Decentralized Data Management Framework for Data Grids

1
Decentralized Data Management Framework for Data
Grids

• Houda Lamehamedi
• Computer Science Department
• Rensselaer Polytechnic Institute

2
Data in Large Scale Computing

• In an increasing number of scientific
  disciplines, large data collections are emerging
  as important community resources
• data produced and collected at experiment sites
  e.g. high energy physics, climate modeling
• processed data and analysis results
• The geographical location of the compute and
  storage resources results in complex and
  stringent performance demands Data Grid

3
Data Grid Requirements

• The Data Grid is a Grid where data is treated as
  a first class citizen
• Scientific collaborations on the Grid generate
  queries involving access to large data sets
• Efficient execution of these queries requires
• careful management of large data caches,
• gigabits data transfer over wide area networks,
• creation, management, and strategic placement of
  replicas

4
Replication in Data Grids

• Globus toolkit is a standard set of services
  supporting Grids and Grid applications
• Data management services offered
• GridFTP offers secure efficient data transfer in
  Grid environments
• Replica Catalog allows users to register files
• Replica Location Service allows users to
  register where data is replicated and locate
  replicas
• The system only provides the users with tools to
  statically replicate data files

5

• Example of a data replication scenario in the
  Compact Muon Solenoid (CMS) experiment. All data
  is collected at CERN, the European Center for
  Nuclear Research located in Geneva Switzerland

6
Example of a Replica Model used by the Sloan
Digital Sky astronomy survey and at the Laser
Interferometer Gravitational Observatory
7
Data Management Issues in Data Grids

• Existing Data Grid frameworks demand extensive
  administrative oversight and management overhead
• Missing support for dynamic and intermittent
  participation on the Data Grid hinders scalable
  growth of collaborative research
• Limited support to replication Data is
  statically replicated under user guidelines

8
Problem Statement

• Ensuring efficient access to huge and widely
  distributed data is a serious challenge to
  network and Grid designers
• An automated system is needed to maximize the use
  of storage, networking and computing resources

9
Proposed Approach

• To address these issues we propose a
  decentralized performance-driven adaptive replica
  management middleware that
• Uses an overlay network to organize Data Grid
  nodes
• Dynamically adapts replica placement to changing
  users and networks needs and behavior
• Dynamically evaluates data access costs vs.
  performance gains before creating a new replica

10
Mechanisms

• Adaptive and scalable data management tools that
  enable users to dynamically join and leave the
  grid
• Replica management services that intelligently
  and transparently place data at strategic
  locations
• Binding data organization to popular and
  commonly used access patterns in data sharing
  environments and data intensive applications

11
Major Components

• A theoretical model of data transfer cost and
  access performance
• Parameterized by the changing computing
  environment
• Data monitoring tools that feed current values of
  resource consumption to the cost function
• Dynamic replica management services
• Offer transparent replication using the cost
  function
• Manage replica placement and discovery

12
Middleware Design

• Distributed Data Grid architecture supported by
  the delegation of management and decision making
  to all member Grid nodes
• Each participating node is responsible for
  managing access to its local and contributed
  resources
• Layered Architecture
• Replica Management Layer
• Resource Access Layer
• Communication Layer

13
Middleware Architecture
Replica Management Layer supports the management
and transferring of data between Grid nodes and
the creation of new replicas. Uses input from
lower layers to track users' access patterns,
monitor data popularity
Resource Access Layer Provides access to
available resources and monitors their usage and
availability. Includes a Replica Catalog to
support transparent access to data at each Grid
node for local and remote users
Communication Layer consists of data transfer
and authen. protocols used to ensure security,
verify users identities, and maintain data
integrity provides support for the overlay
network structure
14
Framework

• Services Offered by the middleware
• Resource Monitoring service
• Replica Creation service
• Replica Location service
• Resource Allocation service
• Routing and Connectivity service

15
Resource Monitoring/Allocation

• The monitoring service is responsible for
• monitoring resource availability at each Grid
  node
• collecting statistics about resource usage and
  data access requests
• The allocation service is responsible for
• allocating space for newly created replicas,
• de-allocating space from the least frequently
  and last accessed locally stored replicas

16
Replica Creation Service

• The service is responsible for creating local
  replicas based on the evaluation of the incurred
  cost of creating a local replica
• A cost function is used to evaluate the cost of
  creating a local replica vs. the cost of
  transferring data based on the
• popularity of the data,
• network resources availability,
• size of data, and
• storage space availability

17
Replica Location Service

• This service is responsible for managing the
  local replica Catalog at each node
• Each newly created file is registered in the
  catalog
• Supports data location and discovery

18
Data Catalog Management
19
Replica Distribution Topology

• Our approach is based on using application level
  overlay networks to enable scalable growth of
  Data Grids and support larger numbers of
  participants
• The overlay network is formed by the set of
  connections between the participating nodes in
  the Data Grid
• Topology means and represents the connectivity
  graph formed by the overlay network

20
Node Addition / Data Model Construction
After a Node joins, it starts developing a list
of preferred neighbors
Requests Flow
Data Flow
21
Data Model Construction

• We use a combination of spanning tree and ring
  topologies
• Grid Node Insertion
• When joining the grid, a node is added through an
  existing grid node by attaching to it as a child
  node or a sibling
• Node Removal
• When a node leaves the tree, it sends a
  notification message to its parent, siblings, and
  children

22

23
Data Search and Replica Location

• The search starts at the local data catalog to
  check if the data is stored and available locally
• If it is not, then the node sends a request
  message to its parent, siblings, and children ?
  flooding mechanism
• The local replica management service chooses the
  best data source

24

25
Cost Model

• Total data transfer cost for object i at node v
  is costv,i (?v,i??v,i)size(i)d(v,r)

The incremental data transfer for placing a
replica at v is costi(N,Ri,v)
-?tv,id(v,c(v,Ri)) ?i(?tr,i-?tv,i)size(i)d(v,c(v
,Ri)))
Tv the partition of nodes which root is v ?tv,i
the total read rate of all nodes at partition
Tv ?tv,i the total write rate of the partition
N the set of all nodes in the system, Ri the
replica set of object i, and c(v,Ri) the
replica of object i closest to node v
Adding r to Ri increases or decreases the read
cost of each node in Tv Our cost function
evaluates the above formula at each node vs. the
cost of data transfer and the storage space
available before replicating any data
26
Cost Function

• The transfer cost of file i at node v from node r
  is
• costi (Ri aWi)size(i)/d(v,r)
• Ri is read rate, Wi write rate and d(v,r) is cost
  of transferring a data unit from v to r
• A replica creation threshold is defined based on
  cost of transferring data from known replica
  sites
• The cost function evaluates the accumulated read
  and write costs at each node vs. the threshold
  before creating a replica

27
Simulation Design

• GridNet is a modular simulator built on top of NS
  (the network simulator) and written in C
• NS provides us with the basic grid specification
  nodes, links, messages
• GridNet introduces application level services
  implemented on top of the existing NS services
  and protocols. It allows us to specify
• different types of nodes (client, cache, and
  server nodes)
• node resources (storage capacity, local data
  files )
• the cost model and its parameters
• the replication strategy used

28
Simulator Architecture

• Each Node in GridNet has three major components
• replica optimizer computes number of read/write
  requests, response time, probes for connection
  bandwidth
• replica routing table parent, sibling, and child
  nodes
• storage element storage space available, and
  files stored

29
Simulation Experiments

• We used 6 file sets with file sizes ranging from
  100Mb to 1Gb
• All files were originally located at the server
• Access patterns
• Recently accessed files are more likely to be
  accessed again
• Files recently accessed by a client are more
  likely to be accessed by neighbor clients

30
Simulation Results
31
Middleware Deployment

• We used two hierarchical distribution models that
  represent the most popular and commonly used
  models
• Bottom Up Multiple collection sites
• Top Down single collection site
• Single collection site data repositories have
  larger storage capacity

32
Access Patterns

• Data access requests are based on patterns
  commonly observed in scientific and data-sharing
  environments
• Files are of similar sizes within an application
• Spikes are generated by new Interesting Files
• Users social organization and interests guide
  the overlay construction
• Interest-based adaptive clustering of users

33
Top Down Model
34
Bottom UP Model
35
Top Down Experiments Results
36
Bottom Up Experiments Results
37
Bottom Up Experiments Results
38
Access Performance Evaluation
39
Conclusions

• Cost guided dynamic replication improves data
  access performance by up to 30 and a minimum of
  10 compared to static user initiated replication
• The combination of parameter selection for cost
  evaluation and resource availability play a key
  role in influencing the performance of the
  system.
• Lower storage availability might lead to race
  conditions where popular data compete for storage
  space.
• The results also show that popular data files
  benefit the most from dynamic replication

40
Contributions

• A solution that combines and adapts approaches
  previously used in different data sharing
  environments the Grid and P2P systems
• Support for the creation of small to medium scale
  data sharing Grids
• A formulation of the replica creation and
  placement problem using a mathematical model
• The use of dynamic and adaptive overlay networks
  and data organization models to connect
  participating nodes in a Data Grid (access
  patterns)
• Generic Data Grid simulator, GridNet
• A new data and replica management middleware that
  scales with the number of users and supports
  dynamic and intermittent user participation

41
Future Directions

• Investigate additional user access patterns and
  overlay structures and study data access
  performance under these settings
• Studying different and new combinations of
  parameters and cost evaluation techniques
• Deploying the middleware in a larger environment
  with larger applications
• Identifying the relation and effect of data and
  replica management services on other services