Morning Session: Petaflop challenges

From ALPS

• Speaker&Chair: Thomas A. Maier "Petscale Computing"

• Discussions on talk
  • Ten years from now: will technical aspects of petaflop computing be standardized?
    • High diversity of algorithms/methods prevents standardization
    • Large teams from different fields are required and actually growing over time
    • Speedup of machines comes at a price of more complicated machines
    • Even now, some applications are not run at teraflop scale
  • Is there always a way to do in-situ data analysis?
    • Filesystems scales to a few hundred nodes at most, even with HDF5
    • Use small runs to explore physics, large-scale runs for big systems when questions are clear
  • Is any method besides petaflop currently living on teraflop?
    • Series expansion: possible
    • Exact diagonalization: has been done
      • Depends on Hamiltonian and geometry
      • Use of symmetries
      • Often O(N^2) communication tasks
    • Tensor network states
      • Terascale possible, petaflop still open
      • Difficult linear algebra operations
    • DMRG
      • Many small matrices, high impact of communication latency
      • Shared-memory very easy & efficient: linear scaling
      • Huge matrix-matrix multiply
      • Using GPUs can help a lot
  • Which fields get most FLOPs at Oakridge?
    • Material science, BEC
    • Industry projects: BOEING
    • Climate