Supercomputing Books

Easy to read overview of the field of supercomputing. Describes technologies (eg, relays to highly integrated circuits) and architectures (eg, scalar to vector, multiprocessor and massively parallel) used to effect supercomputing. Describes applications of supercomputing in physics, in constructing evolutionary trees from sequencing data, in molecular biology, engineering, planetary sciences, weather forecasting, and cosmology.

In the US. the NRC is probably the best qualified to assess where supercomputing is going. The book offers a vendor neutral survey. One that is also largely country neutral. The latest status of various hardware efforts is described. For vector computing, massive clustering, and other approaches.

The latter also includes the ongoing competition between building a supercomputer out of common off the shelf components (COTS), or via designing dedicated hardware. You might wonder at the continued viability of the latter, as design limits keep getting pushed back. Surely a point of diminishing returns will be reached?

A recent grouping of 15 articles published in Scientific American, from 1995-2001. Though computing is still chugging along at Moore's Law's rate, these articles are still quite germane. Written at a level for a broad, educated audience, they give you a good understanding of the various key points in supercomputing.

Issues include how to cluster computers into a functional unit, that can run these all efficiently (load balance). Or, how to make novel arrangements of computers into a supercomputer. Plus, of course, the steady progress of vector supercomputers, which has been the traditional route since Cray pioneered these in the 1970s.

The articles are not just about how to make better supercomputers. They also discuss why you'd want to do so. The problems you can now tackle with better hardware. Meteorology being the classic (unclassified) application.

Reconfigurable computing is a hot topic again, twenty years after the research that went into the book, but means something completely different. Although it has little to contribute to practice in the early 21st century, it offers worthwhile perspective on the field.

The Kartashevs describe a major multiprocessor. It contains some number of coarse-grained function units, each of which can act as an independent 16-bit processor, or which can be ganged to create larger word widths - apparently, an extreme of "bit slice" architecture. Reconfiguration at the next level up means changing the interprocessor network that exchanges results between units. Given that premise, the Kartashevs go on to develop an entire world of compilation techniques, scheduling and conflict resolution algorithms, fault tolerance schemes, and everything else needed to create a supercomputing system from the ground up.

History has not gone in their direction. Off-chip communication became increasingly expensive, making it more attractive to build increasingly capable chips rather than increasingly capable ensembles of chips. The performance bottleneck has generally moved from the CPU and communication systems, which the Kartashevs address, to the memory hierarchy. Interprocessor networks in supercomputers have grown too large for the Kartashev's fine-grained reworking of topology. Application developers could, for adequate reason, give up portability in interprocessor communication to take advantage of the connectivity options, but not for the modest performance gains cited here. Too, the reconfigurable fabric has changed. As chip development costs go through the roof, commodity chips become economically critical. Today, that means FPGAs, and FPGAs mean an enormously finer grain, range of on-chip connectivity options, and degree of control than the Kartashevs imagined. Scheduling and resource sharing have also been recast in terms compatible Unix device handling. And, perhaps most importantly, the driving reasons for reconfigurable computing have changed. Reagan's "Star Wars" defense initiative is long gone; the Soviet Union dissolved two years after this book was published, giving it little to defend against.

Good research remans good, no matter how the world changes around it. A careful reading of some system problems, especially in code analysis and resource scheduling, could offer ideas to today's worker reconfigurable processing. Although this book captures a striking achievement in system design, that system doesn't match technology at the time of this writing.

-- wiredweird

Five stars for thoroughness, but one for relevance to contemporary systems.