Design of high performance networking platform considering mobile environment

1
Design of high performance networking platform
considering mobile environment

• Tai M. Chung
• Internet Management Technology Laboratory
• School of Electrical and Computer Engineering
• SungKyunKwan University300 Chunchun-dong,
  Changan-gu, Suwon, Kyounggi-do,
• 440-746, KoreaTel 82-31-290-7131, Fax
  82-31-299-6673

2
Contents

• Introduction Concept
• Scheduling algorithm
• Implementations
• Conclusions Future Works

3
Introduction Concepts
4
Introduction

• Convergence of mobile and Grid system
• Extend the Grids resource model to wireless
  mobile network
• Allow access to Grid system using the ubiquitous
  mobile devices
• Information gathering using wireless sensor
  networks
• Why wireless?
• Explosive increase of Mobile Device
• New pervasive technology
• CPU Speed, Storage Capacity
• Coverage Ubiquity
• Bandwidth very high-speed
• Mobility management Mobile IP, Ad-hoc routing

5
Mobile Grid
Wireless Grid
Mobile Node
Access Point
Mobile Node
Access Point
Mobile Node

• Powerful machines with large amounts of memory
  and very fast processors
• High-bandwidth links, disconnections are due to
  either explicitly performed or failures
• Static location, hosts can be added, deleted or
  moved

• Limited capabilities slow CPU speed, little
  memory, low battery power and small screen size
• Unpredictable disconnections is considered as a
  part of normal wireless communication
• Hosts may come and leave generally much more
  rapidly
• Bandwidth and quality of the network connection
  may vary greatly

6
Problem Statement

• Frequent changes in the environments
• Variability of network bandwidth, mobility
• Temporary and unannounced loss of network
  connectivity
• Resource Heterogeneity
• Ranging from powerful laptops to human
  WAP-enabled cell phones
• Scarce resources
• slow CPU speed, little memory, low battery power
• Variety of user interface
• Variety of screen size, input scheme

7
Research Objective

• To suggest the mobile grid architecture
  considering wireless network
• To identify and formulate the response time
  function in mobile grid
• To determine the scheduling algorithm of task
  allocation using the response time function
• To establish the mobile grid testbed with the
  PDAs in the wireless LAN network environment
• To show the efficiency of mobile grid
  architecture using the mathematical analysis and
  simulation

8
System Architecture for Mobile Grid

• Challenge Integrating Mobile Wireless Devices
  into the Computational Grid
• Thomas Phan, Lloyd Huang, and Chris Dulan
• Communication Paradigm for Mobile Grid Users
• Dario Bruneo, Marco Scarpa, Angelo Zaia, and
  Antonio Puliafito
• Extending the Condor Distributed Systems for
  Mobile Clients
• Song-Yi Yi and Miron Livny
• SETI_at_Anywhere
• Agile Computing

9
Scheduling Algorithm
10
MG-JSA Model

• MG-JSA Model Mobile Grid Job Scheduling
  Algorithm
• Prediction of task processing time
• Predicting the response time of task processing
  using the mathematical model
• Task allocation and scheduling algorithm
• Partitioning the parallel tasks and allocate the
  subtasks to appropriate mobile nodes to achieve
  efficient performance
• Task processing monitoring until task completed
• Rescheduling the task allocation process
  considering the response time when unexpected
  events are occurred

11
Process Transition of MG-JSA
retrieve
Resource monitoring of mobile nodes
retrieve
resource information gathering
Prediction of task processing time
retrieve
Scheduling algorithm of task allocation
Rescheduling algorithm
distribute the Job in wireless network
Reallocation the subtask
subtask fail
Partition and distribution the job
Task processing monitoring
Task completion
all subtask are completed
12
Task allocation and scheduling algorithm

• Job J arrives in proxy server
• proxy server checks the available mobile nodes
• collect the available resources information of
  mobile nodes
• for (k 1 k lt m k )
• compute the T until Tk gt Tk1
• partition and allocate subtasks to each
  workstation using the prediction of task
  processing time
• monitoring the job processing until job completed
• if subtaskk fails then reallocate the failed
  subtaskk to other available mobile nodes
  considering the early finish time
• if unexpected event is occurred at mobile node k
  then evaluate the processing time between the
  original scheduling and rescheduling
• if original scheduling (time) gt rescheduling
  (time) then rescheduling
• The results of Job J returned

13
Prediction of task processing time

• Prediction of transfer time in wireless network

State transition diagram of mobile node

• Probability of connection state Pc
• Probability of disconnection state Pd
• Disconnection rate a
• Reconnection ß
• Then, the prediction time of data transfer for t
  time unit data is like this

14
Prediction of task processing time

• Prediction of processing time in non-dedicated
  system
• Modeling parameters are like these
• W total demand processing time
• wk demand of the parallel subtask processing
  time on mobile node k
• m number of available mobile nodes
• S the number of interruptions encountered
• ?k rate of the job arrival poisson distribution
  at mobile node k
• µk sequential job service rate at mobile node k
• ?k utilization rate at mobile node k
• Tk parallel task completion time on mobile node
  k
• W can be written like this

15
Prediction of task processing time

• Tk can be written like this

Xi computing time consumed by the parallel
task Yi computing time consumed by the
sequential jobs Z execution time of the last
parallel process that finishes the parallel task

• Assumption for modeling
• Owner job arrival process follows a Poisson
  distribution
• Xi is an exponentially distributed random
  variable
• Owner job processing follows M/G/1

16
Prediction of task processing time

• Then we can acquired the results like these
• Mean of Tk
• Generalization factor
• The mean of subtask completion time at mobile
  node k

17
Performance Evaluation

• Mathematical Analysis
• Network environments are suggested in the paper
  of S. Radhakrishnan et al
• Stable (a 0.003 0.001, ß 0.0030.001)
• Highly disconnective (a 0.027 0.009, ß
  0.0030.001)
• Unstable (a 0.027 0.009, ß 0.027 0.009)
• Highly connective (a 0.003 0.001, ß 0.027
  0.009)
• Parameters used in the mathematical analysis
• Total required time to process the job(W) 2000
  sec
• The number of available mobile node (m) 1 10
  nodes
• time required for transferring input data (tin)
  1 sec
• time required for transferring output data 1
  sec
• ?and CPU speed of mobile nodes
• Node1(0.2, 400Mhz), Node2(0.4, 600Mhz),
  Node3(0.3, 400Mhz)
• Node4(0.4, 600Mhz), Node5(0.3, 400Mhz),
  Node6(0.2, 600Mhz)
• Node7(0.6, 400Mhz), Node8(0.4, 600Mhz),
  Node9(0.5, 400Mhz)
• Node10(0.7, 600Mhz)

18
Performance Evaluation
Highly disconnective
Stable
Highly connective
Unstable
19
Implementations
20
MGA MGP

• Mobile Grid Architectures Mobile Grid Protocol
• Requirements
• Dynamic Resource Management
• Active and efficient discovery management of
  resource and service of mobile node ,considering
  wireless network
• Quality of Service Guarantee
• Reliable request/distribution of job to mobile
  node, accomplishment of jobs, collection of
  results
• Security
• User authentication and encrypted messages
  because wireless network is weak than wired
  network
• Interoperability
• Interoperation between wired-GRID

21
Design of MGA

• Proxy Part

M O B I L E N O D E
Resource Information DB
Job Information DB
Resource Manager
User DB
Job Management
Job Service Manager
Job Scheduler
Parallel Process Management
User Authenticator
Job Process
Survivability Manager
Communicator
Job Information
Survivability Check
Resource Information
22
Design of MGA

• Mobile node Part

User Interface
Super Daemon
Job Execution Module
MGP module
Operation Results Display
P R O X Y
U S E R
System State Management Module
Resource Information Display
Node Controller
Resource Manager
System Kernel
Processor
Power
Memory
23
Design of MGP

• Messages
• Login initialization message
• Periodic resource information message
• Job request/reply message
• Survivability check message
• Special features
• Format
• Header Payload
• Nested TLV
• Periodic resource request
• For job scheduling algorithm
• For decrease operation delay

Total_Length
Rbit
Flag
ID
Version
Header
Payload
Type
Length
Value
Type
Length
Value
Type
Length
Value
24
Operations
Req_Login insert user ID/Password and
initialization
Rep_login_Accept Authentications
Join to Grid network
Req_outArea Check abnormal communication or
disconnection
Rep_outArea mobile node replies when it can
communicate with proxy
Rep_outArea Estimates that mobile node
is capable
Req_Resource Broadcast the message for getting
resource information
Rep_Resource reply information of the mobile to
proxy through Resource Management Module
Rep_Resource receive the resource information
of mobile node and save in DB
Req_Job user can request the job
Job Scheduling
Mobile node reply the results of distributed jobs
to proxy after parallel processing
Parallel processing through MPICH
Rep_Job Reply the job results
Notify_Out_Area If mobile node want to be out
of grid, this message is sent
Initialization
25
Results of Implementation
Mobile node
Proxy
26
Conclusions Future Works
27
Conclusions

• Design and implementation of the mobile grid
  system
• MGA(Mobile Grid Architecture)
• MGP(Mobile Grid Protocol)
• Mobile grid testbed using PDAs in wireless LAN
  networks
• MG-JSA (Mobile Grid Job Scheduling Algorithm)
• Giving the useful information of job partition
• Selection of the appropriate mobile node and its
  number
• Rescheduling the job allocation in time of job
  failure
• Performance Evaluation
• Mathematical analysis
• Simulation in the testbed

28
Future Works

• Extend the mobile grid system considering the
  various mobile nodes and other network
  environments
• Inclusion the security mechanisms for mobile
  communication
• Suggesting the useful applications which can be
  used in mobile grid network
• Suggesting the business model of mobile grid
  network