apache hadoop india summit 2011 talk "adaptive parallel computing over distributed military...

12
Adaptive parallel computing over distributed military computing infrastructures Rituraj Kumar Center of Artificial Intelligence and Robotics – DRDO Lab

Upload: yahoo-developer-network

Post on 13-Dec-2014

1.509 views

Category:

Documents


2 download

DESCRIPTION

 

TRANSCRIPT

Page 1: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

Adaptive parallel computing over distributed military

computing infrastructures

Rituraj Kumar

Center of Artificial Intelligence and Robotics – DRDO Lab

Page 2: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

2

Contents

Introduction Adaptive Parallel Computing Approaches for APC

MPI + DDS Hadoop

Conclusions

Page 3: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

3

Introduction

Net-Centric Paradigm of warfare needs highly compute intensive operations to be performed.

Parallel Computing is the only feasible solution that would guarantee reduced response time.

Tactical deployment would not amenable for – Large Clusters in the field Back-hauling

Intelligent use of the existing spare compute capacity of the computing devices within the cloud is essential.

Page 4: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

4

Adaptive Parallel Computing

MILCOM

Cloud of Heterogeneous Computing Devices

Page 5: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

5

Adaptive Parallel Computing

The Amdahl's law is concerned with the speedup achievable from an improvement to a computation that affects a proportion P of that computation where the improvement has a speedup of S.

Amdahls Law

Where,S = SpeedupP = Parallel fraction of ProgramN = Number of Processors

If 95% of the program can be parallelized, the theoretical maximum speedup using parallel computing would be 20×, no matter how many processors are used.

Page 6: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

6

Cost for parallel Computation = Computation Cost + Serialization Cost

Adaptive Parallel Computing

Page 7: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

7

APC Approaches

MPI Code:• Parallel implementation of

complex mathematical models

MPI Controller:• Execution of parallel MPI code

over distributed network.

DDS:• Reliable Communication over

challenged network.

DMI:• Identification of computing

nodes in the distributed network.

MPI Controller

DDS

Dynamic Membership

Identifier

MPI Code

MPI + DDS

Page 8: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

8

APC Approaches

Advantage:Well-known MPI API Framework. DDS provides assurance of reliable

communication between nodes.

Disadvantage:Needs a wrapper for converting MPI calls to DDS

framework.Performance degradation because of the wrapper.

MPI + DDS

Page 9: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

9

APC Approaches

Map/Reduce can be effectively used for parallelization

Used extensively in varieties of information systems.

Java Based Designed to handle large data sets

HADOOP

Page 10: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

10

APC Approaches

Advantage:Highly ScalableExcellent fault tolerance capabilitiesGood hardware interoperability

Current challenges:Requirement of slight architectural changesCurrently not suitable for resource constraint

hardware

Hadoop

Page 11: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

11

Conclusion

Military domain requires reliable parallel computing infrastructure over disadvantaged communication network.

Dynamic topology poses great challenge for computing infrastructure.

MPI framework has few disadvantages Hadoop is a promising candidate in these

conditions.

Page 12: Apache Hadoop India Summit 2011 talk "Adaptive Parallel Computing over Distributed Military Computing Infrastructures" by Rituraj Kumar

12

Thank You