SuperDuperNetworking Transforming Supercomputing from the point of view of Large Scale Visualization

1
SuperDuperNetworking Transforming
Supercomputingfrom the point of view of Large
Scale Visualization andCollaborative Work

• Jason Leigh
• spiff_at_evl.uic.edu

2
A Typical Data Correlation and Visualization
Pipeline
Data Source ? Correlate/Filter ? Render ?
Display Data Source ? Render ? Display Data
Source ? Render Display Data Source ? Correlate
? Render Display Data Source Correlate ?
Render Display Data Source Correlate ? Render
? Display

• Things to notice
• Pipelines are static for long periods of
  time-NOT like web surfing- so.
• Routing is not crucial.
• Program code is tiny compared to volume of data
  processed. Caching wont help much- so.
• Need to stream lots of data through fast
  concurrent pipelines!
• Need pipelines to be optimized from end to end.

3
Experiment to Use Inexpensive Photonic Switches
as an alternative to traditional million
routers to provide application-controlled
deterministic network paths/pipelines.
Long haul link
The cross connections are application- programmabl
e.
Protocol data rate independent
Calient / Glimmerglass at StarLight EVL
4
In Collaborative Work, Data or Visualization
needs to be Distributed to Collaborating Sites
Data Source ? Correlate ? Render ? Display Data
Source ? Render Display Data Source ? Render ?
Display Data Source ? Correlate ? Render
Display Data Source Correlate ? Render
Display Data Source Correlate ? Render ? Display
5
Photonic Multicast Service
Glimmerglass Reflexion Photonic Multicast-capable
Switch
6
Photonically Multicasting a Visualization

• Challenges
• Need to augment traditional Routing and
  Wavelength Assignment algorithms to consider
  photonic multicast constraints.
• Need extreme speed reliable multicast protocol

7
1st Step Realize Local Area Photonic
Multicasting(Visit Booth R2935 to learn how this
is done on the OptIPuter)
8
Quiz Guess the Mystery Computer with the
enormous bandwidth but tiny caches
2.4GB/s from main memory to graphics (Todays
AGP8X is at 2.1GB/s)
48GB/s! (Todays Quadro FX3000only has 27GB/s)
Tiny caches
Memory 32MB
Graphics Synthesizer 4MB
SIF
IPU
128bit bus
DMA
16K cache
GIF
VU0
300MhZMIPS 3
VU1
FPU
Vector processors with several parallel pipelines
For distributed, collaborative large scale data
visualization, we need a version of this that
extends to wide area environments.