Characterizing Resource Availability for Volunteer Computing and its Impact on Task Distribution Met

1
Characterizing Resource Availability for
Volunteer Computing and its Impact on Task
Distribution Methods
David Toth Professor David Finkel Computer
Science Department Worcester Polytechnic
Institute Worcester, Massachusetts, U.S.A.
2
Volunteer Computing

• Definition Computing using idle time on
  computers volunteered by people.
• Supercomputer for researchers who cant afford
  one
• Projects
• SETI_at_home searches for extraterrestrial
  intelligence
• GIMPS searches for Mersenne primes
• Folding_at_home analyzes effects of protein
  folding
• Grid.org cancer, anthrax research

3
Task Distribution

• Task distribution may affect productivity
• Tasks have deadlines
• Tasks that are late are aborted
• Task distribution methods
• Buffer No Tasks ? Wasted time waiting to get next
  task
• Buffer Multiple Tasks ? Some tasks dont complete
  get reassigned. Part of the time spent on
  uncompleted tasks could have been used to
  complete other tasks.
• Understanding resource availability enables
  analysis of task distribution methods

4
Trace Collection Methodology

• Windows Service
• Runs constantly for 28 days
• Samples screensaver status every 10 sec
• No noticeable impact on Pentium 3, 450 MHz w/ 256
  MB RAM Windows XP (0 CPU)
• Writes , _at_, or time stamp

Start Computer Gets Powered On
TimeStamp
TimeStamp

_at_
S.S. On
S.S. Off
TimeStamp
5
4 Major Categories of Computers

• We expected computer usage might vary between
  different types of users
• Public (68 traces)
• Undergraduate student (38 traces)
• Business (26 traces)
• Home (25 traces)

6
Implications

• Amount of time available for volunteer computing
• lots of time favors buffering more tasks
• little time favors less buffering
• Duration of periods where computers are
  available, unavailable but powered on, and off
• long periods of unavailable but on or off favor
  less buffering
• Usage patterns
• similar usage patterns suggest a single task
  distribution policy may be sufficient
• significantly different usage patterns suggest a
  need for multiple task distribution policies or
  an adaptive policy

7
Results
Indicates buffering likely more effective for
Public computers
8
Results
Short periods of unavailable but on and off (
70 lt 2 hours) indicates buffering likely
effective CDFs look similar, but lots of
variance within classes (some powered on for
entire day, others for a very little time)
suggests need for multiple policies
9
Analysis

• Examined available periods with Rockwell
  Automations Arena program.
• Public computers Weibull for 2 labs (not good
  match), Beta for 3rd (match quality inconclusive)
• Home, Student, Business computers split among
  many distributions (generally not good matches)

10
Conclusions Future Work

• Data collected will be available soon
• Data does not conform to well known
  distribution(s)
• Single method of task retrieval may not be
  adequate given the variance between different
  classes and computers within a class
• Using data in simulations now

11
Questions?

• Thank you!
• Characterizing Resource Availability for
  Volunteer Computing and its Impact on Task
  Distribution Methods
• David Toth Professor David Finkel
• Computer Science Department
• Worcester Polytechnic Institute
• Worcester, Massachusetts, U.S.A.
• toth_at_cs.wpi.edu