Requirements of Programming Models

1
Requirements of Programming Models

• Requirements never change
• Programmability
• Performance
• Satisfying these requirements, of course is a
  function of the underlying platforms
• Hardware platforms, of course, change

2
Challenges from Emerging Platforms

• Scenario 1
• Emerging platforms are no different
• Multicore SMP
• Solution program using MPI/posix
• Scenario 2
• Emerging platforms pose significant challenges
• Scalable multicore cache fragmentation
• Heterogeneous multicore FPU/GPU/FPGAs
• Integrated hardware services
• Network is closer than the memory

3
Scenario2 Unplugged

• How do we achieve programmability and
  performance?
• OpenMP (unless heavily retooled) is certainly not
  the answer
• Performance today results from messaging/posix/CUD
  A hardly a high-productivity environment
• Cache fragmentation (cooperative caching) is an
  algorithmic challenge (with expressiveness and
  support from the model)

4
Emerging Applications

• Upscaling of desktop applications
• Gaming, virtual/synthetic environments
• Enterprise applications
• HPC applications

5
Upscaling Desktop Applications

• Multimedia Becomes a Major Application
• Multiple streams of HD media can challenge most
  systems
• Emphasis here is not simply on the CPU, but on
  management of other resources as well (downstream
  from the memory subsystem)
• Most programming models have limited support for
  such applications (I/O, streaming).

6
Upscaling Desktop Applications

• Gaming and Distributed Gaming
• Tighter integration of graphics hardware,
  networking, and processing
• OpenMP/CUDA is hardly the solution here

7
Servers/Data Centers

• Immediate future sees large scale media servers
• The resources that underly such infrastructure
  are vastly different from those OpenMP is
  targeted to.

8
HPC Applications

• Conventional HPC applications scale down to the
  desktop
• In my opinion, this is a bigger trend than
  pushing the envelope at the top end (because of a
  number of modeling/ analytical issues)
• In the short term, this leads to a cottage
  industry of hand-tuned codes for FPGAs, GPUs,
  heterogeneous multicores.
• In the longer term, application development in
  these environments is entirely unsupported

9
HPC Applications

• Tighter integration of networks is certainly a
  major benefit for data-intensive applications
• Is there a seamless way to integrate services
  deeper into the language?

10
Is there need for anything beyond the very
familiar?

• Very much so Posix and messaging only take you
  so far
• In the short term, much of the development will
  continue to be in Posix/xxx
• In the longer term, we do need something a lot
  more friendly
• In my opinion, OpenMP, in its current form, is
  certainly not the answer, neither is
  posix/MPI/CUDA/....