The history of our message-passing system work at Caltech is interesting in that its motivation departs significantly from that of most other institutions. Since our original goals were problem-oriented rather than motivated by the desire to do parallel processing research, we tended to build utilities that matched our hardware and software goals rather than for our aesthetic sense. If our original machine had had multiplexed DMA channels and specialized routing hardware, we might have started off in a totally different direction. Indeed, this can be seen as motivation for developing some of the alternative systems described in the previous section.
In retrospect, we may have been lucky to have such limited hardware available, since it forced us to develop tools for the user rather than rely on an all-purpose communication system. The resultant decomposition and collective communication routines still provide the basis for most of our successful work-even with the development of Express, we still find that we return again and again to the nearest-neighbor, crystalline communication style, albeit using the portable Express implementation rather than the old rdELT and wtELT calls. Even as we attempt to develop automated mechanisms for constructing parallel code, we rely on this type of technology.
The advent of parallel UNIX variants has not solved the problems of message passing-indeed these systems are among the weakest in terms of providing user-level support for interprocessor communication. We continually find that the best performance, both from our parallel programs and the scientists who develop them, is obtained when working in a loosely synchronous programming environment such as Express, even when this means implementing such a system on top of a native, ``parallel UNIX.''
We believe that the work done by 
is still quite unique, at least
in its approach to problem solving. It is amusing to recall the comment of
one new visitor to Caltech who, coming from an institution building
sophisticated ``parallel UNIXs,'' was surprised to see the low level at
which CrOS III operated. From our point of view, however, it gets the job
done in an efficient and timely manner, which is of paramount importance.