A Distributed Architecture for Multidimensional Indexing and Data Retrieval in Grid Environments

1
A Distributed Architecture for Multi-dimensional
Indexing and Data Retrieval in Grid Environments

• Athanasia Asiki, Katerina Doka,Ioannis
  Konstantinou, Antonis Zissimos and Nectarios
  Koziris
• National Technical University of Athens
• School of Electrical and Computer Engineering
• Computing Systems Laboratory
• e-mail nasia, katerina, ikons, azisi,
  nkoziris_at_cslab.ece.ntua.gr

2
Abstract

• A service-oriented architecture of a generic
  middleware platform, which provides the required
  services for efficient content storage, search
  and retrieval in a distributed environment.
• Algorithms from Peer-to-Peer computing are
  introduced in a grid environment in order
  scalability, fault-tolerance and data
  availability despite nodes arrivals and
  departures to be ensured
• The system consists of heterogeneous resources
  belonging to different Virtual Organizations

3
Outline

• Introduction
• Challenges and requirements
• Overall architecture
• A multidimensional indexing scheme
• The Distributed Replica Location Service
• GridTorrent protocol

4
A brief introduction

• Grid computing
• Remotely located, disjoint and diverse processing
  and data storage facilities are integrated under
  common software architecture (middleware)
• Resources connected to a shared network and
  provide the necessary software-level services to
  be remotely used and administered
• Heterogeneity of resources
• Rules and policies define the sharing of
  resources
• P2P computing
• Oriented towards the sharing of large amount of
  data
• Files are stored in a dynamic set of peers, which
  may join or the leave the network
• Absence of centralized structures

5
Our approach

• Main Idea
• Exploitation of Peer-to-Peer techniques to
  build a service-oriented
• infrastructure for data management and search
• Features of the proposed system
• A powerful metadata search mechanism supporting
  both point and range queries
• A data transfer mechanism for efficient storage
  and retrieval of data in distributed and
  heterogeneous resources
• A distributed Replica Location Service to keep
  track of file locations
• Motivation
• The design of the proposed architecture has
  been largely motivated by the requirements posed
  by the Gredia research project (http//www.gredia.
  eu/?Pagehome)

6
Challenges and requirements

• The data needs to be partitioned among the nodes
  according to a strategy ensuring
• load balance
• effective query processing
• Absence of centralized structures to provide an
  overall view of the system
• Efficient query routing is required so as
• A small number of nodes to process the query
• The number of exchanged messages to remain
  relative small
• Data locality shall be preserved, namely relative
  data to be kept in the same node if feasible
• New advanced features should not affect the
  maintenance cost of the overlay
• The large size of data along with limitations in
  storage capacity and network bandwidth should be
  considered and a more flexible structure should
  be adopted
• ? Metadata and data will be stored in different
  overlays

7
Architecture (1)

• Three different overlays will be implemented
• Metadata overlay
• DRLS overlay
• Storage overlay
• The Metadata overlay and the DRLS overlay are
  implemented with the required extensions to the
  Kademlia DHT
• The Storage overlay comprises a distributed
  repository
• Kademlia DHT
• PING, STORE, FIND_NODE and FIND_VALUE RPCs
• A hash function is used to assign keys to values
  stored in the DHT
• Distance among points in the key space is defined
  by and XOR metric

8
Architecture (2)

• A data file is described by a predefined set of
  attributes included in its metadata file
• Upload a file
• The file is assigned with a unique identifier
• The unique identifier is included in the
  metadata file
• The data file is uploaded to the Storage
  Overlay by the GridTorrent mechanism
• The metadata file is inserted in the Metadata
  Overlay
• The physical location(s) where the file is (are)
  inserted in the DRLS overlay

9
Architecture (3)

• Search procedure
• The search mechanism is applied in the Metadata
  overlay
• A user can search the metadata files according to
  his / her criteria and select the data file(s) of
  its own interest
• The physical location(s) of the data file
  replicas are returned by the DRLS overlay
• The GridTorrent protocol downloads a file
  exploiting sharing properties in order to boost
  aggregate performance

10
Space Filling Curves

• Map continuously a compact interval to a
  d-dimensional space
• Partitioning of the d-dimensional space into 2kd
  cells, which in turn are mapped through the Space
  Filling Curve to 2kd points of a single dimension
  
• Recursive nature (generation of the curve)
• Perseverance of locality, points being close in
  the 1-dimensional space are mapped to points that
  are close together in the d-dimensional space

11
Multidimensional indexing

• The set of d attributes chosen to be indexed form
  a d-dimensional space
• Each combination of attributes values is
  depicted to a point of the d-dimensional space
• The key of a metadata file is produced by the
  Space Filling curve
• Query processing
• Clusters of the Space Filling curve will be
  defined answering the query
• Lookup for the specified clusters
• Load balance
• Virtual servers ? entity that owns an interval of
  the identifier space
• Each physical node contains multiple virtual
  servers
• When a physical node is overloaded by virtue of
  available storage space or bandwidth, it may move
  one ore more of its virtual servers to another,
  underloaded physical node

12
DRLS (1)

• A Distributed Replica Location Service
• A DHT by correlates its inherent key-value pairs
  to the unique identifier of a file to Physical
  File Names (PFNs) mappings
• Problems
• In DHTs read-only files are stored while mutable
  data cannot be handled
• The replication strategy results in key-value
  pairs to be stored in nodes that are close to the
  ID of the key and cached around the network
• The exact location(s) of a key-value pair in a
  given moment cannot be returned and the update of
  all replicas is not ensured
• PFN mappings for a given LFN change frequently

13
DRLS (2)

• Solution
• Every lookup always queries all nodes responsible
  for a specific key-value pair
• The Lookup procedure does not stop to the first
  returned value and peers that do not reply with a
  value, they are considered as not uptodate
• When all available results are returned, the
  query node compares the results based on some
  predefined version vector (indicating the latest
  update of the value)
• Updates are propagated to the nodes it has found
  responsible for storage but not yet up-todate
  with the latest value

14
The data transfer mechanism of the Storage overlay

• Implementation of BitTorrent designed to
  interface and integrate with GridFTP
• A GridTorrent client is able to request file
  fragments from other GridTorrent clients holding
  the file or connect to GridFTP servers
• Effectiveness and scalability, even under extreme
  load and flash crowd conditions
• Optimized replica selection based on the
  rarest-first policy
• BitTorrent
• A peer-to-peer protocol that allows clients to
  download files from multiple sources while
  uploading them to other users at the same time

15
GridTorrents basic components
16
GridTorrent

• Queries the RDLS periodically
• The DRLS returns to the GridTorrent peer
• PFN of replicas identified by the GridTorrent URL
• gtp//site.fully.qualified.domain.name/path/to/fi
  le
• File size, hashes of pieces, size of each piece
• GridTorrent client
• The two invlolved peers initiate communication by
  exchanging the BitTorrent bit field message,
  informing each other of the pieces they possess
• A request message for blocks is issued
• GridFTP server
• The client issues a GridFTP partial get message
  for the data within the specific block it intends
  to download

17
Conclusions

• The paper proposes a service-oriented
  architecture for efficient search and retrieval
  of annotated content
• Different overlays are implemented and
  Peer-to-Peer techniques are introduced in the
  Grid environment
• The design of the platform ensures scalability
  and fault-tolerance
• An extensible architecture is presented favoring
  the integration with other systems and the
  development of Grid applications

18

• Thank you for your attention