Introduction

This paper will survey a selection of issues dealing with software and the user interface, always in the context of parallel computing. The most obvious feature in looking back over the last five to 10 years of parallel computing is that it has been clearly demonstrated that parallel machines can be built. This was a significant issue eight or 10 years ago, particularly in the light of some early experiences. There was a feeling that the systems would be too unreliable, with mean time to failure measured in minutes and with intractable cooling and packaging issues. There are now several examples of real, highly parallel systems, such as Thinking Machine Corporation's Connection Machine CM-2 and the Intel iPSC/860, which are certainly recognized as serious computers and are among the leading supercomputers at present.

It has also been shown that model problems can be solved efficiently on these systems. For example, linear algebra problems, partial differential equation (PDE) solvers, and simple applications such as quantum

chromodynamics have been modeled effectively. In a few cases, more complex models such as oil reservoir simulations and weather models have been parallelized.

Although hardware progress has been dramatic, system software progress has been painfully slow, with a few isolated exceptions that I will highlight in the following section. Another area in which there has been almost no progress is in demonstrating that parallel computers can support general-purpose application environments. Most of what we will present here will be motivated to some extent by these failures.

We could begin by asking what a user would like from a general-purpose supercomputer application environment. In simplest terms, he would like to see a computer consisting of a processor that is as powerful as he wants, a data memory as large as his needs, and massive connectivity with as much bandwidth (internal and external) as desired. Basically, he would like the supercomputer to look like his desktop workstation. This also corresponds to the way he would like to program the system.

It is well known that there are various physical limitations, such as the finite speed of light and cooling issues, that prevent us from designing such a system. As a way around, one has been led to parallelism that replicates processors, memories, and data paths to achieve comparable power and throughput. To complicate the situation further, there are many different ways to actually effect these different connections, as well as several distinctly different choices for control of such systems. Thus, we end up with the current complexity of literally dozens of different topologies and connection strategies.

Yet the user wants to think of this whole system as a monolithic processor if possible. We will focus on several software areas where the user obviously interacts with the system and discuss ways in which the software can help with this issue of focusing on the machine as a single entity.

At the lowest level we will discuss compilers, not from the point of view of producing good machine code for the individual processors but rather from the higher-level aspect of how the compiler can help with the unifying aspect for parallel machines. Then there are the all-important debugging, trace, and simulation phases. People actually spend most of their time developing programs rather than running them, and if this phase is not efficient for the user, the system is likely ultimately to be a failure.

We will briefly discuss several related developments: higher-level languages, language extensions, and libraries. Portable parallel libraries provide a key way in which the user's interaction with systems can be simplified. Graphics and the visualization pipeline are, of course, critical

areas about which we will make several comments. For each of these topics, we will refer to other papers in these proceedings for more coverage.

Finally, we will discuss the development of software for heterogeneous environments, which is in many ways the most important software issue. No one parallel machine is going to turn out to provide the general-purpose computer that the user would really like to have. There will be classes of parallel machines, each well suited to a range of problems but unable to provide a solution environment for the full range of applications. Heterogeneous systems will allow complex algorithms to use appropriate and optimal resources as needed. So ultimately we will be building heterogeneous environments, and the software for those systems is perhaps the greatest challenge in user-interface design in the near future.