CSS434: Parallel & Distributed Computing - PowerPoint

							          CSS434 Grid Computing
       Textbook No Corresponding Chapters
             Professor: Munehiro Fukuda



A portion of these slides were compiled from The Grid: Blueprint for a New
Computer Infrastructure.
                              CSS434 Grid Computing                          1
      Network Infrastructure
                                          Users login their organizational
                                           systems first locally or
                                           remotely.
                                          If they are affiliated with other
                                           organizations,
                                                  They can login from the
     High-speed                                    system of their main use to
Information high way                               some other systems. (They
                                                   are given an opportunity to
                                                   use those resources in
                                                   parallel).
                                          Problems:
                                                  They must orchestrate job
                                                   execution among the
                                                   resources they use.
                                                  Should those resources be
                                                   limited to such a handful
                                                   number of researchers?
                       CSS434 Grid Computing                                   2
Purposes of Computational Grid
   Use computing resource connected to high-speed information
    highway as if we use electric power grid
      Only 30% utilization in academic/commercial environments.

      Many applications have only episodic requirements. So, why
       don’t we share computation resource?
      Computational results and data should be also made
       available to all users.
   Users:
      Computational scientists and engineers

      Experimental scientists

      Association and corporations

      Training and education

      Consumers (E-commerce)




                       CSS434 Grid Computing                  3
                      Grid Applications
Category          Examples                Characteristics


Distributed       DIS and Stellar         Very large problems needing lots of
supercomputing    dynamics                computing resource at a time
High throughput   Chip design and         Harnessing many idle resources to
                  parameter studies       increase aggregate throughput
On demand         Medical                 Allocating special resource dynamically
                  instrumentation
Data intensive    Sky survey              Using distributed data and needing
                                          high-volume data flows
Collaborative     Collaborative design    Support communication or
                  Education               collaborative work


                               CSS434 Grid Computing                            4
  Grid Services Architecture
  from www.globus.org slide
                High-energy         Collaborative           On-line
                physics data         engineering       instrumentation
Applications      analysis     Regional       Parameter
                           climate studies      studies
                Distributed                    Collab.             Remote
 Application    computing                      design               control
Toolkit Layer                     Data-                  Remote
                                intensive                  viz



Grid Services    Information       Resource mgmt         ...
    Layer
                  Security         Data access       Fault detection


                Transport                ...         Multicast
Grid Fabric
  Layer         Instrumentation        Control interfaces      QoS mechanisms


                            CSS434 Grid Computing                             5
            Programming Model
                      Uniform Access

   Paradigm
       Bag of task or master workers (Condor-MW)
       Client server (NetSolve)
       Object oriented (Legion)
       Synchronous applications (Not suited for massively
        parallel computation.)
   Language Support
       MPI-G – message passing (Globus)
       Open MP – shared memory
       Math Library – remote procedure (NetSolve)

                     CSS434 Grid Computing             6
                   Resource Management
                   Discovery, Allocation, and Scheduling
   Centralized resource manager
         Systems        Resource         Front-end         Resource      Job launcher
                       descriptions       process          manager
         Globus        RSL: resource                    Broker and MDS      GRAM
                      spec. language
         Condor        ClassAd and      Schedd Agent     Matchmaker       Sandbox
                         DAGMan                           and startd      (Starter)
         Legion        IDL: interface    Scheduler        Collection       Enactor
                       def. language

       +: easy to manage
       –: a bottleneck
   Decentralized resource manager
       A collection of centralized manager (Condor’s gate flocking)
       A combination of meta and local schedulers.
                                CSS434 Grid Computing                                 7
                  Fault Tolerance
   Check-pointing
       At the master (Condor)
       At each node but collected at the master (Catalina)
       Use a whiteboard (Optimal Grid)
   Re-execution of fault worker jobs from the
    beginning (Bayanihan, Optimal Grid)
   Error code (NetSolve)
       User is responsible to handle errors.


                        CSS434 Grid Computing            8
                        Security
   Resources covered with security layers
       Legion (Message/MayI layers)
       Entropia (Intercepting all system calls)
   A use of commodity tools
       SSL
       Public key
       Security Certificate
       Java sandbox
       Kerberos

                      CSS434 Grid Computing        9
                              NetSolve
                      http://icl.cs.utk.edu/netsolve/
                                             Network of servers             Client
   RPC-based approach
   Clients
       Include a set of APIs called                                    Agent
        as (asynchronous) RPCs
   Agents
       Match client’s requests for                  Agent
                                                               choice            Scalar
        services with servers           Client
                                                     request                     server
   Servers                                            reply
       Encapsulates remotely                                     MPP servers
        accessed numerical libraries

                             CSS434 Grid Computing                               10
                                       Legion
                             http://legion.virginia.edu/

   Legion classes                                     Prog
        Act as managers and make policy                         request
                                                                           Enactor
   Core objects                                     Scheduler
        Provide mechanisms that classes            Converted Legion object ID
         use to implement policies: hosts           By context objects           reserve
         (processors), vaults(memory),      search Converted Logion object address
         context, binding agents, etc.             By binding agents
                                          Resource
   Per-Program Scheduling                database                            Class
                                                                        Host
      Participating sites can assure their      collection
                                                             tty
       local policies.                                             tty     Host Host
                                                                       Resources
      User can choose a scheduling                         Class
       policy.                                                       tty

                                   CSS434 Grid Computing                         11
            Condor
http://www.cs.wisc.edu/condor/




A: User’s local agent
R: Each computer resource        I/O forwarded to
M: Central manager               a user’s home


         CSS434 Grid Computing                      12
AgentTeamwork at UWB
          Architecture




      CSS434 Grid Computing   13
     Paper Review by Students
   Globus
   Legion
   Condor
   Netsolve
   Discussions
       What programming or execution model is each system
        based on?
       What resource allocation and scheduling algorithm does
        each system use?
       Are they fault-tolerant?
       Did they any special security features for their own?


                       CSS434 Grid Computing                     14