The Organic Grid: SelfOrganizing Computation on a PeertoPeer Network

1
The Organic Grid Self-Organizing Computation on
a Peer-to-Peer Network

• J. Chakravarti, G. Baumgartner, and M. Lauria
• Presented by
• Rajamani Sethuram

2
Outline

• Introduction
• Prior Work
• Organic Grid
• Self-Organizing Grid
• Fault Tolerance
• Experiments and Results
• Discussion

3
Introduction

• Massive Computational Power requirement
  internet computing, desktop computing
• Based on Master/worker model ? Problem
• All softwares are not available in all PCs
• Master becomes a bottleneck
• Optimal Computation schedule using global
  knowledge Not scalable
• Hence use a decentralized and adaptive scheduling
  scheme

4
The Problem Statement

• Network is unreliable
• Each node has varying
• computational capability
• Tasks are independent

5
Prior Work

• P2P Internet Computing
• Worm, Condor
• GIMPS, SETI_at_Home, folding_at_home Internet
• Entropia, United Devices Commercial
• Scheduling
• K-Best Neighbor
• G-Commerce
• Kreaseck et al.-Tree Overlays

6
Organic Grid

• An effort to build infrastructure for distributed
  computation from scratch
• Best model for utilization of system
• Use mobile agent approach encapsulate
  computation and behavior into mobile agents. Also
  incorporates dynamism
• Strong mobility in JVM uninterrupted agents
• Forced Mobility Migrating agent using external
  thread

7
Autonomic Scheduling

• Supports an adaptive, decentralized system
• Assumes some knowledge of the network Friend
  list (more scope for research)
• Learns and dynamically changes the tree overlay
  network while ongoing computation

8
Mobile Agent

• Are collection of threads
• Capable of
• Begin execution in the local node
• Receive requests for work from other nodes
• Send requests for work
• Report results to the parent node
• Fault Tolerance

Request
Uncomputed Task PARENT
CHILD
A
B
CLONE
9
Self-organizing

• Each node has c active children ranked
  according to their performance
• Performance measured using the rate at which they
  send results
• Consider results in burst
• p potential children those which are not yet
  evaluated
• Graduated to active child if they show enough
  results
• Grand child becomes a potential child if it
  performs well

10
Improving the Performance

• Self-adjustment of task list
• Node dynamically determines its speed and
  requests for tasks appropriately
• Pre-fetching
• Waiting for new tasks cause delay
• Overcome delay by pre-fetching
• How much to pre-fetch? Depends on the performance
  of current node.

11
Fault Tolerance

• If parent node becomes inaccessible due to link
  failure, the entire tree is disconnected
• Recover from this scenario
• Keep a list of ancestors
• At failure contact all ancestors
• If no ancestor responds, send requests to
  friends, inform the environment and self-destroy

12
Other Issues

• Cycles could create deadlock
• Examine the ancestor list and detect cycle
• Termination
• Root of the tree sends signal to all the leaf
  nodes

13
Experiments

• Network - 18 machines run in different part of
  Ohio running Aglets on top of Linux/Solaris
• To achive heterogeneity, they added delay
• Application NCBIs nucleotide-nucleotide BLAST
  to match 256KB sequence against 320 data chunks,
  each of size 512KB

14
Time required for the first sub-task to reach a
node
15
Effect of Child Propagation
Propagation enabled
Propagation disabled
Running time 2294 sec.
Running time 3035 sec.
16
Effect of varying Result Burst Size
Result Burst 1
Result Burst 8
17
Effect of Task Pre-fetching
18
Discussion

• The Title Organic Grid is not really justified
• If you have to use an ACO framework, what will be
  the approach. i.e. how to abstract pheromones,
  ant movements etc.
• Can we use other biology inspired algorithms such
  as plain diffusion, reactive diffusion,
  proliferation, reinforcement learning etc. to
  solve the problem
• Tasks are assumed to be independent. What if they
  are not?