Distributed Visualization• Our use of this term extends its

traditional meaning •

– Distributed: Still aim to support geographically distributed users and collaborative communities

– Decentralized infrastructure: The infrastructure does not need to be centralized as in “compute” centers

– Shared infrastructure: The comp/storage nodes can be independent Internet computers

Distributed Visualization

Data Replication: exNodes

D1

D2

D3

D4

F1 F2 F3 Physical Layer

IBP

LoRS:runtime system

Applications

L-Bone exNode

Network Functional Unit

Input Allocation

Input Allocation

Output Allocation

IBP

Memory mapping

MEM

PE

RD RDRD/WR

NFU is novel due to:

1. weakened semantics and

2. control of security-sensitive operations.

Scheduling

Server:P2

Server:P3 Server:P4 Server:P5

Server:P6

Server:P1 Workstation B

d9

d1

d3

d4

d11

d12

d6

d10

d7

d8

d2

d5

d4

d12

d1

d2

d8

d9

d5

d6

d10

d11

d3

d7

Task Queues

Task Queues

Workstation A

• Depots: {P1,P2,…,Pm}

Pi described by bw bi & computational power ci

• Partitioned dataset {d1,d2,…, dn}, k-way replication

• Vis only need one copy of each dj

• (Optional) DM tasks

Mij replicates dj on Pi

Key Challenge:

Resource performance varies over time !!!

Dynamic Data Movement• Some data partitions are just “unlucky” to be on slow

or heavily loaded servers• After fast depots are done with local tasks, can

dynamically “steal” some slow “partitions”

Partial multi-stream download /w deadline

Succeed?

Termination

Main multi-streamdata transfer

Information update

Termination

deadline

<Tj, Pi >

yes

no

Fast server selection

Slow task selection

Deadline calculation

<Tj>

<Pi >

Results: the depots• Most depots used running on Planet-Lab• Workloads varied across servers and time• Realistic test of feasibility on shared,

decentralized infrastructure

0

20

40

60

80

100

0 50 100 150 200 250 300 350 400

Server ID

On

e-m

inu

te L

oa

d

0

10

20

30

40

50

0 225 450 675 900 1125 1350

Time (minutes)

On

e-m

inu

te L

oa

d

Results: the data

• Test data: 30 timestep of Tera-scale Supernova Initiative, 75GB in total– Provided by Tony Mezzacappa (ORNL) and John Blondin

(ORNL) under the auspices of DOE SciDAC TSI project

Results: the performance• 800x800 image resolution, 0.5 step size in ray-casting,

per-fragment classification and Phong shading• With 100 depots, the average rendering time: 237 sec