Disk I/O in High-Performance Computing Systems

257 p.Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 1993.There exists an increasing disparity between the performance of high-performance computers and disk I/O systems. This divergence in performance signifies that the I/O system will become an increasing bottleneck in future computer systems. With such a bottleneck, a program's execution time on these systems could approach the associated file access time in the limiting case. We review the evolution of I/O architectures and characterize the access patterns of requests to those I/O systems. We apply this information to investigate the I/O performance of a multiple disk system in a multiprocessor system.Today's disk subsystems typically provide good I/O performance for only sequential access. Our study uses a simulated multiprocessor and multiple disk system architecture. I/O and memory traffic is combined in the system's packet-switched, multistage interconnection network. Our investigation first considers the effects of the I/O traffic on memory access in the multistage network. We then use a set of synthetic access patterns to evaluate the potential I/O performance. We consider the effectiveness of using custom mappings of file data to disks to support non-sequential I/O in this multiprocessor system. Not only do we investigate the more traditional workload of multiple processors accessing separate files, we also study the I/O characteristics of multiple processors accessing a single file.Our experiments show the I/O access pattern, the data placement on disks, the number of disks, and the I/O request rate, all interact to determine the performance of such a multiple disk I/O system. We found no single data placement on the multiple disks can provide good I/O performance for all workload cases. Similarly, a different set of data placement guidelines were identified for multiple processors accessing a single file than those identified for the more conventional access of multiple files.U of I OnlyRestricted to the U of I community idenfinitely during batch ingest of legacy ETD