Upload
roxanne-ball
View
215
Download
1
Embed Size (px)
Citation preview
Preetam GhoshSchool of Computing
Mobile Grid Computing: From Theory to Practice
Preetam Ghosh
School of Computing
The University of Southern Mississippi
E-mail: [email protected] http://www.cs.usm.edu/~pghosh
Preetam GhoshSchool of Computing
Outline Mobile Grid
Mobile Grid Challenges and Applications
Mobile Grid Projects
Our Focus–Investigation of Pricing Model–Cost-effective Job Allocation Schemes
Conclusion
Preetam GhoshSchool of Computing
Computational GridComputational Grid
Resource Resource BrokerBroker
& Trade Server& Trade Server
Resources and Topology Resources and Topology (System)(System)
Tasks and Topology (Workload)Tasks and Topology (Workload)
DCADCA DRBDRB
DCADCA LSLS
GIS : Grid Information ServerDRB: Domain Resource BrokerDCA: Domain Control AgentLS : Local Scheduler
The GridThe Grid
Price Negotiation & Price Negotiation & Mapping StrategyMapping StrategyGISGIS UserUser
• User submits job to User submits job to Resource Resource BrokerBroker•Trade Server negotiates with Trade Server negotiates with the DRB’s (different VO’s) on the DRB’s (different VO’s) on behalf of the user.behalf of the user.
• Price negotiationPrice negotiation results in optimal job results in optimal job distributions among Domains or VO’s. distributions among Domains or VO’s. • Resource Broker optimally maps the jobs Resource Broker optimally maps the jobs within each domain using a Distributed within each domain using a Distributed Mapping StrategyMapping Strategy
Preetam GhoshSchool of Computing
Mobile Grid: From System Level Perspective
Utilize huge resource pool of laptops, PDAs and other mobile devices– reduced CPU performance, small secondary storage, low battery
power, and unreliable low-bandwidth communication
Motivate mobile devices to contribute their resources– negotiation mechanism (pricing strategy)– optimal Job allocation scheme
Computational Grid
Data Grid
Grid Community
Wireless Access Point
A broad view of Grid Community, Wireless Access Point and Wireless devices
Thomas Phan, Lloyd Huang, Chris Dulan “Challenge: Integrating Mobile Wireless Devices Into the Computational Grid ”MOBICOM’02, September 23-26,2002, Atlanta, Georgia, USA.
Preetam GhoshSchool of Computing
Mobile Grid Projects AKOGRIMO (Access to Knowledge through the GRId in a
MObile world)– Seen the integration problem from a business/commercialization
perspective
– Provide support for user mobility through the SIP
– Terminal mobility through Mobile IPv6 techniques
– Does not address incorporation of resource constrained devices
Cyber foraging (Surrogate Computing)– Offload demanding computational tasks to more capable nodes
– Resource discovery and allocation
– Runtime engine/platform to handle the offloading
– Lack of provisioning service in small scale device
A. Messer, I. Greeberg, P. Bernadat, and D. Milojicic, “Towards a Distributed Platform for Resource-Constrained Devices,” ICDCS'2002
B. M. Satyanarayanan, “Pervasive Computing: Vision and Challenges”, IEEE Personal Communications, 2001
Preetam GhoshSchool of Computing
Mobile Grid Projects K*Grid
– A Grid Research project supported and funded by Korean Ministry of Information and Communications
– Make use of idle resources in the mobile community for high performance computing
– Task scheduling
LEECH (Leveraging Every Existing Computer out there)– Proxy-based clustered approach for integrating mobile devices
– Built on top of a customized MPI
P2P Middleware– Convergence of Grid and Peer to Peer techniques (P2P)
1. The K*Grid Mobile Grid project page: http://gridcenter.or.kr/MobileGrid/index.php2. N. Ruiz, “A Framework for Integrating Heterogeneous, Small Scale Devices into Computational Grids and Clusters," M.S. Thesis, University of California, US, 2003.3. Foster and A. Iamnitchi, “On Death, Taxes, and the Convergence of P2P and Grid Computing,” in Proceedings of the 2nd International Workshop on Peer-to-Peer Systems (IPTPS '03), 2003.
Preetam GhoshSchool of Computing
Mobile Grid ChallengesLimited Resources
– Make it difficult to install large software components (e.g. Globus Toolkit, due to s/w dependencies and significant amount of memory and storage capacity)
– Needs a lightweight infrastructure (IBM’s web services Toolkit for Mobile devices)
Increased Dynamicity– Seamless terminal mobility, resource mobility, user mobility
-> session management
– User moves his session from one WAP to another
– Need to take care of various mobile computing related issues (low and fluctuating bandwidth availability, mobility management etc.)
Preetam GhoshSchool of Computing
Mobile Grid Challenges
Increased Heterogeneity– should care for the bridging of a plethora of middleware possibilities
Integration Challenges– big differences in reliability, availability and performance
– Operational failures due to low battery level or because of mobility and roaming
– Unreliable weak link of mobile devices in the application execution chain
– Protection of Grid scheduling and brokering systems from the reduced availability and unpredictability of the mobile resources
Preetam GhoshSchool of Computing
Importance of Mobility Management
User mobility (current and future) is very important in a mobile grid computing paradigm
Efficient resource usage of mobile devices along predicted routes
Guaranteeing accurate (timely) completion of allocated jobs (QoS)
Preetam GhoshSchool of Computing
Efficient Symbolic Representation of node mobility– Movement patterns of a mobile
node is a (piece-wise) stationary, stochastic (ergodic) process
Location update Strategy– Does not use location updates on
every movement of the mobile node
– Updates only on an appropriately determined entropy-minimized subset of this movement sequence
– Estimates number of mobile nodes available (for job allocation) under a particular WAP within the threshold time T
User Mobility Challenges
Grid Controller (GC) , WirelessAccess Points (WAP), Basic ServiceSet (BSS), Extended Service Set (ESS)
Preetam GhoshSchool of Computing
Our Contributions in the System-Level Perspective
Goal is to harness the idle CPU cycles of mobile devices– Requires a pricing strategy that can attract the mobile device
owners to contribute their devices for grid jobs.
Preetam GhoshSchool of Computing
Our Contributions in the System-Level Perspective
Goal is to harness the idle CPU cycles of mobile devices– Requires a pricing strategy that can attract the mobile device
owners to contribute their devices for grid jobs.
The pricing strategy is then used to devise cost-effective grid job allocation schemes to the mobile nodes
Preetam GhoshSchool of Computing
Our Contributions in the System-Level Perspective
Goal is to harness the idle CPU cycles of mobile devices– Requires a pricing strategy that can attract the mobile device
owners to contribute their devices for grid jobs.
The pricing strategy is then used to devise cost-effective grid job allocation schemes to the mobile nodes– consider the processing cost (or delay) at the mobile nodes
Need to consider the internal jobs at the mobile device (e.g. call processing activities)
Preetam GhoshSchool of Computing
Our Contributions in the System-Level Perspective
Goal is to harness the idle CPU cycles of mobile devices– Requires a pricing strategy that can attract the mobile device
owners to contribute their devices for grid jobs.
The pricing strategy is then used to devise cost-effective grid job allocation schemes to the mobile nodes– consider the processing cost (or delay) at the mobile nodes
Need to consider the internal jobs at the mobile device (e.g. call processing activities)
– consider the communication cost (or delay) for transferring the jobs (and hence the dynamic wireless channel bandwidth)
Preetam GhoshSchool of Computing
Our Contributions in the System-Level Perspective
Goal is to harness the idle CPU cycles of mobile devices– Requires a pricing strategy that can attract the mobile device owners to
contribute their devices for grid jobs.
The pricing strategy is then used to devise cost-effective grid job allocation schemes to the mobile nodes
– consider the processing cost (or delay) at the mobile nodes Need to consider the internal jobs at the mobile device (e.g. call processing
activities)
– consider the communication cost (or delay) for transferring the jobs (and hence the dynamic wireless channel bandwidth)
– consider the node mobility as the results of the jobs assigned by a particular WAP to the nodes need to come back to the WAP after completion Requires a mobility management algorithm to track the number of mobile devices
present under a particular WAP within a specific time period (in which the assigned jobs need to complete)
Preetam GhoshSchool of Computing
Our Contributions in the System-Level Perspective
Goal is to harness the idle CPU cycles of mobile devices– Requires a pricing strategy that can attract the mobile device owners to
contribute their devices for grid jobs.
The pricing strategy is then used to devise cost-effective grid job allocation schemes to the mobile nodes
– consider the processing cost (or delay) at the mobile nodes Need to consider the internal jobs at the mobile device (e.g. call processing activities)
– consider the communication cost (or delay) for transferring the jobs (and hence the dynamic wireless channel bandwidth)
– consider the node mobility as the results of the jobs assigned by a particular WAP to the nodes need to come back to the WAP after completion Requires a mobility management algorithm to track the number of mobile devices present
under a particular WAP within a specific time period (in which the assigned jobs need to complete)
– consider dynamic session management techniques i.e., if a particular mobile node goes out of the WAP’s coverage area, how can the completed jobs be transferred back to the original WAP.
Preetam GhoshSchool of Computing
Grid: Existing Pricing Strategies
Market Model Adopted by
Auction Model Spawn and Popcorn
Bargaining Model Mariposa and Nimrod-G
Posted Price Model Nimrod-G
Commodity Market Model Mungi, MOSIX and Nimrod-G
Bid based proportional Resource Rexec and Anemone
Community,Coalition and Bartering SETI@Home,Condor,MojoNation
Tender/Contract-Net Model Mariposa
Table II: Different Distributed Computing Scheduling Systems with the adopted Game Theoretic Approach
• Lack of formulation• Fails to capture competitiveness among the mobile users. • Cooperative game theory solution not suitable
Preetam GhoshSchool of Computing
Motivation: Game theoretic ApproachFigure 1: Dynamics of different Mobile
User groups with different Wireless Access Point
Mobile Users withWAP1
MobileUsers withWAPp
Groups of Mobile Users
WAP1
WAPp
Grid Community
Job assignment
Job Assignment
Resource assignment
Resourceassignment
Pricing strategy implemented using a Game theoretic Model :- Two player non-cooperative bargaining game Efficient, Stable, Simple, Symmetric No central Matchmaker Optimal Static Job Allocation Scheme based on this pricing strategy.
Preetam GhoshSchool of Computing
Needs for Bargaining Model
Alternating-offer bargaining under incomplete information
Updates probabilistic preference list
Bargainers are rationalchoose a strategy leading to Nash equilibrium
Conflict of interestsWAP Servers and Mobile users choose a mutually beneficial agreement
agreement can’t be imposed on either WAP Server or Mobile users without their approval
Preetam GhoshSchool of Computing
Mobile User WAP Server
Proposal
AcceptAgreement
Reject
Counter-offerContinues until agreement/breaks off
Bargaining Protocol
Preetam GhoshSchool of Computing
Cost Optimal Job Allocation Schemes
None of these models consider the dynamic session management requirement !
Schemes Mobile nodes dedicated for grid jobs ?
Considers communication delay (wireless bandwidth) ?
Considers node mobility ?
Grid job class
PRIMAL Yes No No Single
PRIMANGLE No Yes No Single
PRIMULTI No Yes No Multi
PRIBAND No Yes No Multi
PRIPROC No Yes No Multi
PRIMOB No Yes Yes Single
Preetam GhoshSchool of Computing
Summary of the Job Allocation Schemes
PRIMAL
PRIMULTI PRIBAND PRIPROC
PRIMANGLE PRIMOB
M/G/1 Preemptive Priority Model(considers communication delay)
Single-class jobs Multi-class jobs (do not consider node mobility)
Does not consider node mobility
Considers node mobility
Generalized systems
Bandwidth constrained systems
Processing -power constrained systems
Job Allocation Schemes
M/M/1 Model(does not consider communication delay)
Preetam GhoshSchool of Computing
Publications:
1. Preetam Ghosh, Kalyan Basu and Sajal Das, A Game Theory based Pricing Strategy to support Single/Multi-Class Job Allocation Schemes for Bandwidth Constrained Distributed Systems. IEEE Transactions on Parallel and Distributed Systems, 2007, Volume 18, Issue 3, pp. 289-306.
2. Preetam Ghosh, Nirmalya Roy, Sajal Das and Kalyan Basu, A Pricing Strategy for Job Allocation in Mobile Grids using a Non-cooperative bargaining Theory Framework, in Special Issue on Design and Performance of Networks for Super-Cluster and Grid-Computing, JPDC, 2005, Volume 65, Issue 11, pp. 1366-1383.
3. Preetam Ghosh, and Sajal Das, Mobility-aware Cost-efficient Job Scheduling for Single-class Grid jobs in a generic Mobile Grid Architecture. Under 2nd round review at Elsevier Future Generation Computer Systems, 2009.
4. Preetam Ghosh, Nirmalya Roy and Sajal Das, Mobility-based Cost-efficient Job Scheduling in Mobile grids. 1st IEEE International Workshop on Context-Awareness and Mobility in Grid Computing (held in conjunction with CCGrid 2007), 2007, Brazil, pp. 701-706.
5. Preetam Ghosh, Kalyan Basu and Sajal Das, Cost-Optimal Job Allocation Schemes for Bandwidth-Constrained Distributed Computing Systems. 12th Annual IEEE International Conference on High Performance Computing (HiPC), 2005, Goa, India, pp. 40-50.
6. Preetam Ghosh, Nirmalya Roy, Sajal Das and Kalyan Basu, A Game Theory based Pricing Strategy for Job allocation in Mobile Grids. 18th IEEE International Parallel & Distributed Processing Symposium (IPDPS), 2004, USA, pp. 82-91.
Preetam GhoshSchool of Computing