OASCR Usage of High-End Systems

1
OASCR Usage of High-End Systems

• Science computational fluid dynamics, e.g.
  combustion, astrophysics. Adaptive
  finite-difference, finite-element, particle
  methods for time-dependent problems.
• Mathematical algorithm development fundamental
  numerical algorithm design, adaptive methods,
  multiscale / multiphysics modeling.
• Mathematical library development components,
  frameworks, solvers (SciDAC Math ISICs).
• Systems software research and development
  languages / compilers communications, I/O
  libraries performance tools frameworks and
  environments.

2
Hardware Requirements

• More interactive access to parallel systems.
• Mathematics algorithm development, CFD science
  needs faster turnaround on moderate-sized jobs,
  interactive data analysis, visualization,
  debugging.
• Mathematics library developers interactive
  debugging (correctness and performance).
• Systems software developers more
  experimentation, control over environment.
• Possible solution moderate-sized (128-256
  processor) machine, but otherwise identical to
  larger production system, dedicated to
  interactive use (dont allocate for or expect gt
  90 utilization).

3
Hardware Requirements (contd)

• More bandwidth, lower latency (both
  processor-to-memory and network).
• Fast hardware for compilation.
• Lower peak-to-average performance ratio for
  problems not dominated by DGEMM. Same for
  vector-to-average-scalar performance.
• Need enough disk, INODE space (15K INODES is not
  enough for library developers).
• Inclusion of C, adaptive codes in NERSC
  procurement benchmark suite. UPC requirement ?
• Existence of multiple sites requires
  high-performance between them.

4
Software / User Environment

• Vendor compilers must be standards-compliant,
  with greater emphasis on quality control -
  particularly a problem with C. Bring back
  KAI-style aggressive inlining of small objects.
• Support of de-facto standards in I/O (HDF5,
  NetCDF).
• Debugging tools need to be convenient to use for
  100s of processes / threads.
• CPU performance tools support complete
  implementation of PAPI, reliable implementation
  of gprof. Need higher-level tools that scale to
  large C applications (1000s of functions).

5
Software / User Environment

• Network performance analysis tools Quadrics,
  Myrinet, good Lappi bad. Need higher-level
  tools that scale to large number of processes,
  large amounts of message traffic.
• Visualization / data analysis are closely coupled
  problems. Need supported tools and expertise,
  interactive access to high-end platforms, fast
  access to large datasets. Dont need specialized
  graphics hardware.
• More transparency in the batch system. CFD
  research requires long runs.