Rowan believes that despite the power of workstations, there are some problems that require the memory and input-output of mainframes and supercomputers. He says that IBM’s strategy is to have workstations on the desk overlying a compute server that handles data and databases, and the compute servers will be based on parallel or vector-based architectures, or a combination of both. He says that IBM’s approach to supercomputing is different from the classical model. Traditionally, there are both front- and back-end processors and applications are loaded from the front end. Results are sent from the back end to the front and then to workstations, scheduling is normally done on a batch basis, and co-operative processing is difficult. Rowan says it is an expensive process in terms of front end costs, disk, memory requirements and total run time. He claims that IBM’s vector-oriented approach eliminates the need for a front-end processor, there isn’t a dupli-cation of memory and disks, no need for job transfer, and co-operative processing is made easier. With improved vector facilities on the ES/9000 820 and 900 machines, Rowan says thare is a granularity in IBM’s technical computing strategy that is driven from the workstation and across the range, offers a 200-fold performance increase. IBM has just announced enhancements to its supercomputing family and they include extensions to clustered Fortran, three systems extensions that support high-speed real-time visualisation, a disk array subsystem and data conversion between ANSI IEEE and IBM hexadecimal floating point formats (CI Nos 1,555, 1,557). The enhanced clustered Fortran is a combination of hardware and software that enables the use of four ES/9000 or 3090 complexes – four, five or six processor models – representing up to 24 processors operating in parallel, and includes a 4Gb common object memory. The Supercomputer Visualization Enhancement is based on RISC technology and requires a 5081 or 6091 graphics monitor for imaging, and a High Performance Parallel Interface transfer image data directly. The High Performance Parallel Interface-attached Disk Array Subsystem is a large-capacity disk array that attaches to the processor via the Interface extension and is said to facilitate data access at 100Mb per second. The Floating-point Data Interchange Facility is a set of Fortran subroutines for the conversion of floating-point data between IBM’s hexadecimal format and ANSI IEEE standard binary formats, and it facilitates the interchange of application data between workstations and processors. Rowan says that the new extensions take supercomputing into the next generation, but at 9 GFLOPS, it is still a generation away from Steve Chen’s 100 GFLOPS machines. Does anyone need that speed and power? Rowan is unequivocal that there is a demand. He says that weather forecasters, oil companies and aerospace industries need the fastest machines available since simulation and modelling are increasingly important, and these customers are now demanding PetaFLOPS rather than GigaFLOPS or TeraFLOPS. – Janice McGinn
