The authors’ stated aim is twofold: first, to understand the behavior of large-scale, data intensive applications and the impact of I/O subsystems on their performance, and second, to improve the performance of such applications.
The five real-world I/O-intensive applications are as follows:
| Table 1: Real-world I/O-intensive Applications |
| Name | Purpose | Size | I/O | Platform |
| SCF1.1 | self-consistent field computation | 16,500 lines | writes integrals to disk and reads them | Intel Paragon |
| SCF3.0 | self-consistent field computation | 19,000 lines | writes integrals to disk and reads them | Intel Paragon |
| FFT | 2D out-of-core fast Fourier transform | 500 lines | reads and writes two matrices | Intel Paragon |
| BTIO | simulates flow solver I/O | 6,713 lines | periodic array writes | IBM SP-2 |
| AST | simulates gravitational collapses | 17,000 lines | writes arrays for check-pointing and visualization | Intel Paragon |
The I/O optimizations considered were as follows (applications considered are in parentheses):
- collective I/O (BTIO, AST)
- prefetching (SCF1.1)
- file layout optimization (FFT)
- efficient I/O interface (SCF1.1)
- balanced I/O (SCF3.0)
The authors found that although the different applications benefited most from different optimization(s), they were nevertheless able to formulate the following rough general guideline (sequence): first, optimize the global I/O pattern to force processors to make as many accesses to consecutive locations as possible; then, improve individual location access patterns further, using techniques such as file layout transformations and prefetching.
The authors propose further research on increasing the size of both application suite and optimization techniques, as well as more investigation of the interactions between the different optimizations, in order to produce an appropriate ranking order.
