Multicore and Programming Models

1
Multicore and Programming Models

• Nathan Dykman

2
Okay, Quickly

• Our experience with highly parallel programs are
  from the HPC/Scientific Computing worlds
• Complex Fortran, C, C programs, libraries.
• Some Python here and there.
• Complex Algorithms (Data Parallelism big deal)
• Big data, big problems, big machines, big compute
  time, big power bills.
• But, shiny movies err visualizations 8-)

3
Problem (Orick Version)

• Fortran _at_() (they still use that?)
• C and C _at_() (do we have to use it?)
• Therefore MPI, PVM, etc. _at_() (see above)
• OpenMP _at_() (exactly how, Im not sure)
• Complex Algorithms and Data _at_()
• (I cant get them to work)
• Different data, different programs, smaller
  machine, fast response times.

4
Challenges

• Big move to programming in Java, .Net, Scripting
  Libraries (Ruby, Python) (C)
• Managed languages/runtimes in, C, native out.
• So, libraries, benchmarks, optimizations need to
  be targeted for this environment
• One example Microsoft Parallel FX library
• Parallel looping functionality
• Task abstractions (including thread stealing,
  dynamic work assignment)
• C SXM Library (Software Transactional Memory)

5
Things to Think About

• New languages ? New opportunities?
• Increased use of closures, lambdas, late-bound
  query languages.
• Responsiveness good (snappy, ready to go)
• Opportunities for pre-computation, pooling of
  shared resources, pre-loading of critical
  data/code sections.
• Faster context switching, thread and process
  creation.