UC Research Computing Group by RyanTannehill

VIEWS: 0 PAGES: 20

									UC Research Computing Group

     The UC Grid Project
           ITGC Briefing
        September 19, 2006




                             Bill Labate, UCLA
                             UC Research Computing Group
The Challenge
   Siloed resources
       Available to a single campus or research group only
       Difficult to partner
       No method for discovering resources
   Little or no economies of scale
       Hardware or software
         • Vendor partnerships
         • UC license servers
       Systems & application personnel inefficiently used
   Research CI “business continuity” challenges
       Multiple single points of failure
       Each cluster is a custom build
       Backup is a “nicety”
UC Grid Implementation Benefits
   Provides an extensible architecture for providing
    services
   Not the “be all, end all” rather it provides a solid base to
    build on
   Trusted - Certificated authentication
   Robust - Lightweight infrastructure, easy to support
   Discoverable - resources, applications
   Locate resources where it makes sense
   Combine where it makes sense
   Locally managed, virtually available
UC-Wide CI Vision
                            Replicated
                                                            10GB Ethernet Network                      Replicated
                              Data
                                                                                                         Data


          Campus 1
                                                                                                                      Campus 3

                                              Campus Grid                           Campus Grid




                                                                                                       TeraGrid
   Backup                  SDSC                                                                                            Backup
   Services                                                                                                                Services
                                                              UC Grid Portal


“Commercial
Applications”                            Campus Grid
  Cluster                                                                                Campus Grid


    Campus 2                                                                                                             Visualization
                                                                                                                        “Pixel” Cluster


                                                                                                                         Campus 4
                     Replicated
                       Data                                                                              Replicated
                                                                                                           Data
Campus Grid Architecture
                                                                               Head node
                                             Firewall
                                                                               Cluster I

                                                           Credential
                                                           mapped to
                                                           local ID


                         Grid Portal

       https
       Uniform browser
       based interface                                                         Head node
                                                                               Cluster II
                                                            Credential
                                                            mapped to
                                                            local ID



  MyProxy Server    Storage Server     ION Visualization Server
                                                                   Grid Appliance
 UC Grid Architecture
 UC     UC           UC          UC                           UC Portal
 Grid   Register     User        MyProxy
 CA     Service      Database
Single CA is shared among all campuses

                                     A   C
                     UCLA
                     Portal          A   C                     C     A        UCSB
                                                                              Portal
        Storage
        Server
                                             C            C    C     A
                     UCLA
                    MyProxy
                                             A            A          UCSB         Storage
                                                                    MyProxy       Server

                                                  UCI
                                                 Portal
 All appliances talk to both their                             Storage
 campus portal and the UC portal                                Server
                                              UCI
                                             MyProxy
Discoverable Resources
Grid Services Example -
Visualization
Resource Sharing - Grid Pools
   Resource Pool:
        Clusters contribute cycles and/or applications to the resource pool
        TeraGrid cycles purchased by the campus can easily be shared
   Pool Users:
        Students and faculty members who do not have accounts on any
         campus clusters
        Those with low-level or sporadic usage requirements
        Cluster users are also pool users
   Pool use hides all the details
   Pool jobs
        Target cluster selected by the UC Grid Portal
        Applications only in this phase
User Types vs. Portals
                              Portal
                  Campus Portal      UC Portal
U   Pool-only   Can submit jobs to the   Can submit jobs to the
s     User      campus pool              UC pool
e
r
                                         Best choice for those
                Can use those            with access to clusters
T               clusters on campus       on different campuses
y    Cluster    he/she can access        -- can use all clusters,
      User                               UC-wide, he/she can
p                                        Access
                Can submit jobs to
e               the campus pool
                                         Can submit jobs to the
                                         UC pool
To Login to a Portal a User Needs:
• A certificate
• A GridSphere account on that Portal
• Additionally:
   • A Cluster User must be added to the gridmap
   file on the appliance node of each of the
   clusters on which he/she has a login id
   • A pool-only user needs to be assigned a
   storage area on the Grid Portal’s Storage
   Server.
How Pools Work
   Currently applications only
       Can be commercial or custom code
         • Parallel jobs are not efficiently supported
         • Investigating MPICH-G2 for cross-cluster scheduling
   A cluster that contributes to the pool must
    create:
       Guest login id
       Certificate for Grid Portal
       Map a guest to a local account
       Mechanism that allocates resources for pool jobs is
        up to the cluster administrator
         • Local control of resources
Pools -- Now
   Single Pool Concept
      Every campus cluster user is also a pool user
      Pool-only users can submit only to the pool
      Any cluster can contribute resources to the pool
         • Computing power
         • Applications
   When a pool job is submitted, the Portal decides which
    cluster will run it.
       Currently running a simple prediction model for fastest
        completion.
       Collecting data for application of the UCSB prediction
        algorithm to pool jobs.
Futures -- Specialty Pools
   Campuses can form Specialty Pools
       A specialty pool is a granular grouping of
        resources
       Can be arranged according to research group,
        applications, departmental - whatever affinity is
        required
       A user must be approved by the pool admin to join
       Grid Portal database tracks pools and pool users
       Eventually pools will be developed based on
        hardware requirements:
         • Large memory nodes
         • CPU architecture
         • Interconnect
  Specialty Pool Architecture
   UCLA
              UCLA            UC
                                       UCI
           Matlab        Matlab
           Amber         Amber
           Q-Chem        Q-Chem
           Mathematica                 UCSB
                         Mathematica
           Mathematica
           Amber                        *
           Fluent
                          Gromacs
                          Jaguar
                          Vasp   *
     *     Mathematica           *
           Corsica
           Osiris
TeraGrid
           baby b2

           Gromacs       * = CNSI
           Jaguar
                 *
TeraGrid Integration
•   UGP acts as a client to the TeraGrid
•   Globus Toolkit nodes are open to all of their users
•   CA Signing Policy and CA Certificate Public Key for
    each TeraGrid cluster made available to UCLA
                                                             myproxy.ncsa.uiuc.edu
                                  Request for user proxy
                       UCLA                                        User Proxy
                        Grid                                       Certificate
                       Portal
                                User proxy certificate returned because
                                UCLA Grid Portal can authenticate

     user proxy certificate
                                             DataStar at
                                               SDSC


UCLA pool jobs can currently run on two TeraGrid clusters.
Implementation Cost & Suggested
Funding Sources
   UC Grid Equipment - Five systems minimum
        One Apache web server
        Two Tomcat, MySQL, UGP servers
        One MyProxy server
        One Simple CA server
        $25,000-30,000 depending on configuration and vendor
        UC or Campus pooled funding
   Campus Grid Equipment - Three systems minimum
        One Apache web server
        One Tomcat, MySQL, UGP server
        One Storage server
        $12,000-25,000 depending on configuration and vendor
        Campus funded
   Grid Appliance - One per cluster
        $900-1,100 per appliance
        Campus or possibly research group with cluster
    Next Steps
   Bring up UC Grid test bed
       UCG, MyProxy, Simple CA
       Three test campuses with local MyProxy
       Finish writing Grid Registration Service
   Install and test replication and failover configurations
   Perform beta test of UC Grid with three test campuses
   Test with production UCLA Grid Portal
   Add three additional campus Grids
   Continue until fully deployed
   Test and tune during every stage
UC Research Computing Group



         Q&A

								
To top