Defense orders five large Linux clusters
- By Joab Jackson
- Feb 17, 2006
The Defense Department's High Performance Computing Modernization Program has purchased five computer clusters from Linux Networx of Salt Lake City.
The Army Research Laboratory's Major Shared Resource Center (ARL MSRC) will get four of the clusters, which the company will install this summer. The Army's Dugway Proving Ground, Utah, will also receive a cluster as part of the deal.
The ARL will offer use of these clusters to other military services for computing jobs too large to be done in-house, said Charles Nietubicz, director of the ARL's MSRC and director of the High Performance Computing Division in the Computational and Information Sciences Directorate of ARL.
MSRC is run under the Defense Department's High Performance Computing Modernization Program
, which purchased the cluster as part of its 2006 refresh. The program
purchases a small number of supercomputers each year for use departmentwide.
These clusters will tackle scientific computing jobs such as weapons analysis and design, ship hull design, penetration mechanics, aerodynamics of projectiles and the study of fluid dynamics. Such jobs can take anywhere from 15 minutes to two weeks to run.
'These are not spreadsheets, PowerPoint charts or things we can do on a fast personal computer,' Nietubicz said.
Three of the ARL clusters are what Linux Networx brands Advanced Technology Clusters. One will have 1,234 four-processor nodes, 9.4 T of working memory and should provide around 29 trillion floating-point operations (TFLOPs) per second. The second cluster will have 866 four-processor nodes and 7.1 T of working memory, promising 21.5 TFLOPs of computing power. A third 68-node cluster will serve as a test environment for breaking in new software.
All three clusters will use 20-Gbps Infiniband network interconnects and thermally efficient 3.2 GHz Intel Dempsey chips for the core processing nodes.
The large clusters are necessary for grappling with ever-larger data sets, said Tom Kendall, chief engineer of ARL MSRC. Some of the larger jobs the MSRC is tasked with, such as structural mechanics or fluid dynamics, create so much data that researchers can no longer analyze or graphically render the results on a single, separate workstation. So the supercomputer must execute these additional tasks in addition to its data crunching chores.
'The compute clusters are creating so much data that is no longer possible for a post-processing machine to support the graphical analysis,' Kendall said. 'So the whole process from the computation down to the graphical analysis has to be done using multiple processors.'
For smaller visualization tasks, Linux Networx will also deliver to ARL a smaller, ready-to-run, 64-node cluster marketed under the name LS-1 Supersystem. This system will have two Quadro FX 4500 graphics cards for each node from Nvidia Corp. of Santa Clara, Calif. Dugway is also getting the LS-1 model.
The HPCMP tries to procure systems of differing architectures for maximum variance in computing abilities. Clusters are a relatively recent addition to the high-end computing portfolio. A cluster is a collection of interconnected servers running on low-cost commodity processors. With parallel-processing algorithms, a cluster of multiple servers can work in conjunction on a single task.
Nietubicz said the ARL started working with clusters in 2000 with a 16-processor system. It has since procured a series of systems in increasingly larger configurations. Built with standard commodity processors, clusters strike a good balance between performance and cost, Nietubicz said.
'We found out that, even though some people were fearful of those machines, we could get some pretty good scaling,' Nietubicz said.
HPCMP also has just issued
its 2007 call for proposals.
Joab Jackson is the senior technology editor for Government Computer News.