Slides PPT - National e

Scheduling Architecture
and Algorithms within ICENI
Laurie Young, Stephen McGough,
Steven Newhouse, John Darlington
London e-Science Centre
Department of Computing, Imperial College London
Contents
• ICENI
– Scheduling Architecture
• Scheduling Algorithms
– Variety of different algorithms
• Experimental Results
– Different policies
– Different Grid sizes
– Different Application Profiles
ICENI
The Iceni, under Queen Boudicca,
united the tribes of South-East
England in a revolt against the
occupying Roman forces in AD60.
•
•
•
•
•
IC e-Science Networked Infrastructure
Developed by LeSC Grid Middleware Group
Collect and provide relevant Grid meta-data
Use to define and develop higher-level services
Interaction with other frameworks: OGSA, Jxta etc.
Component Applications
•Each job is composed of multiple components.
•Each runs on a different resource
•Each component is connected to at least one other component.
•Data is passed along these connections
Design
Generator
Mesh
Generator
DRACS
Factory
Mesh
Generator
DRACS
Analyser
Mesh
Generator
DRACS
ICENI Scheduling Architecture
ICENI
Launching
Framework
Scheduling
Framework
Condor Launcher
Schedule Evaluation
Globus Launcher
Simulated Annealing
Performance
Repository
Performance Model
Game Theory
Statistical Prediction
ICENI Scheduling Services
Launching Framework
Pluggable Launchers
(SGE, Globus,
Condor, ICENI)
Scheduling Framework
Pluggable Schedulers
(Simulated Annealing, Game Theory
Random, Best of n Random)
Performance Framework
Pluggable Performance
Repositories
(Perf. Models,
Statistical Analysis)
Schedule Evaluation
• Use a Benefit Function.
• Also called a Utility Function or Evaluation Function.
• A Benefit Function maps the metrics we are interested in to a single
Benefit Value.
• Different benefit functions represent different optimisation
preferences.
• Can set benefit to 0 if constraints (e.g. Budget) exceeded.
B  B(b,e,  )
Random / Best of n Random
Random
Best of n Random
Random Scheduler
•Randomly selects a schedule
•Checks schedule can be executed
•Produces schedules very quickly
Best of n Random
•Produces multiple random schedules
•Returns the best one
•Still very fast
•Better results than the random schedules
Simulated Annealing
Random
Simulated Annealing
•
•
•
•
Monte Carlo method
Generate schedule at random
Modify current schedule
Accept new schedule if better
– If worse, accept with probability proportional to
“temperature” and inversely proportional to
benefit change
• Repeat, while reducing “temperature”
• Stop when no modifications to schedule accepted
Game Theory
• Each component is a “Player”
• Each player has to choose best strategy (Grid resource)
• Each strategy has a benefit, depending on the strategy chosen by all
other players.
• Players identify, then remove strategies guaranteed to never be
optimal – “strictly dominated strategies”
Player B
• Produces the “Nash Equilibrium”
Player A
1
2
3
4
1
6,3
6,4
7,3
8,4
2
4,7
3,9
4,7
5,6
3
4,5
5,5
7,4
6,5
4
7,4
5,5
6,4
7,6
Experiments
Grid
Description
Schedulers
Scheduling
Simulated
Application
Scheduling
Description
Policy
Framework
•4
Clusters
resources
•Random
•21
DAG Applications
/ of
Best
of n Random
•Saturn
Produces
usable
Depth
schedules fast.
•Time•Varying
•Consistent
Optimisation
Interface
16 Sparc
III 750
MHz 2Processors
Depth
between
and
7the
Best DAG
Uses
benefit
the same
frominterface
a schedule
as the
with
ICENI
5Gbit
Interconnects
•Game
•Varying
Theory
Complexity
shortest
scheduling
execution
framework
time.allowing
Resultsthe
show
same
•Rhea
Considers
Between
the
8used.
Components
problem
scheduling
schedule
code
time2scheduling
toand
+be
execution
time. as an
8 Sparc
III 900Mhz Processors
economic
problem.
5Gbit Interconnects
component would take 2
•CostAverage
•Repeatability
Optimisation
•Viking
Ton an 2Ghz CPU
•Simulated
minutes
Annealing
Bestthe
As
benefit
underlying
from
adescription
schedule where
files never
the cost
16
node,
2GHz
Pentium
4
Algorithm
forsame
solving
of using the
change
resources
experiment
is optimisation
low.
can Problems
be run
1Gbit
Interconnects
manyAverage
times. communication between
•Viking
C would take 1 minute on a
components
16
node,network
2GHz Pentium 4
100Mbit
100Mbit Interconnects
Results (Cost Optimisation)
Results (Cost Optimisation)
Results (Time Optimisation)
Summary
• ICENI Scheduling Architecture
– Comprised of 3 services, using a pluggable architecture to allow
different implementations to be used
– Launcher implementations allow launching to different underlying
execution environments.
– Performance service enables execution time predictions
– Scheduling service operates on information provided by other
two services
Decouples scheduler from application and environment
Summary
• Scheduling Algorithms
– Four algorithms examined while varying:
• Grid Sizes
• Applications
• Policies
– Simulated Annealing generally the best algorithm tested
– Larger applications take longer to schedule and return
– More choice in resources leads to:
• cheaper computation for users
• Longer return times for applications
Increasing the Grid size can reduce or improve the quality of
service experienced by the user
Acknowledgements
• Director: Professor John Darlington
• Technical Director: Dr Steven Newhouse
• Research Staff:
–
–
–
–
–
Anthony Mayer, Nathalie Furmento
Stephen McGough, James Stanton
Yong Xie, William Lee
Marko Krznaric, Murtaza Gulamali
Asif Saleem, Laurie Young, Gary Kong
• Contact::
– http://www.lesc.imperial.ac.uk/
– e-mail: [email protected]
• Funding:
– PPARC e-Science Studentship (PPA/S/E/2001/03335)