The Organic Grid : Self Organizing Computation on a

Peer to Peer Network

Vikram Negi

Distributed Computing

• Internet Computing– Desktop Grids : Scalability

• SETI@HOME, FOLDING@HOME

• Globus

Comparison

Problem Focus

• Schedule Independent Identical subtasks of an Independent Task Application with data in one location – NCBI nucleotide nucleotide sequence

comparison with BLAST tool.

Approach

• Tree Based Adaptive approach– Machine configuration, connection bandwidths,

network topology (absent)– Start Friends Lists ( Build overlay on fly)– Biology inspired LALI Principle

• Positive feedback - Activation

• Negative feedback - Inhibition

Approach

• Initially recruit Friends List– Encourage close nodes v/s distant nodes– Child nodes give passive feedback to measure

effective performance.– Continuous on fly adaptation

Basic Design

• Computation Agent– Identifies a child machine.– Allocate a subtask(computational process) in

child machine(availability).

• Tree structure– Node seeks more subtasks from parent if free.– Node send results to parents

Design

• Maintenance of child lists– Seek a balanced tree– Rank active children on response times..– Maintain a list of potential children

• Restructuring of the Overlay Network– Maintain a tree– Highest throughput closest to root.– Push low throughput closer to leaves.

Design

• Size of Result Burst– To determine rank – Average over a observation

• Fault tolerance– Maintain list of ancestors– Maintain list of unfinished tasks

• Cycles in the Overlay Network– Check the list of ancestors.

Design

• Termination– Root informs all actual, potential and former

nodes.

• Self adjustment of task list size– High performance machine request more tasks

• Prefetching– Give time and number estimates of future tasks

to parents.

Experiment Set up

• 18 machine

• NCBI BLAST application

• Task : Match– 256KB(sequence) with 320 Data chunk 512KB

size – Subtask match each sequence to chunk

Results

The time required for code and the first subtask to arrive at different nodes.

Autonomic ?

Child Propagation

Burst Size

Higher Burst Size slows experiment

Prefetch – Initial Task Size

Ramp up: Time req for subtasks to reach every single nodeRunning Time increases for > 1 prefetch tasks

Self Adjustment

Number of Children

• Good Initial Configurations no affect on running times.

Discussion

• Autonomic ?

Can we learn from

• Self Organizing :– Bird fly, Fishes swim in pattern

• Biods : simulation