Convex Computer Corp and Hewlett-Packard Co yesterday said that they are hotting up the alliance that saw Hewlett take a 5% stake in Convex last year with a potentially substantial technology exchange to plug the massively parallel applications gap. Convex is building a massively parallel system out of Hewlett’s Precision Architecture RISCs, and under yesterday’s agreement, will swap its parallelising compilers for Hewlett’s HP-UX implementation of Unix so that it should be a relatively straightforward task to adapt the 4,000 HP-UX RISC applications to run on the planned Convex machines. In the context of the agreement, analysts were quoted as suggesting that acquiring a software base for a new supercomputer has traditionally been the biggest drag on sales, and doubting that Convex and Hewlett had cracked it – but up to now there has been no massively parallel machine that is both hardware- and software-compatible with a big existing applications base. Convex’s plans call for a massively parallel machine that will be scalable to TFLOPS performance, in an air-cooled, stand-alone, general purpose system. The company is aiming for three generations of products over the next few years. The first generation will be scalable to 128 processors with up to 25 GFLOPS performance, expected to be available in first half 1994, and the firm aims to improve performance by at least an order of magnitude per generation. The operating system will consist of three components: HP-UX; the ConvexOS environment; and an implementation of the Open Software Foundation’s I/AD Mach microkernel developed specifically for massively parallel systems. The machine will ironically be more compatible with Hewlett-Packard applications than Convex C Series ones, but it will be source code-compatible with the latter – an applications base of 1,300. The Convex Application Compiler will have automatic interprocedural analysis capabilities to provide a high degree of optimisation from existing applications. Critical applications will be able to be further optimised with the advanced parallel environment to be provided with the planned scalable parallel processing systems. The multiple instruction-multiple data scalable parallel processing system will support a shared memory programming model called global shared distributed virtual memory to provide a Fortran and C programming model very similar to current production supercomputers. The system will also support message passing models such as parallel virtual machine and the high performance Fortran extensions of Fortran 90. In September 1992, Convex founded and sponsored the Scalable Computing Working Group of scientists and engineers from industry, academia, and research: it meets quarterly to advance scalable algorithm and application development, and currently has some 40 members. In addition to HP-UX, Hewlett will provide a suite of standards-based libraries, commands, and utilities including Motif, X Window System, software distribution and system administration utilities. Convex says it is also enhancing the Mach 3.0 microkernal with faster input-output capabilities and parallel processing extensions in a scalable architecture.
