The quest for petascale computing

Abstract

Although the challenges to achieving petascale computing within the next decade are daunting, several software and hardware technologies are emerging that could help us reach this goal. The authors review these technologies and consider new algorithms capable of exploiting a petascale computer's architecture. One petaflop per second is a rate of computation corresponding to 10/sup 15/ floating-point operations per second. To be of use in scientific computing, a computer capable of this prodigious speed needs a main memory of tens or hundreds of terabytes and enormous amounts of mass storage. Sophisticated compilers and high memory and I/O bandwidth are also essential to exploit the architecture efficiently. To mask the hardware and software complexities from the scientific end user, it would be advantageous to access and use a petascale computer through an advanced problem-solving environment. Immersive visualization environments could play an important role in analyzing and navigating the output from petascale computations. Thus, petascale computing is capable of driving the next decade of research in high-performance computing and communications and will require advances across all aspects of it.