Laboratory for Advanced Numerical Software
RADIX: Laboratory for Scalable, Parallel Programming Tools
Projects
HIGH-PERFORMANCE I/O AND DATA MANAGEMENT
High-performance I/O has not received the same degree of attention
that high-performance computing has, even though many applications are
performance-limited by the rate at which they can perform I/O (either
saving computations or accessing previous calculations). In addition,
managing this data (on the scale of petabytes of complex data, where
each logical record represents gigabytes of data) remains a major
challenge. These unique needs of DOE are not being met by the
commercial sector, where I/O work is focused on POSIX (or Windows/NT)
compatibility (at the expense of factors of 100 in performance) or on
database applications. Our specific purpose, then, is
to address the obstacles of getting data into and out of teraflops
systems fast enough to avoid severe bottlenecks and
to manage large amounts of data.
PIs:
Rajeev Thakur,
Robert Ross,
William Gropp,
Ewing Lusk
|
ROMIO—a high-performance, portable
implementation that allows MPI-IO to be used on
almost all parallel machines and networks of workstations.
I/O in large Linux clusters—redesign
of PVFS
to enable use of communication mechanisms
(a collaboration with Clemson University)
Java I/O—investigation of I/O issues
associated with development of fast and scalable Web
servers (a collaboratin with the Illinois Institute of
Technology)
|
| TOOLS FOR MASSIVELY PARALLEL PROGRAMMING
The purpose of this project is to
develop the technologies required to dramatically increase the productivity of
scientists developing applications for parallel supercomputers
The focus is fourfold:
(1) integration of portable parallel programming tools,
(2) development of scientific computing tool kits and portable libraries,
(3) research into performace analysis tools for parallel programs
(4) exploration of requirements of future parallel computers.
Our current efforts include
parallel libraries such as MPICH and
performance visualization.
MPICH—MPICH is a portable, high-performance implementation of
the Message Passing Interface.
As part of
the MPICH project, we are exploring tradeoffs between portability
and performance in the context of the MPI Standard
and are investigating interfaces between MPI and tools.
The MPI project also provides a testbed for other research groups working on particular aspects of message passing.
Performance visualization is a
powerful technique for understanding the behavior and performance of parallel programs through the use of trace files (log files).
|
MPD (multipurpose daemon)—a
prototype process manager for fast startup of MPICH programs.
MPD currently provides a runtime
framework that includes management of standard I/O and a simple form parallel
debugging based on existing open source software.
PIs:
W. Gropp,
E. Lusk
ADI-3—a
redesign of the internals of MPICH to
with the objective of making MPICH a significant
component of open source software systems for parallel computers used in the
scientific community
PIs:
W. Gropp,
E. Lusk
MPICH-G2—an
implementation of MPICH integrated with the Globus
metacomputing system, to allow MPI programs to be
started on heterogeneous collections of parallel computers
PIs:
B. Toonen,
N. Karonis
|
|