Hewlett-Packard Co yesterday launched its expected upgrade of the high-end V-Class servers as the V2500, claiming the systems were the first to bridge the gap between scientific and commercial systems. Single V2500 nodes can have up to 32 of HPÆs PA-8500 chips running at 440MHz, running as symmetrical multiprocessing machines. But the V2500 extends that […]
Hewlett-Packard Co yesterday launched its expected upgrade of the high-end V-Class servers as the V2500, claiming the systems were the first to bridge the gap between scientific and commercial systems. Single V2500 nodes can have up to 32 of HPÆs PA-8500 chips running at 440MHz, running as symmetrical multiprocessing machines. But the V2500 extends that to 128 processors by linking up to four nodes using HPÆs low latency 16 Gb/sec Hyperplane crossbar interconnect, which it inherited and developed from its Convex Computer Corp acquisition. The four nodes operate under what HP terms a cache coherent non-uniform memory access architecture, and so the full 128 configuration requires only a single copy of the standard edition of the HP-UX 11 operating system. HP calls the scheme its Scalable Computing Architecture, and says it combines the benefits of SMP and distributed memory (clustering) architectures. It works by implementing a multilevel memory system, with the first level consisting of traditional SMP memory, and the second a tying together of the first level memories through the Hyperplane. Each cabinet contains an SCA HyperLink cache, containing memory references that have been made over the interconnect to other cabinets. This is said to lower the effective system-wide latencies and increases the effective bandwidth of inter-cabinet references. Each cabinet supports up to 32Gb of SDRAM. Traditional clustering can scale the V-class servers up to 2,048 processors. HP says itÆs tested 64-way configurations using the Linpack benchmark as running at over 55 gigaflops, and in a 32-way configuration it delivers over 7,249 SPECint rate 95. Systems are intended to run scientific applications, EDA software and product data management applications, but also datawarehousing. Users can run applications within an SMP node or, if appropriate, across the full four-node system over the Hyperplane. The first V2500s will ship in January, but the full four-node, 128 processor configurations arenÆt expected to be available until mid-year. Entry-level prices start at around $200,000. HP says its primary competition is IBM Corp and Sun Microsystems Inc. But both Silicon Graphics Inc and Sequent Computer Systems Inc offer comparable machines, and both were highly suspicious of HP calling its architecture ccNuma. I think theyÆre mis-using the term said a spokesman from Sequent, the company which claims to have pioneered ccNuma and which offers 64 processor ccNuma boxes. SGI said that shared memory over the whole system was the most critical element of Numa, and said it was doubtful how transparent that memory would be to the applications under HP’s architecture. HP says that SGI systems run only technical software, and Sequent only commercial applications.