In late March, CONVEX Computer Corporation announced a contract with NCSA for the purchase of a 64-CPU Exemplar Scalable Parallel Processor (SPP) system.
"The CONVEX Exemplar is leading the market base toward a new architecture of supercomputing," says NCSA Director Larry Smarr. "It combines the practical scalability of clustering large-memory multiprocessors with the programming advantages of a global shared memory. This creates a new kind of parallel system that is easy- to-program, feels like a workstation, and scales in performance as it scales in size. The Exemplar will host Grand Challenges of academic researchers, while producing a rich set of third party applications for industrial users."
"The Exemplar represents the future of high-performance computing- binary compatibility with the desktop in a scalable system," says Steven Wallach, Convex senior vice-president of technology. "Scalable means that the system performance increases as the system size increases, all the way from the desktop to a system with hundreds of CPUs-as you buy upgrades in CPUs, you get the power you pay for. Only Convex is delivering that capability for high-performance systems."
Based on Hewlett-Packard's (HP's) industry-leading PA-RISC technology, the Exemplar provides supercomputing performance with desktop compatibility at workstation price-performance levels. Compatible with the HP-UX operating system, the Exemplar series provides the capability to run the thousands of applications that are available for HP workstations. The product family is available directly from Convex and from HP, where the Exemplar is sold as an extension of the company's line of PA-RISC-based UNIX workstations.
Convex also announced that NCSA will upgrade the Center's UniTree mass storage system from a CONVEX C220 to a CONVEX C3800 system. The upgraded system will have over 200 gigabytes of high- performance disk and unlimited helical scan storage using a combination of Ampex DD-2 and Metrum tape. Users of NCSA's advanced computing environment can move data into and out of the UniTree system from any of several dozen systems using NCSA's 800 megabits per second (Mbps) HIPPI networks and from hundreds of systems using 100 Mbps FDDI networks.
The installation at NCSA is one of more than 90 Convex DataSeries systems installed worldwide. The Convex DataSeries allows organizations to efficiently manage storage and retrieval of the vast amounts of data created by research or applications such as complex decision support searches, 3D seismic processing, automobile crash analysis, and molecular modeling.
The Monte Carlo equations central to quantum mechanical systems are nearly ideal for the CONVEX Exemplar. They use little memory, but they are computationally very intensive and are innately parallel. As such, they can be run independently or spread across several processors. This latter criterion is enabling the NCSA researchers to test the Exemplar now as a single node; later, when the remainder of the hardware and software is installed, as a hypernode.
So far, Ceperley's group has been running the Exemplar as independent scalar workstations, testing it for classes of errors such as memory management. In preparation for its implementation as a full parallel architecture later this year, they are developing a message passing interface (MPI)-based parallel algorithm so that their quantum codes will run on the Exemplar and be portable to other major parallel machines.
The group is looking forward to exploiting the full capacity of the Exemplar. In the meantime, they are pleased with the machine's performance. Particularly significant has been the ease with which it enables them to transport codes from workstations to a high- performance computing environment because of the compatiblity of the compilers.
"It only took us a few days to get production runs on the Exemplar compared with several months on other machines such as the CM-5," says Ceperley. "When the whole system is up and running, we will be able to use the machine efficiently in any configuration- whether as single workstations, as a computational unit, or as one big parallel node."
Other new and upgraded computational resources at NCSA are listed below.
Holly Korab contributed to this article.
