By William Fellows Sun Microsystems Inc will begin to raise the profile of its HPC high-performance computing servers next week when it unveils a new 3.0 version of its HPC ClusterTools clustering software which will connect 1,024 CPUs – 16 64-way SMP servers – up from 256 CPUs today. Sun has been selling HPC servers […]
By William Fellows
Sun Microsystems Inc will begin to raise the profile of its HPC high-performance computing servers next week when it unveils a new 3.0 version of its HPC ClusterTools clustering software which will connect 1,024 CPUs – 16 64-way SMP servers – up from 256 CPUs today. Sun has been selling HPC servers – essentially its UltraEnterprise line with MPI message passing and other parallel system software technologies – for about two years. It’s estimated by IDC to have done around $500m HPC sales in 1998, giving it around 12% of a $6bn to $8bn market, up from 4.7% in 1997. Sun predicts it could be doing $1bn on HPC systems within two years.
Sun believes the market for very high performance systems is changing with new commercial demand being driven by US government agencies and research labs winning commercial contracts. At the same time, these agencies are beginning to demand more commercial data mining, analysis database and data warehouse applications for use across shared nothing architectures. In addition, a new breed of bioinformatics companies are demanding higher performance servers and software, while others are applying mathematical modeling to business processes. Some companies are also beginning to recognize the use of creating technical supply chain solutions, it claims. Others are seeking to optimize their entire businesses by modeling their complete transaction environments to see what is going on.
The majority of Sun’s HPC business is in the middle market industrial solutions, not at the high-end where Cray vector, NEC, Hitachi and Fujitsu, Intel SGI and IBM SP systems dominate. Industrial solutions are reckoned to account for 40% of the biggest 500 systems in the world; research accounts for 28%, academic 18%; classified (military) 11%; government 1% and vendors’ own systems 14%. SMP represent 24% of the top 500 systems, NUMA accounts for 24%, clusters 3%, MPP 36% and vector systems 13%. By the end of last year Sun had 25% of the market, SGI/Cray 37%, HP/Convex 3%, IBM 21% and Japanese companies 11%.
Sun says the OpenMP cross-industry parallel programming API effort is still underway – although it’s been little heard of lately – and its Dakota Scientific acquisition’s compiler technologies are now being leveraged. Sun now owns the installed base of its recent Maxstrat Corp high storage acquisition and is working on a multi-pronged HPC storage strategy encompassing Maxstrat, and IBM Corp’s HPSS high performance storage system and LCS’ hierarchical storage software which are both being ported to Solaris.
Other cross-industry initiatives include Java Grande a commercial and academic program in which Sun, IBM, Syracuse University and other are participating to put paid to Java’s sluggish performance image by creating techniques for driving numerically intensive computing into third party Java applications. Sun has also making ClusterTools available through its Community Source License Model.
Sun says the cross-industry build-out of ccNUMA distributed shared memories has helped the understanding of distributed application and system technologies. However it thinks its own focus on SMP rather than ccNUMA has only helped its own sales message. It says it will offer ccNUMA systems when it’s good and ready. It has had its Wildfire skunkworks projects out in the field for some time now, as we have previously reported. Node to node latencies across its HPC clusters is still around 10ms.