American HPC User Forum meets European HPC CommunityBristol 23 May 2003 On May 20th, the American HPC User Forum for the first time met European Users in Bristol, UK, and on May 22nd in Annecy, France. In total more than 100 scientists participated. This event starts the dialogue between US and European HPC users and buyers. The European market is very dynamic because of the different countries and languages and not identical to the US market. During the meetings vendors as well as users presented their experiences, application areas and news. (Uwe Harms)
Both venues in Bristol and Annecy are Hewlett-Packard HPC research centres, thus HP hosted both meetings. Annecy has its roots in Compaq and will be moved to Grenoble in July this year. IDC organised these meetings in co-operation with HP. Sponsors have been Cray, HP, IBM, IDC, Intel, SGI and Sun Microsystems in partnership with HPCwire.
Additionally he listed other topics like the future of HPC, of vendors, what to buy, which operating systems and future architectures. The first Forum meeting started with 12 participants, the last included about 100. There are 130 members of the Forum. User organisations have annual dues of 15.000 US$, 3 persons; vendors 35.000 US$, 3 persons; and an individual user, 5.000 US$, 1 person. Earl Joseph discussed new findings in the IDC HPC benchmark. New types of ranking will be added like price/performance and better metrics. SDCS develops their own performance models. Papers and tutorials are at http://www.sdsc.edu/PMaC. Additionally he presented DCs, which contain the presentations of the IDC User Forums in 2001 and 2002.
IDC HPC ForecastChris Willard, IDC, shortly presented the IDC expectations. The overall revenue 2002 versus 2001:
He expects a revenue growth of 6.1% CAGR to 2007. The relation to the peak year of 2000, the CAGR is only 2%. The total revenue in 2007 will be 6.3 Bio. US$. The HPC cluster market will be greater than 2.5 Mio US$ in 2006, compared to 800.000 US$ in 2001 and 1.7 Mio US$ in 2002. He seperated the clusters into bright clusters, delivered by traditional vendors, dim clusters, installed by end users or VAR, dark clusters by end users, and invisible clusters, by end users in departments for example.
European HPC UsersFrom the European user perspective Christine Soleil, EPFL (Swiss Federal Institute of Technology), Lausanne, and Uwe Küster, HLRS (Höchstleistungsrechenzentrum), Stuttgart presented their actual HPC situation and the future planning.EPFL is involved in HPC since its first Cray 1 in 1986. Now Switzerland's HPC community can use a 128 processor SGI Origin 3800 and the Swiss T1 (now replaced), HP systems, MPP clusters, and an IBM SP3 in Lausanne, and a NEC SX5 with 16 processors and 8 IBM Regattas, p690, with 32 processors each, connected with the Colony Switch. She presented some application areas like complex flow simulations, high energy physics, chemical physics and material sciences, imaging computing sciences and networking and earth and environmentalsciences, as well as biosciences and medical imaging. Since 31.10.2002 EPFL has a partnership with HP, which results in 25 SC45 nodes connected with Quadrics and a peak performance of .25 TFlop/s, following Marie-Christine Sawley. Uwe Küster discussed the new HLRS Simulation Workbench, which will be realised within the new procurement - just running. The concept shows a file system in the centre, its name and product are actually not clear. It will be stable longer than the next machines. The selection is of high importance for HPC. Around this file system one finds the supercomputer, several pre/post processing servers for the analysis, and visualisation servers. Thus the huge files need not be moved between the different machines. The next shell is the software core and the applications. All this will be connected to users all over Germany. The machines will be utilised in a single way. In June 2003 there will be the second call for a final proposal, in July HLRS will evaluate the final proposals and start negotiations. In August/September the decision will be published. They expect the first small system for code porting in Q4 2003, in Spring 2004 the main system and in Q4 2005 the second stage of the procurement will be installed. The requirements are a sustained performnance of 2 TeraFlop/s on one application code out of the six in the benchmark.
Vendor TalksEd Turkel, Hewlett-Packard, described the activities in the technical HPC market, MPI (message Passing Interface), mathematical libraries, Compilers. The experiences in the software tuning has impacts on the compiler. Next week, HP will announce the new Itanium 6M, called Madison, based on superdomes with the SX1000 chipset. One can remove the PA RISC cellboards and put the Itanium cellboards in. The system runs with the same operating system, HP-UX, Linux, Windows, Open VMS and probably non Stop Kernel.Addison Snell presented SGI's view, stressing the Numaflex architecture on the MIPS and Itanium based systems. The O3000 contains now 128 CPUs/rack, four times the density of previous models. The Altix 3000, Itanium based, is the first system with 64 CPUs in a single Linux system image. The Altix 3000 supercluster is the first cluster with a global shared memory across multiple nodes. SGI shipped more than 2000 processors, most in high end. he mentioned that Europe is far ahead of the US in grids. William White said that Cray targets in the capability segment. He presented first STREAM results, > 1 TB/s that means twice the performance of the previously highest. He named the products of the road map, X1, X1e, Black Widow, Black Widow +, Black Widow ++. Peter Folkes, HPTC Sun, discussed the new activities. There is a lot experience on Senior Management Team. They bought Cray business servers and Thinking Machines. The Sun HPC cluster allows up to 8 systems, E15K, E12K or 6800.
Supercomputer CentersJim Kasdorf showed details of the Pittsburgh Supercomputer Center, the Terascale system with 750 ES45, 3000 EV68 Alpha processors with 1 GHz. They tried to rebuild the Cray T3E. As the interconnect they realised the Quadrics solution, 2 rail fat tree.Jean Gonnord, CEA/DAM, presented his experience of a classified Terascale computer centre, based on Alphas too. They gained 1.32 TF on 2474 processors. In September 2003, CEA/DAM will assume responsibility for all CEA computing, Classified Defense computing centre (TERA) and Research and Technology computer centre (CCRT, open to collaboration with outside partners: CEA, Electricité de France, Onera, Snecma). He expects a peak performance of one tenth of the classified machine, perhaps higher. The CEA scientific computing complex (CEA/DAM-Ile de France) operates a NEC SX-6 (352 GF = 44 * 8 GF) for climate/weather and HP SC45 (2 TF = 800 * 2.5 GF). For 2005 they have the goal of 10 sustained TF 2005 and 100 sustained TF in 2009. Paul Muzio, U.S. DOD, discussed early Cray X1 experience at the Army High Performance Computing Research Center. AHPCRC consists of two parts: a basic research programme (unclassified) which includes 5 universities and 5 basic research areas and the classified part. The Cray X1 is a capability machine. They also have IBM SPs, clusters, other systems. The X1 is great for vector problems, scalar ones run on a cluster or a T3E. The CFD application is a modern MPI implementation, written in C. They got about 31% of peak, which is pretty good at this stage. Compared with the T3E-1200, on small data set achieved a speedup of 51-53x on production X1. The comparative performance, ~25x faster than Origin 3000, 10-15x faster than IBM SP Power 4. 28 X1 MSPs do the same work as on 1024-processor T3E. In the final discussion, Earl Joseph announced additional meetings after the US Forum meetings in Europe. He wants much interchange between Europe and the US.
Uwe Harms [News on Advanced IT] [Calendar] [Analysis] [IT in Medicine] |
 |