MIDDLEWARE

By Uwe Harms, Harms Supercomputing Consulting -- Last month, the SAVE (Fujitsu Siemens Computers, FSC) working group: supercomputing, met in Stuttgart-Büsnau at Max-Planck-Institute for Metal Research. All topics discussed the new hpcLine Linux-based cluster, new customers, new applications, middleware and support. Partners like Scali, ParTec, scarasoft and science + computing presented their offerings. The second day Intel and AMD disclosed their processor roadmaps. Fujitsu Siemens Computers - hpcLine Sales End of April 2002 FSC had installed 2000 processors in more than 1100 nodes of hpcLine. The first sale outside Germany by Fujitsu Systems Europe (FSE) was in France. CEA (Comm. Energy Atomique) bought a 24 node (CPU) Intel Pentium 4, 1.7 GHz, cluster connected via SCI. The Fujitsu Technical Computing facility itself installed a 16 nodes (32 CPUs) AMD MP1800+ cluster. AUDI AG ordered its third hpcLine, 64 nodes (128 CPUs) Pentium 4 Xeon with 2 GHz for Pamcrash. DaimlerChrysler ordered another 3 clusters, e.g. 17 nodes 18 CPUs, Pentium4 1.7 to 2 GHz for StarHPC in research, motor sports and Trucks. Several others have been delivered to suppliers of car manufacturers, e.g. EDAG in Fulda, 8 + 1 node for LSDyna, TeraPort was the integrator. The Eurocopter system with Pentium 4 for MSC.Nastran was installed together with MSC.Software, running MSC.Linux. T-Systems (former debis Systemhaus) extended the HWW (Joint Supercomputer Center of Academia and Industry) in Stuttgart with a 128 CPU system for CAE applications, LSDyna, StarHPC. This customised system has 64 dual Xeon 2 GHz nodes, connected with switched Ethernet 100 Mbit/1 Gbit/s. The team is evaluating the Myrinet2000 SAN (System Area Network) for LSDyna. At Proteom/University Bochum,the hpcLine is used for protein and DNA analysis. Thea Joint research project with Protagen is funded by BMBF. The customized hpcLine solution: 64 dual processor nodes with 128 AMD Athlon MP1800+ processors, switched Ethernet (100Mbit / 1Gbit). On the software side the main ISVs (Independent Software Vendors) like AEA, AVL, EASi, StarCD, ESI (Pamcrash), Inpro (INDEED), Intes, Magma and Numerical Technologies in the Netherlands have access to a hpcLine Linux Cluster. The FSC Processor Roadmap FSC follows the AMD and Intel processor roadmap. Always the actual processors and nodes are installed: *dual Intel Pentium III, Tualatin, 1.4 GHz up to 4 GB SDRAM *single Pentium 4, Northwood, 2.4 GHz up to 2 GB RD-RAM *dual Xeon Prestonia 2.4 GHz up to 2 GB RD-RAM *dual AMD Athlon MP2000+ with up to 4 GB DDR-SDRAM In development there are: *single Pentium 4, Northwood, > 2.4 GHz *dual Xeon Prestonia > 2.4 GHz. hpcLine Software News EXA Powerflow, a fluid flow simulator, runs on the hpcLine. It needs a high speed network for the communication. A run with Ethernet shows a 50% CPU load compared to 75% using Myrinet. Fluent was measured on faster CPUs, going from Pentium III to Pentium 4. On Pentium III Ethernet is sufficient, but on Pentium 4 one now needs a high-speed network. For more than 8 CPUs a System Area Network is necessary. The Proteom-Center of Ruhr-University Bochum uses a 128 processor AMD MP1800+ hpcLine cluster for sequence analysis and protein research (Primeur 23. March 2002). Their application SEQUEST is now fail save reported Kirsch. This is the first application of that kind on a Linux cluster. News from FSC Partners Scali Scali mentioned they have 175 customers in 20 countries. In 2000 they shipped 800 licensees to 57 customers, in 2001 1842 licensees and 132 customers. The main application areas are oil, gas and life sciences. http://www.scali.com Partec ParaStation 3.1 was released in May this year. The new version 3.1. has additional features for performance, and scaling, reliable data transfer, availability, process control, handling of output, usability - multi user/job/process, standard interfaces and a better price/performance. ParaStation 3 supports Myrinet, and secures the reliable transfer with 32 bit check-sum and the correct sequence of the packets. ParaStation is a factor of 2 faster than GM, the Myrinet software. ParaStation supports LSF and is integrated with LSF parallel. Thus suspending and restart of a parallel job is possible. Additional LSF parallel allows accounting and networking. ParaStation 4 supports additional hardware platforms, Linux IA-64, Intel, AMD, Hewlett-Packard Alpha and Tru64 Unix. http://www.par-tec.com Linux Cluster Installations by science + computing Science + Computing (s + c), Tübingen, have deep experiences in Linux and application software and operate a lot of the Linux clusters in the German Automotive Industry. s + c has own tools like Venus, system administration in a heterogeneous network (Windows, Linux, Unix), FlowGuide, workflow management, and distributes LSF (Load Sharing Facility). For the installation of the operating system and the application software s + c developed its s.cluster concept. They burn a CD with all the necessary software and install it on a master. This distributes it on all the nodes within a short time. With the aid of Venus they have a single point of administration. http://www.science-computing.de Cluster Operating System by Scarasoft The start-up company scarasoft presented its ideas of the Cluster Operating System CLos. It runs on Linux, a kernel extension, and is an automatic, self-configuring and organising system. It recognises automatically the network and nodes and does not need any manual configuration. http://www.scarasoft.com Intel and AMD Processor Roadmaps Intel Intel has in parallel to Itanium 2 (McKinley) 5 to 6 processors in the design cycles. Pacific Northwest Nat. Lab. bought a 1400 processor system based on McKinley and Madison for 24.5 Mio. $ and 8.3 TeraFlop/s with a Quadrics Interconnect. The processor in 2007 for the volume market is supposed to have a 10 GHz clock with about 1 billion transistors. The technology of today will be continued up to 2010. Today there is an ALU (arithmetic logical unit) in the Labs running with 10 GHz. http://www.intel.com/research/silicon AMD Roadmap On the 32-bit architecture AMD will deliver Palomino in the first half of 2002 (July ?). In the second half 2002 Thoroughbred with 0.13 micron and in first half 3003 Barton with 0.13 micron is in the pipeline, front-side bus and Chipset are the same. In the first half 2003 Sledgehammer MP, a 2 to 8 way MP 64-bit system will be delivered. A desktop version, the Clawhammer, a one to 2 way system in November/December this year. Both use SOI (Silicon on Insulator) and 0.13 micron technology. The x86-64 Instruction Set Architecture offers support for 32- and 64-bit computing needs. 32-bit applications and operating systems run with full native performance. Additionally there is full x86 compatibility with BIOS, OS, applications, drivers etc. The x86-64 mode is built on x86. The Integer/Address register files and data paths are native 64-bit. The virtual address space uses 48 bit while the physical address space is 40-bit. Operating systems are Linux (64 bit) or Microsoft .NET, 64 bit. Details and the public specification of the x86-64 can be found at: http://www.x86-64.org Some of the presentations, most of them in English, of this and other SAVE meetings can be found at http://www.hpcline.de.