ABSTRACT
The mid-to-late 1990s have witnessed a dramatic speed-up in the performance of parallel supercomputers. The 1990s started with high performance computing (HPC) based mainly on vector supercomputing hardware offering about one gigaflop of arithmetic computing power (about 10 times more than a decade earlier). Parallel computers existed but were based on many different architectures. They were often not a practical proposition from a reliability perspective; the software was lagging far behind the hardware and most of the programs written in that period will now be either defunct (like the hardware) or else have been rewritten in one of the newly-emerged standard parallel programming languages. Yet by the end of the 1990s it is quite clear that the parallel supercomputing era is here. The nineties decade is ending with the rapid disappearance of HPC based on vector processors and the emergence of mature, reliable, affordable and highly efficient parallel supercomputers offering between 150 and 1000 times more computational power. The early years of the new millennium are confidently expected to witness the appearance of multiteraflop machines that will slowly percolate into the scientific infrastructures accessible to GeoComputation. For those applications that require unbelievably vast amounts of computational power then HPC hardware offers a way forward that is now within the reach of the skill base of many geographers and social scientists. HPC is thus a key tool for those researchers who need to perform large-scale analysis of big datasets, to create numerical experiments of complex processes, or to construct computer models of many physical and human systems. For example, in order to model the largest public domain ‘flow’ dataset that is available in the UK, the journey to work flows from the 1991 census for all wards in Britain, there is a requirement to process a table comprising 10764 columns by 10764 rows. The biggest flow dataset (i.e. telephone flows between houses) that may well exist could well have between 1.7 and 28 million rows and columns. These data can be disaggregated in many different ways (e.g. by time or age or sex or purpose), creating dataset sizes of gigabytes maybe even terabytes. This complements the flood of geo-information being generated by remote sensing devices. It is quite clear that data mining and modelling of these,
and other even larger datasets, can (or could) now be undertaken, and the results analysed and mapped by applying HPC technologies. HPC, however, also has the potential to open up entirely new areas for research.
