apache hadoop india summit 2011 talk "middleware frameworks for adaptive executions and...

24
Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids Sathish Vadhiyar Grid Applications Research Lab Supercomputer Education and Research Centre Indian Institute of Science Bangalore February 16, 2011 Yahoo! Hadoop India Summit, Indian Institute of Science

Upload: yahoo-developer-network

Post on 12-Jun-2015

1.069 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Middleware Frameworks for Adaptive Executions and

Visualizations of Climate and Weather Applications on Grids

Sathish VadhiyarGrid Applications Research LabSupercomputer Education and Research CentreIndian Institute of ScienceBangalore

February 16, 2011

Page 2: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Outline Parallel Simulation and Visualization

◦Resource Constraints◦Impact on Climate Simulations

Adaptive Integrated Framework◦Framework◦Contradictory Objectives

Decision Algorithm Steering the Visualizations Results

◦Progress of Simulation and Visualization◦Adaptation of Parameters

Potential for Cloud Computing

February 16, 2011

Page 3: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Parallel Simulation and Visualization

Critical climate applications like cyclone tracking require High-fidelity high-resolution simulation

◦ High-performance computations◦ Massive amount of output

On-the-fly remote visualization◦ Real-time guidance to policy and decision makers◦ Joint analysis by geographically distributed climate

scientists

High-performancesimulations

Remotevisualization

Figure: Simultaneous simulation and remote visualization using stable storage

Parallel I/O

NetworkDISK

February 16, 2011

Page 4: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Resource Constraints

SIM VIS

Simulation

Process

Visualization Process

Stable Storage

Network

• High computation rate• High I/O bandwidth• Limited network

bandwidth• Limited storage space

Figure: Illustration of resource constraints on simulation

February 16, 2011

Page 5: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Impact on climate simulations

Rapid accumulation of data in the stable storage

Eventual unavailability of storage Stalling of simulation Low temporal resolution Loss of visualization

February 16, 2011

Page 6: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Adaptive Integrated Framework

APPLICATIONMANAGER

APPLICATIONCONFIG

JOB HANDLER

Output Frequency

# Processors

FR

AM

E S

EN

DER

FR

AM

E R

EC

EIV

ER

SIMULATIONPROCESS

ApplicationConfiguration

VISUALIZATIONPROCESS

Stall if no disk space

Network

o Invokes a decision algorithm periodicallyo Reacts to significantly low disk space

o Schedules climate simulation applicationo Starts, stops, restarts simulation process

o Simulates climate across time steps

o Outputs climate data to storage

o Visualizes simulation output

Periodic Invocation

DECISIONALGORITHM

o Adapts to resource and application dynamics

o Determine near-optimal parameters

Storage

February 16, 2011

Page 7: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Decision Algorithm

Objectives1. Maximize rate of simulation2. Maximize temporal resolution3. Enable continuous visualization4. Ensure availability of storage

Contradictory Objectives

February 16, 2011

Page 8: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Decision Algorithm

Input◦ Simulation resolution◦ Network bandwidth◦ Remaining disk space

Output◦ Number of processors for simulation◦ Output frequency

Optimization Based Algorithm

February 16, 2011

Page 9: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Optimization-based Approach

Maximize simulation rate within the constraints related to continuous visualization, acceptable output frequency, I/O bandwidth, disk space and network bandwidth

Causes of faster consumption of storage space Faster execution time Limited network bandwidth High frequency of output

Objectives Optimal processor allocation Best possible output frequency Judicious use of storage

February 16, 2011

Page 10: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Problem FormulationObjective function: minimize t

IO

OS t F T T

b (1)

Time Constraint: Time to solve + Time to output ≤ Time to transfer

t Time to solve one simulation time step

S Number of frames solved in an interval I

F Number of frames output in an interval I

T Number of frames transferred in an interval I

Table: Decision Variables

February 16, 2011

Page 11: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Constraints

LBt T

LB z UB

Bound Constraints: Bounds for t and z

(4)

(5)

.in out

IO

Dn

R R

O F Db

S t F T n

Disk Constraint: Net input to the disk ≤ Remaining disk space

(2)

(3)

February 16, 2011

Page 12: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Experiments Simulation: Weather Research and Forecasting Model v3.0.1 Visualization: VisIt v1.12.0

Climate Application: Tracking Cyclone Aila Modeled area: 32x106 sq. km. from 60ºE - 120ºE and 10ºS - 40ºN Formed: 23th May 2009, Dissipated: 26th May 2009

Figure: Visualization of Perturbation Pressure showing the track of Aila

Pressure (hPa) 995 994 992 990 988 986

Resolution (km) 24 21 18 15 12 10

Table: Resolutions for different Pressure Values February 16, 2011

Page 13: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Experiments

Configuration

Simulation Configuration

Maximu

mCores forSimulatio

n

MaximumDisk

SpaceUsed

AverageSim-Vis

Bandwidth

inter-department

fire: dual-core AMD Opteron 2218 (Gigabit Ethernet)

48 182 GB 56 Mbps

intra-country

gg-blr: Intel Xeon Quad Core Processor X5460 (Infiniband)

90 150 GB 40 Mbps

cross-continent

moria: dual-core AMD Opteron 265 (Gigabit Ethernet)

56 100 GB 60 KbpsTable: Simulation and Visualization Configurations

February 16, 2011

Page 14: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Simulation Progress

Figure: For cross-continent configuration

Simulation stalls in Greedy-Threshold approach

Faster rate of simulation

February 16, 2011

Page 15: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Visualization Progress

Figure: For intra-country configuration

Lags behind in attempt to

visualize every time step initially

Faster rate of visualization

INCREASING LAG

February 16, 2011

Page 16: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Disk Space Utilization

Figure: For intra-country configuration

Higher rate of disk space

consumption

Less than 50% disk

space used

February 16, 2011

Page 17: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Adaptivity

Figure: For inter-department configurationFebruary 16, 2011

Page 18: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

February 16, 2011

Steering the Visualization

Page 19: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

February 16, 2011

Steering Across the Ocean!

Changing number of procs from 96 to 80

Changing Visualization Frequency

Changing Resolution of Simulation

Auto-changing number of procs to maintain QoS

Page 20: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Ship the simulations to a cloud Use resource management services of clouds to

find a “nearby” large storage This will eliminate the storage problem/constraint But new research challenges:

◦ Storage can spill over; Need to maintain metadata of storage repositories

◦ Simulation->Storage->Visualization will now involve multiple hops

◦ Hence added benefits due to large storage-as-service in cloud will have to balanced against loss in performance

February 16, 2011

Potential for Clouds

Page 21: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

The infrastructure has to be expanded to include multiple simultaneous multi-user visualizations of multiple independent simulations

Such independent simulations are natural for executions on clouds.

February 16, 2011

Potential for Clouds

Page 22: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

To minimize lag between simulation and visualization site – choosing representative frames

Multiple visualization-simulation framework Applying for other applications

February 16, 2011

Future Work

Page 23: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

Preeti Malakar (Phd student) Dr. Vijay Natarajan (Co-researcher)

February 16, 2011

Acknowledgements

Page 24: Apache Hadoop India Summit 2011 talk "Middleware Frameworks for Adaptive Executions and Visualizations of Climate and Weather Applications on Grids" by Sathish Vadhiyar

Yahoo! Hadoop India Summit, Indian Institute of Science

February 16, 2011