Business Plan Overview Grid Dynamics

							   Bridging The Paradigms:
Convergence of Enterprise Grids
  with In-Memory Data Grids




           Victoria Livschitz
   Founder and CEO, Grid Dynamics
             About Grid Dynamics
   Grid Dynamics Consulting Services
            Experts in Grid Computing
            Independent, global, pureplay grid engineering firm
            Founded in 2006, headquarters in San Francisco, CA
            Current focus: data-aware enterprise grids
            Clients are grid users (eBay, PayPal, Bank of America) and grid
             vendors (GigaSpaces)

   Speaker Career Highlights
            Founder and CEO, Grid Dynamics
            Principal Architect, Sun Grid division
            Senior Scientist, Sun Labs
            Chief Architect, Sun Financial Services, USA
            HPC Engineer for crash simulations, Ford Labs



Grid Dynamics Consulting Services   2007 IGT Conference Exhibition             1
             Grid in the Enterprise: The Evolution
   Generation One: Compute Grids
            Applications: compute-intensive, processor-bound jobs
            Grid definition: large pool of resources tied by fast interconnects
            Key concepts: resource, job, scheduler, scatter-gather
            Goals: optimize job completion time and job flow through the grid

   Generation Two: Enterprise Grids
            Applications: job-oriented, transactional, batch, interactive
            Grid definition: several large pool of resources tied by fiber pipes
            Key concepts: virtualization, multi-tenancy, SLA, QofS
            Goals: optimize resource utilization, “bet fit: if application needs
             available resources, dynamic right-sizing”


   The New Challenge: data access => system’s bottleneck


Grid Dynamics Consulting Services     2007 IGT Conference Exhibition                2
           A Closer Look at the Challenge
   Reality of Tiered Computing
          Data is stored and managed “far away” from where it is used
          Application performance is bound by data access time:
              Time to find relevant data in large remote databases
              Time to move record sets through the pipes
          Distributed shared-state applications don’t scale

   The Solution: In-Memory Data Grid (IMDG)
          Distributed partitioned, replicated, coherent, persistent, transactional
           application memory
          Data is stored and managed centrally
          Data is accessed locally for processing


  => Preserve physical tiers for ease of management
  => Collapse physical tiers for performance
Grid Dynamics Consulting Services   2007 IGT Conference Exhibition                    3
           In-Memory Data Grids - How They Works
   General Approach (all leading products)
          Data is organized into “partitions” by some principle
          Partitions are distributed in memory of grid compute nodes:
              Read-thru, write-thru, transactional, persistent, lazy load
          Partitions are also replicated
              Coherent writes, distributed locking, back-up for failover

   Data-Aware Applications
          Computing tasks are routed to nodes with the right data
          Without “data affinity” – scatter/gather across all nodes
          With “data affinity” – involve only one right node

   Specific Vendors Create Significant Value-Add
          GigaSpaces: Service Grid
          DataSynapse: FabricServer

Grid Dynamics Consulting Services   2007 IGT Conference Exhibition           4
           Data-Aware Enterprise Grids
   Goal
          Combine the best of two paradigms to achieve next level of
           performance, reliability, scale



   Conceptual Approach
          Use In-Memory Data Grid technology as part of service for grid-
           managed applications
          Use dynamic right-sizing service of Enterprise Grid to optimize
           performance of In-Memory Data Grid



   Practical Implementation
          Integrate commercial products for Enterprise Grid and In-Memory Data
           Grid



Grid Dynamics Consulting Services   2007 IGT Conference Exhibition                5
             Design Patterns
   Five integration scenarios, progressively bringing more benefits
          Higher sophistication, higher benefits
          Some rely on features available from specific products
          For illustration, we’ll use GigaSpaces and DataSynapse

   Using Data Grid to Improve Enterprise Grid
            Scenario 1: Unmanaged Data Grid
            Scenario 2: Managed Data Grid
            Scenario 3: Managed Data Grid with Data-Aware Routing
            Scenario 4: Self-Managed Data Grid (GigaSpaces)



   Using Enterprise Grid to Improve Data Grid
          Scenario 5: On-Demand Data Grid (GigaSpaces, DataSynapse)



Grid Dynamics Consulting Services   2007 IGT Conference Exhibition     6
           “Hello World” Trading Application
             Book – data structure that holds a collection of trades
             Book Segment – a subset of the Book that is stored in a single data partition
              (e.g., range of stock symbols)
             Book Calculation – a computational process over the Book
             Book Calculation Task – a step in Book Computation that requires access to
              one of the Book Segments
             Affinity – ability to map each Book Calculation Task to unique Book
              Segment
             Affinity Key – unique identifier of each Book Segment
             Execution Node – a host machine where Book Calculations occur
             IMDG Node – a host machine where Book Segment is stored
             Scheduler – a component of enterprise grid that decides which Execution
              Node will perform which Book Calculation
             EG Agent – a component of EG that executes Book Calculations


Grid Dynamics Consulting Services    2007 IGT Conference Exhibition                           7
           Scenario 1: Unmanaged Data Grid
   You already have an enterprise
    grid. Now add a data grid.
          Loosely-coupled integration via
           data access API
          No data awareness by grid
           routing algorithms

   Benefits
          Simplest to implement
          Application performance
           improvement due to faster
           data search and retrieval

   Limitations
          Processing still happens
           against remote data
          Data has to be fetched on each
           computation
Grid Dynamics Consulting Services   2007 IGT Conference Exhibition   8
           Scenario 2: Adding Shared Hardware Pool
   Two loosely coupled grids in a
    single operational environment
          Data Grid instances are started
           in the enterprise grid as jobs
          Then like scenario #1


   Benefits
          Single operational grid, not two


   Limitations
          Nothing new




Grid Dynamics Consulting Services   2007 IGT Conference Exhibition   9
           Scenario 3: Adding Data-Aware Routing
  Coordinate between the application
   and two grid managers
         Application knows the affinity key and
          passes that key with the job to Job
          Submitter
         Job Submitter asks Scheduler for book
          segment matching the key
         Scheduler asks the sensors for the
          location of the right segment

   Benefits
          Data-aware routing means data is
           always local to the application

   Limitations
          Requires extensions to traditional products
          Straight-forward to do with DataSynapse
           and GigaSpaces

Grid Dynamics Consulting Services   2007 IGT Conference Exhibition   10
           Scenario 4: Adding Self-Management
  Allow the Data Grid to handle routing
         Scheduler sends the job to ANY node in
          a service grid
         Data Grid maintains implicit data
          awareness and forwards the requests
          where they need to go


   Benefits
          Simplest and most powerful
           scenario all around


   Limitations
          Not supported by all Data Grid
           products




Grid Dynamics Consulting Services   2007 IGT Conference Exhibition   11
           Scenario 5: On-Demand Data Grid
  How big is the Data Grid? Let the
   application workload decide
         Enterprise Grid is notified when the
          Data Grid is overloaded
         If more hardware is available, it will
          schedule another Data Grid container


   Benefits
          On-Demand right-sizing of the Data
           grid


   Limitations
          Not supported by all Enterprise
           Grid products




Grid Dynamics Consulting Services   2007 IGT Conference Exhibition   12
             Summary: Convergence is a Win/Win
  The Sum is Greater Than Its Parts
         Convergence enriches both paradigms


   Convergence is ready for prime time – and happening
            Customers recognize the need
            Case studies & success stories
            Vendors see the opportunity
            Partnerships are forming
            Design patterns and best practices are beginning to emerge




Grid Dynamics Consulting Services   2007 IGT Conference Exhibition        13
Thank You!




    vlivschitz@griddynamics.com