with GPUs and Design Leo Meyerovich (@LMeyerov) CEO 1 is: - - PowerPoint PPT Presentation

Scaling Data Visualization with GPUs and Design

Leo Meyerovich (@LMeyerov) CEO

1

is: Supercharging visual analytics through GPU cloud streaming. (We tricky graphs.)

3

The Future of Visual Analysis

4

Not the Future We Were Promised

5

Ballot Boxes: 100K rows x 30 col CSV

Stack Towns by Voter Turnout

6

0% 25% 50% 75% 100%

Voter Turnout # Towns Most towns had ~40% people vote ballot box stuffing?

7

Opposition Incumbent Tiny square shows town size (area) and vote (color)

8

Filter for towns w/ high turnout

9

Tag suspicious with black

10

11

Analyze suspicious activity in context What parts of the supply chain were hit?

12

A slider is worth a hundred queries.

13

A slider is worth a hundred queries.

Challenge: Tools must keep pace with human ingenuity: interact meaningfully and quickly

DEMO: The Power of Meaningful Layouts

On a small graph (77 nodes), meaningful design adds some clarity

CASE STUDY: TWITTER FRAUD Naïve layout on 1K+ node graphs give impenetrable hairballs.

Gauss-Seidel Force-Directed Graph, O(N^2) n-body, GPU

Node: Twitter account Edge: Friendship

Friends and friend-of-friends of a bot who randomly messaged real people and retweeted them.

With smart layouts, fake account clusters pop out

ForceAtlas2 Layout, O(n log n) n-body, GPU

The spambot is an entrypoint to more bots…

A quiet small business who buys virtual game currency from gamers…

Who somehow got exactly 1 message massively trended & advertised by Twitter

spammer laundering accounts bot retweet network It’s a “retweet laundering” botnet! Tricks Twitter into targeting gamers to check out a cyberfraud site. They steal gamers’ money and identities.

DEMO: GPUs Enable Exploration

22

Uber Trips through SF Start to End

23

Connecting the Dots: OVERPLOTTED!

24

Uber Trips through SF Start to End

Edge Bundling Reveals Arteries

25

Uber Trips through SF Start to End

Edge Bundling Reveals Arteries

… But too slow to filter on time, location, demographics, …

26

Uber Trips through SF Start to End

Edge Bundling Reveals Arteries

DEMO: GPU Acceleration

Under the Hood: Architecting for GPU Cloud Streaming & Benchmarks

Thin/Thick is Dead.

Home: broadband Office: GigE

Build thick/BIG.

multicore + GPU multicore + GPU

Architecting Visual Analytics around thick/BIG (GPU Cloud Streaming)

layout & analytics

cluster, filter, summarize, …

multicore decoder multicore encoder compressed geometry (VBO) rendering engine

zoom, drag, mouseover, …

Portable & predictable. Scalable.

0.1 1 10 100 500K 1.0M 1.5M Frames per second Graph Size: # Nodes + # Edges

Explore 15X More Data, 60X Faster

10

interactivity threshold

(multicore)

60x faster

Graphistry streaming from AWS G2 Gephi on 2014 MacBook Pro Multiple SNAP datasets

(thick/big)

15X+ bigger data

Region by pop. GPUs RAM TFLOPS Cost*

SF 1 4GB 2 $0.06/hr

*Calculated as $0.60/hr AWS G2 instance / 10x timesharing

Region by pop. GPUs RAM TFLOPS Cost*

SF 1 4GB 2 $0.06/hr Bay Area 10 40GB 20 $0.60/hr

*Calculated as $0.60/hr AWS G2 instance / 10x timesharing

Region by pop. GPUs RAM TFLOPS Cost*

SF 1 4GB 2 $0.06/hr Bay Area 10 40GB 20 $0.60/hr California 100 400GB 200 $6.00/hr

*Calculated as $0.60/hr AWS G2 instance / 10x timesharing

Region by pop. GPUs RAM TFLOPS Cost*

SF 1 4GB 2 $0.06/hr Bay Area 10 40GB 20 $0.60/hr California 100 400GB 200 $6.00/hr America 1000 4TB 2 PFLOPS $60.00/hr

*Calculated as $0.60/hr AWS G2 instance / 10x timesharing

… less than even

• ne consultant …

35

Code less, explore more. New era is thick/BIG: GPU cloud streaming.

We’re Hiring! Infoviz & frontend (and contact for info on using/embedding)

36

info@graphistry.com Twitter: @LMeyerov