I/O for TFLOPS Supercomputers

Scalable parallel computers with TFLOPS (Trillion FLoating Point Operations Per Second) performance levels are now under construction. While we believe TFLOPS processor technology is sound, we believe the software and I/O systems surrounding them need improvement. This paper describes our view of a proper system that we built for the nCUBE parallel computer and which is now commercially available. The distinguishing feature of our system is that scalable parallelism is implicit rather than explicit. We did not base our system on new commands, system calls, or languages. Instead, we extended some aspects of Unix® to add parallelism while keeping these aspects unchanged for nonparallel programs. The result is a system that lets one use a future TFLOPS parallel computer without knowing parallel programming. As parallel versions of standard compilers arrive, and large data sets get distributed over multiple I/O devices, then standard Unix commands will run arbitrary mixtures of parallel ...

• Type: Conference
• Year: 1993
• Where: PPSC
• Authors: Erik DeBenedictis, Stephen C. Johnson
• Keywords: Floating Point Operations Per Second, Parallel Computer, PPSC 1993, PPSC 2001, Scalable Parallel Computers