Centralized vs. Decentralized Design for Internet Applications

1
Centralized vs. Decentralized Design for
Internet Applications
I

• Adriana Iamnitchi
• Department of Computer Science
• The University of Chicago

2
Internet Applications

• Components that build the Internet itself (DNS )
• Tools that connect the user to Internet resources
  (browsers, applets, CGIs, ...)
• Services that can be accessed through Internet
  (e-commerce, e-banking, newspapers, e-libraries,
  )

• Applications that run on a collection of
  Internet-connected resources (SETI_at_home, )
• Tools that create new environments over the
  Internet (middleware services)

3
Internet-Connected Resources

• Unreliable communication
• Unreliable resources
• Highly heterogeneous environment
• Potentially very large number of resources
• Potentially highly variable number of resources

4
Centralized or Decentralized?

1. Applications
2. Middleware services

5
Internet-Connected Resources

• Unreliable communication
• Unreliable resources
• Fault-tolerance mechanisms
• Highly heterogeneous environment
• Asynchronous algorithms
• Potentially very large number of resources
• Potentially highly variable number of resources
• Scalability

6
Application Design Decentralized!

• What about
• Distributed management control?
• Fault tolerance in distributed, asynchronous
  systems?
• Termination detection?
• Communication costs?
• Security?

7
Experience with MetaNEOS

• Solving very large optimization problems on
  metacomputing platforms
• Branch-and-bound search algorithms
• Search for optimal solution
• Successive decomposition of the original problem
• Elimination of unpromising subproblems based on
  the best known solution

8
Fully decentralized BB Solution

• Process management group membership based on
  epidemic communication
• Fault-tolerance tree-based encoding of the
  problem space.
• Report completed problems
• Unsolved problems detected/restored based on
  completed problems
• Price redundant work
• Termination detection tree contraction
• Dynamic load balancing

9
Decentralized BB Performance
Processors Execution time (h) BB time Contraction time Communication (MB/h/p)
10 7.93 98.1 0.3 1.0
30 2.91 90.4 5.2 1.4
50 2.00 81.2 11.7 2.3
70 1.37 87.3 2.3 3.1
100 1.04 84.4 1.1 4.5
10
Decentralized BB Fault Tolerance
11
Decentralized BB Fault Tolerance
12
Experience with MetaNEOS

• Decentralized design is wonderful
• Meantime, the centralized implementation produces
  results, because
• Centralized code already exists (master-worker)
• Available resources hundreds resources working
  simultaneously (Condor testbed)
• Centralized code still efficient on relatively
  small collections of resources

13
Centralized or Decentralized?

1. Applications
2. Middleware services

14
Middleware Services for Computational Grids

• Computational Grids hardware and software
  infrastructure that provides access to
  computational capabilities.
• Middleware services responsible for application
  performance
• Information Services
• Service Location Services (Resource Discovery)
• Resource Management
• Security
• Fault tolerance/detection

15
Information Service Resource Discovery

• Information Service
• Resources (networks, computers, applications, )
• Users
• Resource Discovery Give me n resources with
  attribute X
• Input set of resource attributes
• Output set of resources
• Attributes hardware characteristics, current
  load, network connection, existent/available
  software, data, etc.

16
Resource Discovery Requirements

• Scalable
• Increasing number of resources
• Increasing number of users
• Reliable
• Flexible (heterogeneity support)
• Heterogeneity
• Administrative level (policies)
• Technical level (hardware and software)
• Support for changing environment

17
Resource Discovery Requirements

• Efficient
• Accurate
• Secure
• No global hierarchy
• Politically difficult for wide area (impossible?)
• Hierarchical structures are resistant to change

18
Globus

• Toolkit that builds computational grids
• Components
• Metacomputing Directory Service
• Heartbeat Monitor
• Grid Security Infrastructure
• Globus Resource Allocation Manager
• Global Access to Secondary Storage
• Nexus

19
Globus MDS Step 1
CUS, oGlobus, oUC, ouCS
CUS, oGlobus, oANL, ouMCS
CUS, oGlobus, oUSC, ouISI
20
Globus MDS Step 2
CUS, oGlobus, oANL, ouMCS
CUS, oGlobus, oUC, ouCS
CUS, oGlobus, oUSC, ouISI
21
Globus MDS Step 3
oGrid, dcmcs, dcanl, dcgov
Organizational Server
Organizational Server
oGrid, dcisi, dcedu
oGrid, dccs, dcuchicago, dcedu
Organizational Server
22
Decentralized Information Service

• More difficult than the centralized design
• Resource discovery based on attributes
• Rich set of queries to support
• Compound queries
• Static and dynamic data
• Access policies
• Necessary

23
Conclusions

• Applications running on collections of
  Internet-connected resources may be centralized
  or decentralized.
• Middleware services must be decentralized.