12
Looking at the Sky through Clouds Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

Embed Size (px)

Citation preview

Page 1: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

Looking at the Sky through Clouds

Mihai Lucian Cristea, on behalf of SCARIe teamUniversity of Amsterdam

TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

Page 2: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

eVLBI on the Grid: SCARIe

Problems when running SCARIe on Grid

Workflow management: WS-VLAM

Experiments

Conclusions

Overview

Page 3: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

eVLBI

Page 4: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

eVLBI on the Grid:SCARIe

Telescopes

Bring the data from telescopes:Current: 4x256MBpsMid-target: 16x1GbpsFuture scenario: 32x4Gbps

Correlator

Input nodes

Correlator nodes

Output node

Result

Software Correlator Architecture Research and Implementation for e-VLBI

Requirement: constant throughput

Page 5: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

Optimizing the SCARIe application on Grid

Jitter due to network congestion

Telescope

Correlator

Input node

Correlator nodes

Output node

80% 80% 95% 80%

Jitter due to network overload at ingress

2

3

1

NE

Page 6: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

Specific services to applications:◦ Only the App knows how to optimally use the

resources

Solutions to meet the specific network demands:◦ Schedule network resources (e.g., parallelize the

link usage, not only the CPU usage, tradeoffs link connectivity vs. energy budget)

◦ Application controls the network resources

Network as a service in Grid

Page 7: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

NetBroker-Topology-Bandwidth

NetBroker-Topology-Bandwidth

Grid NGrid 1

GridBroker-CPU-Storage

GridBroker-CPU-Storage

WS-VLAM

Actuator Profiler

Scheduler

experiment

2

13

4 5

0

WS-VLAM

App.2App.1

ACsNEs

App.4App.3

Workflow management: WS-VLAM

Page 8: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

Workflow management: WS-VLAM

Page 9: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

Experiments: Testbed

DAS1

DAS3

DAS5

DAS7

10.1.0.27

10.1.0.29

10.1.0.31

10.1.0.33

10.1

.0.x

100M

bps S

witc

h

DAS2

DAS4

DAS6

DAS8

10.1.0.30

10.1.0.32

10.1.0.34

10.1.0.28

IXDP2850

1Gbps 10.10.0.32

10.10.0.30

10.10.0.28

10.10.0.34

10.10.0.31

10.10.0.29

10.10.0.27

10.10.0.33

Network Broker

Page 10: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

DAS1

DAS3

DAS5

DAS7

10.1.0.27

10.1.0.29

10.1.0.31

10.1.0.33

10.1

.0.x

100M

bps S

witc

h

DAS2

DAS4

DAS6

DAS8

10.1.0.30

10.1.0.32

10.1.0.34

10.1.0.28

IXDP2850

1Gbps 10.10.0.32

10.10.0.30

10.10.0.28

10.10.0.34

10.10.0.31

10.10.0.29

10.10.0.27

10.10.0.33

Network Broker

W1

W2

W3

W4

R1

R2

R3

R4

A

B

C

D

E

Experiments

Page 11: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

Results

Playback Demos:http://staff.science.uva.nl/~gvlam/wsvlam/demos/wsvlam-dynamic-bw.htmlhttp://staff.science.uva.nl/~gvlam/wsvlam/demos/wsvlam-vlc-demo.html

Page 12: Mihai Lucian Cristea, on behalf of SCARIe team University of Amsterdam TERENA CONFERENCE ‘10, Vilnius, 1 June 2010

Close interactions between applications and networks enables better usage of resources

We support it in Grids by enabling networks as a service

When network resources are not transparent to applications, the interfaces between sensors, networks, and computational resources in the Grid can be managed in order to achieve an optimal interworking

Conclusions