High-performance scientific computations have become more and more I/O bound.
because of the increasing mismatch between I/O and CPU speeds;
this presents a bottleneck to scientific computation.
For many scientific applications, data decompositions and their corresponding
mappings have been used to increase I/O scalability
in MPP systems.
Most of those mappings have the power-of-2 limitation - the block size, the
number of processors and the number of disks are all assumed to be powers of 2.
As a result, users are forced to limit the numbers of processors and disks
participating in the execution of applications and thus achieve only limited
performance.
In an effort to alleviate this limitation and to improve the overall parallel
I/O performance,
we devise a scheme, called Resident-Guest distribution, and formulate its
mapping function.
We also design a high bandwidth file system for a flexibly configurable parallel
architecture implementing our proposed distribution strategy.
A congestion study for bandwidth intensive applications was done on an Intel
Paragon to help in the design of the file system.
We present the concepts, the methods and the issues in designing our
distribution strategy and the file system together
with the experimental results of the Resident-Guest mapping on an nCUBE and the
simulation data for the proposed file system.