P2P-Tuple: Towards a Robust Volunteer Computing Platform

. . . . . .

Volunteer ComputingP2P-Tuple Architecture

EvaluationFuture

.

.. ..

.

.

P2P-Tuple: Towards a Robust VolunteerComputing Platform

Emmanouil Dimogerontakis

FIB UPC

March 13, 2012

1 / 27

. . . . . .


EvaluationFuture

Importance of VCVC Infrastructures

.. .1 Volunteer ComputingImportance of VCVC Infrastructures

.. .2 P2P-Tuple ArchitectureOverviewPeer Design

.. .3 EvaluationSimulationResults

.. .4 Future

2 / 27

. . . . . .


EvaluationFuture


.. Grid-y for Science

3 / 27

. . . . . .


EvaluationFuture


.. BOINC

4 / 27

. . . . . .


EvaluationFuture


.. P2P-Tuple Contribution

5 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design




.. .4 Future

6 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design

.. Base technologies

7 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design


8 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design


9 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design

.. Overall Operation

Step 1: Peers join the overlay and a SCRIBE group.Step 2: Group owners submit their task to the Tuple Space using aclient program.(Tasks can also be pipelined)Step 3: Peers periodically broadcasts task notificationmessage.Step 4: Peers in the same group start pulling jobs related to thattask.Step 5: Notified peers spread the word.Step 6: Once the job is finished the result is put encoded back onthe Tuple Space.

10 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design

.. Peer Design

11 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design

.. Peer Daemon

12 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design

.. Job Executor Daemon

13 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design

.. Measures to Achieve Fault Tolerance

Measure: Provide high availability, face high churn rate, remainreliable.Effect: Redundancy of data, Replication, use of Erasure Code.

Measure: Loss of data from high churn rateEffect: Communicating Mirrored Replicas of stored objects.Measure: Untrusted peers, possible low job completion probability.Effect: Job Duplication using a randomized job scheduling scheme.Measure: Disconnecting Client Application.Effect: Checkpoints.

14 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design


Measure: Provide high availability, face high churn rate, remainreliable.Effect: Redundancy of data, Replication, use of Erasure Code.Measure: Loss of data from high churn rateEffect: Communicating Mirrored Replicas of stored objects.

Measure: Untrusted peers, possible low job completion probability.Effect: Job Duplication using a randomized job scheduling scheme.Measure: Disconnecting Client Application.Effect: Checkpoints.

15 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design


Measure: Provide high availability, face high churn rate, remainreliable.Effect: Redundancy of data, Replication, use of Erasure Code.Measure: Loss of data from high churn rateEffect: Communicating Mirrored Replicas of stored objects.Measure: Untrusted peers, possible low job completion probability.Effect: Job Duplication using a randomized job scheduling scheme.

Measure: Disconnecting Client Application.Effect: Checkpoints.

16 / 27

. . . . . .


EvaluationFuture

OverviewPeer Design


Measure: Provide high availability, face high churn rate, remainreliable.Effect: Redundancy of data, Replication, use of Erasure Code.Measure: Loss of data from high churn rateEffect: Communicating Mirrored Replicas of stored objects.Measure: Untrusted peers, possible low job completion probability.Effect: Job Duplication using a randomized job scheduling scheme.Measure: Disconnecting Client Application.Effect: Checkpoints.

17 / 27

. . . . . .


EvaluationFuture

SimulationResults




.. .4 Future

18 / 27

. . . . . .


EvaluationFuture

SimulationResults

.. Impact of mirrors

.Result.... ..

.

.Using 1 or 2 mirrors extends the duration of the data availability.

19 / 27

. . . . . .


EvaluationFuture

SimulationResults

.. Impact of Redundancy Level

.Result..

.. ..

.

.

The increase of the duration of data availability is proportionate tothe increase of Redundancy level.

20 / 27

. . . . . .


EvaluationFuture

SimulationResults

.. Behavior in high churn environment

.Result.... ..

.

.In a high churn rate environment the availability decreases easily.

21 / 27

. . . . . .


EvaluationFuture

SimulationResults

.. Results

A Volunteer Computing infrastructure:DecentralizedFault Tolerant

that provides:High AvailabilityReliability

22 / 27

. . . . . .


EvaluationFuture

SimulationResults

.. Results

A Volunteer Computing infrastructure:DecentralizedFault Tolerant

that provides:High AvailabilityReliability

23 / 27

. . . . . .


EvaluationFuture




.. .4 Future

24 / 27

. . . . . .


EvaluationFuture

.. Future

Optimizing current infrastructures and ensure system stability”MPI-like” support”MapReduce-like” support

25 / 27

. . . . . .


EvaluationFuture

26 / 27

. . . . . .


EvaluationFuture

.. Bibliography & Useful Links

Lei Ni, Aaron Harwood, P2P-Tuple: Towards a RobustVolunteer Computing Platform. International Conference onParallel and Distributed Computing, Applications andTechnologies, 2009.Hao Ning, Robust Overlay networks for Volunteer Computing.Master thesis on Chalmers University of Technology, 2010.http://ww2.cs.mu.oz.au/~lni/p2ptuple/.http://www.freepastry.org

27 / 27

http://ww2.cs.mu.oz.au/~lni/p2ptuple/

http://www.freepastry.org