GAMING

BALTIMORE, MD - DataDirect Networks, an established leader in high performance, high scalability storage networking, is demonstrating its SwiftCluster (tm) clustering framework at SC 2002. SwiftCluster is a technological breakthrough which combines DataDirect's S2A8000 parallel block level storage controller, its scalable serial PCI I/O node interface, Direct Memory Access optimized protocol layer, all tying into a highly efficient parallel file system, to create a blueprint for the future of clustering. With SwiftCluster, customers will be able to build very efficient computer clusters scaling to tens of teraflops, while benefiting from Object Based Disk (OBD) file system technology and DataDirect's proven S2A8000, the fastest storage controller on the market today. SwiftCluster is an extremely compelling solution due to its linear performance scalability (to levels which are expected to exceed 10GB/second in aggregate bandwidth), and its highly cost-effective model. With SwiftCluster, the majority of the traditionally expensive CPUs, disks and disk enclosures become off the shelf components. SwiftCluster relies today on the open source Linux Lustre file system (an additional advantage for highly scalable implementations), but also plans to support other operating systems such as Windows and Unix for broader heterogeneous market appeal. "As a company, we have been thinking long and hard about the challenges of building a large scale massively scalable cluster, which would not only be extremely fast, but also very cost-effective," Alex Bouzari, chief executive officer, DataDirect Networks, said." SwiftCluster was developed over several years with significant input from some of the world's largest supercomputer customers, and we believe that it will enable the deployment of high performance network clusters for less than $1 Million/Teraflop, including server processors, switching, host bus adapters, storage controllers, disks and infrastructure." The DataDirect SwiftCluster framework leverages the capabilities of the Lustre cluster file system -- now being deployed at governmental Department of Energy and bioengineering laboratories -- by adding a layer of performance optimization for direct memory transfers from the S2A8000 block level storage controller to the GigE card residing in the computer node. This approach is expected to yield 700MB/second of bi-directional sustained block level sequential read/write performance per pipe, and 500MB/second of bi-directional sustained file level bandwidth out of the computer node's GigE front end. Performance is expected to scale linearly and very cost effectively to above 10GB/second of bandwidth and 10 Teraflops, as computer and storage nodes are added. "The SwiftCluster architecture meets our demanding I/O requirements for next generation scalable file system components," Mark Seager, deputy department head, advanced technology, Lawrence Livermore National Laboratory, said. "This type of device provides us with the ability to share data between multiple clusters at native file system speeds that scale with the size of the application." "DataDirect has followed a unique new approach to integrate their leading storage controller technology with our Lustre file system, and with SwiftCluster they are fully leveraging both their technology and the vast resources of Linux networking and file system software," Peter Braam, chief executive officer, Cluster File Systems, Inc., said. "With SwiftCluster, performance will be higher and TCO lower." With the high performance, cost-effective clustering technology enabled by SwiftCluster, processing-intensive applications such as scientific simulations, real-time modeling, and three dimensional analysis become substantially easier to perform, freeing scientists from many of the burdens imposed by the limitations of earlier computer tools. SwiftCluster is equally suited for smaller high performance cluster implementations. Using a Quadrics switch and Elan cards for client/server interconnect, and dual Intel® Xeon® 2.4GHz compute nodes, SwiftCluster is able to demonstrate in excess of 400 MB/second across just two clients. "Modern day cluster-based supercomputing systems rely not only on the low-latency, high bandwidth communication but also on seamless access to vast repositories of on-line storage. In fact, similar traits exist in many commercial sectors," Drazen Stilinovic, general manager, Quadrics Ltd., said. "Our QsNet supports not only the necessary user-level communications for supercomputing applications but can also support fast and scalable kernel messaging services, and the combination of QsNet with SwiftCluster is now providing the class of environment for such production capability. We are very excited with the collaboration with DataDirect Networks." A SwiftCluster implementation will typically utilize the following building blocks: * Parallel S2A8000 high performance storage controllers with virtualization capability * A DMA engine which closely couples block level and file level elements of the cluster * Serial PCI I/O node cards * A high performance switching infrastructure * Performance-optimized TOE-enabled GigE host bus adapters * Standard computer node front end with off the shelf Intel Xeon or Itanium processors * Storage enclosures with high volumetric efficiency * Fibre Channel or SATA disk drive technology SwiftCluster is being demonstrated in the DataDirect Networks suite and Quadrics booths (735) at SC2002 from Nov. 19-21. Access to DataDirect's technology suite can be scheduled at the DataDirect Network booth (1603) on the SC2002 show floor.