Report for Scalable I/O, July 1995
Contact: EOSDIS
Project. We have also developed a suite of programs for our NSF/ARPA
Grand Challenge project on Land Cover Dynamics as
well as an out-of-core sparse factorization code. The Grand Challenge
programs are undergoing a process of optimization and refinement.
The applications effort is also supported by funds from EOSDIS and
the Land Cover Dynamics NSF/ARPA Grand Challenge project.
Library Development:
The University of Maryland has developed two different I/O libraries.
The first library is a two-phase I/O library oriented towards loosely
synchronous programs whereas the second library is simpler and is
designed to handle coarse-grain independent requests. The two-phase
I/O library was completed approximately 8 months ago; this library has
been extensively tested on the SP-2.
As we analyzed the EOSDIS sensor data analysis programs and developed
our own programs for land cover dynamics, we came to the conclusion
that two-phase I/O was not necessarily the best approach to optimizing
parallel I/O performance. This prompted us to develop our second
library which appears to be better suited to this class of problems.
We have implemented this library on the 16 processor IBM SP-2 at the
University of Maryland. Multi-threading allows both the application
and the library to use the high performance switch on the SP-2 in high
bandwidth user-space mode. In joint work between Maryland and Argonne,
we are making use of both Maryland I/O libraries, along with Argonne's
Nexus software to find the most effective strategies for optimizing
parallel I/O performance for programs that exhibit various classes of
I/O behavior.
Compilers:
We have developed an Interprocedural Partial Redundancy Elimination
(IPRE) algorithm that will be used to carry out optimized placement of
asynchronous I/O routines. The algorithm will be presented in a paper
that will be presented at Supercomputing '95.
1996 Plans:
Libraries:
We plan to carry out a comprehensive performance
study of the effectiveness of various parallel I/O optimizations on a
broad range of application codes. This study will lead to a the
Compiler and Runtime group agreeing on a parallel I/O library API, and
to reference I/O library implementations.
Compilers:
The Interprocedural Partial Redundancy
Elimination (IPRE) algorithm has been successfully implemented as part
of the Fortran D System framework but IPRE has not yet been used for
parallel I/O optimization (IPRE has been used to place collective
communication calls). We will use the Interprocedural Partial
Redundancy Elimination (IPRE) algorithm to carry out optimized
placement of asynchronous I/O calls.
Applications:
At Supercomputing '95, we will use the IWAY
to demonstrate our suite of sensor data query, analysis and image
processing modules. The sensor database, the database query
infrastructure and the image processing routines will reside on our
SP-2 at the University of Maryland. We will develop a suite of sensor
data analysis based parallel I/O templates; these templates will be
made available to the parallel I/O consortium and to the community at
large.
Technology Transfer:
We are working closely with EOSDIS;
we are using EOSDIS applications to motivate our scalable I/O work
and we are making use of our Scalable I/O libraries in porting and
benchmarking EOSDIS codes.