Delivering Parallel Computing Power to the Desktop

1
Delivering Parallel ComputingPower to the Desktop
1
2
Interactive Supercomputing
MATLAB
Desktop Applications
Sparse Matrix Manipulation
TCP/IP Network
Finite Elements
Large Data Set Manipulation
Visualization
3
A New Way To Do HPC

• Problem Sizes Growing Everywhere
• More Machines Now
• Hard To Program (MPI not good enough anymore)
• Community Needs Interactive Parallel Environment
• The Star-P Solution Focuses On
• Huge problems
• Interactive environment
• Seamless integration of libraries -- convenience
• Distributed global array, sparse/dense data
  structures

4
Sparse Matrix Principles

• Problems Get Bigger à Sparse Matrices
• Networked Systems à Sparse Matrices
• Interrelated Data à Sparse Matrices
• Substantial Increased Use at MIT!
• Design Goals
• Sparse matrices never created automatically, once
  created propagate
• Performance important -- usability, simplicity,
  completeness, robustness more important
• Storage O (nonzeros)
• Time O (flops)

5
Distributed Sparse Data
P0
P1

• Each processor stores
• of local nonzeros
• range of local rows
• CSR form

P2
Pn
6
The Star-P Vision

• Star-P Bridges Serial Desktop Applications to HPC
• Enables high level simulation tools to operate in
  parallel with high interactivity
• Computation, visualization, data exploration
• APIs for OEMs -- differentiate offerings for
  hardware software uniqueness
• Delivers Material Gains
• Programming time
• 10-100X faster
• Parallelism propagates automatically to 90 of
  code
• Execution time
• 100x demonstrated heavily app-dependent
• Problem size limited only by memory available
  SGI now 24 tb
• Interactive problem solving
• Programming visualization

7
Star-P Solution
Master Server
PBlas
FFTW
Scalapack
PBase
SuperLU
Star-P
Matrix 1
Sparse Dense
Matrix 2
Matrix 3

• Six Years of Evolution
• Power of Abstraction

8
Star-P Solution
Master Server
PBlas
FFTW
Scalapack
PBase
SuperLU
Library Manager
Matrix 1
Sparse Dense
Data Manager
Matrix 2
Matrix 3
Server 1
Server 2
Server 3
Server 4
9
Star-P Solution
Master Server
PBlas
FFTW
Scalapack
PBase
SuperLU
Library Manager
Client Manager
Client
Matrix 1
MATLAB
Sparse Dense
Data Manager
Server Manager
Matrix 2
Matrix 3
Server 1
Server 2
Server 3
Server 4
10
Star-P Solution
Master Server
PBlas
FFTW
Scalapack
PBase
SuperLU
Library Manager
Client Manager
Client
Matrix 1
MATLAB
Sparse Dense
Data Manager
Server Manager
Matrix 2
Other (simulation Modeling) Applications
Matrix 3
Server 1
Server 2
Server 3
Server 4
11
DEMOS

• Real Live Demos With The SGI Altix On The Floor

12
Product Release Stages

• Current Initial Use Partners
• Ohio supercomputing center
• Large government agency
• Commercialization Customers
• Release to Market

13
Conclusion

• If In Your Use Of MATLAB You Need
• Solutions for huge problems (not easily
  decomposable)
• Interactive environment
• Seamless integration of libraries for
  convenience
• Distributed global array, sparse/dense data
  structures
• Star-P Is A Must-Have Contact Us Now
• If You Need Parallel Performance But MATLAB Is
  Not Your Preferred Environment, Please Talk To Us!

14
For more information

• www.interactivesupercomputing.com
• or
• edelman_at_interactivesupercomputing.com
• ppeterson_at_interactivesupercomputing.com

15
Star-P Solutions

• Parallel Computing Needs
• Serial to Parallel (Holy Grail)
• New code development made easier
• Server Support
• Flexible connections to Client
• Flexible connections to Libraries
• Enables high-end computing
• Enables large problem sizes
• Prototypes à Production