laurent castanié (earth decision / paradigm – inria lorraine – project alice) christophe mion...

20
Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine – Project ALICE) Bruno Lévy (INRIA Lorraine – Project ALICE) Distributed Shared Memory for Roaming Large Volumes

Upload: victoria-dowd

Post on 27-Mar-2015

223 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

Laurent Castanié(Earth Decision / Paradigm – INRIA Lorraine – Project ALICE)

Christophe Mion(INRIA Lorraine – Project ALICE)

Xavier Cavin(INRIA Lorraine – Project ALICE)

Bruno Lévy(INRIA Lorraine – Project ALICE)

Distributed Shared Memory for Roaming Large Volumes

Distributed Shared Memory for Roaming Large Volumes

Page 2: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

2

OutlineOutline

Introduction• Large volumes in the Oil and Gas EP domain• Previous work: single-workstation cache system

COTS cluster solution: DHCS• Distributed volume rendering• Distributed data management

Real-time roaming in gigantic data sets with DHCS Conclusions

Page 3: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

3

OutlineOutline

Introduction• Large volumes in the Oil and Gas EP domain• Previous work: single-workstation cache system

COTS cluster solution: DHCS• Distributed volume rendering• Distributed data management

Real-time roaming in gigantic data sets with DHCS Conclusions

Page 4: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

4

Reservoir scale volume300x400x400 (~50 MB)100-300 Km2

Targeted ROI1000x1000x1000 (~1 GB)

Targeted volume4000x5000x5000 (~100 GB)

Reservoir scale ROI100x200x200 (~4 MB)

Targeted ROI~ 250x typical

reservoir scale ROI

REGIONAL SCALE10000-30000 Km2

IntroductionInterpretation scales in Oil and Gas EPIntroductionInterpretation scales in Oil and Gas EP

Page 5: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

5

IntroductionOOC visualization on a single workstationIntroductionOOC visualization on a single workstation

8% 6%

0.5%

100 GB

Data Volume(Disk)

512 MB

Graphics Card(V-RAM)

8 GB

Workstation(RAM)

Page 6: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

6

IntroductionOOC visualization on a single workstationIntroductionOOC visualization on a single workstation

Bhaniramka and Demange, IEEE VolVis 2002, OpenGL Volumizer

Plate et al., VISSYM 2002, Octreemizer

Castanie et al., IEEE Visualization 2005VolumeExplorer (coupling OOC visualization and data processing)

Probe-based roaming systems with LRU volume

paging

Page 7: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

7

Efficient solution up to 20-30 GB, however:

ROI size is limited to the amount of graphics memory available; Performance decreases rapidly when the size of data on disk

increases over 30 GB.

=> How to SCALE our solution up to 100-200 GB ?

Distributed Hierarchical Cache System (DHCS)on COTS cluster

IntroductionOOC visualization on a single workstationIntroductionOOC visualization on a single workstation

Page 8: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

8

OutlineOutline

Introduction• Large volumes in the Oil and Gas EP domain• Previous work: single-workstation cache system

COTS cluster solution: DHCS• Distributed volume rendering• Distributed data management

Real-time roaming in gigantic data sets with DHCS Conclusions

Page 9: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

9

SLAVE SLAVE NODENODE

SLAVE SLAVE NODENODE

SLAVE SLAVE NODENODE

SLAVE SLAVE NODENODE

SLAVE SLAVE NODENODE

SLAVE SLAVE NODENODE

SLAVE SLAVE NODENODE

SLAVE SLAVE NODENODE

MASTER MASTER NODENODE

1. Segmentation

2. Distribution

3. Rendering

4. Composition

Distributed volume renderingSort-last parallel volume renderingDistributed volume renderingSort-last parallel volume rendering

Page 10: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

10

Distributed volume renderingPipelined binary-swap compositingDistributed volume renderingPipelined binary-swap compositing

P0 P1 P2 P3

Result on MASTER

node

Ma et al., IEEE CG&A 1994Binary-swap compositing

Cavin et al., IEEE Visualization 2005Cavin et al., Eurographics PGV 2006DViz pipelined implementation

Page 11: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

11

ROI size several GBs => ~15-20 fps

One virtual graphics

card

8 GB

DViz

Master

Distributed volume renderingPipelined binary-swap compositingDistributed volume renderingPipelined binary-swap compositing

16 nodes with GeForce 6800 ULTRA - 512 MB

Page 12: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

12

OutlineOutline

Introduction• Large volumes in the Oil and Gas EP domain• Previous work: single-workstation cache system

COTS cluster solution: DHCS• Distributed volume rendering• Distributed data management

Real-time roaming in gigantic data sets with DHCS Conclusions

Page 13: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

13

Very low disk to memory bandwidth Faster transfers through the network?

?

Distributed data managementLimited disk to memory bandwidthDistributed data managementLimited disk to memory bandwidth

Graphics Card(V-RAM)

Workstation(RAM)

Data Volume(Disk)

Gigabit EthernetNetwork

~1 GB/s~50 MB/s

Page 14: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

14

0 100 200 300 400 500 600

Bandwidth (MB/s)

1xGigabit Ethernet2xGigabit EthernetInfiniband 4xSATA 150SCSI U320 2x RAID0 Maxtor AtlasSATA 150 2x RAID0 WD Raptor

4x faster transfers through the network

50 MB/s

Network

Disk

220 MB/s

Distributed data managementDisk Vs Network bandwidthDistributed data managementDisk Vs Network bandwidth

120 MB/s

500 MB/s

Page 15: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

15

Fully dynamic memory state that must be kept up-to-date

~50 MB/s

~220 MB/s1 1

2

Distributed data managementOur fully dynamic implementation (DHCS)Distributed data managementOur fully dynamic implementation (DHCS)

8 GB8 GB8 GB8 GB

Page 16: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

16

OutlineOutline

Introduction• Large volumes in the Oil and Gas EP domain• Previous work: single-workstation cache system

COTS cluster solution: DHCS• Distributed volume rendering• Distributed data management

Real-time roaming in gigantic data sets with DHCS Conclusions

Page 17: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

17

30 copies of the Visible Human data set= 5580x5400x3840 ~ 107 GB

ROI 1000x1000x1000 ~ 1 GB

=> Real-time rendering and volume roaming at full resolution at 12 fps on average on a 16-node cluster

ResultsReal-time rendering and volume roamingResultsReal-time rendering and volume roaming

Full resolution volume roaming

Full resolution volume rendering

Page 18: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

18

OutlineOutline

Introduction• Large volumes in the Oil and Gas EP domain• Previous work: single-workstation cache system

COTS cluster solution: DHCS• Distributed volume rendering• Distributed data management

Real-time roaming in gigantic data sets with DHCS Conclusions

Page 19: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

19

Volume visualization of ~100 GB Volume roaming a ROI of

several GBs Cluster-based hierarchical cache

system• Distributed volume rendering• Distributed data management

Compression techniques Better load balancing of the communications on the network Pre-fetching strategies to hide disk access Other use cases:

• Combination of multiple attributes of 100 GB each• Real-time full resolution volume slicing at 20-30 slices per second

ConclusionsConclusions

Page 20: Laurent Castanié (Earth Decision / Paradigm – INRIA Lorraine – Project ALICE) Christophe Mion (INRIA Lorraine – Project ALICE) Xavier Cavin (INRIA Lorraine

20

AcknowledgementsAcknowledgements

This work involved and has been supported by:• Earth Decision (now part of Paradigm) http://

www.earthdecision.com• LORIA / INRIA – Project ALICE http://alice.loria.fr• DViz http://www.dviz.fr• Region Lorraine (CRVHP) • GOCAD consortium http://www.gocad.org