Introduction (Chapter 1)

[bollig]

• Need motivation for research
• Background of the problem/study
  • High level; will get into detail in related work section
• Statement of the problem
• High level research questions
  • Possible to prepare for the transition to exa-scale computing? What is necessary?
  • What algorithms or data-structures would be necessary?
• Importance of the study (why should the reader care)?
• Scope of the study?
  • What is not within the scope?
  • we don't focus entirely on the problem of optimizing GPU kernels as our problem leverages higher level matrix algebra primitives (spmv, saxpy, gemm)
  • we don't worry about domain partitioning (leave it to METIS), but we do worry about load balancing communication (overlap comm and comp).
  • Do worry about node ordering
  • Do worry about node indexing
  • Do worry about explicit/implicit solvers
  • Do worry about applications
• Definition of Terms
  • node
  • stencil
  • center
  • weight
  • rbf
  • gpu
  • partition
• Limitations
  • Limited GPU kernels (solver only takes advantage of the GPU)
  • Balanced MPI, OpenCL, GPU Library, and Application studies.
  • Not as advanced as PETsc, Deal.ii, Trilinos, etc. which all have 10+ years of development