments
• Closing Remarks

34
Future Work

• Extend execution context optimizations to other
  resources for energy or performance (memory,
  wireless, alternative storage)
• Identify contexts where optimizations have
  negative impact and selectively disable
• Expand user study to other populations
• Implement adaptive context transitions

35
Summary

• Disk optimization techniques should consider
  multiprogramming models
• Language extensions lead to simple inter-process
  communication framework
• Execution context optimizations offer significant
  energy reduction at low performance cost