Hallucinations about Petaflops

1
Hallucinations about Petaflops

• Or was it visionary comments?
• Paul Messina
• U.S. DOE/NNSA
• March 9, 2000

2
What do these have in common?

• Streets in Oxnard, California
• System software
• and
• Pots and pans from Pompeii
• Mathematical software libraries
• e.g., LINPACK, LAPACK, SCALAPACK

3
Types of Petaflops machines

• Special purpose
• Grape VI
• Blue Gene
• General purpose from COTS (clusters on steroids)
• ASCI platform of year 2009
• Grid-based
• distributed, heterogeneous

4
There will be Petaflops machines with new
architectural features

• The ITR program of FY2004 - FY2010 will fund
  research on computer architectures -- and
  associated system software and tools -- for
  petaflops machines
• recognizing that exaflops systems will need to
  have new features to be usable
• instead of only funding research on ideas that
  will not be usable for another 20 - 30 years

5
There will be Petaflops machines with new
architectural features

• Peta- and exa-flops machines will have
  designed-in subsystems for
• I/O
• internal connection networks (plural)
• visualization and other data manipulations

6
Software for Petaflops computers

• Petaflops computers will have open source system
  software
• There will be libraries of scalable, adaptable,
  fault-tolerant algorithms
• and applications programs will use them!
• Application programs will have a software
  architecture that facilitates integration of new
  modules

7
Open Source through the ages

• Stu Feldmans f77 compiler
• PVM
• vs. Parasofts Express
• MPICH
• all initial MPI implementations based on this
• Linux

8
More hallucinations

• Developers of applications for petaflops
  computers will use runtime libraries, algorithmic
  modules, etc., developed by skilled experts
• instead of claiming that it is already too late
  to wait for that software, therefore the
  developers must devise their own
• and dont have time to look at previous work

9
Summary

• Its my dream and I can put whatever I want in
  it!
• We should be more scholarly in our RD on
  computing