Grid Forum Korea 2002

1
Grid Forum Korea 2002
Adaptive Resource Allocation Policy for
Computational Grid
July 11, 2002 Chan-Hyun Youn Information and
Communications University
2
Contents

• Introduction to GMC
• Architecture of adaptive scheduling policy
• Analysis of proposed scheduling policy
• Experimental results
• Conclusions

3
Grid Middleware Center
MIC ITRC Program KIPA
Overseas Collaboration
National Institute
Universities
Industry

• ICU
• Korea Univ.
• Kyunghee Univ.
• Hanyang Univ.
• Daejon Univ.
• Kumoh National
• Institute of
• Technology

• ImpressTek
• VERYTECH

• MIT
• Univ. of Tokyo
• Univ. of Lecce
• NASA Ames
• Lab.
• GGF

• ETRI

Information and Communications University
(ICU) Grid Middleware Center
4
Collaboration of participants

• Resource Manager
• Resource allocation monitoring
• Support user QoS

• Artificial Heart Application
• Hemodynamics model
• Parallel program for analysis

• Grid Security Infrastructure
• Grid Security Protocol
• Globus based Grid security
• service

Cluster System
Linux Cluster System

• LDAP based Resource Searching
• LDAP service platform
• LDAP data consistency

• MPICH-G2 Optimization
• Object oriented middleware
• MPICH-G2 performance improvement

5
Overview

• To present an effective way to simulate the
  hemodynamics of the KTAH (Korean Total Artificial
  Heart)
• To propose an adaptive resource scheduling policy
  to optimize applications performance
• Discuss QoS constrained resource scheduling
  scheme required in collaborative work for
  artificial heart modeling
• Propose Delay Adaptive Resource Allocation Policy
  (DARAP) in order to improve the execution time of
  collaborative Grid applications

6
Background and Motivation

• Traditional grid resource scheduler
• Optimize the execution time of computation
  intensive Grid applications
• Communication capability among resources is not
  considered
• Problem of traditional grid resource scheduler
• In some application, both computational and
  communication performance affect the execution
  time of the application
• Resource scheduling scheme without identifying
  network status may not guarantee the user
  requirements especially QoS

7
Schematic of the blood sac in the KTAH(Korean
Total Artificial Heart)
8
Hemodynamics of the KTAH
9
Computational aspect
Computer simulation
10
Subdivision for parallel computing
11
Flowchart for parallel finiteelement method
12
Grid testbed for artificial heart application
STAR-TAP / Abilene
MIT
penelope
Globus / MPICH-G2 Platform
ICU
lilly
KNUT
KOREN/KREONET2
ICU
fluid001
rose
13
Architecture of proposed adaptive resource
scheduling policy (1)
Application characteristics
Policy Server
Task Analyzer (Parameters for task
characteristics)
Application A (Master) in Resource 1
Information Manager (Generate the list
of candidate resources)
Instability Analyzer (Compute network status)
Policy Generator (Determine the candidate
resources)
BGP Information Server
Resource Manager
MDS Server
Backbone Router
Application B (Sub-job) in Resource N
Network
? MDS Metacomputing Directory Server
14
Architecture of proposed adaptive resource
scheduling policy (2)

• Policy server responsible for managing and
  selecting the candidate resources to execute user
  tasks
• Information manager generates the list of
  candidate resources that may be allocated for
  user tasks by using MDS
• Instability analyzer computes the network
  status among grid resources in each domain based
  on BGP routing information
• Task analyzer takes the parameters that
  describe the characteristics of user tasks
  whether they are computation intensive or
  communication intensive tasks
• Policy generator determines the candidate
  resources for user tasks by using the proposed
  policy rule (Delay Adaptive Resource Allocation
  Policy DARAP)

15
Policy rule for selecting grid resources

• PPref i???1?Wi????2?1/Di
• Wi the performance of grid resource Ri in MIPS
• Di the average delay of all links adjacent to
  Ri in ms
• ? and ? weighting factors for computation
  intensity and communication intensity,
  respectively (0? ? ?1 and 0? ? ?1)
• ?1 and ?2 normalizing factors for adjusting the
  scale of resource performance and communication
  performance between any two pairs of resources,
  respectively (experimentally determined)
• Decision policy grid resources are selected
  according to the value of PPref i

16
Proposed DARAP algorithm

• For a user task that need K resources and divided
  into K sub-tasks,

Start
Get the list of available resources
Get BGP information of each available resource
Determine performance and delay of each resource
For each available resource Ri
Compute Di and PPref i
Sort list of Di
Allocate K resources
End
17
Evaluation of proposed DARAP algorithmwith
example (1)

• R1 , R3 , R4 were modeled according to the
  specification of our Linux cluster system
• Since it is common that submitting user tasks to
  remote high performance resources in grid, we
  assume that one high performance resource R2 is
  located in the distance and has relatively high
  link delay

18
Evaluation of proposed DARAP algorithmwith
example (2)

• Assume that each user task consists of 0.1 MI and
  each instruction has same computation time
  (?0.3, ?0.7, ?11/50, ?21)
• Conventional deadline scheduling policy (no
  identification of network delay)
• Resources are selected based on the performance
  of resource (W) according to largest W first
  scheduling in order to minimize the execution
  time
• In this scheme, R1 , R2 and R4 are selected
• Execution Time T 0.021ms
• DARAP algorithm (identification of network delay)
• resources are selected based on policy rule
  according to largest PPref first scheduling
• In this scheme, R1 , R3 and R4 are selected
• Execution Time T 0.017ms

19
Simulation result of Grid basedartificial heart
Interface between two domains
20
Velocity contour of artificial heart
at time 0.1 sec
at time 0.7 sec
21
2D simualtion of the blood sac in the KTAH
Pressure contour
Velocity magnitude contour
22
Concluding remarks

• Optimization approaches to develop QoS guaranteed
  real-time visualization and collaborative works
• Distributed parallel processing technologies for
  high-end applications
• Dynamic resource allocation model for
  computational Grid applications
• Future work
• Development of 3D distributed simulation code for
  the hemodynamics of the KTAH
• Real time visualization using OpenGL
• Fairness based Scheduling Mechanism