Quick Search:    advanced search

GSW Home   GeoRef Home   My GSW Alerts   Contact GSW   About GSW   Journals List   Help

This Article Figures Only Full Text Full Text (PDF) Submit a response Alert me when this article is cited Alert me when eLetters are posted Alert me if a correction is posted Services Email this article to a friend Similar articles in this journal Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via Google Scholar Google Scholar Articles by Klepacki, D. J. Search for Related Content

High performance computing at IBM

The Bluegene/L supercomputer

IBM T. J. Watson Research Center, Yorktown Heights, N.Y., U.S.

Corresponding author: klepacki@us.ibm.com

The first 20% of the full text of this article appears below.

The current approach to handling very large compute-intensive and data-intensive applications is to deploy the cumulative effect of a large number of processors in parallel, either via a networked cluster of computers or via a more tightly coupled massively parallel system. As simulation models evolve in complexity together with the accelerated proliferation of modeling and imaging data, this approach to computing quickly becomes limited by issues of power consumption, cooling, and physical size, not to mention cost and performance. The challenge is to construct a dense air-cooled machine with high performance and low cost.

One resolution is to adopt the concept of cellular computer architecture. Such architectures integrate processor, memory, and interprocessor communication fabric into a single computational "cell" that can be replicated, in principle indefinitely. In other words, these cells can be used as building blocks to construct arbitrarily large computing systems, without the need for a network switch. Furthermore, system-on-a-chip techniques can be used with modest processor designs to fabricate the cell onto a single ASIC. This results in overall reduction in power, cooling, and physical size, while retaining high system performance in a scalable manner. The use of commodity technology with high volume yields keeps the costs low. The trade-off for these advantages is that the processor connectivity cannot be "flat." That is, a node cannot be equidistant to all other nodes.

Historically, this cellular methodology evolved from efforts to build special-purpose machines. In particular, the QCDSP machine, an effort led by Columbia University, was constructed to optimize quantum chromodynamic calculations. It ran with a peak aggregate speed of 600 GF, and won the Gordon Bell Prize in 1998 for the most cost-effective computer. (IBM is currently working with Columbia University on a successor to this machine, known as QCDOC, which is based on nodes using IBM's . . . [Full Text of this Article]