Parallel computing currently offers the best price/performance ratio, and this advantage will increase in the future. Along with improved systems software, language standards, improved compilers, a new generation of computer users educated in parallel computing, and the entrance of large computer vendors such as Digital, Hewlett-Packard and IBM into the field, parallel computing will in the next few years generate enough ``headroom'' to become the dominant technology.
Much progress is being made on high-level parallel languages such as
High Performance Fortran, which will
accelerate the transition to parallel computing.
HPF can express the majority (perhaps 
) of computational science
problems, including many irregular problems. More than half of applications
are synchronous or embarrassingly parallel, and these problems should be
mapped very efficiently to different architectures by an HPF compiler.
The challenge for compiler writers will be to generate efficient code
for irregular problems, which may be difficult or impossible for vector or
SIMD architectures.
Even for MIMD architectures, the expression of certain types of irregular
loosely synchronous problems in HPF, and the generation of efficient MIMD
code by an HPF compiler, are still the subject of much research.
However if this cannot be done using standard HPF,
there is an escape route whereby sections of the code may be called as
EXTRINSIC functions, which may be implemented more efficiently using
explicit message passing.
We also have a challenge for computational scientists: is your application expressible in HPF? If not, we would like to hear from you!
One final point to note is that scientific applications are certainly important, but are only a limited market for a computer vendor. Most commercial uses of computers involve information processing, decision support, economic (and other complex system) modeling, network simulation, scheduling, manufacturing, education and entertainment. [32] Parallel computing needs to make an impact in these areas, since it will become the dominant technology if and only if it can benefit industry. Industrial applications typically have much larger codes than scientific applications (sometimes on the order of millions of lines), which will only be ported to parallel computers when improved, portable and maintainable software, such as HPF, becomes available.