| |
| |
Mpp2 at PNNL
Quadrics worked very closely with HP and HP sub-contractors in architecting a Linux/Supercomputing solution. Quadrics interconnect (QsNet) and software-stack (QsRMS) have been ported to Linux and are now being deployed in production environments at both US National Laboratories and Industrial organization where performance and reliability both play demanding roles in guaranteeing time-to-solution.
Quadrics together with Hewlett Packard Corporation and Battelle embarked on the provision of a production Supercomputing environment combining IA-64 technology with the Quadrics QsNet products based around a Linux environment. The system will be used by the US Government's Department of Energy at the Pacific Northwest National Laboratory for various scientific research and development activities.
Quadrics is delighted to be working with HP and Battelle in this illustrious research environment, the Linux cluster market is now maturing very quickly and there is now an imperative to provide performance coupled with the necessary support infrastructure to guarantee that the price/performance benefits of Linux and the commodity are transferable to the Supercomputing environment. Quadrics in its product portfolio of high performance interconnect and cluster-wide resource management software provides a unique combination for vendors to perform this integration. | |  |
Intel Testimonial:
PNNL specified the Quadrics QsNet interconnect, which is specifically designed for high performance clusters and provides peak bandwidth of 340 MB/sec in each direction. Most important, QsNet offers vanishingly low latencies, ranging between 2 us and 5 us. By comparison, Ethernet latencies are typically measured in the tens and even hundreds of microseconds.
QsNet improves server efficiency by offloading communication tasks from the processor. The interface cards include a processor that manages protocol handling and enables write operations directly into system memory without tasking the system processor. The result: Faster writes, fewer system state changes and markedly lower processor overhead. These capabilities are vital to achieving a scalable cluster architecture that delivers performance benefits beyond 950 nodes. |
PNNL Testimonial
The prototype HP supercomputer utilizes Itanium-2 processors, HP's high-performance ZX1 chipset and a Quadrics interconnect. When running a key kernel of NWChem (a matrix multiply), the CPU sustained 96 percent of peak theoretical performance. The prototype HP supercomputer's memory bandwidth was sustained at 3.8GB/sec for a single CPU and the Quadrics Elan3 interconnect achieved 5 microsecond latency between nodes. The interconnect has demonstrated greater than 320MB/sec bandwidth, which is downright impressive.
"This is the fastest system we've ever run our NWChem codes on, and the results are even more amazing because this is just the prototype," said Scott Studham, operations technical group leader for the Molecular Science Computing Facility. "This type of performance will take us to the next level in addressing the complex environmental challenges facing the nation. Once the HP supercomputer is completely operational at more than nine teraFLOPS, we expect these key performance metrics to jump even higher than on our prototype."
|
> Legal
| |
|