The communication and load-balancing algorithms were implemented on
Caltech/JPL hypercube Mark II. The need for interactive GUIs emerged for
the first time during our work on parallel implementation of a
neurophysiological model for olfactory cortex
[Furmanski:87a] and, in the next, step, in the machine vision
research [Furmanski:88c]. At that time (1988), we were using the
nCUBE1 system at C
P and also the ``personal hypercube'' system
based on IBM AT under XENIX with the 4-node nCUBE1 add-on board. The
graphics support in the latter environment was nonexistent and we
constructed from scratch the GUI system based on the interpreted
language g [Furmanski:88a], custom designed and coupled with the
regular parallel computing software components. The language g was
80286 assembly-coded and XENIX kernel-based and hence very fast.
However, it couldn't be ported anywhere beyond this environment, which
clearly became obsolete before the g-based implementation work was
even fully completed. Some design concepts and implementation
techniques from this first experiment survived and are now part of the
MOVIE System, but the major lesson learned was that GUIs for HPC must
be based on portable graphics standards rather than on custom-made or
vendor-specific models.
It was at about the same time (late 1980s) that the major multivendor effort started towards standardizing computer graphics in the UNIX environment. We were participating actively in this process, experimenting with subsequent models such as SunView, X10, X11, NeWS, XView, OpenLook, Motif, DPS, and finally PHIGS/PEX/GL. It was very difficult to build a stable graphics-intensive system for HPC during the period of the last four years, when the standardization efforts were competing with the vendor-specific customization efforts. However, certain generic concepts and required features of such a system gradually started to emerge in the course of my experiments with subsequent standard candidates.
For example, it became clear, with the onset of network-extensible server-based graphics models such as X or NeWS, that any solid HPC environment must include the distributing computing model as well and to unify it with the SIMD- and MIMD-parallel models. Also, to cope with portability issues in the emerging heterogeneous HPC environments, such a system must include appropriate high-level software abstraction layers, supporting Virtual Machine-based techniques. A network of compute servers, tightly/loosely coupled in MIMD-parallel/distributed mode, with each server following the high-level Virtual Machine design, appeared to be the natural overall software architecture. Modern software techniques such as preemptive multithreading and object orientation are required to assure appropriate dynamics and functionality of such a server in diverse tasks involving graphics, computation, and communication. Among the emerging standards, the design closest to the above specification was offered by the NeWS (Network-extensible Window System) [Gosling:89a] server, developed by Sun. Following NeWS ideas, we adopted PostScript [Adobe:87a] syntax for the server language design, extending it appropriately to support object-oriented techniques and enhancing substantially its functionality towards the HPC domain.
The resulting system was called MOVIE, due both to its adequate acronym and its stress on the relevance of interactive graphics in the system design. The server language, integrating computation, graphics, and communication, was called MovieScript. The first implementation of the MOVIE Server was done at Caltech [Furmanski:89d] on a Sun workstation and then the system was ported to the DEC environment at Syracuse University [Furmanski:90b]. Here, we learned about virtual reality [Furmanski:91g] which we now consider the ultimate GUI model for MOVIE, ``closing'' the system design. In the fall of 1991, the MOVIE group was formed and the ``individual researcher'' period of the system development is now followed by the team development period [Furmanski:93a].