TorusVis ND : Unraveling High- Dimensional Torus Networks for Network Traffic Visualizations Shenghui Cheng, Pradipta De, Shaofeng H.-C. Jiang* and Klaus.

TorusVisND: Unraveling High-Dimensional Torus Networks for Network Traffic Visualizations

Shenghui Cheng, Pradipta De, Shaofeng H.-C. Jiang* and Klaus Mueller

Visual Analytics and Imaging LabComputer Science Department

Stony Brook University and SUNY Korea

*Computer Science Dept. The University of Hong Kong

Motivation

Measuring performance on large scale computers is important

Visualizing these measurements can boost understanding

Exploration is best done in the context of the network

Complexity of the interconnections makes exploration difficult

Obstacles

Would like to visually explore processor state and occupancy communications between processors do this over time

Would be fairly easy for a 2D mesh network

Unfortunately we have a mesh with D>2

How to Visualize an ND Mesh

Provide a large number of projections this can be overwhelming to the user

Create an optimized 2D layout of the nodes could use MDS but the interconnections would clutter the display

Need a display that separates the nodes from the links

Our Strategy

Basic concept form a circle equally space nodes on it interconnect in circle interior

Problem is line clutter in interior

Overcome with edge bundling

Mapping the Nodes onto the Circle

Need a node serialization scheme

Naïve serialization sequential numbering/indexing increment node indices in modulus order has an uneven degree of locality

How can locality be improved? space-filling curve Hilbert curve proven to have the best locality generate for any D self-similar fractal structure

Locality

Locality metric

sequential L = 1,822 Hilbert L = 1,414 (20% better) for N= 45 = 1.024

Average distance of neighbors

( ( ), ( ))1

( , )N

d C i C j i ji

L w Dist V V

sequential Hilbert

Zooming In

Looking at a local group of 6 processors sequential and Hilbert indexing clearly expressed Hilbert appears more local than sequential

sequential Hilbert

Interaction

Key to deal with data deluge number of processors number if interconnections types of performance metrics time

We allow user to select processor groups of interest time slices of interest performance metric selector not yet implemented

Node Selection Interface

Based on parallel coordinates all nodes

selecting a single node and showing its links in context of others

Node Filtering and Bracketing

Isolate a group of processors in a certain address range

Certain processors and links might be more important for example: a larger number of

messages in a certain time interval other importance metric

Use Case

Simulated the wake-up of a processor network random processor wakes up and sends a message to a

random neighbor neighbor sends the message to its won random neighbor all processors awake after half the simulation time

First visualization track a single message over time lighter color is older

heavier traffic more emphasized

Time Slice Selection

Features width of stream maps to number of messages in a time interval

Observations each node has times of no messages sent, but also burst periods there are also quiet and burst times overall

time

selected time slice

Selected Time Interval

Time slicer, node selector and network display are tightly coupled

links with messages links with and without messages

Conclusions

TorusVisND follows one of the classic paradigms of information visualization

overview, filter, and detail on demand the “Visual Information Seeking Mantra” puts the user into the loop of steering the data

exploration operations like selection, filtering, and brushing

Brushing tools can be used in two ways selection and filtering visualization of network performance allow users to interact with large and complex data

TorusVisND not restricted to torus networks any highly-connected network can benefit in principle

Future Work

Make interface more scalable introduce multi-resolution capabilities into the network

display to allow it to handle larger numbers of network nodes

introduce multi-perspective lenses to the network display interior to allow users to zoom into multiple areas of interest.

Work with domain experts and real data this will truly optimize our framework and system make it more practical inspire new work

Questions?

Funding: NSF grant IIS 1117132 MSIP (Ministry of Science, ICT and Future Planning),

Korea, under the "IT Consilience Creative Program (ITCCP)" (NIPA-2013-H0203-13-1001) supervised by NIPA (National IT Industry Promotion Agency)