Munich Systems Fair presents various noveltiesMunich 22 October 2001 During the Systems Fair in Munich, 15. to 19. October, Uwe Harms discussed some HPC topics with presenting vendors like Fujitsu Siemens Computers, Hewlett-Packard, IBM, NEC and Sun Microsystems. Compaq did not exhibit.
In cooperation FSC and science + computing (s + c), Tübingen, demonstrated the performance of the hpcLine computer - a cluster computer - and s + c's Linux expertise in engineering applications. As Eric Schnepf, FSC, mentioned, hpcLine now is available with AMD processors too. Thus FSC can deliver systems that are optimal for specific applications, Intel and AMD processors, SCI, Myrinet and FastEthernet connection. s + c developed an "engineering Linux", which is stripped down for these applications. Additionally they offer tools for a fast implementation of Linux and the application software on multinode clusters.
HP groups its servers and systems into tower, rack, superscalable servers, new are blade servers - extra thin, blade storage will appear and appliances. L- and N-class will have new names, e.g. rp7400 as L-class. The "p" in rp means PA-RISC processor. The PA-RISC 8700 was announced recently with 650 and 750 MHz clock rate. The PA-RISC 8800 will be delivered in products end of 2002, the 8900 in 2003. Then the PA-RISC history will probably end.
The rx4610 and rx9610 are based on Itanium, which one can recognise in the "x" in the product name. They have up to two and up to 4 processors respectively. In discussions HP expects a
70% performance improvement with the advent of McKinley compared to Itanium. At their booth HP demonstrated MSC.Software Nastran running on Itanium. The performance of this processors compares
in this application to PA-RISC 8700. Nastran is about 20% faster using HP-UX operating system compared to Linux, a result of better compilers.
Here I had the chance to discuss the topics with Francis Kuhlen, IBM Vice President Systems Sales Central Region and Hans-Jörgen Rehm, Public Relations Germany. First we stressed the topic autonomic computing. Autonomic computing is a visionary approach to building self-governing computing systems that require a minimum of human interference. The term derives from the body's autonomic nervous system, which controls key functions without conscious awareness or involvement. This project started Monday, October 15th, and will bring together IBM's and academic resources. IBM funds it with more than one billion US$ within the next three years. It addresses the problem of the big parallel systems with thousands of nodes and ten thousands of processors and the probability of a processor or memory failure too.
IBM will use its chipset Summit, which will be extended, in Intel server architectures. An other important project is Grid computing. Here IBM won several projects, e.g. the British Grid.
Some weeks ago, IBM announced Regatta, the eServer pSeries 690. It has a maximum of 32 POWER4 processors. Two processors are on a chip, accessing a shared cache. Four of these POWER4-Chips are
assembled on a so called Multi-Chip-Module. This connects the chips in a way that on a space of 85 x 85 Millimeters one can find an 8-way POWER4 SMP-unit with a shared 6 Megabyte Level2-
and 128 Megabyte Level3-Cache. The POWER4 processor is clocked with 1.1 and 1.3 GHz.
Francis Kuhlen mentioned that there is a HPC version of the p690. In this case one processor is deactivated per chip. The reason is the memory bandwidth. Now the processor can access the whole cache and therefore deliver a higher application performance. The smallest, 8-processor version, of the p690 will cost about US$ 450.000.
For the future 3 p690 can be connected using NUMA-technology to a 96 processor system. The p690 can be connected as a node with the SP-Switch. Then they look like an IBM SP parallel computer with external nodes. The p690 can be partitioned down to one per processor. On mainframes a partition can be a fraction of a processor, thus one can have several partitions per processor, which share the peripherals. I expect that on the pSeries the same will be realised in the future.
By the way, Hans-Jörgen Rehm mentioned that one thousand zSeries (Mainframes) have been sold within one year.
Although NEC did not present any supercomputer in their Systems booth, I had a talk with Ralf Wolf, Manager Marketing and Communications NEC. He told me that NEC had won the DKRZ race,
German Climate Research Center in Hamburg. Until end of October this year they will use the 1995 installed Cray C90 with 16 processors, 16 GFlop/s. Now the DKRZ can acquire a NEC SX-6. In
the first step - as an interim solution - they will use two SX-4. In February 2002 DKRZ installs 8 NEC SX-6 nodes, 64 processors, followed by another 8 nodes in August 2002. In April 2003 the final configuration will be installed, approximately 192 CPUs, 1.5 TByte memory and 1.5 TeraFlop/s peak performance.
In typical climate application programs DKRZ expects a sustained performance of 500 GFlop/s and a performance improvement of a factor of 100 compared with the Cray C90. The disc capacity will
grow to more than 60 TeraByte and more than 3.4 PetaByte tape memory archive.
As Sun announced its new Sun Fire E15K on September 24 in Munich, there is no more news. They showed an E15K running on their booth.
Uwe Harms |
 |